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.1002111 | A Dynamic Landscape for Antibody Binding Modulates Antibody-Mediated Neutralization of West Nile Virus | Neutralizing antibodies are a significant component of the host's protective response against flavivirus infection. Neutralization of flaviviruses occurs when individual virions are engaged by antibodies with a stoichiometry that exceeds a required threshold. From this “multiple-hit” perspective, the neutralizing activity of antibodies is governed by the affinity with which it binds its epitope and the number of times this determinant is displayed on the surface of the virion. In this study, we investigated time-dependent changes in the fate of West Nile virus (WNV) decorated with antibody in solution. Experiments with the well-characterized neutralizing monoclonal antibody (MAb) E16 revealed a significant increase in neutralization activity over time that could not be explained by the kinetics of antibody binding, virion aggregation, or the action of complement. Additional kinetic experiments using the fusion-loop specific MAb E53, which has limited neutralizing activity because it recognizes a relatively inaccessible epitope on mature virions, identified a role of virus “breathing” in regulating neutralization activity. Remarkably, MAb E53 neutralized mature WNV in a time- and temperature-dependent manner. This phenomenon was confirmed in studies with a large panel of MAbs specific for epitopes in each domain of the WNV envelope protein, with sera from recipients of a live attenuated WNV vaccine, and in experiments with dengue virus. Given enough time, significant inhibition of infection was observed even for antibodies with very limited, or no neutralizing activity in standard neutralization assays. Together, our data suggests that the structural dynamics of flaviviruses impacts antibody-mediated neutralization via exposure of otherwise inaccessible epitopes, allowing for antibodies to dock on the virion with a stoichiometry sufficient for neutralization.
| Neutralizing antibodies are a critical aspect of protection from flavivirus infection. The primary targets of neutralizing antibodies are the envelope (E) proteins incorporated into virions. The neutralizing activity of antibodies is determined by the affinity with which they interact with the virion, and the total number of sites available for binding. In this study, we investigate the impact of dynamic motion of the viral E proteins on antibody-mediated neutralization. Using panels of monoclonal antibodies and immune sera, we demonstrate that the dynamic motion of virions significantly impacts antibody-mediated neutralization of West Nile and dengue viruses by modulating epitope accessibility. Increasing the length of the interactions between antibody and virus resulted in increased neutralization reflecting engagement of epitopes that are not exposed on the surface of the virion in its average state, but instead become accessible through the dynamic motion of E proteins. While examples of the impact of structural dynamics on antibody binding have been described previously, our data suggests this phenomenon plays a role in neutralization by all antibodies that bind the array of E proteins on the virion. Our data identifies epitope accessibility as a critical, yet dynamic, factor that governs the neutralizing activity of anti-flavivirus antibodies.
| Flaviviruses are a group of ∼70 RNA viruses that cause morbidity and mortality on a global scale, with greater than 100 million human infections annually [1]. Viruses within this genus of medical concern include yellow fever virus, tick-borne encephalitis virus, Japanese encephalitis virus, dengue virus (DENV) and West Nile virus (WNV). WNV is a mosquito-borne flavivirus maintained in nature in an enzootic cycle with birds. WNV infections of humans result in a spectrum of clinical symptoms depending, in part, on the age and immune status of the individual. While most infections are sub-clinical, symptomatic cases range from self-limiting fever to acute flaccid paralysis and encephalitis [2]. Since its introduction into North America in 1999, as many as three million people have been infected by WNV [3], with ∼1000 severe infections occurring in the United States annually (www.cdc.gov). To date, there are no WNV-specific treatments or vaccines licensed for use in humans.
Flaviviruses are small spherical virions that encapsidate an ∼11 kb genomic RNA of positive-sense polarity [1]. This RNA is translated as a single polyprotein that is processed by viral and host cell proteases into ten functionally distinct proteins. Flaviviruses encode three structural proteins that comprise the virus particle and seven non-structural proteins that function to process the viral polyprotein, replicate the viral genome, and antagonize the host's protective response to infection [1], [4], [5], [6], [7], [8], [9], [10], [11]. Flaviviruses bud into the endoplasmic reticulum as immature viruses that incorporate 60 heterotrimeric spikes of the envelope (E) and precursor to membrane (prM) proteins [12], [13]. Maturation of virions during egress from the cell is associated with a pH-dependent change in the arrangement and oligomeric state of the E protein and cleavage of prM by a host cell furin-like serine protease [14], [15], [16]. While prM cleavage is a required step in the formation of an infectious virion [17], several lines of evidence suggest that a significant fraction of infectious virions retain some uncleaved prM [18], [19], [20], [21], [22]. The E protein is a class II fusion protein composed of three structurally distinct domains (domains I-III; E-DI-III) connected to the viral membrane by a helical stem region [23]. On mature virions, 180 copies of the E protein are arranged as anti-parallel dimers in an unusual herringbone T = 3 pseudo-icosahedral array [24], [25], [26]. In this configuration, E proteins are not quasi-equivalent, but exist in three distinct chemical environments defined by their proximity to the two-, three-, or five-fold symmetry axis. This arrangement adds complexity to the antigenic surface of virus particles as antibodies may not be able to bind all 180 E proteins due to steric constraints on antibody binding or occlusion of the epitope itself [27], [28].
Humoral immunity is a critical aspect of protection against flavivirus infection (reviewed in [29]). Eliciting a protective antibody response is a primary goal in the development of safe and effective flavivirus vaccines [30]. The majority of neutralizing antibodies against flaviviruses are directed against the E protein; neutralizing monoclonal antibodies (MAbs) have been identified that bind to all three E protein domains (reviewed in [29], [31]). While antibodies that bind the prM protein have been detected in humans, they possess limited neutralizing activity [32], [33], [34], [35], [36].
Antibody-mediated neutralization of WNV is a “multiple-hit” phenomenon that occurs when virions are engaged by antibody with a stoichiometry that exceeds a required threshold [29], [37]. Our estimate of this threshold is ∼30 antibody molecules per virion [38]. Two factors principally govern the neutralizing activity of an antibody. Antibody affinity (or avidity) controls the fraction of accessible epitopes on the intact virion occupied at a particular concentration of antibody. From a vaccination standpoint, eliciting antibodies that bind viral antigens with high affinity is desirable because they can reach the stoichiometric threshold required for neutralization at lower concentrations. However, high-affinity interactions do not always translate into significant functional potency. Because of steric constraints arising from the dense icosahedral arrangement of E proteins on the mature virion, not all epitopes are displayed on intact virions equivalently. Epitope accessibility is a second independent parameter that defines the circumstances that allow for neutralization, and differs markedly between antibodies that recognize structurally distinct epitopes [21], [28], [38], [39], [40]. In fact, many epitopes, including those recognized by antibodies commonly elicited in vivo, are not displayed on the surface of the mature virion enough times to allow for neutralization even when fully occupied [21], [38], [41].
An important concept of the “multiple-hit” model of neutralization is that infectious virions may be decorated with non-neutralizing quantities of antibody. Evidence in support of this includes increases in the neutralizing activity of antibodies observed in the presence of anti-IgG secondary antibodies or complement [37], [42], [43], and the phenomenon of antibody-dependent enhancement (ADE) of infection [44]. Because the rate of antibody-virion interactions is orders of magnitude faster than the rate of virus binding to cells, viruses are likely engaged by antibody for significant periods prior to productive interactions with target cells. Here, we investigated the fate of virions decorated by antibody with a stoichiometry that does not exceed the threshold required for neutralization. Our data revealed time- and temperature-dependent increases in neutralization of WNV that would not be predicted for static virions bound by antibodies under steady-state conditions. Because both viral proteins and intact virions are dynamic [45], [46], [47], [48], the increased neutralization can be explained by changes in epitope accessibility occurring as virions sample different states of a dynamic ensemble of conformations. That kinetic aspects of neutralization were observed for every WNV- and DENV-reactive MAb examined, as well as WNV-polyclonal immune sera, suggests a widespread impact of the dynamic motion of virions on epitope accessibility and antibody-mediated neutralization.
To investigate whether prolonged exposure of virions to “non-neutralizing” quantities of antibody has functional consequences, we performed a series of kinetic experiments with the WNV E-DIII-lateral ridge specific MAb E16 at a concentration sufficient to neutralize the infectivity of half the virus particles (the EC50). At this concentration, half the virions in a population are bound by antibody with a stoichiometry that exceeds the neutralization threshold, whereas the other half are bound by fewer antibody molecules. WNV reporter virus particles (RVPs) encoding a GFP reporter gene were incubated with 10 ng/ml E16 for one hour at room temperature to allow for steady-state binding (Figure S1). After the room temperature incubation, RVP-antibody complexes were shifted to 37°C for the indicated periods prior to the addition of target Raji-DC-SIGNR cells. The infectivity of RVPs in the presence and absence of antibody was determined by flow cytometric analysis, and is expressed relative to the level of infectivity observed when cells were added to viral immune complexes immediately after the initial room temperature incubation (Figure 1a). Using this approach, we observed an antibody-independent decrease in RVP infectivity with prolonged incubation at 37°C (intrinsic decay), consistent with our previously published findings [49]. Incubation in the presence of E16 resulted in a 4.7-fold reduction in the apparent half-life of WNV RVPs relative to the intrinsic rate of decay observed in the absence of antibody (n = 23; p<0.0001) (Figure 1b). A similar 5.6-fold change in the rate with which virions lost infectivity in the presence of E16 was observed using fully-infectious WNV (Figure 1c). Incubation of RVPs or virus with high concentrations of the DENV-specific MAb 3H5 did not impact WNV infectivity, indicating a requirement for virus-reactive antibody (Figure 1b and c). These results were not explained by increased binding of virions to tissue-culture plastic during prolonged incubation with antibody (p = .32, n = 5; Figure S2a), the action of complement (p = .17, n = 5; Figure S2b), or antibody-mediated aggregation (E16 Fab experiments discussed below; Figure 2b). Additionally, the rapid reduction in RVP infectivity in the presence of virus-specific antibody was still observed in the presence of a broad spectrum of protease inhibitors (p = .74, n = 2; Figure S2c), suggesting that the loss of infectivity cannot be explained by physical destruction of RVPs due to contaminating proteases.
We next evaluated the impact of kinetics on neutralization by E16 in a series of dose-response studies. WNV RVPs were incubated with serial four-fold dilutions of E16 at room temperature for one hour, followed by a shift to 37°C for the indicated times prior to the addition of target cells (Figure 2a). The dose-response profile of E16 obtained when cells were added immediately after the room temperature incubation required to achieve steady-state binding revealed the expected relatively steep sigmoidal curve (Figure 2a; reference curve) [38]. Dose-response curves obtained from samples incubated for varying lengths of time at 37°C prior to the addition of cells differed from the reference curve in several respects. First, the percentage of cells infected in the presence of low concentrations of antibody (the top of the sigmoidal dose-response curve) decreased as a function of increased incubation at 37°C, corresponding to the intrinsic decay of virus particles observed in the kinetic studies described above (Figure 1, Figure 2a; left panel). Second, increasing the length of incubation at 37°C shifted the EC50 to lower concentrations of antibody (Figure 2a; right panel (n = 19, p<0.0001)). Analysis of changes in infectivity over time revealed a 5.8-fold decrease in the apparent half-life at ∼10 ng/ml E16 relative to the antibody-independent intrinsic decay for RVPs alone (n = 18 independent dose-response experiments, p<0.0001), similar to the results obtained using the single antibody-concentration studies described in Figure 1. Time-dependent changes in dose-response curves were also observed with Fab fragments of E16 (∼3.8-fold decrease in EC50 after 24 hr incubation, n = 2) (Figure 2b), indicating the phenomenon does not reflect antibody-mediated virus aggregation.
ADE describes the dramatic increase in infection of Fcγ-receptor-expressing cells in the presence of non- or weakly-neutralizing quantities of antibody and has been linked to severe clinical outcomes following secondary DENV infections of humans (reviewed in [50]). Antibody-mediated neutralization and ADE are two phenomena related by the number of antibodies bound to the virion [38], [51], therefore kinetic changes in the neutralization activity of an antibody should have a corresponding impact on the ADE profile as well. To explore this, kinetic dose-response curves were set up as described above, except using FcγIIb-expressing K652 target cells. Infection of K562 cells occurs through Fcγ-receptor-mediated uptake of antibody-virus immune complexes; in the absence of antibody, these cells are refractory to infection due to an inability to efficiently bind WNV. When immune complexes were added to cells immediately following the room temperature incubation required to achieve steady-state binding (Figure 3; reference curve), the expected bell shaped dose-response profile was observed. Incubation of antibody-virion complexes at 37°C prior to the addition of cells resulted in a marked change in the shape of the ADE profile. A reduction in the magnitude of ADE was observed over time, consistent with a decrease in the number of infectious virions in solution due to intrinsic decay (Figure 3; left panel). Notably, prolonged incubation resulted in a reduction in the concentration of E16 at which maximal ADE was observed (Figure 3; right panel), corresponding to the reduction in the EC50 observed above (Figure 2a; right panel).
Because the kinetic neutralization and enhancement experiments presented above were designed to allow for steady-state binding of the virion by antibody prior to incubation at 37°C (validated in Figure S1), changes in virus infectivity over time should not reflect continued engagement of individual virions by increasing numbers of antibody molecules. However, an important assumption of this model is that the number of epitopes on the virion does not change.
To explore the role of dynamic motion of the virion in WNV neutralization, we took advantage of the fact that the neutralizing activity of several classes of antibodies is significantly limited by epitope accessibility [21], [39], [40]. MAb E53 is a high affinity DII-fusion loop-reactive antibody that has a limited capacity to neutralize mature virions because its epitope is buried on the surface of the virus particle. Even in the presence of saturating concentrations of antibody, E53 does not bind mature virions enough times to exceed the stoichiometric requirements for neutralization [21]. In support, structural studies revealed that E53 Fab fragments were only capable of binding E proteins of the heterotrimeric spikes present on immature virions [52]. For E53 to engage a mature virion with a stoichiometry that exceeds the neutralization threshold, changes in the accessibility of its otherwise cryptic epitopes would be required.
In agreement with our previous findings [21], MAb E53 failed to significantly neutralize a homogeneous population of mature WNV RVPs when antibody-virion complexes were added to cells following a one hour incubation at room temperature (Figure 4a and b, reference curves). Incubation at 37°C for increasing time intervals prior to infection revealed a gradual increase in antibody potency, with more than half the virions incubated at 37°C for 26 hours becoming susceptible to neutralization (Figure 4a). Because the E53 epitope is not accessible on the highly ordered surface of the mature virion [52], the increase in potency requires changes in exposure of the fusion loop epitope over time as the E proteins on the surface of the virion sample different conformations at equilibrium. Binding of E53 has the potential to stabilize the epitope in an exposed conformation that occurs only transiently in the absence of antibody. That significant neutralization requires considerable time at 37°C likely reflects the relatively large number of epitopes that must become accessible on the average virion in order to exceed the neutralization threshold (Figure S3). If increased neutralization by E53 reflects dynamic movement of the virion, one would predict it would be both time- and temperature-dependent. To test this, we performed additional experiments with longer incubation times and a range of temperatures. Increasing the temperature to 40°C enhanced neutralization as compared to 37°C (Figure 4b, compare green to orange curves), whereas decreasing the temperature to 33°C reduced neutralization capacity (Figure 4b, compare green to blue curves). Similar results were observed for the E-DI specific MAb E121 shown previously to bind an epitope that is poorly accessible on the mature virion (Figure 4c) [21], [39].
To investigate whether kinetic effects on antibody-mediated neutralization are a common phenomenon, we performed additional studies using a panel of 12 WNV MAbs. In each case, we observed time- and temperature-dependent increases in neutralization, similar to our results with E16, E53, and E121 (Figure 5). Collectively, the 15 WNV MAbs used in the current study bind epitopes distributed across all three domains of the E protein and are characterized by varying degrees of functional potency [39], [53]. We next performed experiments using previously characterized immune sera from five recipients of a phase I clinical trial of a live attenuated WNV vaccine [21] (A. Durbin, S. Whitehead, and colleagues, unpublished data). After a one hour room temperature incubation to allow for steady-state binding, virus-sera mixtures either were used to immediately infect cells (Figure 6, reference curves) or incubated an additional 23 hours at 37°C or 40°C before infection (Figure 6, blue and orange curves, respectively). We found that the polyclonal mixtures of antibody present in immune sera behaved in a manner similar to MAb of defined specificity.
To extend our observations to another flavivirus, we performed analogous studies with DENV serotype 1 (DENV-1) RVPs and a panel of previously characterized DENV-1 E protein-reactive MAbs [53], [54]. After incubating RVPs and antibody at 37°C for one hour to allow steady-state binding to occur, RVP-antibody complexes were either immediately used to infect cells (Figure 7, reference curves) or further incubated for two or seven additional hours at 37°C and 40°C before infection (Figure 7, dotted curves). Because it was established previously that some antibodies do not bind DENV efficiently at low temperatures [46], incubation at 37°C was used to establish steady-state binding (Figure S1). As observed in our studies with WNV, time- and temperature-dependent increases in neutralization potency were observed (Figure 7). Interestingly, shorter incubation times were necessary as compared to experiments with WNV, due to a more rapid intrinsic decay rate of DENV as compared to WNV, which has previously been reported [49]. This was also demonstrated in parallel studies of WNV and DENV-1 RVPs with the cross-reactive MAb E60; significant increases in neutralization were observed more rapidly with DENV than WNV (Figure S4). Altogether, these results suggest that DENV exists in a more dynamic state than WNV.
The existence of high-resolution structures of the E proteins of flaviviruses, and their arrangement on the surface of the virion at different stages of the viral lifecycle, have been a powerful tool for studying how these viruses interact with the host (reviewed in [6], [26]). Interpretation of these structures in the context of biological systems is complicated by the fact that they represent the average state of what is likely a very dynamic ensemble of conformations sampled by virions at equilibrium. The structural dynamics of non-enveloped viruses have been studied extensively (reviewed in [45], [55]). Limited proteolysis studies of the Flock house virus strongly suggest the capsid proteins “breathe”, allowing proteases access to internal structures predicted to be inaccessible on an intact and non-dynamic virion [47]. Furthermore, drugs that prevent the dynamic movement of the picornavirus human rhinovirus 14 have been shown to effectively inhibit infectivity [48]. While the dynamics of enveloped viruses have been studied less extensively, examples of temperature- and time-dependent antibody reactivity have been described for several viruses [46], [56], [57]. The DENV group-reactive neutralizing MAb 1A1D-2 binds an epitope on the A-strand of E-DIII that is poorly accessible on proteins proximal to each of the three symmetry axes of the mature virion. Binding of this MAb is temperature-dependent (does not occur at 4°C) and appears to trap E proteins on the virion surface in a conformation distinct from the herringbone icosahedral arrangement of the mature virion [46].
Overall, prior studies implicating a kinetic component of antibody-mediated neutralization have focused on either a single MAb (DENV MAb 1A1D-2, influenza MAb Y8-10C2) or a panel of MAbs that recognizes a similar epitope (MAbs specific for the membrane-proximal external region (MPER) of HIV gp41), with the implication that such antibodies were atypical in their ability to bind buried or inaccessible epitopes [57], [58], [59]. In this study we provide functional evidence identifying the widespread impact of the dynamic movement of flaviviruses on neutralization by all antibodies that bind the E protein; a kinetic aspect of neutralization appears to be the rule rather than the exception.
Our data suggest that the structural dynamics of virions has the potential to modulate the potency of all antibodies that bind E proteins arrayed on the surface of the virion through changes in epitope accessibility. This is illustrated most dramatically by antibodies that bind poorly accessible determinants on the mature virion, such as the WNV DII-fusion loop-reactive MAb E53. Neutralization of mature WNV by E53 is restricted by the number of times the antibody can bind the virion. E53 binds the E protein only when associated with prM as heterotrimeric spikes that project off the surface of the immature virus particle [21], [52]. While homogeneous populations of mature WNV are not efficiently neutralized by E53 when assayed using conventional approaches, increasing the time the virus is incubated with antibody resulted in a marked increase in neutralization activity (>100-fold) (Figure 4a and b). Because the virus cannot revert to an immature configuration once prM cleavage occurs and the virion is released from the cell [15], this dramatic increase in neutralization can only be explained by exposure of the DII-fusion loop epitope through dynamic motion of viral E proteins.
The neutralization activity of every monoclonal (n = 20) and polyclonal (n = 5) antibody assayed in this study was enhanced by increasing the time the virion was exposed to antibody prior to the addition of target cells, including antibodies previously demonstrated to be non- or weakly-neutralizing using standard assays. This reflects the fact that as the number of accessible epitopes on the individual virion increases as a consequence of dynamic motion, the fraction of them that must be bound in order to exceed the stoichiometric threshold (percent occupancy) is reduced; neutralization can then occur at lower concentrations of antibody. However, the magnitude of this kinetic effect was not uniform among antibodies localizing to different epitopes. This may reflect differences in the number of times an antibody binds the average state of the virion relative to the threshold number of antibodies required for neutralization, as well as the rate at which additional epitopes are made available for binding (Figure S3). In contrast to the cryptic nature of the DII-fusion loop epitope recognized by E53, the potently neutralizing MAb E16 is specific for a relatively accessible determinant displayed on the lateral ridge of DIII [53]. Cryo-electron microscopy studies indicate this antibody binds 120 out of 180 E proteins incorporated into mature virions; the remaining 60 E proteins proximal to the five-fold symmetry axis of the virion cannot be bound due to steric conflicts among the tightly clustered DIII epitopes [27], [28]. An increase in the accessibility of these additional epitopes on the virion through dynamic motion would translate into a modest reduction in the occupancy requirements for neutralization by E16, in agreement with the 4.0-fold increase in antibody potency (n = 11, range 2.3–6.8) observed in our studies after ∼24 hours incubation. A similar 3.8-fold increase after ∼24 hours was observed when E16 Fab fragments were used (n = 2, range 3.2–4.5).
The dynamic motion of virions has the potential to increase antibody potency by providing access to otherwise cryptic antibody-binding determinants. Of note, mapping studies suggest that many epitopes on the mature virion are poorly accessible for antibody binding [38], [39], [40], [46]. Antibodies do not induce viral breathing, but rather stabilize conformations of the E protein that exist as part of the ensemble of conformations sampled by the virion at equilibrium. The longer the virion remains exposed to antibody, the more opportunities exist for engagement of an otherwise inaccessible epitope, allowing for time-dependent increases in the stoichiometry with which antibodies decorate virions. If changes in epitope accessibility are the underlying mechanism of the kinetic aspects of neutralization, there should be a limit to the increase in potency observed over time. Eventually, dynamic virion structures should expose all potential epitopes, and these will become fixed in place by antibody binding, yielding a neutralization profile determined by the relationship between antibody occupancy and the stoichiometric threshold. In support of this, increases in neutralization and changes in the ADE curves for E16 no longer occurred when incubations longer than 24 hours were performed (Figure S5).
In addition to exposing more epitopes for antibody binding, time-dependent changes in the antigenic surface of the virus particle may also allow engagement of the virion with increased affinity, via bivalent interactions among E proteins in conformations not present in the average state, as well as cooperative effects during antibody binding. That the kinetic impact on neutralization by E16 was observed using both intact antibodies and Fab fragments incapable of cross-linking virions indicates that increases in antibody potency do not reflect antibody-mediated aggregation among virions. Importantly, all of the experiments included in our study were performed using conditions of antibody excess, and yielded results that were independent of the concentration of virus in the assay. In contrast, the aggregation of antigens by antibodies is dependent on the antibody-antigen ratio.
Our results suggest that changes in the antibody-mediated neutralization of DENV occur more rapidly than with WNV. One interpretation of this result is that DENV virions are more dynamic than those of WNV, allowing more rapid access to otherwise inaccessible determinants. In the absence of antibody, preparations of DENV become less infectious at a faster rate than observed with WNV, consistent with prior studies [49]. Additionally, kinetic changes in neutralization with the cross-reactive MAb E60 occur at a faster rate with DENV-1 than WNV RVPs when compared in parallel studies (Figure S4). While we do not yet understand, in molecular terms, why viruses lose infectivity over time, one possibility is that the intrinsic decay of flaviviruses is a consequence of structural dynamics. Viruses sampling multiple conformations in dynamic equilibrium may not always return to the average state because moving backwards may no longer be the most energetically favorable path. Additional evidence of time-dependent structural changes to the virus population is demonstrated by differences in the intrinsic decay rate of WNV observed when different cell types are used to measure infectivity. The rate of decay of viruses was ∼2.7-fold more rapid when assayed on Raji-DC-SIGNR cells as compared to a K562 cell line expressing the same attachment protein (n = 6, p<0.0001). Thus, the observed intrinsic decay cannot be attributed solely to the physical destruction of the virus, and suggests the additional possibility that not all conformations in a heterogeneous ensemble of virions are equally infectious on different cell types. Of interest, E proteins on individual virions in conformations that may no longer contribute functionally to fusion may also stably expose a different array of epitopes.
Our data suggest that the circumstances of antibody-virion interactions may significantly impact the fate of the virion immune complex. Standard in vitro neutralization assays for flaviviruses generally include a short pre-incubation (∼1 hr) of antibody and virus prior to infection of target cells; this incubation presumably allows the binding reaction between antibody and cognate epitope to reach steady-state. However, depending on the extent to which a virion is structurally dynamic (which controls the rate at which epitopes may become transiently accessible), the target cell type, and the volume of infection in vitro, this presumption may be inaccurate. Because increases in the neutralization activity of DENV-reactive antibodies that bind dynamically exposed epitopes occur rapidly (within two hours) (Figure 7), the interaction of DENV with antibodies may never truly reach steady-state. From this perspective, the length of time antibody is incubated with DENV is a variable that cannot be ignored. While antibody-mediated neutralization activity measured in vitro using standard plaque reduction neutralization tests (PRNT) generally correlates with protection in vivo [39], [53], [54], [60], [61], [62], this is an imperfect relationship. Antibodies with limited neutralization activity have been shown to protect in animal models of flavivirus infection [39], [54], [62]. While this may reflect the direct contributions of effector functions of antibodies in vivo, it is also possible that existing assays of the functional properties of antibodies have limitations. Considering the contribution of the structural dynamics of the virion when designing neutralization studies merits a systematic evaluation.
The impact of the dynamic exposure of viral epitopes in vivo remains uncertain. Virtually nothing is known about the relevant concentrations and volumes that govern antibody-virion interactions in the tissues where many of the key events in the pathogenesis of these viruses occur. Kinetic changes in neutralization occur gradually over time as dynamic motion provides new opportunities for engagement of virions with a stoichiometry sufficient for neutralization (Figure S3). Thus the rate of virus entry in vivo is also an important, yet unknown, parameter that defines the extent to which this phenomenon will contribute to protection of the host. Of interest, the kinetics of WNV binding to target cells in vitro occurred rather slowly (with maximal binding requiring ∼3 hours) even in the presence of the high affinity attachment factor DC-SIGNR (Figure S6). As DENV appears to be extremely dynamic, with kinetic effects on neutralization observed almost immediately, the impact of viral breathing on neutralization in vivo cannot be discounted. Because the kinetics of neutralization are increased by an elevated temperature, it is interesting to speculate that certain classes of antibodies, such as those recognizing the fusion loop epitope commonly observed in infected individuals, may function better than previously anticipated in the context of the febrile response. Resolving this question awaits the development of approaches to quantitatively and directly measure antibody-mediated neutralization in vivo.
Neither proteins, nor intact virions, are static structures. Our findings are consistent with a model in which the dynamic motion of flaviviruses provides an opportunity for antibodies to engage virions at otherwise inaccessible epitopes to reach a stoichiometry sufficient for neutralization. Given time, all of the E protein-reactive antibodies investigated were able to block virus infection, even those described originally as non-neutralizing using conventional assays [39]. These results add to the complexity of our understanding of the functional properties of antibodies and suggest new avenues of investigation and analysis into the widespread and unappreciated impact of the dynamic motion of virions as moving targets for antibody recognition.
Cell lines were maintained at 37°C in the presence of 7% CO2. HEK-293T cells were grown in complete Dulbecco's modified Eagle medium (DMEM) containing Glutamax (Invitrogen, Carlsbad, CA) and supplemented with 10% fetal bovine serum (FBS; (HyClone, Logan, UT)) and 100 U/ml penicillin-streptomycin (PS; (Invitrogen, Carlsbad, CA)). K562 and Raji cell lines were maintained in RPMI-1640 medium containing Glutamax (Invitrogen, Carlsbad, CA) and supplemented with 10% FBS and 100 U/ml PS.
Neutralization studies were performed using sera obtained from recipients of a candidate WNV vaccine. Sera from five participants of a Phase I double-blinded, placebo-controlled study designed to evaluate the safety and immunogenicity of a single dose of live attenuated WNV/DENV-4 vaccine were obtained and characterized previously [21]. The clinical study was conducted at the Center for Immunization at the Johns Hopkins Bloomberg School of Public Health under an investigational new drug application reviewed by the United States Food and Drug Administration.
Reporter virus particles (RVPs) are pseudo-infectious virions produced by genetic complementation of a WNV sub-genomic replicon with the structural genes in trans. Flavivirus replicons do not encode intact genes for the three structural proteins of the virus, thus RVPs that encapsidate these sub-genomic RNAs are capable of only a single round of infection. WNV and DENV RVPs have been used extensively to characterize the functional properties of anti-flavivirus antibodies [21], [38], [39], [42], [63], [64].
RVPs were produced by transfection of HEK-293T cells with DNA plasmids encoding the structural genes and a sub-genomic WNV replicon as described previously [38], [65]. Standard preparations of WNV RVPs were produced using plasmids encoding a GFP-expressing replicon, WNV C-prM-E, and pcDNA3.1 with a 1∶3∶0.5 ratio by mass. Because standard preparations of WNV RVPs retain detectable amounts of uncleaved prM, homogeneous populations of mature WNV RVPs were produced using the plasmids above except that a plasmid encoding human furin was used in place of pcDNA3.1 to promote efficient cleavage of prM [21]. Over-expression of furin in this context has been shown to reduce the amount of uncleaved prM protein in populations of RVPs to levels that are no longer detectable by Western blot [21], [66]. Standard RVPs composed of the structural genes of the DENV Western Pacific strain (serotype I: genotype IV) were produced as described above by substituting a plasmid encoding DENV structural genes in place of the WNV construct [49]. All transfections were performed using Lipofectamine LTX (Invitrogen, Carlsbad, CA) in accordance with the manufacturer's instructions. RVPs were harvested at 48 h post-transfection, filtered through a 0.22 µM filter, and frozen in aliquots at −80°C.
Because the relationship between infected cells and virus dose is typically not linear for flaviviruses, the titers of all RVP stocks were measured by serial dilution. Serial two-fold dilutions of RVP-containing supernatants were used to infect Raji cells that express the attachment factor DC-SIGNR. Infection was measured 48 h post-infection by flow cytometry. Only data obtained from the linear portion of the resulting virus dose-response curves was used for analysis.
The genomic RNA content of RVP populations was measured using a modification of a previously described protocol [68]. Briefly, RVP containing supernatants were treated with 100 U recombinant DNase I, followed by RNA isolation using the QiaAmp Viral RNA Mini kit per the manufacturer's instructions (Qiagen, Valencia, CA). Amplification of viral genomic RNA was accomplished using the Superscript III One-Step RT-PCR system (Invitrogen, Carlsbad, CA) and primers specific for the 3′ untranslated region of the WNV lineage II replicon [69].
Standard preparations of WNV RVPs were incubated with Raji-DC-SIGNR cells at 37°C in the absence or presence of the MAb E16 for incremental lengths of time to allow virus-cell binding to occur. Target cells were incubated in the presence of 20 mM NH4Cl to prevent virus fusion and genomic RNA replication. At the indicated times, cells were washed extensively with media and lysed. Total RNA was isolated using the QIAshredder and RNeasy Mini kits in accordance with the manufacturer's instructions (Qiagen, Valencia, CA). The relative amount of bound WNV RNA was enumerated by quantitative real-time PCR using a DNA molecular clone plasmid as a standard [49], [68]. The resulting kinetic data was fit to a one-phase association model using GraphPad Prism (GraphPad Software, La Jolla, CA).
|
10.1371/journal.pbio.1000572 | Laminar Analysis of Excitatory Local Circuits in Vibrissal Motor and Sensory Cortical Areas | Rodents move their whiskers to locate and identify objects. Cortical areas involved in vibrissal somatosensation and sensorimotor integration include the vibrissal area of the primary motor cortex (vM1), primary somatosensory cortex (vS1; barrel cortex), and secondary somatosensory cortex (S2). We mapped local excitatory pathways in each area across all cortical layers using glutamate uncaging and laser scanning photostimulation. We analyzed these maps to derive laminar connectivity matrices describing the average strengths of pathways between individual neurons in different layers and between entire cortical layers. In vM1, the strongest projection was L2/3→L5. In vS1, strong projections were L2/3→L5 and L4→L3. L6 input and output were weak in both areas. In S2, L2/3→L5 exceeded the strength of the ascending L4→L3 projection, and local input to L6 was prominent. The most conserved pathways were L2/3→L5, and the most variable were L4→L2/3 and pathways involving L6. Local excitatory circuits in different cortical areas are organized around a prominent descending pathway from L2/3→L5, suggesting that sensory cortices are elaborations on a basic motor cortex-like plan.
| The neocortex of the mammalian brain is divided into different regions that serve specific functions. These include sensory areas for vision, hearing, and touch, and motor areas for directing aspects of movement. However, the similarities and differences in local circuit organization between these areas are not well understood. The cortex is a layered structure numbered in an outside-in fashion, such that layer 1 is closest to the cortical surface and layer 6 is deepest. Each layer harbors distinct cell types. The precise circuit wiring within and between these layers allows for specific functions performed by particular cortical regions. To directly compare circuits from distinct cortical areas, we combined optical and electrophysiological tools to map connections between layers in different brain regions. We examined three regions of mouse neocortex that are involved in active whisker sensation: vibrissal motor cortex (vM1), primary somatosensory cortex (vS1), and secondary somatosensory cortex (S2). Our results demonstrate that excitatory connections from layer 2/3 to layer 5 are prominent in all three regions. In contrast, strong ascending pathways from middle layers (layer 4) to superficial ones (layer 3) and local inputs to layer 6 were prominent only in the two sensory cortical areas. These results indicate that cortical circuits employ regional specializations when processing motor versus sensory information. Moreover, our data suggest that sensory cortices are elaborations on a basic motor cortical plan involving layer 2/3 to layer 5 pathways.
| Sensation in the rodent vibrissal system relies on active whisking for interactions with the environment [1],[2]. Motor circuits control whisker movement, while sensory afferents collect information about contact with objects. Interactions between motor and sensory systems are necessary for object localization and identification [3]–[5].
Ascending sensory and descending motor pathways interact at multiple levels including the brainstem [6], thalamus [7], and cortex [8]. Three areas in the cerebral cortex are activated by whisker stimulation. Primary somatosensory cortex (vS1) responds with short latencies [9], whereas secondary somatosensory cortex (S2) and vibrissal motor cortex (vM1) respond 10–20 ms later [10]. These areas are also strongly interconnected in a bidirectional manner [8],[11].
In rodents, some of the cytoarchitectonic features of vM1, vS1, and S2 are area-specific, such as the presence of “barrels” in layer (L) 4 of vS1, and others are not, such as the presence of most cortical layers, including L1, L2/3, L5A, L5B, and L6 [12]. Here, to explore the synaptic organization of cortical circuits in these three areas, we used glutamate uncaging and laser scanning photostimulation (LSPS) to map the local sources of excitatory synaptic input to individual excitatory neurons in vM1, vS1, and S2. We recorded from postsynaptic neurons distributed across L2–6 (i.e., all the cortical layers that contain excitatory neurons) and, for each one, stimulated presynaptic neurons also distributed across L2–6. The collection of synaptic input maps for each area was analyzed to extract a laminar connectivity matrix representing the local pathways between excitatory neurons in each area [13],[14]. These connectivity matrices provide a quantitative survey of the interlaminar organization of local excitatory networks in each of these three cortical areas.
We identified vibrissal motor cortex (vM1), primary somatosensory (barrel) cortex (vS1), and secondary somatosensory cortex (S2) based on anatomical coordinates, cytoarchitectonic features, anatomical labeling experiments, and in the case of vM1, optical microstimulation mapping. vM1 (Figure 1A) was located in the posteromedial part of frontal agranular cortex, anteromedial to the barrel cortex [10],[15]–[17]. When anterograde tracers were injected into vM1, fluorescently labeled axons were observed in brainstem nuclei involved in whisker motor control (Figure S1) [18]. Furthermore, microstimulation mapping using channelrhodopsin-2 (ChR2) revealed that vM1 had the lowest thresholds for whisker movements (14) [19],[20]. vS1 (Figure 1B) was identified by the presence of cytoarchitectonic “barrels” in L4 [21]. S2 (Figure 1C) was located in dysgranular cortex, lateral to the barrel cortex [8],[10],[22]. Axons projected from vS1 to S2, and from S2 to vM1 and vS1 (Figure S1). These experiments enabled us to target our mapping experiments to specific cortical locations corresponding to vM1, vS1, and S2.
We prepared coronal brain slices containing vM1, vS1, or S2 (Figure 1A–C) and used LSPS with glutamate uncaging [23]–[25] to map excitatory inputs to excitatory neurons (Figure 1D–G). We excited small clusters of neurons at each site in an array of locations while recording from individual excitatory neurons (Figure 1D,E), obtaining maps of local intracortical sources of excitatory input (Figure 1F,G).
To calibrate LSPS, we recorded in cell-attached mode from excitatory neurons, while uncaging glutamate on a grid around the cell (Figure 2A,B). The spatial distribution of action potentials (APs) evoked by uncaging (the “excitation profile”) provides a measure of the effective spatial resolution of photostimulation (Figure 2A,B). These data were used to estimate neuronal photoexcitability (Figure 2C) and the spatial resolution of LSPS (Figure 2D) for photostimulating neurons in different cortical layers and areas. Photostimulation-evoked APs always occurred in perisomatic regions (Figure 2B, Figure S2) with short latencies, and almost always as singlets. Stimulation of strong synaptic pathways, such as L4→L3 in vS1, did not cause APs in the target location (Figure S2), indicating that synaptic activity did not cause APs in neurons that were not directly photostimulated. Ultraviolet (UV) attenuation in scattering tissue causes photoexcitation to decline as a function of depth in the slice; consistent with this, excitation was not observed for neurons deeper than 100 µm (Figure S3) [26]. The total number of neurons excited per stimulus, estimated from the excitation profiles and measured densities of neurons (Figure 2 and Figure S4), was in the range 50–200, consistent with previous results [13],[26]. Only a small fraction of these neurons were synaptically connected to the recorded postsynaptic neuron [27].
An input map represents the aggregate functional synaptic connectivity between small clusters of presynaptic excitatory neurons at the stimulus locations and individual postsynaptic neurons. Pixels in input maps do not represent the strengths of unitary connections; rather, they measure average monosynaptic excitatory responses to a single uncaging event (see Text S1, Equations 1–4) [26]:(1)where ρcell is the neuronal density at the point of uncaging (neurons/µm3), Vexc is the volume of excited neurons (µm3), and SAP is number of APs fired per presynaptic neuron (AP/neuron). The average strength of a synaptic connection (qcon) is calculated from equation (1). The collection of qcon for different neuronal populations defined by laminar location is the basis of connectivity matrices. We first present the mapping data for each area in the more familiar form of average input maps. In subsequent sections we summarize connectivity in laminar connectivity matrices, which take into account the parameters in equation (1).
Unlike vS1, vM1 lacks a distinct granular L4. The superficial layers L2/3 and L5A are compressed, and deeper layers L5B and L6 are expanded, consistent with vM1's location at the crest of a cortical convexity [28]. In addition, L1 was thicker than in the other areas (Table 1). Both superficial and deep L5 neurons had dense basal dendrites and a single apical dendrite extending to L1, and L6 neurons had apical dendrites that did not extend to L1; in some cases, these were inverted pyramids (Figure 3A).
We recorded from 95 excitatory neurons located in all layers (i.e., from upper L2 to lower L6) and mapped the local sources of excitatory synaptic input with LSPS using a stimulus grid that spanned vM1 (Figure 3B; Figure S5). We pooled neurons into groups by dividing the cortex into 10 equal distance bins; the top-most bin was empty, because L1 lacks excitatory neurons. We averaged the maps in each bin (Figure 3C). The strongest pathway was a descending projection, L2/3→ upper L5. Weaker ascending projections, within L5 and L5A→L2/3, were also found (Figure 3C). On average, neurons in the lower one-third (0.7–1.0) of vM1 showed weak inputs. However, individual neurons in this deeper range received strong inputs, but these tended to be spatially dispersed and sparse (Figure S5).
We recorded from 80 excitatory neurons in vS1, using a different stimulus grid matched to the cortical thickness (Figure 4; Figure S6). In vS1, laminar boundaries were distinct, allowing pooling of cytoarchitectonically defined groups for binning (Table 1; Figure S8). The ascending L4→L3 pathway and the descending L2/3→L5 pathway were both prominent (Figure 4; Figure S6). Similar to vM1, L6 neurons had relatively weak inputs (mainly from L4). L4 neurons also showed little intracortical interlaminar input [29]. In addition, we further distinguished sub-layers within L2/3 and L5B based on patterns of connectivity observed in the input maps. For example, L2 constituted a narrow superficial layer of neurons lacking strong input from L4, but with input from L5A [30]. Binning with a simple three-layer scheme (‘supragranular-granular-infragranular’; Figure 4) conveyed the main feedforward local excitatory connections in vS1.
S2 abuts the lateral edge of vS1, where the barrel pattern terminates (Figure S1). The cytoarchitectonic layers appeared similar in S2 and vS1, except that the cortex was thinner and L5A thicker. L4 included neurons with a sparse apical dendrite, and neurons lacking an apical dendrite (Figure 5A). L5 neurons had many basal dendrites and an apical dendrite that ramified in L1; L6 neurons' apical dendrites did not extend above L4.
We recorded input maps for 100 excitatory neurons in S2 (Figure 5B,C; Figure S7). Similar to vS1, an ascending pathway to more superficial layers (L4→L3) was present but was not the strongest projection. Instead, the descending projection L2/3→L5 was predominant. L5 also received substantial ascending input from L6.
Connectivity matrices represent local circuits in a compact manner [13],[14],[27],[31],[32]. Each element (i, j) in the matrix (Wi,j) corresponds to the strength of a connection (qcon; Equation 1) from the jth presynaptic location (along the rows) to the ith postsynaptic location (along the column). Distance is measured in normalized units along the radial directions (pia, 0; white matter, 1). Because of the curvature of vM1 at the cortical flexure (Figure 6A,B), we converted map data from the coordinates of the slice image (x, y) to coordinates corresponding to an unfolded cortex (h, r), where h is the horizontal distance along the laminar contour and r is the distance along the radial axis. Figure 6 provides a graphical illustration of the process of converting the pixels in an input map from x-y coordinates (Figure 6A), using a spatial transform defined on the basis of the radial structure of the cortex (Figure 6B), into r-h coordinates (Figure 6C,D).
This approach allowed us furthermore to convert input maps into vectors, by averaging input across the horizontal dimension (h) at a given presynaptic radial distance (r) into bins (Figure 6E–G; a similar analysis in the horizontal dimension is given in Figure S9). This is identical to averaging along the rows of input maps, except that it takes into account the curvature of the cortex. One neuron's input vector (Figure 6G) thus represents the inputs to one neuron from different laminar locations; i.e., the horizontal dimension has been collapsed. Each neuron was also assigned a postsynaptic radial distance. This allowed us to group all the input vectors and then sort them by the postsynaptic neuron's depth in the cortex (Figure 6H). Stacking the vectors on top of each other, sorted by depth, provided a raw connectivity matrix, Wraw(rpost, rpre), describing connectivity between neurons at different locations along the radial axis (Figure 6H, Figure 7A). The rows in such a connectivity matrix represent synaptic input to a particular laminar location, and the columns represent synaptic output from that laminar location. Intralaminar connections lie along the main diagonal. We note that intralaminar connectivity was undersampled because of direct excitation of the postsynaptic neurons' dendrites.
In addition to deriving matrices based on the collections of input vectors (Figure 7A,D,G), we further analyzed the data in terms of the excitation parameters given in equation (1). To compute the average connectivity matrix at the level of individual neurons (Wneuron), we binned the data and applied correction factors to derive the strength of input per presynaptic neuron per AP. We divided the connection strength in the raw connectivity matrix by the mean number of APs per uncaging event at the presynaptic region (Figure 7B,E,H; Figure S10; Text S1) and the number of presynaptic neurons stimulated. The number of stimulated neurons was obtained from measurements of ρcell (Figure S4) and Vexc. To compute the connectivity matrix at the level of cortical layers (Wlayer) we multiplied the neuron→neuron connections by the number of presynaptic and postsynaptic neurons per layer (Figure 7C,F,I; a detailed calculation is illustrated in Figures S11 and S12). Values for all connectivity matrices are provided in Table S1 and Dataset S1.
We used glutamate uncaging and LSPS to map local synaptic connections among excitatory neurons in mouse vM1, vS1, and S2, three cortical areas centrally involved in vibrissa-based somatosensation. From single cell input maps recorded at different cortical depths, we derived connectivity matrices that compactly describe the local network. Our main findings were that vM1 contains a strong pathway from L2/3 to upper L5; that vS1 and S2 contain two strong pathways, corresponding to L4→L3 and L2/3→L5; and that S2 contains these plus pathways between L6 and L5B.
The connectivity matrix description allows us to directly contrast local circuits in different cortical regions. The elements (pixels) in the neuron→neuron connectivity matrices, Wneuron (Figure 7B,E,H), represent the mean strength of postsynaptic response in a single neuron extrapolated to a single presynaptic AP in a single cell of the indicated layer (qcon). Pixel values were 10–100 times lower than typical unitary EPSCs, reflecting both the generally low probability of connections between excitatory neurons in cortical circuits (typically 0.1–0.2) [27],[33]–[35], and the fact that the current amplitude in the maps represents a mean over 50 ms rather than the peak of the EPSC.
In contrast, the elements in the layer→layer connectivity matrices, Wlayer (Figure 7C,F,I), represent the average strength of connections extrapolated to the entire projection from one layer to another. The Wlayer matrices differ from the Wneuron matrices in that they enhance thicker and more neuron-dense layers and diminish thinner and less neuron-dense layers. For example, because in vS1 the L5A is thin (Table 1) and both L5A and L5B are low in neuronal density (Figure S4), the projections to and from L5, such as L5A→L2/3 and L2/3→L5B, are relatively strong at the level of neuron→neuron connectivity (Figure 7E) but relatively weak at the level of layer→layer connectivity (Figure 7F). Interestingly, in rat vS1 the L4→L2/3 projection is functionally weak compared to the structural density of L4 axons and L2/3 dendrites, while the converse holds for the L5A→L2/3 projection [36]. Our results here show how weak neuron→neuron connections may be strong in aggregate at the layer→layer level. Further structure-function analyses will be required to determine whether it is generally the case that larger and more neuron-dense layers have weaker neuron→neuron but stronger layer→layer projections.
The connectivity matrix representations of vM1 show strong descending projections from L2/3→upper L5 (Figure 7A–C), similar to the forelimb area of mouse M1 [13],[14],[37]. This input straddled the L5A/B border. L5B received an additional hotspot from itself, which appeared strong when considered as an entire layer (Figure 7C). The deepest one-third of vM1 (consisting mostly of L6) had weak inputs and outputs.
The vS1 excitatory circuits were more complex (Figure 7D–F). The major ascending pathway from L4→L3 was paralleled by an ascending component from L5A. The high cell density in L4 made the L4→L3 connection prominent in the laminar analysis (Figure 7F). Another prominent projection was from L2 and L3 to L5A and L5B; inputs originating in more superficial regions of L2/3 targeted relatively more superficial regions of L5A/L5B (note the diagonal shape of the L2/3→L5 hotspot in Figure 7E). On a neuron→neuron basis, the L3→L5B connection was stronger than L4→L3, although the layer→layer analysis showed a reduction in cell density relative to L4. L2 received input from L3, and weaker input from L5A. However, L2 was thin and thus contributed little to Wlayer. As in vM1, deep layers had weak inputs and outputs.
In S2 (Figure 7G–I), an ascending L4→L2/3 pathway and descending L3→L5 pathway were present. Neurons on the L5A/L5B border also showed strong intralaminar connections. The L6 output evident in the input maps (Figure 5; Figure S7) also supplied potent input to L5B. Although not as strong at the single cell level, the entire L6 excited L5B as much as L3 (Figure 7I). L6 was enhanced in S2 relative to other regions as both a source of synaptic output and a recipient of synaptic input, due to the relatively high density of neurons (Figure S4) and their relatively low photoexcitability (Figure 2C–E). The functional connectivity in the local excitatory circuits of all three regions is simplified into quantitative laminar wiring diagrams (Figure 8).
LSPS with glutamate uncaging simultaneously excites a group of presynaptic neurons, while the postsynaptic response is measured. To derive average connection strength per neuron (qcon), the number of excited neurons needs to be estimated, based on the excitability (SAP), neuron density (ρcell), and excitation volume (Vexc) at the uncaging location (Equation 1). The accuracy of the estimate of qcon is limited by our measurement of ρcell and neuronal excitation (SAP, Vexc; Text S1 Equations 3–4): Measurements of neuronal density vary by a factor of two [27],[38],[39]. Although excitation profiles give a direct measure of evoked APs in brain slices under the relevant recording conditions (Figure 2), excitation varies across neurons and somewhat across cortical areas, and decreases with depth in the slice; these effects together introduce uncertainty roughly on the order of a factor of two (Figure S3). Despite these uncertainties, our estimates of qcon are broadly consistent with those derived from pair recordings (Figure S13).
Because LSPS excites many neurons, this strong stimulus allows weak pathways to be detected. However, the average connection strength, qcon, reflects both the connection probability and unitary connection strength:(2)It is therefore not possible to separate connection probability and unitary connection strength directly. Furthermore, pcon is inversely related to the horizontal separation between cell pairs [34]. LSPS averages inputs from a range of presynaptic locations with varied horizontal offset. For each cell class, a broad distribution of pcon values contributes to LSPS maps.
In addition, by computing the average connection strength, we average out the underlying distribution of unitary connection strength, which is a skewed distribution of numerous weak and a few strong connections [27],[33]. This inherent averaging also makes LSPS insensitive to certain non-laminar aspects of cell-type specificity in cortical connectivity [14],[33],[35],[40]–[43].
Comparison of our neuron→neuron connectivity matrix with a pair-recording study [27] reveals qualitative similarities (Figure S13). After both methods are corrected to similar units (peak amplitude in pA/AP), the general shape of the connectivity matrix and values for neuron→neuron connectivity are similar. The major interlaminar pathways are L2/3→L5 and L4/5A→L2/3. However, local intralaminar connections are underestimated in our data set due to direct responses to uncaging. Furthermore, descending projections from L4→L5A and from L5A→L5B may be underestimated in LSPS relative to pair recording due to exclusion of direct responses along the apical dendrite of the postsynaptic neuron (see L5A and L5B maps in Figure S6). Under-sampling of connected pairs in low-pcon pathways, such as L4→L6, may account for differences from LSPS, where many L4 neurons are excited during each L6 recording. Lastly, L2 connectivity differs in part because of differences in the definition of this layer.
We compared the matrices for the four areas so far studied, vM1 (present study), the forelimb region of somatic M1 [13], vS1 (also the present study) [27], and S2 (present study). Overall, the main differences are attributable to the presence of a distinct granular layer in somatosensory cortex. Specifically in vS1, L4 outflow contributed strongly to the connectivity matrix. L4→L2/3 is also a major pathway in rodent V1 [44]. In S2, the local excitatory circuit differs from vS1 most prominently in that the L4→L3 pathway is reduced. LSPS analyses of auditory cortex circuits have found L4→L2/3 inputs [45],[46], which is adjacent to S2. However, ascending pathways were not unique to vS1, as a similar but weaker L3/5A→L2/3 pathway was prominent in forelimb M1, and present but weaker still in vM1 (Figure 3C and Figure S5C, leftmost panels). The upward compression of layers in vM1, typical of cortical convexities [28], may be why L3/5A→L2 was less distinct in vM1 than in forelimb M1 (e.g., it was more prone to masking by dendritic responses of L2 neurons). However, inspection of individual maps and traces (Figure S5C) showed that these ascending pathways were present for some L2 neurons.
A second main interlaminar hotspot in vS1 was the descending pathway(s) L2/3→L5, which was the predominant hotspot in the two motor areas. We noted that this pathway was present in all three cortical regions studied here and was similarly prominent in somatic M1 [13]. Indeed, it was the predominant pathway in S2. Thus, a strong supragranular to infragranular descending connection emerged as a common element of local cortical circuits examined here. Superficial L5B neurons and deep L5A neurons at the laminar border were most strongly activated, suggesting that the cytoarchitectonic boundaries identified do not correspond well with functional gradient within L5. Perhaps an alternative molecular marker, such as Etv1 (Figure S8), better denotes this functional division.
In three of the four areas, L6 neither received nor sent strong projections (but vS1 neurons in L6 received a weak projection from L4). L6 output is provided by an ascending connection to L4 in cat visual cortex [31],[32],[47] but was absent or reduced in all vibrissal areas we studied. L6→L4 projections studied in mouse somatosensory and auditory cortical areas have “modulator” rather than “driver” properties, including paired pulse facilitation [48]. Although deeper neurons tend to have relatively small dendritic arbors [49], which may account for a reduction (but not absence) of inputs, this difference in arbor size is not of sufficient magnitude to account for the paucity of inputs. Similarly, the paucity of outputs was not due to lack of photoexcitability of these neurons. Channelrhodopsin-assisted circuit mapping experiments [50] have shown that the supragranular layers indeed connect preferentially to upper rather than lower infragranular neurons. Thus, the lack of inputs was not due simply to slice-related artifacts such as severing of pathways. Consistent with weak local inputs, in vivo recordings in cat motor cortex suggest that a large number of L6 neurons are virtually silent, even during motor activity [51]. Thus, the sources and modes of excitation for L6 neurons remain to be determined [49],[52]. However, L6 was more engaged in local circuits in S2, supplying a measurable output to L5A and L5B and to other L6 neurons. In addition to input from L5B, L6 neurons in S2 collected inputs from a wide horizontal distance, sometimes >300 µm (Figure S7B,C at right). Thus, S2 may be better suited for studying L6 function.
One major difficulty in making a comparison of connectivity between two cortical areas is selecting the laminar position of pre- and postsynaptic neurons for the comparison. Is it better to compare identical relative laminar depths between cortical areas, not accounting for the decreased thickness of superficial layers, and increased thickness of deep layers, in motor areas? How shall we treat the presence or apparent absence of a distinct layer 4? We present a direct quantitative comparison of three major areas identified in our study, based on cytoarchitectonic laminar divisions (Table 1, Figure S8) (Figure 9). In vS1 [53] and vM1 (Tianyi Mao, BMH, GMGS, KS, unpublished observations) these layers correspond to distinct cell types with different projection patterns. The descending projection from L2/3→L5A/B was prominent in all areas, but the strength of the pathways at a neuron→neuron level varied by a factor of four between the areas. Ascending projections from middle layers to superficial ones (L4→L2/3 in vS1 and S2; L5A→L2/3 in vM1 for comparison) were also present in all regions but were the least prominent in agranular vM1. Lastly, the L6→L5 projection identified in S2 was more than twice as strong at the neuron→neuron level than in vS1 (and the difference was greater with vM1). Our approach provides a defined framework for measuring similarities and differences between cortical microcircuits in a quantitative manner.
We use the term radial to refer to the axis defined by the apical dendrites of pyramidal neurons; this axis is approximately normal to the cortical surface. Normalized radial distance is along the radial axis, bounded by the pia and the white matter, where pia = 0 and the L6/white matter border = 1. Vertical is synonymous with radial. Horizontal, or lateral, refers to planes normal to the radial axis, approximately parallel to layers, or laminae (Figure 6). Oblique refers to off-axis interlaminar connections.
Mice were decapitated at postnatal day 20–25 under isofluorane anesthesia, and the brain rapidly placed in ice cold choline solution (in mM: 110 choline chloride, 25 NaHCO3, 25 D-glucose, 11.6 sodium ascorbate, 7 MgCl2, 3.1 sodium pyruvate, 2.5 KCl, 1.25 NaH2PO4, 0.5 CaCl2). Coronal brain slices (300 µm) were cut (Microm HM 650V), incubated 30 min at 37°C in oxygenated ACSF (in mM: 127 NaCl, 25 NaHCO3, 25 D-Glucose, 2.5 KCl, 2 CaCl2, 1.25 NaH2PO4, 1 MgCl2), and maintained in a holding chamber at 22°C for up to 5 h during recording. For vM1 slices, the brain was pitched upward ∼10° to optimize alignment with the radial axis of vM1, and slices ∼0.7–1.3 mm anterior to bregma were used; for vS1 and S2 slices, the brain was cut coronally, and slices ∼1–2 mm posterior to bregma were used (Figure 1A,B). To determine the optimal slice angle for each area, we used the appearance of the intact apical trunk at high magnification to select slices for recording and avoided those sections where the apical dendritic trunk was at an angle with respect to the slice plane. Thus, only one or two sections per animal could be used for recording. Separate experiments in our laboratory using the photostimulation methods in vS1 [42] and vM1 (unpublished data) measure input to L1 dendrites of L5 pyramidal neurons, confirming the apical trunk is intact using this slice angle. We added biocytin to visualize a subset of dendritic arbors, some of which are reconstructed in Figure 3 and Figure 5. These neurons appeared radially symmetric, with arbors ranging from 300–500 µm in diameter. Since the neurons were 50–100 µm deep in the slice, a portion of the apical and basal dendrites are truncated by slicing and the deep half of the arbor is intact.
Recordings were performed at room temperature (22°C) in ACSF. Neuronal excitability was reduced by increased divalent ions (4 mM CaCl2 and 4 mM MgCl2), and NMDA receptor blockade with 5 µM 3-((R)-2-carboxypiperazin-4-yl)-propyl-1-phosphonic acid (CPP; Tocris). Patch pipettes were fabricated from borosilicate glass with filament (4–6 MΩ). Intracellular solution contained (in mM): 128 K-gluconate, 10 HEPES, 10 sodium phosphocreatine, 4 MgCl2, 4 Na2ATP, 3 sodium L-ascorbate, 1 EGTA, and 0.4 Na2GTP (pH 7.25; 290 mOsm). To visualize dendritic arbors, 20 µM Alexa 594 or 488 (Molecular Probes) was added to the internal solution. In some cases, 2–3 mg/mL biocytin was included. Electrophysiological signals were amplified with an Axopatch 700B amplifier (Molecular Devices), filtered at 4 KHz, and digitized at 10 KHz. Data I/O were controlled by Ephus, a suite of custom Matlab-based (Mathworks) software tools available online (https://openwiki.janelia.org [54]).
Neurons were selected based on pyramidal appearance, or in the case of L4 recordings in vS1, either pyramidal or stellate appearance. In vS1, recordings were generally made in the middle of the barrel field and not a specific whisker barrel. Following patching, a family of current steps was presented to determine firing properties. Neurons with narrow APs and high firing rates were rejected for analysis as presumed interneurons.
Methods followed published procedures [13],[26]. MNI-glutamate (0.2 mM; Tocris, MO [55]) was added to a recirculating bath. Photolysis was performed by shuttering (1.0 ms pulse) the beam of an ultraviolet (355 nm) laser (DPSS Lasers, San Jose, CA), ∼20 mW in the specimen plane, set by a combination of a gradient neutral density filter wheel and Pockels cell (electro-optical modulator; Conoptics). A 16×16 standard stimulus grid for input maps had row and column spacing of 110 µm for vM1 and 90 µm for vS1 and S2 recordings. Maps were recorded in voltage clamp at −70 mV. Inhibitory input amplitude was minimized by recording near the chloride reversal potential. The 256 grid sites were visited in a sequence that optimized the spatiotemporal separation between sites [25]. The sequence was repeated 2–4 times per neuron. In vS1 and S2, the map was aligned to the top of the pia and centered on the soma. In vM1, the medial edge of the map was aligned to the interhemispheric fissure, and the top to the dorsal-most edge of pia.
To convert each map's set of traces into an array of pixels that represent response amplitudes, we calculated the average current over a 50 ms post-stimulus window. Direct dendritic responses were excluded on the basis of temporal windowing [56], rejecting traces with events (detected as >3 SD above baseline) with onset latencies of <7 ms. At locations where some maps had direct responses at a given pixel while others did not, the average of the non-direct responses was used; pixels were excluded from the average raw input map for a given neuron if all traces had direct responses.
We measured excitation profiles using loose-seal recordings with the amplifier in voltage-follower mode, to gauge the efficacy of photostimulation for neurons in the different layers in the three areas. Excitation profiles were recorded and analyzed following previous methods [13],[25],[26],[30]. To characterize the size of the excitation profile, we calculated the mean weighted distance from the soma of AP generating sites as: Σ(APs × distance from soma)/Σ(APs).
Procedures build on [13]. A transformation step was added, to account for cortical curvature, which was especially strong in vM1. As described in Results, we assigned each point in the stimulus grid a normalized radial distance and horizontal offset (Figure 6). Individual recorded neurons were also assigned a postsynaptic radial distance based on the same criteria. Individual input maps for a given neuron could then be averaged together based on postsynaptic radial distance. Furthermore, when computing the input to a given neuron for the purpose of determining the connectivity matrix, a presynaptic point would be averaged into a bin appropriate to its location. Most aspects of local connectivity were robust to changes in binning. Subsequent corrections to the connectivity matrices were made to account for variations in excitability between layers, and the number of neurons in pre- and post-synaptic layers; these were then presented as neuron→neuron connectivity matrices and layer→layer connectivity matrices (see Results and Text S1).
To make quantitative comparisons between the strength of pathways in different areas, we determined both the average strength of pathways and their variability using a bootstrap-based analysis (Figure 9). After selecting the pre- and postsynaptic neuron populations by relative laminar depth, the strength of corresponding pixels in the input map (limited to maps from neurons in the postsynaptic layer, and pixels in the presynaptic layer within 300 µm horizontal distance of the soma) were averaged for each map. We resampled the individual map averages 10,000 times with replacement and resampled other factors contributing to the individual neuron→neuron strength (number of APs from cortical area's excitation profiles and neuron density). Pathways were presented with the average strength and SD from the bootstrap analysis.
In vS1 and S2, we performed morphometric measurements of cortical landmarks in video images of brain slices. Along a radial line, we marked the locations of the soma, pia, white matter, and major laminar boundaries and calculated the absolute and normalized radial distances of these locations. The bottom extent of cortex was marked at the border between L6 (including the subplate zone) and white matter [57]. The distances to lower borders of layers (±SD) are given in Table 1. The division between L2 and L3 in vS1 was drawn between groups of neurons that did not receive appreciable L4 input (L2 [58]) and those that did. Since this functional division was not clear in S2, L2/3 was divided in half. These values are bracketed in the table. vM1 appearance was similar to somatic motor cortex, with a prominent clear zone in the upper middle part of the cortex, corresponding to L5A [13]. Thus, landmarks indicating the border between L1, a compressed L2/3, and the bottom of L5A were apparent in video images and used to measure laminar boundaries in vM1. The division between L5B and L6 was estimated as the radial distance where cell density increased (Figure S4; Table 1), as a clear border was not apparent based on image contrast. Alternative methods of determining cortical layers in motor and sensory cortex were performed on images of gene expression patterns from the Allen Brain Atlas (Figure S8).
|
10.1371/journal.pbio.1001115 | A Novel Sperm-Delivered Toxin Causes Late-Stage Embryo Lethality and Transmission Ratio Distortion in C. elegans | The evolutionary fate of an allele ordinarily depends on its contribution to host fitness. Occasionally, however, genetic elements arise that are able to gain a transmission advantage while simultaneously imposing a fitness cost on their hosts. We previously discovered one such element in C. elegans that gains a transmission advantage through a combination of paternal-effect killing and zygotic self-rescue. Here we demonstrate that this element is composed of a sperm-delivered toxin, peel-1, and an embryo-expressed antidote, zeel-1. peel-1 and zeel-1 are located adjacent to one another in the genome and co-occur in an insertion/deletion polymorphism. peel-1 encodes a novel four-pass transmembrane protein that is expressed in sperm and delivered to the embryo via specialized, sperm-specific vesicles. In the absence of zeel-1, sperm-delivered PEEL-1 causes lethal defects in muscle and epidermal tissue at the 2-fold stage of embryogenesis. zeel-1 is expressed transiently in the embryo and encodes a novel six-pass transmembrane domain fused to a domain with sequence similarity to zyg-11, a substrate-recognition subunit of an E3 ubiquitin ligase. zeel-1 appears to have arisen recently, during an expansion of the zyg-11 family, and the transmembrane domain of zeel-1 is required and partially sufficient for antidote activity. Although PEEL-1 and ZEEL-1 normally function in embryos, these proteins can act at other stages as well. When expressed ectopically in adults, PEEL-1 kills a variety of cell types, and ectopic expression of ZEEL-1 rescues these effects. Our results demonstrate that the tight physical linkage between two novel transmembrane proteins has facilitated their co-evolution into an element capable of promoting its own transmission to the detriment of organisms carrying it.
| Natural selection typically favors only those genetic variants that increase the overall fitness of the organism. Occasionally, however, variants arise that are able to increase their representation in future generations, while simultaneously reducing the fertility or fecundity of their hosts. Although such variants occur in a wide variety of taxa, their genetic bases and molecular mechanisms remain poorly understood. Here we demonstrate that one such variant in the roundworm C. elegans is composed of two adjacent genes: a sperm-delivered toxin and an embryo-expressed antidote. The toxin protein is expressed in sperm and delivered to the embryo upon fertilization. In the presence of the toxin, embryos that don't inherit the antidote gene die during late embryogenesis, whereas those that do develop normally. Both the toxin and the antidote genes encode transmembrane proteins, and both are evolutionarily novel. Our results imply that the tight physical linkage between these two novel genes has facilitated their evolution into a co-adapted gene complex capable of promoting its own transmission to the detriment of host fitness.
| The evolutionary fate of an allele ordinarily depends on the reproductive fitness of the organisms carrying it. In some cases, however, alleles are able to increase their representation in future generations while being neutral or detrimental to the fitness of their bearers. These elements, sometimes termed “selfish” or “parasitic” genes, influence transmission probabilities in a variety of ways. Some self-replicate and insert themselves into new genomic locations (e.g., transposons) [1]. Others act during meiosis to preferentially segregate into the oocyte [2]–[4] or to reduce the viability of sperm or spores inheriting alternate alleles [5]–[8]. Still others act at the level of the zygote to destroy progeny not inheriting them. Medea-factors in Tribolium destroy non-carrier animals through a combination of maternal-effect killing and zygotic self-rescue [9]. An analogous phenomenon occurs in organisms infected by the maternally transmitted bacteria, Wolbachia or Cardinium [10]–[12], which modify the sperm of infected males to cause lethal defects when paired with the oocytes of uninfected females.
Previously, we discovered a nuclear-encoded element in Caenorhabditis elegans that kills non-carrier animals in a novel way [13]. This element, referred to as the peel-1/zeel-1 element, is polymorphic within the species, and when animals carrying the peel-1/zeel-1 element are crossed to animals lacking it, the peel-1/zeel-1 element acts in the F1 heterozygote via paternal effect to kill F2 or backcross embryos not inheriting it. This lethality acts independently of maternal genotype and causes the peel-1/zeel-1 element to become over-represented among the viable progeny of heterozygous sires, even while incurring a substantial fitness cost to these animals.
Paternal-effect loci are extremely rare in all of biology [14],[15], and the observed combination of nuclear-encoded, paternal-effect killing and zygotic self-rescue is unprecedented. In C. elegans, moreover, a paternal-effect locus whose effects can be rescued zygotically is mechanistically surprising because in this species, sperm-supplied factors are thought to act only during fertilization and first cleavage [16], whereas zygotic transcription does not begin until the four-cell stage [17].
Although the peel-1/zeel-1 element is capable of promoting its own transmission, it rarely has the opportunity to do so in natural populations. C. elegans is an androdioecious species that reproduces primarily through self-fertilizing hermaphrodites. Because inbreeding is high [18], the peel-1/zeel-1 element normally exists in the homozygous state, where no opportunity for self-promotion exists.
High inbreeding notwithstanding, out-crossing events in C. elegans between hermaphrodites and males do occur, albeit rarely [18]–[20]. And because the peel-1/zeel-1 element is globally distributed and confers no apparent fitness disadvantage in the homozygous state [13], this element is expected to drive itself to fixation faster than a neutrally evolving locus. Consistent with this prediction, in laboratory populations where out-crossing is forced, the peel-1/zeel-1 element fixes rapidly [13]. In natural populations, however, the peel-1/zeel-1 element has remained polymorphic for an estimated 8 million generations [13], much longer than expected under neutrality. One likely explanation for this paradox is that the peel-1/zeel-1 element is tightly linked to a polymorphism maintained by balancing selection, and the tightness of this linkage maintains the peel-1/zeel-1 element in the polymorphic state [13].
Given the unusual genetics of the peel-1/zeel-1 element, we sought to understand its mechanism of action. We previously identified one component of the peel-1/zee1-1 element as the gene zeel-1 (Y39G10AR.5), which acts zygotically to rescue the paternal-effect killing [13]. Here we demonstrate that zeel-1 is fully separable from the paternal-effect killing, and that this killing activity is encoded by a second gene, peel-1 (Y39G10AR.25). We show that PEEL-1 acts as a sperm-supplied toxin, and ZEEL-1 an embryo-expressed antidote. We characterize the developmental defects caused by sperm-supplied PEEL-1, and we report a dose-dependent relationship between the severity of these defects and the quantity of PEEL-1 delivered to the embryo. We analyze the phylogenetic origins and functionality of each domain of zeel-1, and we test the tissue-autonomy of zeel-1 rescue. Finally, in order to determine whether peel-1 and zeel-1 can function outside of embryogenesis, we express both genes ectopically in adults.
The genetics of the peel-1/zeel-1 element are consistent with it being composed of two interacting loci: a dominant-lethal, paternal-effect “toxin,” peel-1, and a zygotically acting “antidote,” zeel-1 [13]. The activities of peel-1 and zeel-1 are present in the reference strain, Bristol (N2), and in approximately two-thirds of wild isolates [13]. These strains are said to carry the peel-1/zeel-1 element. The commonly used wild strain, collected from Hawaii (CB4856), and all but two of the additional wild strains lack the activities of both peel-1 and zeel-1 [13]. The two remaining strains, one collected from Germany (MY19) and one collected from Utah (EG4348), exhibit the activity of zeel-1 but are unable to induce the paternal-effect killing ([13], Figure S1A).
We previously mapped the peel-1/zeel-1 element in the Bristol strain to a 62 kb interval on the left arm of chromosome I [13]. Within this interval, we identified a single gene capable of providing antidote activity. This gene, which we named zeel-1, encodes a 917-amino acid protein whose N-terminus is predicted to form a six-pass transmembrane domain and whose C-terminus exhibits sequence similarity to ZYG-11, a substrate-recognition subunit of an E3 ubiquitin ligase [21]. The Hawaii strain carries a 19 kb deficiency (niDf9) spanning zeel-1, and this deficiency is shared by all other wild isolates lacking the activities of both peel-1 and zeel-1 [13]. This deficiency is not shared by wild strains carrying the peel-1/zeel-1 element, nor by MY19 or EG4348 ([13], unpublished data).
The phenotype of MY19 and EG4348 demonstrates that the zeel-1 gene is not sufficient for peel-1 activity. Conversely, a deletion allele of zeel-1 in the Bristol background demonstrates that zeel-1 is also not required for it. This deletion, tm3419, removes 221 base pairs spanning the start codon of zeel-1 (Figure 1A). As expected, this deletion abolishes antidote activity (Figure S1B); however, this deletion does not abolish the paternal-effect killing (1B). zeel-1 is therefore genetically separable from a second, paternal-effect locus, peel-1. As a consequence of this separability, zeel-1 deletions in the Bristol and Hawaii backgrounds have opposite phenotypic effects: The niDf9 deficiency is perfectly viable, whereas the tm3419 deletion behaves as a conventional recessive-lethal mutation.
In MY19 and EG4348, absence of peel-1 activity is tightly linked to the 62 kb peel-1 interval ([13], Figure S1C), suggesting that these strains carry loss-of-function alleles of peel-1, rather than extra-genic suppressors. In addition, sequence analysis of the peel-1 interval in MY19 [13] and EG4348 (see Materials and Methods) indicates that absence of peel-1 activity in these strains is not caused by recombination breaking apart the peel-1/zeel-1 element. We hypothesized, therefore, that MY19 and EG4348 carry secondary, loss-of-function mutations in peel-1 We reasoned that by identifying these mutations, we would be able to identify peel-1 itself.
To accomplish this goal, we crossed MY19 and EG4348 to a strain of the Bristol background and generated recombinant chromosomes across the peel-1 interval. Using these recombinants, we mapped the causative mutations in MY19 and EG4348 to regions of less than 10 kb (Figure 1A). We then sequenced these regions to identify all sequence changes relative to Bristol. After excluding those sequence changes shared by one or more wild strains having intact peel-1 activity (Figure 1A, Figure S2), we defined six candidate mutations in MY19 and a single candidate mutation in EG4348.
The candidate mutations in MY19 and EG4348 reside in the intergenic interval immediately downstream of zeel-1 (Figure 1A). We searched for genes in this interval using targeted RT-PCR on Bristol animals, and we discovered a previously unannotated transcript. This transcript encodes a 174-amino acid protein (Figure 1A,D), and three observations confirm this gene to be peel-1. First, the single candidate mutation in EG4348 and one of the candidate mutations in MY19 produce nonsense mutations in this transcript, consistent with the phenotype of these strains (frameshift in MY19; glycine to stop codon in EG4348) (Figure 1A,D; Figure S2). Second, this gene resides within the 19 kb deficiency carried by the Hawaii strain and in all other wild isolates lacking the activities of both peel-1 and zeel-1 (Figure 1A). Third, when we expressed the Bristol allele of this gene transgenically in a strain carrying the nonsense mutation found in EG4348, peel-1 activity was restored (Figure 1B–C).
PEEL-1 is a hydrophobic protein containing four predicted transmembrane helices (Figure 1D). Neither the peptide nor the nucleotide sequence of peel-1 has any detectable sequence similarity to any other gene in C. elegans or in the GenBank sequence database. Although peel-1 and zeel-1 are located adjacent to one another in the genome and oriented in the same direction, we were unable to recover transcripts carrying both genes (unpublished data), demonstrating that peel-1 and zeel-1 are not isoforms of a single transcript or cistrons in an operon. We conclude that the peel-1/zeel-1 element is composed of a 19 kb insertion carrying two distinct genes: peel-1, which kills embryos via paternal-effect, and zeel-1, which acts zygotically to rescue this lethality.
Henceforth we refer to the Bristol alleles of peel-1 and zeel-1 as peel-1(+) and zeel-1(+) and the Hawaii alleles as peel-1(Δ) and zeel-1(Δ). We use the term “peel-1-affected embryos” to refer to zeel-1(Δ) embryos fathered by a peel-1(+) animal.
As expected for a paternal-effect gene, peel-1 is expressed exclusively in sperm. In both males and hermaphrodites, a GFP reporter driven by the peel-1 promoter was expressed strongly in spermatocytes but not in any other tissue (Figure 2A). In fem-1(ts) mutants, which lack sperm [22], peel-1 expression via quantitative RT-PCR was undetectable in hermaphrodites and over 100-fold reduced in males (Figure 2C). Residual expression in males probably reflects incomplete penetrance of the fem-1(ts) allele, because fem-1(ts) males occasionally produce a small number of sperm [22], and another sperm-specific gene, spe-9 [23], also showed residual expression in males (Figure 2C).
The sperm-specific expression of peel-1 suggested that the paternal-effect killing occurs through delivery to the embryo of either sperm-supplied peel-1 transcripts or sperm-supplied PEEL-1 protein. To distinguish between the two, we searched for peel-1 transcripts in mature sperm via single-molecule fluorescence in situ hybridization (FISH). peel-1 transcripts were observed in spermatocytes but not in mature sperm (Figure S3), thus excluding such transcripts as the likely mediators of peel-1 toxicity.
Next, we searched for PEEL-1 protein both by tagging PEEL-1 with GFP and by staining sperm with an antibody raised against the C-terminal 15 amino acids of PEEL-1. Both experiments demonstrate that PEEL-1 protein is packaged into sperm, and this packaging is mediated by localization of PEEL-1 to sperm-specific vesicles called fibrous body-membranous organelles (FB-MOs) (Figure 3A). PEEL-1::GFP was visible in mature sperm (Figure 3E–F), and at each stage of spermatogenesis, its localization matched the pattern expected for a protein located in the membranes of FB-MOs: In early spermatocytes, PEEL-1::GFP localized to cytoplasmic puncta (Figure 3B), but after the pachytene stage, these puncta dissociated into a mesh-like web (Figure 3B–D); after the completion of spermatogenesis, PEEL-1::GFP re-condensed into puncta located at the spermatid cortex (Figure 3E); and after sperm activation, these puncta localized opposite the newly formed pseudopod (Figure 3F). This localization pattern was replicated by the anti-PEEL-1 antibody (Figure 3G), and staining with this antibody overlapped perfectly with a marker of FB-MOs (Figure 3H–I).
We also discovered that the leader peptide of PEEL-1 can act as a sperm-localization signal. Our Ppeel-1::GFP reporter, which expressed untagged GFP, showed diffuse localization in spermatocytes and was excluded from sperm (Figure 2B). This exclusion was not surprising because GFP is a heterologous protein, and trafficking of cellular components into sperm is tightly regulated [16]. Nevertheless, when we tagged GFP with the N-terminal 12 amino acids of PEEL-1 (MRFDFQNLKFSM), its localization changed dramatically. The tagged version of GFP localized to a reticulated structure within spermatocytes, and this structure was trafficked into sperm (Figure 2B). To our knowledge, this result represents the first identification of a sperm localization signal in C. elegans.
Unlike other known paternal-effect genes [14],[15], sperm-supplied PEEL-1 does not cause defects until late in development. In peel-1-affected embryos, early embryogenesis, gastrulation, epidermal enclosure, and early elongation occur normally (Figure 4C). Then, at the 2-fold stage of elongation, when all major tissues have begun differentiating and nearly all embryonic cell divisions have already occurred, the majority of peel-1-affected embryos arrest elongation and fail to begin rolling within their eggshells (Movies S1–S3). Shortly thereafter, the bulk of the embryo compresses inward, towards the mid-embryo bend, and the head and tail become flaccid and thin (Figure 4D, Movies S1–S2). Approximately 2 h later, cytoplasm begins leaking from the external epidermis, and the lumen of the excretory cell distends to form large vacuoles (Figure 4D, Movies S1–S2).
The defects observed in peel-1-affected embryos indicate severe malfunction of muscle and epidermal tissue. The phenotype of paralysis and 2-fold arrest is characteristic of a complete absence of the function of body-wall muscle [24],[25]. The shape changes observed after the 2-fold arrest indicate detachment of body-wall muscle fibers from the overlying epidermis [26]. Epidermal leakage and distention of the excretory cell, the only epidermal cell located in the interior of the animal, indicate further deterioration of both external and internal epidermis. These four defects—paralysis and 2-fold arrest, muscle-epidermal detachment, epidermal leakage, and excretory cell distention—are not known to occur as side-effects of one another [25],[27], suggesting that sperm-supplied PEEL-1 may affect each tissue independently.
Paralysis and 2-fold arrest have only two known causes: defective sarcomere assembly and lack of muscle contraction [24],[25]. To distinguish between the two, we examined (i) the localization of perlecan, a basement membrane protein required for sarcomere recruitment [25], and (ii) the structure of actin and myosin myofilaments, which assemble downstream of all other sarcomere proteins [24]. In peel-1-affected embryos, perlecan localized normally (Figure 4I–J). Likewise, actin and myosin filaments assembled correctly, except for slight abnormalities in regions of muscle detachment (Figure 4K–N). We conclude that in peel-1-affected embryos, the phenotype of paralysis and 2-fold arrest results from a defect in muscle contraction, not sarcomere assembly.
Next, to determine the cause of muscle-epidermal detachment, we examined trans-epidermal attachments, the specialized structures that span the epidermal syncytium and anchor it to underlying muscle [27]. A weakening of these structures is known to cause muscle-epidermal detachment [28], and consistent with this causality, in peel-1-affected embryos these structures were highly disorganized. As visualized by VAB-10A and intermediate filaments, trans-epidermal attachments did not organize into evenly spaced, circumferentially oriented bands. Instead, these bands were clumpy, disordered, and non-uniform in width (Figure 4O–R). This disorganization occurred even in areas where muscles remained attached, suggesting it to be the cause of muscle detachment, rather than an effect of it. In addition, in areas of highest stress, such as the mid-embryo bend, staining in post-arrest embryos (but not pre-arrest embryos) was often absent altogether (Figure 4P). This absence implies that trans-epidermal attachments in peel-1-affected embryos are so weak that in areas of highest stress, they rupture entirely.
Although the majority of peel-1-affected embryos display the aforementioned defects in muscle and epidermal tissue, the severity of these defects is variable and depends on the sex [13] and age of the sperm parent. Male-sired embryos always arrest paralyzed at the 2-fold stage, always exhibit epidermal leakage, and never hatch (n>2,000). Some hermaphrodite-sired embryos, on the other hand, elongate past the 2-fold stage (Figure 4E) or do not exhibit epidermal leakage. Occasionally, hermaphrodite-sired embryos even hatch, and the hatched progeny range from severely deformed larvae that die soon after hatching (Figure 4F) to morphologically normal larvae that develop into viable, fertile adults. In addition, among hermaphrodite-sired embryos, the proportion of peel-1-affected embryos arresting at the 2-fold stage and the proportion failing to hatch decreased dramatically with parental age (Figure 5A–B, Figure S4). Among male-sired embryos, parental age had no effect (Figure 5C).
One explanation for the decreased phenotypic severity of hermaphrodite- versus male-sired embryos is PEEL-1 dosage. Male sperm are up to 5-fold larger than hermaphrodite sperm [29], and as such, they may deliver more PEEL-1 protein to the embryo. In support of this hypothesis, we were able to alter the phenotype of peel-1-affected embryos, independent of sperm origin, by varying peel-1 dosage. Among hermaphrodite-sired embryos, doubling peel-1 copy number resulted in earlier onset of epidermal leakage (Figure 5D). Among male-sired embryos, expression of peel-1 exclusively from extra-chromosomal arrays, which are subject to germline silencing [30], had the opposite effect. For three of the peel-1 arrays shown in Figure 1B, 3%–10% of male-sired embryos elongated past the 2-fold stage and hatched into deformed larvae (n>150 embryos per array). When the same peel-1 transgene was expressed from a single-copy genomic insertion, which is not silenced, no hatching was observed (Figure 1C).
Given the relationship between PEEL-1 dosage, sperm size, and phenotypic severity, we suspected that the age-related decrease in phenotypic severity among hermaphrodite-sired embryos might reflect underlying size differences and size-based competition among hermaphrodite sperm. Hermaphrodite sperm vary approximately 2-fold in size ([29], personal observations) and are produced in a single bout of spermatogenesis at the onset of adulthood. Larger sperm in C. elegans experience a competitive advantage because they are able to crawl faster to reach the oocyte [29],[31]. One explanation for the age effect, therefore, is that larger-than-average sperm monopolize fertilization events early in life, leaving smaller, less toxic sperm to dominate fertilizations later on. In support of this hypothesis, we were able to reduce the age effect among hermaphrodite-sired embryos by delaying the hermaphrodite's use of self-sperm via partial mating to a male (Figure 5B). This result demonstrates that the age effect arises from a correlation between the competitive ability of each sperm and its toxicity to the embryo. Given the known biology of C. elegans sperm, the most parsimonious explanation for this correlation is that larger hermaphrodite sperm both are more competitive and carry more PEEL-1 protein. This correlation might also arise from PEEL-1 having a direct effect on the competitive ability of each sperm, although we have no evidence for such an effect.
To our knowledge, the above results represent the first evidence of competition among hermaphrodite sperm in vivo, as well as the first evidence of naturally occurring differences in sperm size affecting embryonic development. Insofar as PEEL-1 levels scale with sperm size, the wide phenotypic variability among hermaphrodite-sired embryos implies that for low levels of PEEL-1, phenotypic severity is acutely sensitive to PEEL-1 dosage. By the same logic, however, the phenotypic homogeneity among male-sired embryos implies that above a certain threshold level of PEEL-1, phenotypic severity ceases to increase. In support of this threshold effect, doubling or tripling peel-1 copy number among male-sired embryos did not produce a more severe phenotype, at least as measured by the onset of epidermal leakage (Figure 5D).
Consistent with its function as an antidote to sperm-supplied PEEL-1, zeel-1 is expressed in the embryo. Yet its expression is transient. By single-molecule FISH, zeel-1 expression begins at the eight-cell stage, peaks at approximately the 150-cell stage, and then turns off (Figure 6A). Transient expression was also observed for a GFP-tagged version of ZEEL-1, whose levels peaked during mid-embryogenesis (Figure S5), but whose expression was not observed in late-stage embryos, nor larvae or adults (unpublished data).
Within embryos, ZEEL-1::GFP was expressed in all or almost all cell types (Figure 6B–C, Figure S5). The protein localized most strongly to cell membranes (Figure 6B–C), consistent with ZEEL-1 having an N-terminal transmembrane domain. In some tissues, such as the developing pharynx and intestine, ZEEL-1::GFP appeared more concentrated at the apical face (Figure 6C).
Ubiquitous zeel-1 expression is consistent with zeel-1's ability to rescue the seemingly independent muscle and epidermal defects observed in peel-1-affected embryos. To test the tissue-autonomy of zeel-1 rescue, we expressed zeel-1 only in muscle and only in epidermis. We used the promoters of hlh-1 and lin-26, respectively, [32],[33], which initiate expression at the 80- to 100-cell stage [34]. Consistent with sperm-supplied PEEL-1 affecting muscle and epidermis independently, tissue-specific expression of zeel-1 produced tissue-specific rescue. In male-sired embryos, expression of zeel-1 only in muscle rescued the muscle defect of paralysis and 2-fold arrest, but it did not rescue epidermal leakage (n = 120 embryos; Figure 4G). (Muscle detachment and excretory cell distention could not be assayed because muscle movement, combined with epidermal leakage, ripped embryos apart entirely.) Conversely, expression of zeel-1 only in the epidermis rescued the epidermal defects, but it did not rescue paralysis and 2-fold arrest (n = 69 embryos; Figure 4H). In hermaphrodite-sired embryos, both constructs fully rescued a subset of embryos, and rescue activity increased with hermaphrodite age (Figure S6). This result is consistent with the age effect among hermaphrodite-sired embryos and the fact that hermaphrodite-sired embryos do not always exhibit defects in both muscle and epidermal tissue.
The structure of zeel-1—a C-terminal region (∼700 amino acids) predicted to be soluble and homologous to the highly conserved gene, zyg-11, and an N-terminus (∼200 amino acids) predicted to form a six-pass transmembrane domain—is highly unusual, and phylogenetic analysis indicates that this structure arose during a recent expansion of the zyg-11 family. Most metazoan genomes contain one to two zyg-11 orthologs, but in C. elegans, C. briggsae, and C. remanei, the zyg-11 family has expanded such that these species carry 19 to 37 zyg-11-like genes each (Figure S7). Most of these additional family members probably arose after the split with out-group C. japonica, because the genome of C. japonica contains only three zyg-11-like genes (including Cja-zyg-11 itself).
Aside from zeel-1, only two other members of the zyg-11 family—paralogs Y71A12B.17 and Y55F3C.9—contain predicted transmembrane domains (Figures 7A, S6). The transmembrane domains of these three genes are homologous to one another, but these domains show no detectable sequence similarity to any other gene in C. elegans or in the GenBank sequence database. This pattern, combined with the closest-paralog relationship between the C-terminal domains of zeel-1, Y71A12B.17, and Y55F3C.9 (Figures 7A, S6), implies that their shared transmembrane domain is evolutionarily novel and originated after the split between C. elegans and the C. briggsae/C. remanei lineage.
Analysis of gene order and sequence data suggests that zeel-1 arose through duplication of the Y71A12B.17 locus. zeel-1 and Y71A12B.17 are one another's closest paralogs (Figures 7A, S6), and the two genes are 55% identical at the amino acid level. Y71A12B.17 is located 12 Mb from zeel-1 in a tandem array of three additional zyg-11 family members, none of which contain the N-terminal transmembrane domain. Assuming that Y71A12B.17 and its neighbors originated in their current genomic location via repeated tandem duplication, then two scenarios for the origin of zeel-1 and Y71A12B.17 are possible. First, Y71A12B.17 may have originated via duplication of another gene in the tandem array, gained its transmembrane domain during or after duplication, and later been duplicated again to produce zeel-1. Alternatively, the tandem array may have arisen through partial duplication of Y71A12B.17. The second scenario is less parsimonious than the first, because it requires secondary loss of the transmembrane domain during creation of the tandem array. However, the second scenario still implies that the Y71A12B.17 locus predates zeel-1, because if the opposite were true, then zeel-1 would form an out-group to the tandem array, and it does not (Figure 7A).
Given the chimeric structure of ZEEL-1, we tested whether either domain alone could rescue the lethality of peel-1-affected embryos. The C-terminal ZYG-11-like domain, ZEEL-1SOL, provided no rescue (Figure 7B). The transmembrane domain, ZEEL-1TM, provided full rescue to hermaphrodite-sired embryos, but only partial rescue to male-sired embryos (Figure 7B–C). In contrast, the positive control transgene, full-length ZEEL-1 tagged with GFP, provided full rescue to both male- and hermaphrodite-sired embryos (Figure 7B–C). We conclude that the transmembrane domain of ZEEL-1 is required for antidote activity, and that this domain alone is able to neutralize the low doses of PEEL-1 delivered by hermaphrodite sperm but not the higher doses delivered by male sperm.
The partial rescue activity of ZEEL-1TM demonstrates that ZEEL-1SOL does contribute to the antidote activity of full-length ZEEL-1. To examine this contribution more carefully, we tested whether ZEEL-1SOL could rescue the lethality of peel-1-affected embryos when fused to the transmembrane domain of zeel-1's closest relative, Y71A12B.17. Like ZEEL-1SOL, the chimeric transgene, Y71A12B.17TM::ZEEL-1SOL, provided no rescue (Figure 7B). Assuming that this transgene was stably expressed, this result demonstrates that ZEEL-1SOL cannot confer antidote activity to a related transmembrane domain. Additionally, this result demonstrates that the transmembrane domains of zeel-1 and Y71A12B.17 have diverged functionally since their common ancestor. The molecular basis of this divergence remains unclear, however, because the transmembrane domains of zeel-1 and Y71A12B.17 are only 35% identical at the amino acid level, with substitutions distributed throughout their length (Figure S7).
To determine whether PEEL-1 can function as a toxin outside of embryos, we expressed peel-1 ectopically in larvae and adults. pee1-1 was expressed using each of two heat-shock promoters, Phsp-16.2 and Phsp-16.41 [35]. For each promoter construct, we generated both single-copy genomic insertions and extra-chromosomal arrays, which typically contain tens to hundreds of copies of a transgene [36]. Both types of animals grew normally under standard laboratory conditions, but a 1-h heat shock at 34°C was lethal to all: array-carrying adults died within 2 h after the start of heat-shock, and insertion-carrying animals within 4.5 h (Figure 8A). Faster killing of array-carrying animals is consistent with their higher peel-1 dosage, and similar results were observed for heat-shocked larvae (unpublished data). In addition, aside from the gross phenotype of death, the heat-shocked animals showed defects in most, if not all, tissues. Beginning approximately 30 to 45 min before death, the body-wall and male-tail muscles hyper-contracted; vacuoles formed in many tissues (Figure S8K); the lumen of the excretory cell distended (Figure S8L); the gonad appeared to disintegrate (Figure S8M); and in hermaphrodites, the gonad and intestine occasionally exploded through the vulva. We conclude that PEEL-1 is a nearly universal toxin, affecting many developmental stages and cell types.
Next, we tested whether heat-shock expression of zeel-1 could rescue the lethality caused by heat-shock expression of peel-1. We generated five extra-chromosomal arrays and one single-copy insertion of Phsp-16.41::zeel-1, and we tested these against one array and one insertion of Phsp-16.41::peel-1. Heat-shock expression of zeel-1 was able to rescue the lethality caused by Phsp-16.41::peel-1, but only when Phsp-16.41::peel-1 was expressed from insertions, not arrays (Figure 8B). The ability of Phsp-16.41::zeel-1 to rescue Phsp-16.41::peel-1 even when both were expressed from single-copy insertions (Figure 8B) indicates that insofar as these two transgenes produce equivalent levels of protein, zeel-1-mediated rescue does not require levels of ZEEL-1 to be higher than levels of PEEL-1.
The fact that peel-1-affected embryos do not exhibit defects until late in development—at 2-fold stage—is surprising because sperm-supplied factors are thought to act only during egg-activation and first cleavage [16]. One explanation for this paradox is that the late-occurring defects might be a downstream manifestation of a cryptic defect earlier in development. Alternatively, sperm-supplied PEEL-1 might persist long after fertilization but only become toxic at the 2-fold stage. While we have been unable to visualize PEEL-1 after fertilization, using either PEEL-1::GFP or the anti-PEEL-1 antibody (presumably because PEEL-1 becomes too diffuse), three observations are consistent with PEEL-1 acting directly during the 2-fold stage.
First, pre-2-fold embryos were able to develop normally even when exposed to more PEEL-1 protein than is delivered by sperm. We heat-shocked pre-2-fold embryos, aged 3 to 7.5 h after the four-cell stage, carrying either an array or insertion of Phsp-16.41::peel-1. Embryos were heat-shocked for 20 min at 34°C. Longer and earlier heat-shock were not possible because even in wild-type embryos, such conditions cause premature arrest (personal observations). Except for occasional subtle shape defects during early elongation, more than 97% of array- and insertion-carrying embryos developed normally to the 2-fold stage (n≥200; Figure S8A). This result is consistent with sperm-supplied PEEL-1 being able to persist until the 2-fold stage without manifesting a visible defect earlier in development.
Second, heat-shock expression of peel-1 as late as 30 min before the 2-fold stage phenocopied the defects observed in peel-1-affected embryos. In the heat-shocked embryos described above, embryos carrying an array of Phsp-16.41::peel-1 uniformly arrested at the 2-fold stage and showed muscle detachment, epidermal leakage, and excretory cell distention (Figure 8C, Figure S8B–J). These defects were also observed among insertion-carrying embryos, although their occurrence required earlier induction of the transgene (Figure 8C). In addition, consistent with heat-shock treatment exposing embryos to more PEEL-1 protein than is delivered by sperm, the defects induced by heat-shock were often more severe than those observed in peel-1-affected embryos, and even among insertion-carrying embryos, these defects could not be rescued by endogenous expression of zeel-1 (Figure 8C). These results demonstrate that as long as peel-1 is expressed at or just before the 2-fold stage, presence of PEEL-1 in the early embryo is dispensable for the 2-fold arrest.
Finally, rescue of peel-1-affected embryos did not require early expression of zeel-1. We induced zeel-1 expression in male-sired, peel-1-affected embryos by heat-shocking embryos carrying either an array or an insertion of Phsp-16.41::zeel-1. Heat-shock treatment rescued 53%–100% of array-carrying embryos (n = 17−43), as long as heat-shock treatment occurred at least 1 h before the 2-fold stage (Figure 8C). (Here rescue is defined as elongation past the 2-fold stage. See Figure S8A for the proportion of embryos that hatched.) Similar results were observed for insertion-carrying embryos, although rescue activity required earlier induction of the transgene (Figure 8C). These results imply that ZEEL-1 can neutralize sperm-supplied PEEL-1 at any time before the 2-fold stage. This scenario is temporally discordant with PEEL-1 causing a cryptic defect early in development.
To test the cell-autonomy of peel-1 killing, as well as the utility of peel-1 as a tool for cell-specific ablation, we expressed peel-1 under the control of each of two cell-specific promoters: Punc-47, which expresses in the GABA-ergic neurons [37], and Pexp-3, which expresses in the egg-laying muscles and the anal depressor muscle (C. Frøkjœr-Jensen, personal communication). For each promoter construct, we examined the presence or absence of the corresponding cell types for four independent extra-chromosomal arrays.
In both muscle cells and neurons, cell-specific expression of peel-1 produced cell-specific ablation, although the efficacy of ablation varied among arrays. Three of the Punc-47::peel-1 arrays killed 94.2%–99.8% (n = 241−453) of GABA-ergic neurons, although the fourth array killed only 28.6% (n = 350) of them. Each of the Pexp-3::peel-1 arrays killed 6%–27% (n = 83−90) of anal depressor muscles and 74%–94% (n = 140−180) of egg-laying muscles. Of the egg-laying muscles that remained alive, all were severely atrophied (Figure 8E). Lower toxicity to the anal depressor muscle may have been caused by selection bias among our arrays, because animals lacking the anal depressor muscle were very severely constipated and therefore more slow-growing than others. In addition, even among animals in which the anal depressor remained alive, constipation was prevalent (Figure S9), indicating that function of this muscle was impaired.
Aside from the defects caused by ablation of the corresponding cell types, animals carrying either type of construct were morphologically and behaviorally normal, consistent with PEEL-1 acting cell-autonomously. In addition, with one exception, no defects were observed outside the ablated cells types (Figure 8F). The exception was that the three “high kill” Punc-47::peel-1 arrays were lethal to the animal when inherited somatically along the lineage of the four RME neurons: For these three arrays, embryo and early larval lethality was very high (33.0%–57.7%; n = 327−1,095), and the only animals surviving to adulthood were those that lost the arrays somatically in the four RME neurons. Given that the RME neurons are not required for survival [38], this lethality implies that either (i) expression of peel-1 in the RME neurons kills a neighboring cell nonautonomously or (ii) expression of peel-1 is leaky and kills one or more essential cells along the RME lineage. While we did not distinguish between these possibilities, we note that the sister cell to one RME neuron is the excretory cell, which is essential for survival.
We have shown that the peel-1/zeel-1 element in C. elegans is composed of two, tightly linked genes: a sperm-delivered toxin, peel-1, and an embryo-expressed antidote, zeel-1. peel-1 and zeel-1 are located adjacent to one another in the genome, and both genes encode transmembrane proteins. peel-1 is expressed in the male germline, and its product is delivered to the embryo via fibrous body-membranous organelles. In the absence of zeel-1, sperm-supplied PEEL-1 causes dose-dependent, late-occurring defects in muscle and epidermal tissue. zeel-1 is expressed transiently in the embryo, and tissue-specific expression of zeel-1 produces tissue-specific rescue. The transmembrane domain of zeel-1 is required and partially sufficient for function, and like peel-1, this domain is evolutionarily novel and does not occur outside C. elegans. Finally, although PEEL-1 and ZEEL-1 normally function in embryos, peel-1 is lethal when expressed ectopically in adults, and this lethality is rescued by ectopic expression of zeel-1.
Given the evidence that sperm-supplied PEEL-1 may persist throughout embryogenesis and act directly during the 2-fold stage, PEEL-1 must be a remarkably potent toxin. Sperm are tiny in size compared to the oocyte, roughly 1% as large by volume [14], so PEEL-1 concentrations in the embryo are necessarily low. Moreover, assuming that PEEL-1 localizes to plasma membranes in the embryo, as might be expected for a FB-MO protein, then with each cell division, PEEL-1 will become more and more dilute relative to the total membrane component of the embryo.
Equivalently, the 2-fold stage of development must be remarkably sensitive to the toxic effects of PEEL-1. While the cause of this hypersensitivity remains unclear, we emphasize that the morphogenetic processes occurring at the 2-fold stage involve changes in cell shape and cell adhesion that are vastly more dramatic than those in earlier development. In addition, the two tissues most affected by PEEL-1—muscle and epidermis—are also the two tissues in which these morphogenetic changes are most pronounced.
The high potency of PEEL-1, combined with its widespread toxicity to a variety of cell types, highlights an unusual aspect of sperm cell biology: sperm are able to function normally, despite high concentrations of PEEL-1. While the mechanism of sperm protection remains unclear, we note that sperm differ from other cell types in three ways. First, sperm contain only a nucleus, some mitochondria, and FB-MOs; all other organelles and all ribosomes are excluded [39],[40]. Second, sperm lack an actin-based cytoskeleton [41], and instead crawl using polymers of the Major Sperm Protein [42]. Third, sperm sequester PEEL-1 in FB-MOs. Such sequestration is not possible in other cell types because FB-MOs are sperm-specific. In addition, although FB-MOs fuse with the plasma membrane upon sperm activation, they persist as permanent fusion pores [43]. This morphology prevents at least some FB-MO proteins from diffusing into the plasma membrane [44], and it may prevent diffusion of PEEL-1 as well.
The fact that pre-2-fold embryos are able to develop normally even when peel-1 is induced by heat-shock indicates that pre-2-fold development is less sensitive than the 2-fold stage to the toxic effects of PEEL-1. However, given the hypersensitivity of the 2-fold stage, it remains unclear whether pre-2-fold embryos are fully resistant to PEEL-1 (like sperm), or whether PEEL-1 levels in the heat-shocked, pre-2-fold embryos were too low to produce general cytotoxic effects. While we cannot discount the possibility of full resistance, we note that in the heat-shocked embryos, the time interval between heat-shock and the 2-fold stage was 5 h, at most. Five hours was sufficient for necrosis to develop in heat-shocked adults, but the two heat-shock experiments are not directly comparable because the heat-shock response in adults and embryos may not be equivalent, and the duration of heat-shock was shorter in embryos than in adults.
Because PEEL-1 has no sequence similarity to any other protein, the PEEL-1 sequence cannot be used to infer the mechanism of its toxicity. The fact that PEEL-1 is toxic even in extremely tiny amounts suggests that PEEL-1 might act catalytically—for example, by nucleating aggregation events or by acting as a transmembrane protease. The muscle hyper-contraction observed in heat-shocked adults, as well as the paralysis and 2-fold arrest observed in peel-1-affected embryos (which may in theory result from too much muscle contraction rather than too little), suggests that PEEL-1 might act by releasing intracellular calcium, perhaps by generating a membrane pore. It remains unclear, however, how calcium release alone can account for the epidermal defects observed in peel-1-affected embryos, because increased calcium signaling alone does not cause embryonic arrest [45] and is even known to suppress certain defects in epidermal morphogenesis [46]. In addition, it remains unclear how sperm might be protected from increased calcium, given the role of calcium in sperm activation [44].
Given the uncertain mechanism of PEEL-1 toxicity, there are also many possible mechanisms of ZEEL-1-mediated rescue. ZEEL-1 might promote degradation of PEEL-1, or it might prevent PEEL-1 from binding to its target, either by acting as a competitive inhibitor or by neutralizing PEEL-1 through direct interaction. While we have been unable to demonstrate a direct physical interaction between PEEL-1 and ZEEL-1, several observations are consistent with it. First, both PEEL-1 and ZEEL-1 are transmembrane proteins. ZEEL-1 localizes to cell membranes, and PEEL-1 localizes to FB-MOs. Assuming that FB-MOs do not endocytose during fertilization, localization to these organelles should deliver PEEL-1 to the plasma membrane of the zygote, where it should have the opportunity to encounter ZEEL-1 later in development. Second, the transmembrane domain of ZEEL-1 is required and partially sufficient for function, consistent with this domain binding directly to PEEL-1. Third, tissue-specific expression of zeel-1 produces tissue-specific rescue, consistent with ZEEL-1 being able to neutralize PEEL-1 only when both proteins are present within the same cell. Fourth, ZEEL-1 can neutralize PEEL-1 toxicity even in adults, demonstrating that the genetic interaction between peel-1 and zeel-1 does not require any intermediaries specific to embryogenesis. Fifth, in both embryos and adults, ZEEL-1 is able to neutralize small but not large doses of PEEL-1. This dose-dependence implies that the genetic interaction between peel-1 and zeel-1 requires a minimum ratio of ZEEL-1 to PEEL-1.
Like nearly all other “selfish” genetic elements whose genetic basis is known [2],[9],[47], the peel-1/zeel-1 element experiences a suppression of recombination between component parts: The insertion/deletion of peel-1 and zeel-1 removes both genes at once, so the two genes cannot be separated by homologous recombination. This genomic organization has undoubtedly allowed peel-1 to persist in spite of its toxic effects, because recombination breaking apart peel-1 and zeel-1 would have generated haplotypes carrying peel-1 alone, and such haplotypes are effectively suicidal.
The peel-1/zeel-1 element's mode of action is similar to that of Wolbachia, in that both types of elements act through paternal-effect killing [10],[11]. Wolbachia's molecular mechanism is very different, however, because Wolbachia does not load sperm with an extra-nuclear toxin, but instead modifies the sperm pronucleus to undergo a chromatin condensation defect during the first mitotic division [48],[49]. In addition, Wolbachia is an intracellular bacterium, not a nuclear-encoded locus, and rescue of Wolbachia-mediated killing depends upon the contents of maternal ooplasm, not zygotic transcription of a nuclear-encoded gene. The peel-1/zeel-1 element's mode of action is also similar to the maternal-effect killing and zygotic self-rescue of Medea-factors [50], although the extent of this similarity at the molecular level is unclear because the mechanism of Medea-factor killing is unknown [9].
Previously, we demonstrated that haplotypes carrying the peel-1/zeel-1 element and haplotypes lacking it are maintained by balancing selection [13]. We hypothesized that the target of selection may be a linked polymorphism, rather than the peel-1/zeel-1 element itself [13]. Under this scenario, peel-1 may represent an unprecedented case of “inverted” sheltered load. Sheltered load refers to the incidental maintenance of deleterious alleles tightly linked to sites under balancing selection [51]. Ordinarily, sheltered load occurs when deleterious recessives arise on haplotypes maintained in persistent heterozygosity, such as those of major histocompatibility complex loci in vertebrates [52] or self-incompatibility loci in plants [53]. peel-1 is like these deleterious recessives in that although peel-1 has the potential to impose substantial genetic load on the species, its effects are rarely visible to natural selection. In the case of peel-1, however, sheltering is inverted because peel-1 is only visible when heterozygous, and in C. elegans, heterozygosity is the exception rather than the norm.
Like any locus promoting its own transmission to the detriment of the rest of the genome, the peel-1/zeel-1 element creates a selective environment favoring its own suppression. From a genic perspective, loci unlinked to the peel-1/zeel-1 element suffer a fitness cost every time they are transmitted to a peel-1-affected embryo. As a consequence, mutations unlinked to peel-1 and zeel-1 that either suppress the activity of peel-1 or mimic the activity of zeel-1 will be favored by natural selection. Insofar as such alleles are accessible in mutational space, their absence further attests to the sheltering of the peel-1/zeel-1 element by near-perpetual homozygosity. (The peel-1 mutations in strains MY19 and EG4348 do not represent favored alleles because they do not arise on haplotypes suffering a fitness cost.)
The insertion/deletion polymorphism of peel-1 and zeel-1 raises the following question: Did this indel polymorphism arise by an insertion event or by deletion of preexisting sequence? With respect to zeel-1, this polymorphism probably arose by a deletion event, because the divergence between zeel-1 and its presumed ancestor, Y71A12B.17, predates allelic divergence at the peel-1/zeel-1 locus. zeel-1 and Y71A12B.17 are 45% diverged at the amino acid level, and divergence at synonymous sites is saturated (see Materials and Methods). In comparison, the Bristol and Hawaii alleles of genes surrounding the indel polymorphism of peel-1 and zeel-1 are roughly 2% diverged at the amino acid level and 10%–16% diverged at synonymous sites [13], and this level of divergence is representative of the divergence between all haplotypes carrying the peel-1/zeel-1 element and all haplotypes lacking it [13].
It is reasonable to suppose that the peel-1/zeel-1 element originated as a weak toxin-antidote pair and then co-evolved into its current form. Yet given the low selective pressure for transmission ratio distortion in a self-fertilizing species, it is unlikely that peel-1 and zeel-1 co-evolved within C. elegans as result of this type of selective pressure alone. One possible solution to this paradox is that peel-1 and zeel-1 co-evolved in the out-crossing ancestor of C. elegans, where the selective pressure for transmission ratio distortion would have been much stronger. Another, non-mutually exclusive hypothesis is that peel-1 was originally favored because it aided in another cellular process, such as sperm competition, and its toxicity to the embryo was initially mild and incidental. Under this scenario, zeel-1 would have arisen to counteract the toxicity of peel-1, and once zeel-1 became established, the presence of zeel-1 would have allowed for stronger toxicity on the part of peel-1 and, eventually, lethality in zeel-1's absence. Regardless of the initial selective pressures favoring peel-1 and zeel-1, however, the fact that both peel-1 and the transmembrane domain of zeel-1 are evolutionarily novel indicates that the self-promoting activity of the peel-1/zeel-1 element arose fundamentally from the co-evolution of two novel proteins.
Strains were maintained at 19–23°C on NGM plates spotted with E. coli strain OP50. In all age-effect experiments, strains were strictly maintained at 20°C. For all transgenes described in this publication, only one array and/or one insertion is given in the strain list, although unless otherwise specified in the Results section, multiple independent arrays or insertions were examined.
CB4088: him-5(e1490) V.
CB4856: Hawaii natural isolate carrying niDf9 I. niDf9 designates the 19 kb deficiency spanning peel-1 and zeel-1.
EG1285: oxIs12[Punc-47::GFP; lin-15(+)] lin-15B(n765) X.
EG4322: ttTi5605 II; unc-119(ed3) III.
EG4348: Utah natural isolate carrying peel-1(qq99) I. EG4348 was collected by M. Ailion from Salt Lake City, Utah (this publication). qq99 designates the naturally occurring nonsense mutation in peel-1.
EG5389: qqIr7[peel-1(qq99)] I; oxIs494[Ppeel-1::GFP, Cbr-unc-119(+)] II; unc-119(ed3) III.
EG5655: qqIr7[peel-1(qq99)] I; oxSi19[peel-1(+), Cbr-unc-119(+)] II; unc-119(ed3) III.
EG5766: qqIr7[peel-1(qq99)] I; oxSi77[Ppeel-1::peel-1::GFP, Cbr-unc-119(+)] II; unc-119(ed3) III.
EG5801: oxSi87[Ppeel-1::peel-112a.a.::GFP, Cbr-unc-119(+)] II; unc-119(ed3) III.
EG5955: qqIr7[peel-1(qq99)] I; ttTi5605 II; unc-119(ed3) III; oxEx1462[Phsp-16.41::peel-1, Cbr-unc-119(+), Pmyo-2::mCherry, Pmyo-3::mCherry, Prab-3::mCherry].
EG5958: qqIr7[peel-1(qq99)] I; oxSi186[Phsp-16.41::peel-1, Cbr-unc-119(+)] II; unc-119(ed3) III.
EG5960: qqIr7[peel-1(qq99)] I; oxSi188[Phsp-16.2::peel-1, Cbr-unc-119(+)] II; unc-119(ed3) III.
EG5961: qqIr7[peel-1(qq99)] I; ttTi5605 II; unc-119(ed3) III; oxEx1464[Phsp-16.2::peel-1, Cbr-unc-119(+), Pmyo-2::mCherry, Pmyo-3::mCherry, Prab-3::mCherry].
EG6297: qqIr5[niDf9] I; oxSi298[Phsp-16.41::zeel-1::tagRFP, Cbr-unc-119(+)] II; unc-119(ed3) III.
EG6298: qqIr5[niDf9] I; ttTi5605 II; unc-119(ed3) III; oxEx1501[Phsp-16.41::zeel-1::tagRFP, Cbr-unc-119(+), Pmyo-2::GFP].
EG6301: qqIr5[niDf9] I; ttTi5605 II; unc-119(ed3) III; oxEx1504[Pexp-3::peel-1, Cbr-unc-119(+), Pmyo-2::mCherry, Pmyo-3::mCherry, Prab-3::mCherry].
EG6306: qqIr5[niDf9] I; ttTi5605 II; unc-119(ed3) III; oxEx1509[Punc-47::peel-1, Cbr-unc-119(+), Prab-3::mCherry].
MT1344: bli-3(e767) lin-17(n677) I.
MT3301: fem-1(hc17) IV; him-5(e1490) V.
MY19: German natural isolate carrying peel-1(qq98) I. MY19 was collected from Roxel, Germany [54]. qq98 designates the naturally occurring nonsense mutation in peel-1.
N2: Laboratory reference strain, Bristol.
PD4790: mIs12[myo-2::GFP, pes-10::GFP and gut::GFP].
QX1015: niDf9 I; qqIr8[N2 = >CB4856, unc-119(ed3)] III.
QX1197: qqIr5[CB4856 = >N2, niDf9] I. qqIr5 is an 140–370 kb introgression from CB4856 into N2. This strain was used in some experiments instead of CB4856, in order to reduce the genetic variation segregating in the background.
QX1257: niDf9 I; qqIr8[unc-119(ed3)] III; qqIs2[zeel-1genomic::GFP, unc-119(+)].
QX1264: niDf9 I; qqIr8[unc-119(ed3)] III; qqEx2[zeel-1genomic::GFP, unc-119(+)].
QX1319: zeel-1(tm3419)/hT2[qIs48] I; +/hT2[qIs48] III.
QX1320: qqIr6[EG4348 = >N2, peel-1(qq99)] I; unc-119(ed3) III.
QX1384: niDf9 I; qqIr8[unc-119(ed3)] III; qqEx6[Pzeel-1:: zeel-1SOL, unc-119(+)].
QX1392: qqIr6[peel-1(qq99)] I; unc-119(ed3) III; qqEx3[peel-1(+), unc-119(+)].
QX1409: qqIr7[EG4348 = >N2, peel-1(qq99)] I; ttTi5605 II; unc-119(ed3) III.
QX1577: qqIr5[niDf9] I; qqEx1[Pzeel-1::zeel-1cDNA::GFP, Pmyo-2::RFP].
QX1589: qqIr5[niDf9] I; qqEx4[Pzeel-1::Y71A12B.17TM:: zeel-1SOL, Pmyo-2::RFP].
QX1605: qqIr5[niDf9] I; ttTi5605 II; unc-119(ed3) III.
QX1607: qqIr5[niDf9] I; qqEx5[Pzeel-1:: zeel-1TM, Pmyo-2::RFP].
QX1618: qqIr5[niDf9] I; qqEx7[Plin-26::zeel-1, Pmyo-2::RFP].
QX1619: qqIr5[niDf9] I; qqEx8[Phlh-1::zeel-1, Pmyo-2::RFP].
QX1624: qqIr5[niDf9] I; oxSi186[Phsp-16.41::peel-1, Cbr-unc-119(+)] II; unc-119(ed3) III.
QX1650: oxSi19[peel-1(+), Cbr-unc-119(+)] II.
QX1772: qqIr5[niDf9] I; ttTi5605 II; unc-119(ed3) III; oxEx1462[Phsp-16.41::peel-1, Cbr-unc-119(+), Pmyo-2::mCherry, Pmyo-3::mCherry, Prab-3::mCherry].
SJ4157: zcIs21[Phsp-16::clpp-1(WT)::3xmyc-His tag+Pmyo-3::GFP] V.
In all experiments except the age-effect experiment, embryo lethality from self-fertilizing hermaphrodites was scored by isolating hermaphrodites at the L4 stage and singling them to fresh plates the following day. After laying eggs for 8–10 h, the hermaphrodites were removed and embryos were counted. Unhatched embryos were counted ∼24 h later.
To score embryo lethality from mated hermaphrodites, three or four L4 hermaphrodites were mated to six to ten L4 or young adult males for 24–36 h. Hermaphrodites were then singled to fresh plates and embryo lethality was scored as above. Broods were examined for the presence of males 2–3 d later, and any broods lacking males were excluded. To allow male sperm to age within the reproductive tract of the hermaphrodite, mated hermaphrodites were removed from males, and lethality was scored among embryos laid 3 d after removal.
In the age-effect experiment, 91 zeel-1(tm3419)peel-1(+)/niDf9 hermaphrodites were singled at the L4 stage and transferred every 12 h to fresh plates. Hermaphrodites were discarded after the first 12-h period in which they failed to lay fertilized embryos. Total embryos were counted at the end of each laying period, and unhatched embryos were counted ∼24 h after each laying period had ended.
To score embryo lethality from partially mated hermaphrodites, 130 zeel-1(tm3419)peel-1(+)/niDf9 hermaphrodites were mated at the L4 stage to an equal number of PD4790 males, which carry an insertion of the fluorescent marker, Pmyo-2::GFP. After 24 h, hermaphrodites were singled and transferred every 12 h to fresh plates until day five. Embryo lethality was scored as above, except that after unhatched embryos were counted, hatched and unhatched progeny were classified as self- or cross-progeny according to presence of pharyngeal GFP. Hermaphrodites laying 100% self-progeny or more than 95% cross-progeny were excluded. The remaining hermaphrodites, which we define as “partially mated,” laid ∼10%–50% self-progeny. In these broods, we calculated the portion of self-progeny, laid during days 3 to 5, that failed to hatch.
Absence of the paternal-effect in EG4348 was mapped relative to bli-3(e767), a visible marker located ∼10 cM from the peel-1 interval. Mapping was performed as described [13]. Briefly, EG4348 males were crossed to MT1344 hermaphrodites, and F1 hermaphrodites were mated to CB4856 males. The resulting hermaphrodite progeny were allowed to self-fertilize, and their broods were scored for embryo lethality (i.e., presence of peel-1 activity) and presence of Bli animals. Directionality with respect to bli-3 could be inferred because bli-3 is located at the left-hand tip of chromosome I.
Preliminary sequence analysis of the peel-1 interval in EG4348 was performed by genotyping EG4348 with a subset of the markers listed in Table S2 of [13]. These markers tile across the peel-1 interval, and they distinguish all haplotypes carrying an intact copy of the peel-1/zeel-1 element from all haplotypes lacking it [13]. In other words, the Bristol-like alleles of these markers are in perfect linkage disequilibrium with presence of the peel-1/zeel-1 element. At all markers we assayed, EG4348 carried the Bristol-like allele.
Fine-mapping in MY19 and EG4348 was performed by crossing each strain to MT1344 and collecting Lin Non-Bli and Bli Non-Lin recombinants in the F2 generation. Recombinant animals were genotyped (via a portion of their F3 broods) at each of two markers flanking the peel-1 interval. The right-hand marker for the MY19 cross was a BstCI snip-SNP amplified with primers 5′-GTA TTC CGA CGA TTC GGA TG-3′ and 5′-CAT TGA GAA CAC AAA AAC AAA CG-3′. The right-hand marker for EG4348 cross was an AfeI snip-SNP amplified with primers 5′- GAC ATA TTT CCC GCA ACC TG-3′ and 5′- GTG ACG AGG CTT GAG GAT TC-3′. The left-hand marker for both crosses was a BanI snip-SNP amplified with primers 5′-CGC CAA ATA TGT TGT GCA GT-3′ and 5′-CAC CAC GTG TCC TTT CTC ATT-3′.
Recombinants breaking within the peel-1 interval were homozygosed for the recombinant chromosome, and the resulting homozygotes were phenotyped for peel-1 activity. Phenotyping was performed by crossing each line to CB4856 and scoring embryo lethality from self-fertilizing, F1 hermaphrodites, and from F1 males backcrossed to CB4856 hermaphrodites. Recombinants were classified as having peel-1 activity if these crosses produced ∼25% and ∼50% embryo lethality, respectively. Next, the locations of recombination breakpoints were mapped more finely by sequencing six to ten sequence polymorphisms, located throughout the peel-1 interval, that distinguish MY19 or EG4348 from Bristol. The MY19 polymorphisms were determined from the MY19 sequence described in [13], and the EG4348 polymorphisms were determined by amplifying and sequencing arbitrary fragments from this strain. Some polymorphisms are shared between MY19 and EG4348, and these were used in both crosses. For the most informative recombinants, we later sequenced across the entire breakpoint region in order to map these breakpoints to the level of adjacent polymorphisms. This approach mapped the peel-1-disrupting mutations to regions of 5 kb in MY19 and 8 kb in EG4348. These intervals were then sequenced in the corresponding strains, and all sequence polymorphisms were identified. Finally, we genotyped these polymorphisms in a panel of 38 wild strains previously identified as having intact peel-1 activity [13]. These strains, as well as the primers used for genotyping them, are given in Figure S2. The MY19 sequence was deposited in GenBank previously [13], and the EG4348 sequence was deposited under accession number HQ291558.
RNA was collected from mixed-staged Bristol animals by freeze-cracking and extracting in Trizol (Invitrogen) according to the manufacturer's protocol. Reverse transcription-PCR (RT-PCR) was performed using pairs of primers flanking each candidate mutation in MY19 and EG4348. For each pair of primers, the forward and reverse primers were located ∼100 bp apart, and two reactions were performed, one using each of the two primers as the RT primer. Product was observed for only one pair of primers, and for that pair, only in one direction. These primers were 5′-ACA TGT ATC TTG ATC TGC CTG A-3′ (forward) and 5′-AAA AAT TAA CCA CAA TGA AGC AA-3′ (reverse), and product was only observed using the reverse primer as the RT primer. To recover the remainder of this putative transcript, 3′ and 5′ RACE were performed using standard methods [55]. For 3′ RACE, the RT reaction was performed using 5′-GTT TTC CCA GTC ACG ACT TTT TTT TTT TTT TTT TT-3′, and PCR was performed using the gene-specific primer, 5′-ACA TGT ATC TTG ATC TGC CTG A-3′ (forward), and the adaptor primer, 5′-GTT TTC CCA GTC ACG AC-3′ (reverse). For 5′ RACE, the RT reaction was performed using a gene-specific primer that spanned the putative stop codon, 5′-TCA ATT TCA TGG ATT TTC AAC A-3′, and PCR was performed using 5′-GGC CAC GCG TCG ACT AGT ACG GGI IGG GII GGG IIG-3′ (forward) and a nested, gene-specific primer, 5′-AAA AAT TAA CCA CAA TGA AGC AA-3′ (reverse). Then, a second round of PCR was performed using the adaptor primer, 5′-GGC CAC GCG TCG ACT AGT AC-3′ (forward), and another nested, gene-specific primer, 5′-AGA GCA ATA ACA TGC GCA AA-3′ (reverse). SuperScript III (Invitrogen) was used in all RT reactions, and PlatinumTaq (Invitrogen) was used for all PCR reactions. The peel-1 transcript did not contain a splice leader sequence and was deposited in GenBank under accession number HQ291556. More recently, the peel-1 transcript was identified independently by WormBase curators and assigned the identification number, Y39G10AR.25.
To search for transcripts carrying both peel-1 and zeel-1, an RT reaction was performed using the peel-1-specific primer, 5′-AAA AAT TAA CCA CAA TGA AGC AA-3′, and PCR was performed using a forward primer located in the 3′ end of zeel-1 (5′-CCA TCC GAG ATA ACC GAA AA-3′) and a reverse primer located in the 5′ end of peel-1 (5′-AGA GCA ATA ACA TGC GCA AA-3′). No product was observed.
CB4088 and MT3301 animals were grow at 15°C and synchronized at the L1 stage by bleaching and hatching overnight in M9. L1s were split into two populations, and one population was shifted to 25°C. When animals had reached young adulthood, hermaphrodites and males were separated by hand, and RNA was collected as above. Real-time PCR of peel-1, spe-9, and rpl-26 was performed in triplicate, for 40 cycles, on an ABI 7900HT using the QuantiTect SYBR Green Kit (Qiagen). Relative expression levels of peel-1 and spe-9 were calculated separately for males and hermaphrodites, using the 2−ΔΔCt method, with rpl-26 as the endogenous control and the 15°C MT3301 sample as the reference sample. Primers used to amplify peel-1 were 5′-TAC ACC CGT CAC ACC AAC TG-3′ and 5′-TCC GAC TAT GAT GTT CCA CAA-3′; primers for spe-9 were 5′-CGG CTT GCA TAC ACA ATG AG-3′ and 5′-ACG CCA TGA CTC TTG CTC TT-3′; and primers for rpl-26 were 5′-TCC AAT CAG AAC CGA TGA TG-3′ and 5′-GTG CAC AGT GGA TCC GTT AG-3′.
Among the hermaphrodite samples, relative expression levels of peel-1 and spe-9 were roughly equivalent, except for the 25°C MT3301 sample, where expression of peel-1 and spe-9 was undetectable. That is, in this sample, signal for peel-1 and spe-9 failed to rise above the detection threshold, even after 40 cycles, despite rpl-26 amplifying normally.
Single molecule FISH of was performed as in [56], with the embryos and hermaphrodites squashed down to ∼9 µm thickness for imaging. Automated counting of nuclei in embryos was performed using software developed in [56],[57].
Transgenic animals carrying peel-1(+) were generated by two methods: bombardment [58] and Mos1-mediated, single-copy insertion [59]. For bombardment, a fragment containing the Bristol allele of peel-1, along with ∼2.8 kb of upstream sequence and ∼1 kb of downstream sequence, was excised from fosmid WRM0633bE09 (Bioscience LifeSciences, Nottingham, UK) using AhdI and NgoMIV. This fragment was cloned into the yeast shuttle vector, pRS246 (ATCC, Manassas, VA), via yeast-mediated ligation [60] of the fragment's ends. The resulting plasmid, pHS11, was bombarded into QX1320, along with the unc-119(+) rescue vector, pDP#MM016B [61]. Bombardment was performed as in [62], although only extra-chromosomal arrays were recovered. Nine independent transgenic lines were tested for peel-1 activity by crossing them to CB4856 and scoring embryo lethality from self-fertilizing, F1 hermaphrodites (self-cross) and F1 males backcrossed to CB4856 hermaphrodites (backcross).
For Mos1-mediated insertion, the peel-1 fragment from pHS11 was amplified by PCR, using primers having NheI cut sites, and this amplicon was cut with NheI and ligated into pCFJ151 [59] linearized with AvrII. The resulting plasmid, pHS26, was injected into QX1409 along with the vectors needed to generate single-copy insertions [59]. Insertion-carrying animals were recovered by the direct insertion method [59], and five independent insertion-carrying lines were tested for peel-1 activity as above. For one of the two insertions that did exhibit peel-1 activity, the self-cross and backcross were repeated, and hatched progeny were collected and genotyped for a PCR-length polymorphism located less than 1 kb from niDf9. The primers used to amplify this polymorphism were 5′-TGG ATA CGA TTC GAG CTT CC-3′ (forward) and 5′-CCC CCT AAT TTC CAA GTG GT-3′ (reverse).
For three of the peel-1 array lines, a small number of severely deformed L1s were observed in the backcross, similar to the “escapers” typically observed among peel-1-affected embryos sired by hermaphrodites. We suspected that these L1s had “escaped” the paternal-effect due to partial germline silencing of the peel-1 arrays. Consistent with this hypothesis, we genotyped 13 of these animals, using the PCR-length polymorphism described above, and all were zeel-1(niDf9) homozygotes. We then calculated the frequency of these escapers relative to the total number of peel-1-affected progeny (i.e., relative to the total number of dead embryos and deformed L1s).
ZEEL-1::GFP was generated by amplifying GFP from PD95.75 and inserting it into pHS4.1, a genomic subclone of zeel-1(+) described previously [13]. pHS4.1 was linearized with AhdI, and yeast-mediated ligation [60] was used to insert GFP just upstream of the zeel-1 stop codon. Later, a second ZEEL-1::GFP construct was generated using the cDNA of zeel-1, instead of the genomic locus. This construct was generated by first cutting pHS4.1 with EcoNI and BglII, in order to remove the entire coding region of zeel-1, and then inserting a full cDNA of zeel-1, followed by GFP. The cDNA of zeel-1 was cloned previously [13], and this replacement was performed using yeast-mediated ligation [60]. Both constructs showed full rescue of peel-1-affected embryos, and data from the two constructs were combined.
To generate ZEEL-1SOL, pHS4.1 was cut with EcoNI and KpnI, and the fragment containing zeel-1 codons 5 to 205 was removed. The remaining fragment was then re-circularized, using yeast-mediated ligation [60], to fuse codon 4 to codon 206. To generate ZEEL-1TM, the entire coding region of zeel-1 was excised from pHS4.1 using EcoNI and BglII, and this fragment was replaced with a partial cDNA of zeel-1 encoding the first 205 amino acids of the protein. This replacement was performed using yeast-mediated ligation [60].
To generate Y71A12B.17TM::ZEEL-1SOL, the coding region of zeel-1 was excised from pHS4.1, as above, and yeast mediated ligation [60] was used to replace this fragment with a partial cDNA of Y71A12B.17, followed by a partial cDNA of zeel-1. The resulting construct contained the N-terminal 207 codons of Y71A12B.17 fused to the C-terminal 712 codons of zeel-1. The junction of this fusion was chosen to overlap a string of seven amino acids (KNERKEG) that are perfectly conserved between the two proteins. The Y71A12B.17 cDNA was cloned by reverse transcribing RNA from the Bristol strain using primer 5′-TTG AAC AAA AAC AAT GGA TAT GTA A-3′, and then performing PCR using primers 5′-GGG GAC AAG TTT GTA CAA AAA AGC AGG CTT CAT GTC GGA TTT CGA CTC AGA-3′ (forward) and 5′-GGG GAC CAC TTT GTA CAA GAA AGC TGG GTC ATT TAT TAA CTC CAA CAA TGA TTC G-3′ (reverse). This PCR product was then cloned into the vector, pDONR221 (Invitrogen), using the Gateway cloning kit (Invitrogen). The Y71A12B.17 cDNA differs slightly from the WormBase gene prediction and was deposited in GenBank under accession number HQ291557.
All constructs were bombarded [62] into QX1015 along with the unc-119(+) rescue vector, pDP#MM016B [61], or they were injected [63] into QX1197 at ∼80ng/µl, along with the fluorescent marker, Pmyo-2::RFP at 3 ng/µl. To test each transgene for its ability to rescue peel-1-affected embryos, transgenic animals were crossed to the Bristol strain, and lethality was scored among embryos derived from two crosses: self-fertilizing F1 hermaphrodites (self-cross) and F1 males backcrossed to hermaphrodites of the original transgenic line. To calculate hatch rates among peel-1-affected embryos inheriting ZEEL-1::GFP and ZEEL-1TM, transgenic animals were crossed to QX1319, and embryos were collected from (i) transgenic, self-fertilizing, F1, zeel-1(tm3419)peel-1(+)/niDf9 hermaphrodites and (ii) transgenic, F1, zeel-1(tm3419)peel-1(+)/niDf9 males backcrossed to non-transgenic, niDf9/niDf9 hermaphrodites. Inheritance of ZEEL-1::GFP and ZEEL-1TM was inferred by expression of the co-injection marker, Pmyo-2::RFP, which can be scored even in arrested embryos.
All other transgenes were generated using the three-site Gateway system from Invitrogen. This method allows three separate DNA fragments to be joined together and inserted into pCFJ150, which contains Cbr-unc-119(+) and the sequences needed for Mos1-mediated insertion at the ttTi5605 Mos site on chromosome II [59]. In most cases, this method was used to join together a promoter of interest, a coding sequence, and a 3′ UTR.
For Ppeel-1::GFP, we joined together the peel-1 promoter, GFP, and the peel-1 3′UTR. For the peel-1 promoter, we used all intergenic sequences between the peel-1 start codon and last coding segment of zeel-1 (i.e., 2,473 bp of sequence). The peel-1 3′ UTR was determined empirically and extended 86 bp downstream of the peel-1 stop codon. For GFP, we used a variant containing S65C and three internal introns (identical to the variant in pPD95.75).
For Ppeel-1::PEEL-112a.a.::GFP, the PEEL-1 leader peptide was added by extending the promoter fragment to include the first 12 amino acids of PEEL-1. This signal peptide was discovered while we were investigating a possible regulatory role of the first intron of peel-1. We had generated a GFP reporter driven by the peel-1 promoter and the first intron of peel-1, and this construct also happened to carry the first 12 amino acids of PEEL-1. GFP driven by this construct was packaged into sperm (unpublished data), and in order to confirm that sperm packaging was caused by the leader peptide, rather than the intron, we generated Ppeel-1::PEEL-112a.a.::GFP (which excludes the first intron). Conversely, we also generated a reporter carrying the first intron and a randomized leader peptide, and for this construct, no sperm packaging was observed (unpublished data).
To tag PEEL-1 with GFP, the promoter fragment was extended even further to include the entire peel-1 gene, up to (but excluding) the stop codon.
All three constructs were injected into EG4322 or QX1409, and single-copy insertions were obtained using the direct insertion MosSCI method [59]. Three to six independent insertions were analyzed for each construct, and no differences were observed among insertions of the same construct. We note that although PEEL-1::GFP appears to localize normally, it failed to exhibit peel-1 activity (unpublished data), presumably because the GFP tag inhibited function.
For Plin-26::zeel-1 and Phlh-1::zeel-1, the promoters of lin-26 or hlh-1 were joined to the cDNA of zeel-1 and the 3′ UTR of let-858. For the promoters of lin-26 and hlh-1, 7,122 bp of sequence and 3,037 bp of sequence upstream of the respective start codons were used. For the let-858 3′ UTR, 434 bp of sequence downstream of the stop codon was used. Transgenic animals were generated by injecting Plin-26::zeel-1 and Phlh-1::zeel-1 into QX1197 at ∼80 ng/µl, along with the fluorescent marker, Pmyo-2::RFP at 3 ng/µl.
To evaluate rescue among male-sired embryos, transgenic animals were crossed to QX1319, and transgenic, F1, zeel-1(tm3419)peel-1(+)/niDf9 males were backcrossed to non-transgenic, niDf9/niDf9 hermaphrodites. Embryos were dissected from these hermaphrodites and imaged every 10–20 min, starting before the 2-fold stage and ending at least 6 h after the 2-fold stage. To evaluate rescue among hermaphrodite-sired embryos, transgenic lines were crossed to QX1319, and embryo viability was scored among embryos collected from transgenic, self-fertilizing, F1, zeel-1(tm3419)peel-1(+)/niDf9 hermaphrodites. Among both male- and hermaphrodite-sired embryos, inheritance of the transgene was inferred by expression of Pmyo-2::RFP.
Phsp-16.41::peel-1, Phsp-16.2::peel-1, Pexp-3::peel-1, and Punc-47::peel-1 were generated by joining the peel-1 cDNA downstream of the appropriate promoter and upstream of the tbb-2 3′ UTR. We describe the promoter and 3′UTR fragments in terms of length of sequence upstream or downstream of the appropriate start or stop codons: Phsp-16.41 (501 bp), Phsp-16.2 (493 bp), Pexp-3 (2,877 bp), Punc-47 (1,251 bp), and tbb-2 3′ UTR (331 bp). Phsp-16.41::peel-1 and Phsp-16.2::peel-1 were injected into QX1409 at 25 ng/µl, and arrays and MosSCI insertions were recovered as in [59]. The arrays carry co-injection markers Pmyo-2::mCherry, Pmyo-3::mCherry, and Prab-3::mCherry. Pexp-3::peel-1 and Punc-47::peel-1 were injected into QX1605 at 25 and 10 ng/µl, respectively, along with co-injection markers Prab-3::mCherry, Pmyo-2::mCherry, and Pmyo-3::mCherry (for Pexp-3::peel-1) and marker Prab-3::mCherry (for Punc-47::peel-1). In all cases, Pmyo-3::mCherry and Prab-3::mCherry were injected at 10 ng/µl, and Pmyo-2::mCherry was injected at 5 ng/µl.
Phsp-16.41::zeel-1 was generated using the hsp-16.41 promoter described above, the zeel-1 cDNA, and the let-858 3′UTR fused downstream of tagRFP. tagRFP was added to confirm expression of zeel-1 after heat-shock. Phsp-16.41::zeel-1 was injected at 10 ng/µl into QX1605 and the arrays and the MosSCI insertion were recovered as in [59], except that a GFP-based co-injection marker (Pmyo-2::GFP injected at 2.5 ng/µl) was used in order to distinguish these arrays from the Phsp-16.41::peel-1 arrays.
Imaging of fixed embryos and live imaging of ZEEL-1::GFP embryos was performed on a PerkinElmer RS3 spinning disk confocal. All other imaging was performed on a Nikon 90i equipped with a CoolSNAP HQ2 camera and a X-Cite 120 Series fluorescent light source. Images were acquired and background subtracted with either Volocity (PerkinElmer) or NIS Elements (Nikon), and (in some cases) multiple channels were overlaid in Adobe Photoshop. To image dissected gonads, spermatocytes, and sperm, adult males or mated hermaphrodites were dissected into sperm media containing dextrose (50 mM Hepes, 1 mM MgSO4, 25 mM KCl, 45 mM NaCl, 5 mM CaCl2, 10 mM dextrose).
To measure the onset of epidermal leakage in peel-1-affected embryos, pre-arrest embryos were dissected from the following crosses. Hermaphrodite-sired embryos were dissected from (i) self-fertilizing, zeel-1(tm3419)peel-1(+)/niDf9 hermaphrodites and (ii) self-fertilizing, zeel-1(tm3419)peel-1(+)/zeel-1(tm3419)peel-1(+) hermaphrodites. Male-sired embryos were dissected from niDf9/niDf9 hermaphrodites mated to three types of males: (i) zeel-1(tm3419)peel-1(+)/niDf9; (ii) zeel-1(tm3419)peel-1(+)/zeel-1(+)peel-1(+); and (iii) zeel-1(tm3419)peel-1(+)/zeel-1(+)peel-1(+); oxSi19[peel-1(+)]/+. The self-fertilizing hermaphrodites were aged 24 h post-L4 at the time of dissection, and mated hermaphrodites were aged 24–48 h at the time of dissection. After dissection, embryos were imaged every 10 min, starting before the 1.5-fold stage and ending 7 or more hours after the 1.5-fold stage. The onset of epidermal leakage was calculated as the time between the 1.5-fold stage and the first frame in which leakage was observed. Calculations were truncated at 7 h past the 1.5-fold stage because this represents 1 h after the average hatching time of wild-type embryos. Finally, embryos were binned into 30-min intervals in order to generate the inverted histograms shown in Figure 4A.
To calculate the percentage of peel-1-affected embryos elongating past 2-fold, zeel-1(tm3419)peel-1(+)/niDf9 hermaphrodites were isolated at the L4 stage and allowed to age for 24, 48, 60, and 72 h. Embryos were then dissected and imaged every 20 or 30 min for at least 10 h.
Anti-PEEL-1 is a rabbit polyclonal generated against the C-terminal 15 amino acids of PEEL-1. This antibody was generated and purified by GenScript, Piscataway, NJ. 1CB4 is a mouse monoclonal used to stain FB-MOs [64]. 1CB4 was a gift from Steven L'Hernault. To stain sperm, adult males were dissected into sperm media on charged slides, freeze-cracked in liquid nitrogen, and fixed overnight in −20°C methanol. Slides were washed with PBST (PBS+0.1% Triton-X 100), blocked for 30 min with PBST+0.5% BSA, and incubated for 4 h with anti-PEEL-1 (1/100) and 1CB4 (1/2,000), diluted in PBST+0.5% BSA. Slides were then washed three times in PBST and incubated for 2 h with Alexa568-labeled anti-mouse (1/500) and Alexa488-labeled anti-rabbit (1/500) (Invitrogen), diluted in PBST+0.5% BSA. Slides were washed again three times in PBST and mounted in Vectashield mounting media with DAPI.
To visualize actin filaments in peel-1-affected embryos, embryos were stained with Alexa568-labeled Phalloidin (Invitrogen) according to Protocol 7 in [65]. To visualize all the other proteins, embryos were stained with monoclonal antibodies MH2 (perlecan), DM5.6 (myosin heavy chain A), MH5 (VAB-10A), and MH4 (intermediate filaments). All monoclonals were obtained from The Developmental Studies Hybridoma Bank, Iowa City, Iowa. For these experiments, embryos were fixed for 10 min in 3% paraformaldehdye, freeze cracked in liquid nitrogen, and incubated for 5–10 min in −20°C methanol. Embryos were then washed three times in PBST and incubated overnight with the primary antibody diluted in PBST+1% BSA. MH2, MH4, and MH5 were diluted 1/150, and DM5.6 was diluted 1/1,000. Embryos were washed three times in PBST and incubated overnight with Alexa488-labeled anti-mouse (1/500) (Invitrogen), diluted in PBST+1% BSA. Embryos were washed again three times in PBST and mounted in Vectashield mounting media with DAPI.
Separate phylogenetic trees were built for (i) zyg-11 and all zyg-11 homologs in C. elegans and (ii) zyg-11 and all zyg-11 homologs in C. elegans, C. briggsae, C. remanei, and C. japonica. zyg-11 homologs were defined as all genes carrying the zyg-11-like leucine-rich repeat region. After removing the predicted transmembrane domains of ZEEL-1, Y71A12B.17, and Y55F3C.9, all protein sequences were aligned using MUSCLE [66]. The alignments were performed using the BLOSUM30 substitution matrix, a gap open penalty of −10, and a gap extend penalty of −1. The C. elegans–only alignment was trimmed to exclude residues having gaps in more than 90% of sequences, and the multi-species alignment was trimmed using the heuristic method, automated1, from TrimAL [67], which is optimized for maximum likelihood tree construction. Finally, phylogenetic trees were constructed using PhyML [68], using the LG substitution model [69], zero invariant sites, and four substitution rate categories. Branch support was determined using bootstrap sampling with 100 replicates.
Divergence between zeel-1 and Y71A12B.17 was determined by aligning the two proteins with MUSCLE [66], trimming the alignment of gaps, and using PAML [70] to calculate synonymous site divergence on the corresponding nucleotide sequences. (The total length of gaps was less than 0.1% of the length of the total alignment.) The summary statistics are as follows: number of synonymous sites = 715.5; number of non-synonymous sites = 2,002.5; synonymous substitutions per site (dS) = 1.0709; non-synonymous substitutions per site (dN) = 0.3267.
Adults and larvae were heat-shocked by submerging sealed agar plates in a 34°C water bath for 1 h. Embryos were heat-shocked by mounting embryos on an agar pad, incubating the slide at 19–20°C for the prescribed number of hours before placing the slide on the floor of a sealed, 1 cm×8 cm×8 cm plastic box, and submerging the box in a 34°C water bath for 20 min. After heat-shock, embryos were imaged every 20 min for at least 10 h. Initially, embryos were staged directly by collecting and mounting four-cell embryos. Later, when it became clear that the vast majority of heat-shocked embryos developed to the 2-fold stage without defects or delay, throughput was increased by collecting mixed stage embryos and mounting, incubating, and heat-shocking as above. These embryos were staged relative to the time at which they initiated elongation and by comparing their morphology before heat-shock to images of embryos that had been staged using the direct method.
To quantify peel-1-mediated killing of the egg-laying muscles and the anal depressor muscle, the Pexp-3::peel-1 arrays were crossed to SJ4157, which carries an integrated array of the muscle marker, Pmyo-3::GFP. In 1-d-old, F1 hermaphrodites, two of the four egg-laying muscles and the single anal depressor muscle were observed and classified as live or dead based on expression of GFP. Live egg-laying muscles were classified as morphologically normal or atrophied, and both live egg-laying muscles and live anal depressor muscles were then classified as mCherry+ or mCherry−, indicating that they had or had not inherited the Pexp-3::peel-1 array. For each cell type, the percent of cells killed by the arrays was calculated assuming that all dead cells had inherited the array. In addition, each F1 animal was classified as constipated if it contained bacteria in the posterior intestine and as egg-laying defective if it contained 3-fold embryos in the uterus.
To quantify peel-1-mediated killing of GABA neurons, the Punc-47::peel-1 arrays were crossed to EG1285, which carries an integrated array of the GABA-neuron marker, Punc-47::GFP. GABA-neurons were observed in F1 hermaphrodites, and peel-1-mediated killing was quantified as above. Typically, the DVB neuron and five to nine ventral cord neurons were scored per hermaphrodite. In addition, each animal was classified as mosaic or non-mosaic based on expression of the co-injection marker, Prab-3::mCherry.
|
10.1371/journal.pntd.0002322 | rKLO8, a Novel Leishmania donovani – Derived Recombinant Immunodominant Protein for Sensitive Detection of Visceral Leishmaniasis in Sudan | For effective control of visceral leishmaniasis (VL) in East Africa, new rapid diagnostic tests are required to replace current tests with low sensitivity. The aim of this study is to improve diagnosis of VL in East Africa by testing a new antigen from an autochthonous L. donovani strain in Sudan.
We cloned, expressed and purified a novel recombinant protein antigen of L. donovani from Sudan, designated rKLO8, that contains putative conserved domains with significant similarity to the immunodominant kinesin proteins of Leishmania. rKLO8 exhibited 93% and 88% amino acid identity with cloned kinesin proteins of L. infantum (synonymous L. chagasi) (K39) and L. donovani (KE16), respectively. We evaluated the diagnostic efficiency of the recombinant protein in ELISA for specific detection of VL patients from Sudan. Data were compared with a rK39 ELISA and two commercial kits, the rK39 strip test and the direct agglutination test (DAT). Of 106 parasitologically confirmed VL sera, 104 (98.1%) were tested positive by rKLO8 as compared to 102 (96.2%) by rK39. Importantly, the patients' sera showed increased reactivity with rKLO8 than rK39. Specificity was 96.1% and 94.8% for rKLO8- and rK39 ELISAs, respectively. DAT showed 100% specificity and 94.3% sensitivity while rK39 strip test performed with 81.1% sensitivity and 98.7% specificity.
The increased reactivity of Sudanese VL sera with the rKLO8 makes this antigen a potential candidate for diagnosis of visceral leishmaniasis in Sudan. However, the suitability at the field level will depend on its performance in a rapid test format.
| Visceral leishmaniasis (VL) is an infectious disease caused by the Leishmania donovani complex including Leishmania donovani in East Africa and India and by Leishmania infantum in the Mediterranean area and Latin America. Clinical diagnosis of VL in East Africa is difficult as maladies with similar symptoms are endemic. For this reason, reliable diagnosis of VL is extremely important. However, tests based on antibody reaction with rK39 are not sensitive in East Africa most likely due to the genetic diversity of different Leishmania species. In this study, we cloned and expressed a new antigenic protein (rKLO8) of L. donovani strain originating from Sudan. Sequence analysis confirmed that KLO8 differs from other kinesin proteins of Leishmania. We thus tested and compared the performance of rKLO8 with rK39 and other commercial tests for VL diagnosis in Sudan. Our data show that sera of VL patients reacted stronger with rKLO8 than rK39, suggesting improved diagnosis of patients with low antibody titres.
| Visceral leishmaniasis (VL) is a protozoan parasitic diseases caused by members of the Leishmania donovani (L. d) complex that includes L. d. donovani in East Africa and the Indian subcontinent, L. d. infantum in Europe and North Africa and L. d. chagasi in Latin America [1], [2]. However, recent molecular and enzymatic studies revealed that L. chagasi is synonymous with L. infantum [3], [4]. Visceral leishmaniasis is still a major health problem with approximately 0.2–0.4 million new cases annually [5]. The majority of infections (90%) occur in countries like India, Bangladesh, Sudan, South Sudan, Brazil and Ethiopia with East African countries having the second highest disease burden after the Indian continent [5].
Sudan has the highest number of reported cases in East Africa [5] where the disease is endemic since the early 1900s, in particular in eastern and central regions [6]. The diagnosis of VL in Sudan is difficult, as the disease is endemic in rural areas with no or little access to medical facilities. Detection of Leishmania amastigotes in tissue aspirates is still used for confirmation of the disease in Sudan although it is invasive and of low sensitivity. Diagnosis is further hindered, as the disease sometimes appears with atypical clinical pictures [7] which need confirmation by laboratory tests. Due to high fatality and toxicity of commonly used drugs [8], [9], diagnostic tests have to be of high accuracy.
Commercially available rapid tests are either based on the rK39 of L. infantum (synonym. L. chagasi) [10] or rKE16 of L. donovani [11]. Field tests based on rK39 are used in several countries with high reliability [12]–[17]. However, the low sensitivity in Sudan limits its use in this region [7], [18], [19], [20]. New rapid tests based on the recombinant protein rKE16 from an Indian strain of L. donovani have shown similarly high sensitivity and specificity compared to the classically used rK39-based tests in India [11], [21]. A multiregional study with five different rapid tests based on either rK39 or rKE16 demonstrated equal performance with high sensitivity (92.8–100%) in India [20]. However, sensitivity was significantly lower (36.8–92%) in Brazil and East Africa [20].
The direct agglutination test (DAT), which detects antibodies against whole L. donovani promastigotes, has proven to be a useful tool for diagnosis of VL in several countries including Sudan [22]–[27]. The stability of DAT was improved by using freeze-dried and glycerol preserved antigens which does not require storage at 4°C thus making the test suitable for field application [28], [29]. However, the test procedure and the need for overnight incubation give limitations for the field use.
Here, we report identification, expression and testing of a new immunodominant protein (rKLo8) from an autochthonous L. donovani strain in Sudan. For detection of VL antibodies in patients and controls from Sudan, an rKLo8 ELISA was developed and compared with the rK39 ELISA and two commercial kits.
The L. donovani reference strain Lo8 was kindly provided by Prof. Bernhard Fleischer, Bernhard Nocht Institute for Tropical Medicine (BNITM), Hamburg. The strain was originally isolated in Sudan from a confirmed case of visceral leishmaniasis. The parasite was maintained in RPMI-1640 supplemented with L-glutamine, NaHCO3 (Sigma-Aldrich) and 10% (v/v) fetal calf serum (Sigma-Aldrich).
Sera used in this study were collected in the rural hospital of Doka, Eastern State of Sudan [30], [31]. All patients and controls have given consent for participation in the study. Tests and experiments with patients' sera and Leishmania strains were anonymized. The study was approved by the Ethical Review Committee of the Federal Ministry of Health in Sudan and by the Regierungspräsidium Gießen, Germany.
A total number of 183 human serum samples were obtained from the serum bank at the Biomedical Research Laboratory, Ahfad University (Omdurman-Sudan). The majority of samples (106) were from VL patients with confirmed lymph-node aspiration, 30 from healthy individuals resident at Doka village (an endemic area for VL) and 20 from healthy people living in the non-endemic area of Omdurman city, 11 from confirmed malaria cases, 10 from patients with diagnosed pulmonary tuberculosis and 6 from leukaemic patients. Diagnosis of VL was done at the rural hospital in Doka by two expert laboratory technicians and sera were collected only from patients older than 2 years. Sera of patients and diseased controls were collected before administration of treatment. All sera were stored frozen (−20°C) at the Biomedical Research Laboratory. Sera from healthy controls were used to determine the ELISA cut off value. Sera were tested blindly without knowing their clinical status or results of lymph node smears.
The partial gene fragment encoding the immunodominant repeats of L. donovani, designated KLO8, were amplified from promastigote genomic DNA using the forward (5′-GAGCTCGCAACCGAGTGGGAGG–3′) and reverse (5′- GCTCCGCAGCGCGCTCC–3′) primers, designed according to the published L. chagasi gene for kinesin-related protein (GenBank: L07879.1). PCR reaction was performed using Phusion High-Fidelity DNA Polymerase (FINNZYMES OY, Finland) in a total volume of 50 µl, containing 3% (v/v) DMSO, 10 µl HF buffer, 10 mM dNTPs mix-OLS (OMNI life science) and 100 ng genomic DNA. PCR conditions were as follows: denaturation at 98°C for 30 s followed by 25 cycles of denaturation at 98°C for 10 s, annealing at 65°C for 20 s, and extension at 72°C for 20 s. Amplified products revealed multiple bands with sizes equivalent to 117 bp repeats. The largest amplification product (883 bp) was gel purified, digested with EcoRV and cloned (according to the manufacturer's instructions) into the plasmid vector pcDNA3.1(+) (Invitrogen life technologies, USA) generating the non-tagged KLO8 construct, pcDNA/KLO8. The sequence was confirmed by restriction digestion with BamHI and XbaI (Fermentas GmbH, Germany) and by sequence analysis at Seqlab-Sequence Laboratories, Göttingen GmbH. Each insert was sequenced at least twice.
For expression and purification of the recombinant protein, KLO8 was subcloned into the His-tag vector pQE41 (Qiagen GmbH, Germany). The DNA construct pcDNA/KLO8 was used as template with the forward (5′-GTGGAATTCTGCAGATGGATCCATGGAGCTCGCAACC–3′) and reverse (5′-GCCGCCACTGTGCTGGATGTCGACGCTCC–3′) primers, designed to introduce restriction sites for the enzymes BamHI and SalI (underlined). Amplification was performed using Phusion Hot Start II DNA Polymerase (Thermofisher Scientific, USA) as recommended by the manufacturer. Amplified DNA fragments were digested with the same restriction enzymes and cloned in-frame and down stream of 6× His-tag into the corresponding sites of the vector pQE41 to generate the plasmid construct carrying the target gene, named as pQE41/KLO8. The recombinant plasmid was verified by DNA sequencing and restriction analysis and then transformed into competent M15 E.coli cells (Qiagen GmbH, Germany). E. coli were grown at 37°C in Luria-Bertani (LB) medium containing 100 µg/ml ampicillin (Sigma-Aldrich, Germany) and 25 µg/ml kanamycin (Sigma-Aldrich, Germany) to a optical density (OD600) of 0.8. Recombinant protein expression was induced by adding 1 mM isopropyl β-D-thiogalactoside (IPTG; Roth, Germany) for 4 hours. E. coli cells were harvested by centrifugation at 3340 g for 10 min at 4°C. Bacterial pellets were then lysed in PBS (pH 7.4) containing 0.25 mg/ml lysozyme (Roth, Germany), 25 U/ml benzonase nuclease (Novagen, Germany), 10 mM imidazol (Roth, Germany), 1 mM PMSF (Sigma, USA) and 2 µM β-mercaptoethanol (Sigma, USA). Subsequently, bacterial lysates were sonicated 6 times (Bandelin Sonorex, Germany) on ice for 10 seconds each with >10 seconds rest and stored at −20°C. The rKLO8 was expressed as 6× His-tagged His-rKLO8 fusion protein and was recovered in the soluble fraction of the bacterial lysate by SDS-PAGE. Purification was carried out using nickel nitrilotriacetic (Ni-NTA) columns (Qiagen GmbH, Germany). The supernatant was loaded into a Ni-NTA column, which was pre-equilibrated with PBS, pH 7.4, containing 10 mM imidazole, 1 mM PMSF and 2 µM β-mercaptoethanol. The recombinant protein was eluted with the same buffer containing 400 mM imidazole. Salts and imidazole were removed by dialysis in PBS buffer. Protein concentration was determined using the Bradford assay compared to bovine serum albumin (BSA) as standard. Protein aliquots were kept frozen at −80°C.
The deduced amino acid sequence of the plasmid insert, determined with the ExPASy Proteomics Server of the Swiss Institute of Bioinformatics (http://web.expasy.org/translate/), was compared with published sequences obtained from the National Centre for Biotechnology Information (http://www.ncbi.nlm.nih.gov/). Immunodominant repeats of KLO8 (294 AA), K39 (252 AA) and KE16 (155 AA) were aligned using the ClustalW2-Multiple Sequence Alignment program (http://www.ebi.ac.uk/Tools/msa/clustalw2/). Homology search was performed with BLASTP 2.2.1 in 25.07.2012. Two different clones were analysed and found to contain the same insert (883 bp). Tandem Repeats Finder (http://tandem.bu.edu/trf/trf.html) [32] was used to locate and display tandem repeats in DNA sequences.
The recombinant protein rKLO8 was loaded on a 12% SDS-PAGE under denaturing conditions [33] using fractions of bacterial cell lysates or the purified protein and stained with Coomassie Brilliant Blue G250 (Merck KGaA, Germany). Proteins were transferred to a nitrocellulose transfer membrane (Whatman GmbH, Germany) using the Bio-Rad Semi-dry Trans-Blot at 200 mA for 1 hr. The membrane was blocked with 5% BSA (w/v) in 100 mM NaCl, 0.05% Tween 20 (v/v) and 10 mM Tris-HCl, pH 7.4 (blocking buffer) and subsequently incubated for 18 hrs at 4°C with sera from patients – or healthy controls, diluted 1∶1000 in blocking buffer. After washing, blots were incubated for 1 hr at room temperature (R/T) with Peroxidase-conjugated Donkey Anti-Human IgG (H+L) (Jackson Immunoresearch Laboratories, USA) diluted 1∶10000. The protein bands were revealed with Maximum Sensitivity Substrate system (Thermo Scientific, USA).
The recombinant lipoprotein antigen rK39 of L. infantum (synonymous L. chagasi) was purchased from Rekom Biotech, S.L., Granada Spain. It contains repetitive immunodominant epitopes of kinesin-related protein. It was expressed as 6× His-tagged His-rK39 fusion protein at the C-terminus of the kinesin-related protein of L. chagasi with 100% identity with the accession number AAA29254.1. Upon receipt, the protein concentration was verified with the same method used to measure the recombinant protein rKLO8 (Bradford). Aliquots were kept at −80°C.
The optimal protein concentration and serum dilutions were determined using pooled sera from 10 VL patients from Sudan and 10 control sera from non-endemic areas in Sudan. To select conditions which best discriminate between positive and negative sera, different protein concentrations were titrated against serial dilutions of positive or negative sera. High protein-binding capacity polystyrene 96 ELISA plates (NUNC TM Serving Life Science, Denmark) were used. Protein concentrations of 5 ng/well to 50 ng/well were tested for coating ELISA plates overnight at 4°C in 0.1 M NaCO3 buffer, pH 9.6. Plates were washed with PBS containing 0.05% (v/v) Tween-20 and then blocked with 3% (w/v) BSA, in the same buffer, at R/T for 1–2 hours. After additional washes, 50 µl diluted positive or negative serum samples were added to each well, and plates were incubated at R/T for 45 minutes. After washing, 50 µl/well Peroxidase-conjugated AffiniPure Donkey Anti-Human IgG (H+L) (Jackson Immunoresearch Laboratories, USA), diluted 1∶10000, were added to each well and plates were incubated at R/T for further 1 hr. The reaction was visualised with hydrogen peroxide and tetramethylbenzidine (R&D Systems, USA). The reaction was stopped with 2N sulfuric acid after 10 minutes incubation in the dark. The optical density (OD) was measured at 450 nm using an ELISA microreader (FLUOstar Omega, BMG LABTECH). Each sample was tested in duplicates and the mean was taken. Samples reaching invalid or inconsistent results were repeated. As control, the pooled positive and negative sera were included in each plate, when testing individual sera.
Individual IT LEISH dipstick kits using the recombinant K39 antigen [10] for detection of human visceral leishmaniasis antibodies were purchased from Bio-Rad, France. The test was performed and interpreted as recommended by the manufacturer. Sera were considered positive when a dark purple control band appeared. Samples with invalid results were repeated.
The DAT (ITMA-DAT/VL) kits (Lot 11D1B1) were purchased from the Institute of Tropical Medicine, Antwerp-Belgium (ITMA). The antigen is a freeze-dried suspension of trypsin-treated, fixed and stained promastigotes of L. donovani strain 1-S [34], [35]. The test was performed in 96 V-shape microplates (Greiner Bio One, Germany) according to the manufacturer's instructions. Besides internal controls, positive and negative pooled sera were included in each plate and results were read after overnight incubation at R/T. Samples with titres of 1∶≥3200 serum dilutions were considered positive, whereas samples with titres of 1∶800 and 1∶1600 serum dilutions were considered as borderline and were repeated.
Data were analyzed using the GraphPad Prism software (GraphPad Prism Inc., San Diego, Ca). Significance of antibody responses was assessed using Student t test or one way ANOVA test. Differences of p values <0.05 were considered significant. Cut off values for each recombinant protein were defined as mean absorbance values of 30 sera of healthy controls from Sudan plus 3 standard deviations (SD). Sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were calculated to assess usefulness of the diagnostic assays at 95% confidence intervals [36].
Nucleotide and amino acid sequences of L. donovani KLO8 have been deposited in GenBank under the accession numbers KC788285 and AGL98402, respectively. Accession numbers of other Leishmania kinesins were L07879.1, L07879.1, AAA29254.1, AAT40474.1, ABI14928.1, ADR74368.1 and AY615886.1.
A plasmid encoding a partial gene of the immunodominant kinesin protein L. donovani (KLO8) was constructed and subsequently subcloned to generate the expression vector pQE41/KLO8. Sequence analysis of the cloned fragment revealed a partial open reading frame (lacking the ATG) of a 883-basepair product. Tandem repeat analysis identified one repeat of 117 bp with 6.3 copies encoding 39 AAs. The DNA construct was further confirmed by restriction analysis using BamH I and Sal I resulting in an approximately 883-bp product. KLO8 encodes a protein of 294 amino acids with a predicted molecular mass of 32.4 kDa and an isoelectric point (IP) of 4.39. Homology search in the protein database showed that KLO8 contains putative conserved domains of high similarity with kinesin proteins of Leishmania. KLO8 exhibited 93% and 88% amino acid identities with the kinesin proteins K39 of L. infantum (synonym. L. chagasi) strain BA-2 from Brazil (GenBank: AAA29254.1) and KE16 of L. donovani strain KE16 from India (GenBank: AAT40474.1), respectively. Of interest, KLO8 exhibited 97% identity with the kinesin protein Ldk39 of L. donovani 1S-CL2D from Sudan (GenBank: ABI14928.1), which however was never processed for development of a diagnostic test. Moreover, BLAST analysis showed 79% identity with the K28 fusion protein (GenBank: ADR74368.1), a synthetic protein construct derived from L. donovani [37]. The 756 bp immunodominant repeats of K39 (GenBank L07879.1) contains 6.4 copies of 117 bp encoding 252 AA. In contrast, KE16 (GenBank AY615886.1) showed only 4 copies of the 117 bp repeat encoding 155 AA. To identify differences in the immunodominant epitopes in the three Leishmania antigens, KLO8 was aligned with the AA repeats of K39 and KE16 using ClustalW2-Multiple Sequence Alignment (http://www.ebi.ac.uk/Tools/msa/clustalw2/. As shown in Fig. 1, immunodominant epitopes in the 3 antigens display presence of non-conservative amino acids. Differences in AA composition were highlighted in black and identical regions were left unmarked. These results confirm the variability in immunodominant repeats of Leishmania kinesin-related proteins.
The KLO8 was expressed as His-tagged recombinant protein in M15 E. coli and expression was confirmed by SDS-PAGE. As shown in Fig. 2A (lane 2 & 3), the apparent molecular weight of the His tagged fusion protein was 35 kDa. The reactivity of purified recombinant protein rKLO8 was assessed in Western blot analysis using pooled sera from 10 VL patients or 10 healthy controls. As shown in Fig. 2B, the positive sera recognized the recombinant protein (lane 2 & 3), while the negative sera did not (lane 1). These results demonstrate that the recombinant protein rKLO8 is suitable for specific detection of Leishmania antibodies.
We next established an indirect IgG ELISA system using the recombinant protein rKLO8 for detection of Leishmania-specific antibodies in patients' sera. As shown in Fig. 3, all tested protein concentrations (50-5 ng/well) were recognized by pooled VL sera and did not cross-react with pooled sera from the healthy individuals. ODs of positive sera were at least 4 fold higher compared to negative sera, however this ratio changed to much higher values with more diluted sera. Coating ELISA plates with a concentration of 5 ng rKLO8 protein was sufficient for positive detection of sera from VL patients diluted up to 1∶25600. As a result of these titrations, a protein concentration of 5 ng/well and serum dilutions of 1∶800 were selected as standard conditions in subsequent experiments. In some experiments, VL sera with negative results at 1∶800, were re-tested at a serum dilution of 1∶100.
Reactivity of the two recombinant proteins rKLO8 or rK39 was evaluated in ELISA using individual human VL (n = 106) and control (n = 77) sera from Sudan. The recombinant protein rK39, obtained from Rekom Biotech, was expressed as 6× His-tagged fusion protein in E. coli. To ensure similar conditions, rKLO8 was also expressed as 6× His-tagged protein in E. coli. Sera of patients were diluted at 1∶800 and tested on a protein concentration of 5 ng/well (Fig. 4A). Quantitative analysis of antibodies in VL sera to both recombinant proteins demonstrated significantly higher antibody levels than those of control subjects (P<0.0001) although absorbance values among the patients' sera varied depending on the recombinant proteins. In general, sera tested on rKLO8 yielded higher OD values than on rK39 with a mean value of 1.12±0.97 for rKLO8 and 0.93±0.77 for rK39. In addition, sensitivity of rKLO8 was also increased with 92.5% (98/106) for rKLO8 versus 86.8% (92/106) for rK39. Notably, none of the healthy or diseased controls (n = 77) showed cross-reaction with either of the recombinant proteins (Fig. 4A). VL sera that were negative on rK39 or rKLO8 (n = 14) were then compared to control sera (n = 77) at 1∶100 serum dilution. As shown in Fig. 4B, re-testing on rKLO8 yielded increased positive detection of VL patients (12/14) as compared to rK39 (10/14) at cut-off values of 0.41 and 0.32 for rKLO8 and rK39, respectively. In addition, control sera tested on rKLO8 revealed less cross-reactivity as compared to rK39. Both proteins showed cross-reactivity with 3 sera from malaria patients and in addition rK39 showed false positivity of one healthy endemic individual.
Using the same panel of VL and control sera, the results obtained from the rKLO8 and rK39 ELISA were next compared with two commercial diagnostic kits, the rK39 strip test (Bio-Rad) and a freeze-dried version of DAT (ITMA-DAT/VL). As shown in Table 1, the overall sensitivities of rKLO8 (98.1%) and rK39 (96.2%), measured by ELISA, were higher than those of the rK39 strip test (81.1%) and DAT (94.3%). With respect to specificity, the rKLO8- and rK39 ELISA showed equally high performance (96.1% and 94.8%, respectively) but was slightly lower than DAT (100%) and rK39 strip test (98.7%). Accordingly, the PPVs and NPVs were 97.2% and 97.4% for the rKLO8 ELISA, 96.2% and 94.8% for the rK39 ELISA, 98.9% and 79.2.9% for the rK39 strip test and 100% and 92.8% for the DAT, respectively. Interestingly, results of the four tests showed some discrepancies. Although tested positive in the rKLO8 ELISA, 6 (5.7%) sera of the confirmed VL patients were negative (1∶<1600) in DAT (Fig. 5A, Table 1). In addition, sera of 6 patients had weak DAT titres (1∶3200–1∶6400) (Fig. 5A). On the other hand, while being positive in DAT, 4 (3.8%) or 2 (1.9%) sera of VL patients were not detected by rK39 or rKLO8, respectively (Table 1). Those 4 cases were also negative in the strip test. However, VL sera with positive or negative DAT results reacted similarly with rKLO8 (Fig. 5B), suggesting that rKLO8 or DAT monitor different immune reactivities. Thus, a combination of both rKLO8 ELISA and DAT provides 100% sensitivity for detection of VL. In addition, the rKLO8 ELISA detected all VL sera that were positive in the rK39 strip test (Fig. 5C), but sera negative in the strip test displayed still low antibody reactivity when tested with rKLO8 (p<0.0001) (Fig. 5D).
Despite the availability of several recombinant proteins for serodiagnosis of VL, commercially available rapid tests are still based mainly on the rK39. While these tests are quite effective in diagnosing VL in Brazil and Indian subcontinent, their use in East-Africa is not satisfactory. Despite the development of freeze dried DAT test based on Sudanese L. donovani, there is little interest in developing new rapid tests for VL in East Africa [38]. Improving VL diagnosis in these countries requires identification and testing of new antigens from autochthonous strains of Leishmania. Here, we aimed to clone, express and test a new recombinant protein of Sudanese L. donovani termed rKLO8 that shows homology with kinesin proteins of Leishmania. As expected, rKLO8 shows high sequence identity with the LdK39 protein of L. donovani 1S, a strain from Sudan which however has never been further used for a diagnostic procedure. Sequence analysis of KLO8 confirmed that the AA compositions of the immunodominant kinesin proteins of Leishmania show variability even among strains from the same region [39], [21]. The heterogeneity of kinesin immunodominant epitopes may explain why the use of rK39 and rKE16 is not sufficient to provide reliable diagnosis in the different endemic regions. Recently, the genetic diversity in the immunodominant kinesin repeats in strains of L. donovani and L. infantum has been documented [40].
Diagnostic methods with improved sensitivity for VL in Sudan are needed to replace low sensitive tests based on the kinesin of L. chagasi (rK39). Thus, a new rapid test based on the recombinant K28 protein has been developed and first data show promising results concerning serodiagnosis of VL patients in Sudan [37]. Our results with rKLO8 show also increased reactivity with patients sera as compared to rK39 ELISA. In addition, we (unpublished data) and others [37] have shown that some VL patients from Sudan have decreased immune responses to rK39, which explains the low sensitivity of rK39-based diagnostic tests in this region. This is in accordance with our finding that VL sera with negative rK39 strip test results show low but significant immune responses to rKLO8 (Fig. 5C and 5D). Thus, the increased reactivity of rKLO8 may provide enhanced detection of VL sera with low antibody titres. Here, we also show that the rKLO8 ELISA is more sensitive than the DAT (94.3%) and rK39 strip test (81.1%) confirming the low sensitivity of rK39 strip test in Sudan. Our data also show that difficulties of VL diagnosis in certain geographical areas might only be overcome, if the detection system is based on antigens derived from autochthonous parasites originating from the same endemic area [41], [42].
Antigenic variation due to parasite diversity has been proposed to be the cause for the low diagnostic sensitivity of VL diagnostics based on rK39 [20], [21]. In addition, the complexity and specificity of the humoral immune response against Leishmania parasites plays a crucial role in determining the outcome of serological assays. As a result, specific immune responses against Leishmania may be lost completely when tested against parasites isolated from different endemic area [43]. However, it cannot be generalized that antigens of endemic VL strains always result in improved diagnostic sensitivity, as sera of VL patients from Bangladesh reacted equally well with rK39 and rKRP42 derived from L. donovani from Bangladesh, despite marked heterogeneity between the two proteins [44].
An ideal diagnostic test should identify all positive sera without cross-reacting with negative sera. Our data show that none of the serological test used was able to detect all VL cases from Sudan. Only the combination of rKLO8 ELISA and DAT resulted in 100% diagnostic sensitivity. As antibodies of different specificities are detected, we recommend to combine rKLO8 ELISA and DAT to overcome the lower sensitivity in Sudan. Combination of different tests has also been suggested to overcome the problem of low sensitivity in East Africa [20], as detection of immune responses directed against different antigens is expected to improve sensitivity of an assay [43].
We have to be aware that recently infected persons have elevated IgM responses but not yet mounted an IgG response. Sera of such patients would give false negative results, if tested in an ELISA based on detection of IgG antibodies. This could explain the results of 2 patients that were tested negative in the rKLO8 ELISA and strip test despite strong positivity in DAT, which detects different antibody subclasses. More difficult to interpret are those 6 confirmed VL cases which were negative in DAT despite detectable antibody responses to rKLO8. Again, antibody specificities and parasite diversity could play a role. In addition, low antibody titers of these patients may result from coinfection with HIV, which complicate VL diagnosis by serological tests [45].
Unfortunately, malaria and other diseases are common in VL endemic region of Africa and Asia [46], [47]. Thus, a good test system needs robust discrimination between VL and potential co-infections. Indeed, malaria is known to be a major cause of cross reactivity to rK39 [48]. Cross reactivity to rK39 has also been reported with healthy sera of endemic and non-endemic controls from Sudan [37]. Our data indicate that a serum dilution of 1∶800 provides optimal specificity and sensitivity for rKLO8. Sera of malaria patients did not give a signal in the rKLO8 ELISA. This is in accordance with a previous study from Sudan, where 100% specificity of VL detection has been shown with sera tested at dilutions of 1∶1600 [30]. However in general, results with low antibody titres should be interpreted with caution. Notably, the DAT kit showed no cross reaction with any of the control sera tested and thus provides best specificity.
In conclusion, rKLO8 is a novel recombinant protein of L. donovani with increased reactivity to VL sera from Sudan. To finally evaluate its performance, rKLO8 has to be formulated as rapid test and assessed in a comparative field study.
|
10.1371/journal.ppat.0030176 | Multiple Infections by the Anther Smut Pathogen Are Frequent and Involve Related Strains | Population models of host–parasite interactions predict that when different parasite genotypes compete within a host for limited resources, those that exploit the host faster will be selected, leading to an increase in parasite virulence. When parasites sharing a host are related, however, kin selection should lead to more cooperative host exploitation that may involve slower rates of parasite reproduction. Despite their potential importance, studies that assess the prevalence of multiple genotype infections in natural populations remain rare, and studies quantifying the relatedness of parasites occurring together as natural multiple infections are particularly scarce. We investigated multiple infections in natural populations of the systemic fungal plant parasite Microbotryum violaceum, the anther smut of Caryophyllaceae, on its host, Silene latifolia. We found that multiple infections can be extremely frequent, with different fungal genotypes found in different stems of single plants. Multiple infections involved parasite genotypes more closely related than would be expected based upon their genetic diversity or due to spatial substructuring within the parasite populations. Together with previous sequential inoculation experiments, our results suggest that M. violaceum actively excludes divergent competitors while tolerating closely related genotypes. Such an exclusion mechanism might explain why multiple infections were less frequent in populations with the highest genetic diversity, which is at odds with intuitive expectations. Thus, these results demonstrate that genetic diversity can influence the prevalence of multiple infections in nature, which will have important consequences for their optimal levels of virulence. Measuring the occurrence of multiple infections and the relatedness among parasites within hosts in natural populations may be important for understanding the evolutionary dynamics of disease, the consequences of vaccine use, and forces driving the population genetic structure of parasites.
| Infections of one host individual by multiple genotypes of a parasite occur in many natural systems and have major consequences on the evolution of disease severity. Under such multiple infections, the parasite genotypes compete for the host's limited resources, and the faster exploiters will be advantaged over more prudent genotypes, selecting for parasites that cause greater damage and mortality, i.e., having higher virulence. However, when different parasite genotypes within a host are related, a reduction of competitive conflicts between them should lead to more cooperative host exploitation, and thus to lower severity of disease. The occurrence of multiple infections and the relatedness among parasite genotypes within hosts therefore are important to our understanding of diseases, but studies that assess these parameters in nature remain scarce. We investigated multiple infections by the fungus Microbotryum violaceum, responsible for the anther smut disease on the plant Silene latifolia. Multiple infections in natural populations were extremely frequent, with many different genotypes within single host plants. The fungal genotypes found in the different stems of single plants were, however, more related than expected by chance. Together with previous artificial inoculation experiments, these results suggest that M. violaceum actively excludes dissimilar genotypes while tolerating closely related competitors.
| It is generally considered that a parasite's optimal level of virulence, i.e., the decrease in host fitness induced by disease that maximizes parasite transmission, depends on several factors, such as mode of transmission, dormancy ability, host availability, and the frequency of occurrence within single hosts of multiple infections by different parasite genotypes [1,2]. The occurrence of multiple infections is theoretically predicted to be an important determinant of virulence evolution because when different parasites compete for limited resources within a single host, more rapid exploitation strategies should win over prudent strategies, thus selecting for parasites that cause greater damage and mortality [2]. However, when related parasites share a host, kin selection should reduce between-genotype competition, resulting in shared host exploitation that is more cooperative [3]. This predicted consequence of kin selection among parasites has been supported by some empirical data [4–7]. The relatedness among parasites can also affect the prevalence of within-host competition if there is competitive exclusion that is dependent upon the genetic distance among parasites [8]. Our understanding of parasite virulence evolution would therefore benefit from measures of the frequency of multiple infections and the relatedness among parasite genotypes within hosts in natural populations. Multiple infections have been reported for several plant parasites [9], including a few diseases causing systemic infections [10,11], for which the competition for resources should be stronger because they invade their whole host. Multiple infections have also been reported for a few animal and human parasites [7,12–14]. In contrast, the relatedness between naturally infecting parasites within plants and animals in natural populations remains largely unexplored (but see [12,14] for human parasites).
In this study, we surveyed several natural populations of the Caryophyllaceae Silene latifolia infected by the anther smut fungus Microbotryum violaceum. The M. violaceum–S. latifolia system is an important biological model for the study of host–parasite evolution [8,15–18]. The fungus sterilizes the hosts by destroying ovaries and replacing pollen with fungal spores that are then dispersed by natural pollinators of the plant (Figure 1). Multiple infections have been observed in S. latifolia after experimental inoculation by several genotypes of M. violaceum [18,19] and in experimental populations [20], with different fungal genotypes segregating in different stems of the plant. Even more interestingly, experimental inoculations showed that sequential inoculations with more related genotypes led to a higher frequency of multiple infections than sequential inoculation with less related genotypes even though the different strains inoculated alone yielded similar levels of infection [8]. These experimental results strongly suggested that a mechanism of active exclusion exists whereby resident genotypes exclude challenge by distantly related parasites. However, studies in natural populations are required to assess whether this experimental phenomenon was relevant to natural infections.
We collected one flower bud per branch of every diseased plant in 12 roadside populations of S. latifolia in order to determine the frequency of multiple infections and to assess the degree of parasite relatedness within plants with multiple infections. For each bud, the parasite genotyped from one sporulating anther was tested by using six microsatellite markers and amplified fragment length polymorphism (AFLP).
Multiple infections were extremely frequent: 70% of the 190 diseased plants analyzed carried distinct parasite genotypes in different stems. On average, we detected 2.33 ± 0.5 genotypes per plant, and some plants yielded up to nine different parasite genotypes (Figure 2). In contrast, only a single genotype was detected per flower bud. In most buds, only a single allele was detected at each locus, in accordance with the high degree of homozygosity known to occur in M. violaceum [21], and we never detected more than two alleles. However, multiple genotypes within buds might have been missed because we only genotyped a single anther per bud, and experimental dilutions had shown that the limits of detecting a mixtures of two alleles by our PCR method was ratio of 1:10. We therefore also genotyped all the anthers of two flower buds from each of three stems from four infected plants in which several genotypes had been detected in different stems. Again, only a single genotype was detected in each bud. Thus, if multiple genotypes are present within buds, the minority genotypes can only account for less than 10% of the teliospores in all the anthers within a bud.
We found a significant correlation between the number of genotypes and the number of stems per plant (R2 = 0.37, p < 0.00001), indicating that plants with more stems are more likely to carry multiple fungal genotypes. This correlation is consistent with the findings that plants with more flowers are visited more often by pollinators [22] and that different branches from a given plant appear to be infected independently by distinct parasite genotypes in experimental field populations [19,20]. The distribution of the number of plants with a given number of genotypes deviated significantly from a Poisson distribution (p < 0.0001), with an excess of plants carrying many genotypes (Figure 2). Given that the number of genotypes per plant was positively correlated to the number of stems, this may reflect the distribution of the number of stems per plant, which also exhibited significantly more plants carrying many stems than would be expected from a Poisson distribution (p < 0.0001).
Regression estimates based on shared alleles [23] were used to characterize the relatedness among multiple genotypes infecting a given plant. Relatedness (r) is an estimate of the probability of gene sharing among individuals beyond the baseline probability set by the gene's frequency in the population. Values of r that are significantly greater than zero are considered to represent significant relatedness, whereas negative values result when individuals are less similar than expected by chance. Such measures of relatedness are typically used in all studies on social insects within social groups where kin selection is expected to act. A few studies have shown that animal parasites were more similar within hosts than between hosts by using Fst [12]. Regression estimates based on shared alleles [23] have two main advantages over Fst for estimating relatedness among parasites: (1) r estimates exactly the proportion of shared alleles between individuals compared to the probability of sharing alleles by chances given the allelic frequencies in the population. It is therefore a direct measure of exactly what is important for kin selection, namely, the probability that an altruistic individual shares the altruistic allele with the individual towards whom it interacts in an altruistic manner [23]. (2) Hamilton showed that the relatedness for diploids is equal to 2Fst/(1 + Fit) [24]. Fst used alone can be misleading because any additional similarity to self due to inbreeding (Fit) would be neglected.
Parasite genotypes within plants were significantly related, both when tested over all 12 populations (r = 0.38 ± 0.06) and when analyzed within ten of the 12 individual populations (Figure 3). One possible explanation for such high relatedness could be isolation by distance (i.e., spatial substructure of parasite genetic diversity, which results in correlations between genetic and spatial distances). Isolation by distance has been detected within roadside populations of M. violaceum on S. latifolia [21] and was also found again in this study (not shown). We therefore tested the possibility that the significant relatedness detected here was due to isolation by distance. To this end we calculated the relatedness of genotypes present in the geographically closest pairs of diseased plants that were only infected by a single genotype. The relatedness found between strains from neighboring singly infected hosts was significantly higher than zero (r = 0.06 ± 0.04) but much lower than r in multiply infected plants, which was 0.38 ± 0.06. This indicates that isolation by distance cannot be the sole phenomenon explaining the high relatedness among strains within plants.
A second possibility was that relatedness with multiply infected plants reflected expansion of nearly identical genotypes by selfing during parasite transmission from one branch to additional branches within the same plant, or due to crosses between multiple genotypes within plants. However, visual inspection of the microsatellite genotypes revealed that neither of these explanations were the case. Instead, most pairs of genotypes within single plants yielded three alleles for at least one locus, and the individual genotypes within a plant were most often homozygotic for different alleles at least at one locus. These observations are incompatible with selfing or crosses between resident strains, unless the crosses involved related strains. Of the 357 genotypes, only 41 could have been derived from selfing from another genotype in the same plant or from crosses between two other genotypes within the same plant. After excluding these 41 genotypes from the analysis, r between strains from multiply infected plants was still significant and very high (r = 0.35 ± 0.07). Thus, neither selfing nor crossing between resident strains seems to play a major role in the high relatedness found among parasites within plants.
We speculated that the high relatedness among genotypes within plants was due to active exclusion of genetically divergent parasite genotypes. This explanation is also strongly supported by previous experiments, which showed that sequential inoculation with more related genotypes led to a higher frequency of multiple infections than did sequential inoculation with more distant genotypes, even though similar levels of infection were obtained when strains were inoculated alone [8]. Other mechanisms, however, cannot be excluded and may contribute in addition to active exclusion, such as a higher mortality of plants carrying more divergent genotypes or a high fitness variance of the genotypes. If dissimilar genotypes had a higher average variance in fitness, one genotype might indeed predominate more rapidly than when the genotypes are more similar. These hypotheses will be tested in future work following experimental populations.
No significant correlation was found between the mean number of genotypes per plant and the size of the population. Instead, and somewhat surprisingly, we found a significant negative correlation over all 12 plant populations for the mean number of genotypes per plant versus the genetic diversity of the parasite population (Figure 4; r = −0.62, p < 0.03). This observation was surprising because a naive assumption might have been that multiple infections would be more frequent in populations with higher parasite genetic diversity. We argue instead that multiple infections in more diverse parasite populations occur less frequently due to active exclusion. In populations with higher numbers of parasite genotypes, the frequency of multiple infections is reduced due to a higher probability that a strain deposited on an infected plant is sufficiently genetically different from the resident genotype to be excluded. The negative correlation between the mean number of genotypes per plant and the genetic diversity of populations thus provides further support for relatedness-dependent competitive exclusions by M. violaceum.
Active genotype exclusion might result from the production of specific toxins against competitors [25] or from mechanisms of vegetative incompatibility, such as the death of hyphal cells after fusion of two genetically different genotypes [26]. Unfortunately, these mechanisms are difficult to test experimentally in Microbotryum, because hyphae cannot be cultured in vitro and hyphal interactions are not readily observable in host tissues. However, it is very likely that different strains within a plant come in close contact and can compete, even if we only detected them in different stems. When single strains are inoculated onto a plant, disease is only evident during the first year on the branch receiving the inoculation, and sometimes the branches next to the inoculated flower. However, all the flowers are usually diseased once new branches and flowers grow the next year. This is thought to reflect invasion of all the meristematic tissues of the plant by the fungus during the winter [27]. When a plant is infected by multiple strains, different genotypes are thus likely to compete directly for colonization of meristems during the following winter.
Virulence is expected to be higher in systems where multiple infections are frequent and faster rates of host exploitation lead to a transmission advantage [2]. In sterilizing parasites such as M. violaceum, whose major impact on host fitness is the prevention of reproduction, castration is also expected to be maximal [28]. Therefore, it does not appear surprising that M. violaceum exhibits very high virulence, and plants are usually completely sterile one year after the first infection. The frequent occurrence of multiple infections in this system may have been one of the factors selecting for high virulence. In contrast, relatedness-dependent competitive exclusion in M. violaceum appears to restrict multiple infections to more closely related genotypes and should thus lessen the severity of evolutionary conflicts between competing parasite genotypes. Such contrasting selective forces have resulted in the variable virulence dependent on within-host dynamics in other systems [29]. The host plants for M. violaceum can recover with substantial probabilities [30,31], which would be expected to result in variable virulence if the probability of dying before recovering varied between plants with single or multiple infections. For example, if the strains in multiple infections were to produce more spores, they might drain host resources more rapidly and make the plant die faster. Variability in teliospore production among strains has been reported [32]. The presence of multiple genotypes might also directly affect the probability of plant recovery. We will follow individual plants in experimental populations in future experiments to investigate whether the probability of dying and/or recovering depends on the number of genotypes within a host plant and/or their relatedness.
This study shows that there is a high and significant relatedness among fungal parasite genotypes in multiply infected host plants. Together with a previous experimental study [8], our results strongly suggest that the significant relatedness of parasites within hosts is due to an active mechanism of exclusion. Microgeographic structure, selfing, and crosses on the host probably also contribute to the high relatedness of parasite genotypes within host plants, but to a lesser extent. Theoretical models and experimental studies have shown that competition between parasites in multiple infections should be weaker when parasites within a given host are more closely related [3–7]. Our results show that the biased occurrence of related parasite genotypes within hosts can indeed be found in natural populations, thus providing empirical support for previous theoretical considerations.
Teliospores of M. violaceum were collected in the fall of 2004 from 12 roadside populations of S. latifolia in the Essonne region (France). These were also included in the populations sampled by Giraud [21], who analyzed one single flower bud per plant. Here, we collected one bud from each of the stems of every diseased plant in the populations. Sampled populations contained a mean ± standard error (SE) of 18.3 ± 3.0 diseased plants, having a mean ± SE of 5.7 ± 0.3 stems. As far as we could assess, the populations grew in very similar environments (roadsides), with similar levels of disturbance.
To estimate the number of different genotypes per plant, we collected and genotyped all the infected individuals in each studied population. A single bud was genotyped per branch. To test for the presence of more than one genotype within single floral buds, we separately genotyped all the ten anthers for each of two buds belonging to each of three stems. This test was performed on a total of four plants that had been previously scored as having multiple fungal genotypes in different stems. Samples were stored and DNA extracted as in Giraud [21]. We scored six microsatellite loci whose polymorphism is sufficiently high in the studied populations for distinguishing genotypes: L14, L17, L18 [33], SL8, SL12, and SL19 [34], using the same protocol as Giraud [21]. All these markers are nuclear and can yield diploid genotypes in Microbotryum teliospores. For ten plants with identical microsatellite genotypes, we confirmed that their genotypes were identical by AFLP that was performed as in López-Villavicencio et al. [17].
When multiple samples of an identical genotype were found within a plant, a single sample was kept for the subsequent genetic analyses. This provides conservative estimates of relatedness because multiple identical genotypes could potentially represent the same fungal individual present in multiple stems. The unbiased genetic diversity of Nei [35] was computed using FSTAT [36].
Relatedness estimates were computed in a given population based on the allelic frequencies estimated within that population. Standard errors of r were obtained by jack-knifing over loci. Relatedness calculations were performed using RELATEDNESS 5.0 (http://www.gsoftnet.us/GSoft.html).
|
10.1371/journal.pmed.1002424 | Cardiovascular disease (CVD) and chronic kidney disease (CKD) event rates in HIV-positive persons at high predicted CVD and CKD risk: A prospective analysis of the D:A:D observational study | The Data Collection on Adverse Events of Anti-HIV Drugs (D:A:D) study has developed predictive risk scores for cardiovascular disease (CVD) and chronic kidney disease (CKD, defined as confirmed estimated glomerular filtration rate [eGFR] ≤ 60 ml/min/1.73 m2) events in HIV-positive people. We hypothesized that participants in D:A:D at high (>5%) predicted risk for both CVD and CKD would be at even greater risk for CVD and CKD events.
We included all participants with complete risk factor (covariate) data, baseline eGFR > 60 ml/min/1.73 m2, and a confirmed (>3 months apart) eGFR < 60 ml/min/1.73 m2 thereafter to calculate CVD and CKD risk scores. We calculated CVD and CKD event rates by predicted 5-year CVD and CKD risk groups (≤1%, >1%–5%, >5%) and fitted Poisson models to assess whether CVD and CKD risk group effects were multiplicative. A total of 27,215 participants contributed 202,034 person-years of follow-up: 74% male, median (IQR) age 42 (36, 49) years, median (IQR) baseline year of follow-up 2005 (2004, 2008). D:A:D risk equations predicted 3,560 (13.1%) participants at high CVD risk, 4,996 (18.4%) participants at high CKD risk, and 1,585 (5.8%) participants at both high CKD and high CVD risk. CVD and CKD event rates by predicted risk group were multiplicative. Participants at high CVD risk had a 5.63-fold (95% CI 4.47, 7.09, p < 0.001) increase in CKD events compared to those at low risk; participants at high CKD risk had a 1.31-fold (95% CI 1.09, 1.56, p = 0.005) increase in CVD events compared to those at low risk. Participants’ CVD and CKD risk groups had multiplicative predictive effects, with no evidence of an interaction (p = 0.329 and p = 0.291 for CKD and CVD, respectively). The main study limitation is the difference in the ascertainment of the clinically defined CVD endpoints and the laboratory-defined CKD endpoints.
We found that people at high predicted risk for both CVD and CKD have substantially greater risks for both CVD and CKD events compared with those at low predicted risk for both outcomes, and compared to those at high predicted risk for only CVD or CKD events. This suggests that CVD and CKD risk in HIV-positive persons should be assessed together. The results further encourage clinicians to prioritise addressing modifiable risks for CVD and CKD in HIV-positive people.
| Access to combination antiretroviral therapy has increased the life expectancy of HIV-positive people.
Despite this success, there is evidence that HIV-positive people may experience an earlier onset of comorbidities (e.g., cardiovascular disease [CVD] and chronic kidney disease [CKD]) in comparison with their HIV-negative peers.
The D:A:D study has developed specific models to predict the risk for CVD and CKD events in HIV-positive people.
This study was designed to investigate whether HIV-positive people at high risk for both CVD and CKD are at even greater risk for CVD and CKD events, and may therefore warrant particularly close clinical attention and management.
The D:A:D study is a prospective observational study that combines 11 cohorts of >49,000 HIV-positive people followed prospectively since 1999.
We performed a study that included the 27,215 participants with the requisite data to calculate their D:A:D CVD and CKD risk scores. We calculated the CVD and CKD event rates by predicted 5-year CVD and CKD risk groups (≤1%, >1%–5%, >5%) and analysed whether CVD and CKD risk group effects were multiplicative.
We found that people at high predicted risk for both CVD (>5%) and CKD (>5%) events have substantially greater risks for both CVD and CKD events compared with those at low predicted risk for both outcomes and those at high predicted risk for CVD or CKD events alone.
HIV-positive people not uncommonly have high predicted CVD and/or CKD risk; these interact to create substantial risks for future morbid events.
Our results suggest that CVD and CKD risk in HIV-positive people should be assessed together.
The results should further encourage clinicians to prioritise addressing modifiable risks for CVD and CKD in HIV-positive people.
| Combination antiretroviral therapy has transformed the lives of HIV-positive people. Over the past 20 years in high-income countries, rates of opportunistic diseases have declined and life expectancy has reached levels similar to that of the HIV-negative population [1–4]. However, there is evidence suggesting that people living with HIV experience greater and earlier onset of comorbidities compared with their HIV-negative peers [5–7]. The relative extent to which this observation relates to a higher prevalence of behaviours and risks associated with such comorbidities, increased levels of immune activation and coagulopathy despite sustained virological suppression, antiretroviral treatment, and/or other mechanisms remains unclear [8,9].
In the general population, it is well established that chronic kidney disease (CKD) is an important and independent risk factor (covariate) for cardiovascular disease (CVD) [10]. It has been argued that CKD should be considered a coronary risk factor, and that aggressive risk factor reduction for other CVD risk factors should be part of standard therapy for patients with CKD [11].
CVD in turn is associated with CKD. This may be mediated by the development of atherosclerosis, or may result as a consequence of shared risk factors with CKD such as hypertension and diabetes mellitus. It is well recognised that the effects of comorbidity may be greater than the effects of the sum of risk of each disease, and that their co-existence may lead to more severe illness, poorer prognosis, and premature death [11].
In HIV-positive people, a similarly strong relationship between CKD and CVD has been reported [12]. The Data Collection on Adverse Events of Anti-HIV Drugs (D:A:D) collaboration has developed HIV-specific predictive risk models for CVD [13] and CKD [14] in an attempt to help guide care in these individuals. We hypothesized that participants in D:A:D at high (>5%) predicted risk for both CVD and CKD would be at even greater risk for CVD and CKD events, and may therefore warrant particularly close clinical attention and management.
The D:A:D study is a prospective observational study that combines 11 cohorts of HIV-positive people. Its objective is to establish whether the use of combination antiretroviral therapy is associated with an elevated risk of CVD, end-stage renal disease (ESRD), liver disease, cancer, and death. The 11 cohorts contribute data on 49,717 participants enrolled at 212 clinics in Europe, the US, Argentina, and Australia. The standardised dataset includes information on socio-demographics, AIDS events and death, known CVD risk factors, height and weight, laboratory markers, antiretroviral treatment history, estimated glomerular filtration rate (eGFR) (Cockcroft—Gault equation), and treatments including those prescribed for CVD risk.
The main analyses in this study were performed according to a prospectively established statistical analysis plan (S2 Text). The non-prospectively defined analyses were those investigating the covariates of the CVD risk score as predictors for CKD events adjusted for CKD risk group, the covariates of the CKD risk score as predictors for CVD events adjusted for CVD risk group, and the prevalence of diabetes at baseline according to predicted CKD and CVD risk group.
All participating cohorts in the D:A:D study followed local national guidelines/regulations regarding patient consent and ethical review.
In this study we included individuals for whom a complete set of risk covariate data was available after 1 January 2004 as required by both the D:A:D CVD and CKD risk equations, except for missing data for exposure to viral hepatitis C and family history of CVD, in which case these were imputed to be negative (thus slightly underestimating CVD risk in these individuals). All those included in the analysis had an eGFR > 60 ml/min/1.73 m2 at baseline and an eGFR ≤ 60 ml/min/1.73 m2 observed subsequently on at least 2 consecutive occasions at least 3 months apart. We used definitions of endpoints indicating CVD and CKD as defined in previous D:A:D publications [13,14], but note that CVD events were centrally validated clinical events (myocardial infarction, stroke, and invasive cardiac procedures), whereas CKD events were based on confirmed laboratory criteria.
Patient follow-up was censored at the earliest of the following: last eGFR, last visit plus 6 months, or 1 February 2015. We excluded participants with a previous history of CVD, baseline eGFR ≤ 60 ml/min/1.73 m2, or a CVD or CKD event that occurred within 30 days of baseline. We included CVD events and follow-up after the definition of a CKD event had been met and CKD events after the definition of a CVD event had been met.
We calculated the 5-year predicted risk for CKD and CVD using the D:A:D risk equations [13,14]. The CVD equation includes age, sex, diabetes, CVD family history, current and former smoking, total cholesterol, high-density lipoprotein (HDL) cholesterol, systolic blood pressure, current CD4 count, current receipt of abacavir, and cumulative exposure to protease inhibitors (PIs) and nucleoside/nucleotide reverse transcriptase inhibitors (N[t]RTIs). We censored PI and N(t)RTI at 3 and 7 years, respectively, as this reflects the range of data on which the models were developed and so avoids over-predicting risk. The CKD equation includes age, sex, exposure to HIV through injecting drug use, hepatitis C coinfection, baseline eGFR, nadir CD4 count, hypertension, prior CVD and diabetes, adjustment for current receipt of unboosted atazanavir or ritonavir-boosted lopinavir, and a higher adjustment for tenofovir disoproxil fumarate, ritonavir-boosted atazanavir and any other PI.
We stratified the 5-year predicted CKD and CVD risk into groups as ≤1%, >1%–5%, and >5%. We did this rather than use predicted risk as a continuous measure for 2 reasons. First, this reflects how clinicians and their patients use these risk equations, categorising predicted risk into groups that then indicate possible interventions. Second, using risk groups in this fashion allows clear presentation of the rates of CKD and CVD events in each predicted risk group. We deliberately chose not to use the low, moderate, and high CKD risk groups defined by Mocroft et al. [14] in order that the CVD and CKD risk groups could be interpreted in the same fashion. We performed a secondary analysis of events using the Framingham equation, recalibrated to the D:A:D study as described in a previous publication [13], to assess whether the results would be similar compared to the primary analysis. We also examined CKD and CVD events using predicted risks as continuous covariates as a sensitivity analysis.
We calculated CKD and CVD event rates by CKD and CVD risk group. We fitted Poisson models to assess whether CKD and CVD risk group effects are additive or multiplicative. In an attempt to assess how the predicted CVD risk score was contributing to prediction of CKD events, we investigated predictors of CKD events, fitting the CKD risk score group and then each of the covariates in the CVD risk score that are not included in the CKD risk score. We also investigated predictors of CVD events, fitting the CVD risk score group and then each of the covariates in the CKD risk score that are not included in the CVD risk score.
Of the 49,717 individuals enrolled in the D:A:D study, 27,215 (55%) had the required complete covariate data after 1 January 2004 and were included in this analysis, contributing 202,034 person-years (pyrs) of follow-up.
Characteristics of the 27,215 individuals included at baseline, defined as the first time each individual had complete CVD and CKD risk factor data available after 1 January 2004, are shown in Table 1. The median (IQR) baseline year of follow-up was 2005 (2004, 2008). The cohort was 74% male, with a median (IQR) age of 42 (36, 49) years. Fifty percent were smokers at baseline, and 3.8% had a diagnosis of diabetes mellitus. The median (IQR) 5-year predicted CKD risk was 1.1% (0.6%, 3.7%), and the median (IQR) 5-year predicted CVD risk was 1.6% (0.8%, 3.3%).
Of the 1,415 people with a CKD event and 918 people with a CVD event, 154 (10.9%) experienced both types of event. Eighty-six (56%) of those who had both events experienced the CVD event first, and 67 (44%) the CKD event first. In 1 person the CKD and CVD event occurred on the same day. The majority of the CKD and CVD events (117, 76%) occurred 1 or more years apart.
The numbers of participants in each predicted risk group combination are expressed in Table 2. Of the 3,560 people with a 5-year predicted CVD risk > 5%, 3,331 (94%) were men and 229 (6%) were women. In these 3,331 men there were 528 CKD events (rate 23.0 per 1,000 pyrs), and in the 229 women there were 53 CKD events (rate 37.3 per 1,000 pyrs).
We observed 1,415 CKD events during follow-up, an overall rate of 7.00 per 1,000 pyrs (95% CI 6.6–7.7 per 1,000 pyrs). CKD event rates by predicted CKD and CVD risk group are shown in Fig 1, indicating that within each CKD risk group, the CKD event rate increases with higher predicted CVD risk.
This multiplicative effect of CKD and CVD predicted risk group on CKD events is confirmed by the Poisson regression models (Table 3), which show a highly statistically significant association between predicted CVD risk group and CKD events, after adjustment for CKD risk group. In these models there was no statistical evidence of an interaction between the predicted CKD and CVD risk groups, suggesting that the effects are multiplicative (i.e., additive on a log scale). This means, for example, that for an individual with a high predicted 5-year risk of both CKD and CVD, the incidence rate ratio (IRR) for CKD events would be 13.81 multiplied by 5.63, which equals 77.75, compared with an individual at low predicted risk for both CKD and CVD.
The CKD event rates by predicted Framingham CVD and CKD risk strata were similar to the findings in the primary analysis (S1 Table).
There were 918 CVD events documented during follow-up, an overall rate of 4.5 per 1,000 pyrs (95% CI 4.2–4.8 per 1,000 pyrs). CVD event rates by predicted CVD and CKD risk groups are shown in Fig 2. This figure shows large increases in CVD event rates with increasing predicted CVD risk, as would be expected. But there is also some evidence of increasing CVD events with increasing predicted CKD risk.
The Poisson regression models did confirm a statistically significant association between higher predicted CKD risk group and CVD events after adjustment for predicted CVD risk (Table 3). But the magnitude of this association was smaller than was seen for the association between predicted CVD risk group and CKD events. Again, there was no statistical evidence of an interaction between predicted CVD and CKD risk groups.
The CVD event rates by predicted Framingham CVD and CKD risk strata were also similar to the findings made in the primary analysis (S2 Table).
The relationship between the covariates of the CVD risk score and prediction of CKD events (adjusted for CKD risk group) is summarised in Table 4. We found that total plasma cholesterol level was associated with risk of a CKD event (IRR 1.48 per unit log total cholesterol, 95% CI 1.20, 1.83, p < 0.001), as were baseline CD4 count (IRR 0.90, 95% CI 0.86, 0.95, p < 0.001) and cumulative exposure to a PI (IRR 1.11 per year exposure, 95% CI 1.06, 1.15, p < 0.001) or to an N(t)RTI (IRR 1.05 per year exposure, 95% CI 1.03, 1.08, p < 0.001). We did not find statistically significant relationships for having a family history of CVD, being an ex- or current smoker, HDL cholesterol, or currently receiving abacavir.
In a multivariate analysis, after adjusting the covariates for each other as predictors of CKD events, we found similar associations as for the analysis adjusted only for CKD risk group, although the association for receiving a PI lost statistical significance (IRR 1.04, 95% CI 0.99, 1.10, p = 0.149).
After adjustment for CVD risk group, we found 1 covariate in the CKD risk score that was associated with CVD events (Table 5). Higher nadir CD4 count was associated with a reduced risk of CVD events: IRR 0.94 per 100 cells (95% CI 0.091, 0.98, p = 0.002). We did not find associations between predicted CKD risk and HIV exposure through injecting drug use, baseline eGFR, or being coinfected with viral hepatitis C.
Diabetes is a key covariate in both the CKD and CVD risk scores. Therefore, we were interested in exploring the prevalence of diabetes in the CKD and CVD predicted risk groups (Table 6). The prevalence of diabetes at baseline increased markedly with higher predicted CVD risk group. The association was less clear with CKD risk group, but of the 1,585 people at >5% predicted risk for both CVD and CKD, 366 (23.1%) had been diagnosed with diabetes.
We conducted sensitivity analyses assessing the predicted CKD and CVD risk scores as continuous measures rather than in groups. After adjustment for 5-year predicted CKD risk, 5-year CVD predicted risk was significantly associated with CKD events: IRR 1.065 per 1% increase (95% CI 1.058, 1.072, p < 0.001). Five-year CKD predicted risk was not significantly associated with CVD events after adjustment for 5-year CVD predicted risk: IRR 1.005 per 1% increase (95% CI 1.000, 1.011, p = 0.052). These analyses suggest that the multiplicative effects of predicted CKD and CVD risk groups on CKD and CVD events are not simply due to a loss of precision by categorising the predicted risks as ≤1%, >1%–5%, and >5%.
In this analysis we found that HIV-positive people with high predicted CVD or CKD risk were at significant risk for a future CVD or CKD event, and that this risk was multiplicative for those with greater degrees of risk. For instance, we observed a far higher CKD event rate for those at high risk (>5%) for both CVD and CKD (44.0 per 1,000 pyrs) compared to those at low risk for both (0.5 per 1,000 pyrs) and those at intermediate risk for both (5.1 per 1,000 pyrs). We found that the CKD event rate in those with a high CKD risk was highly sensitive to the degree of CVD risk, with a CKD event rate of 7.1, 21.9, and 44.0 events per 1,000 pyrs for a CVD risk of ≤1%, >1%–5%, and >5%, respectively. For CVD events we found that event rates were multiplicative in that there was a gradient from those at low risk for both events (0.45 per 1,000 pyrs) to those at high risk for both events (16.50 per 1,000 pyrs), with those at intermediate risk in between.
It is widely acknowledged that there are complex relationships between CVD and CKD, and that the risks that mediate both of these conditions can interact, leading to more severe illness and poorer prognosis [10]. Our results are consistent with this understanding in that we found that both chronic diseases were associated with an elevated risk for the other. Overall our results suggest that among people living with HIV, the association between CVD risk and a future CKD event is stronger than the association between CKD risk and a future CVD event, and that the strength of the association may be particularly marked in those who are at high risk for both CVD and CKD. This is a novel finding, as the opposite has consistently been noted in the general population, where those with CKD are at markedly elevated risk for CVD. For the general population, for instance, when adjusted for traditional cardiovascular risk factors, impaired kidney function and raised concentrations of albumin in urine increase the risk of CVD 2- to 4-fold [11]. In interpreting these findings, it should be noted that this difference may be a function of the difference in the definition of the 2 disease endpoints used in the study; CVD is a clinical event adjudicated centrally according to specific criteria, whereas CKD is defined by the presence of impaired renal function as indicated by a laboratory biomarker (eGFR) observed over 2 consecutive occasions at least 3 months apart. This laboratory-based definition is thereby a sensitive measurement that captures people with true CKD but may also include people with transient decreases in eGFR not indicative of irreversible renal function deterioration.
In terms of the risk factors for development of CKD events in those with CVD risk, we found that total cholesterol level and cumulative PI and N(t)RTI use raised the risk of CKD events, and that a higher baseline CD4 count was associated with a lower rate of future CKD events. These findings are consistent with results previously reported in the D:A:D cohort [15–18]. We found that cumulative exposure to N(t)RTIs predicts CKD events, despite tenofovir exposure being included as an adjustment in the CKD predicted risk score. This may be because tenofovir exposure was included in the CKD predicted risk model as an acute effect rather than an effect that increases with increasing exposure.
For CVD events according to CKD risk, we did not find an association between eGFR and CVD risk. This is not consistent with findings in the general community. However, the failure to find such an association may be due to effective management of renal insufficiency, with a resultant decrease in the risk of CVD in this population. This finding also seems inconsistent with previous D:A:D analyses [12]. This apparent inconsistency is likely explained by the fact that the present analysis was based on a single baseline eGFR with participants selected with eGFR in a relatively healthy range (>60 ml/min/1.73 m2), whereas previous analyses were time updated and included far lower eGFR values.
It is recognised that both CKD and CVD are largely preventable conditions and that modifying and controlling their risk factors will reduce the risk of onset of both diseases. Despite this evidence and much effort over the past decade to focus attention on reducing CVD risk in HIV-positive people, a lack of attention to the management of CVD in people living with HIV has been documented [19]. A recent modelling analysis within the ATHENA cohort examined the impact of several interventions, such as earlier antiretroviral treatment, smoking cessation, and anti-hypertensive or lipid-lowering treatment, on future CVD events among HIV-positive people [20]. The results suggested that all the selected interventions, but especially smoking cessation and blood pressure and serum lipid control, would have a positive effect in reducing CVD.
We performed an analysis that assessed the extent to which the presence of diabetes in the participants may provide at least a partial explanation of the multiplicative risks we observed in those with higher degrees of risk, and particularly in those with a high risk for both events. We found that nearly a quarter of those with a high 5-year CKD and CVD risk had been diagnosed with diabetes, compared to very few individuals diagnosed with diabetes (0.4%) in the low CKD and CVD risk groups. This confirms that diabetes is a powerful risk factor for poor outcomes in HIV-positive people, as in the general community, and it behoves all clinicians to adequately screen, diagnose, treat, and, above all, prevent this complication to facilitate the optimal longevity and quality of life in HIV-positive people.
Our study has some limitations. Prediction models are limited by restrictions in the available data, and some variables that may affect CKD and CVD risk were not available for analysis (e.g., inflammatory markers). In this analysis, the risk equations were applied to the datasets that were largely used to develop them. It may be, therefore, that the equations predict more accurately in D:A:D data than they would in other independent datasets, and this may limit the generalisability of our findings. Both equations, however, were validated: the CVD equation using internal—external cross-validation, and the CKD equation on independent datasets. The CVD and CKD endpoints used in this analysis are quite different. CVD is a serious clinical event, subject to central validation, whereas CKD is an earlier, less serious event based on a confirmed laboratory marker. As noted above, compared with the more specific adjudicated CVD endpoint, the CKD definition may be capturing a broader set of events that might include true episodes of chronic and progressive CKD but also transient events in people who may be temporarily systemically unwell, hospitalised, and/or undergoing a medical procedure, all of which may lead to episodic decreases in eGFR without this necessarily resulting in persistent renal dysfunction. We attempted to minimise the possibility of miscategorising people with CKD by requiring there to be at least 2 consecutive eGFR measurements ≤ 60 ml/min/1.73 m2 to make the diagnosis. However, the ‘softer’ definition of CKD may at least partially explain the finding that the magnitude of association was greater when analysing CKD events by CVD risk than when analysing CVD events by CKD risk. It is uncertain how the results might have differed if a hard clinical endpoint for renal disease, such as ESRD, had been used. Such an approach, however, was not possible in our analyses, as a prediction tool for ESRD in HIV-positive people is not available, and the number of ESRD events in D:A:D is not large. In the D:A:D database, we are unable to distinguish cases of either CKD or CVD that were a direct result of a previous event.
In conclusion, we have shown that HIV-positive people not uncommonly have high predicted CVD and/or CKD risk, and that these interact to create substantial risks for future morbid events. We found that people at high predicted risk for both CVD and CKD have substantially greater risks for both CVD and CKD compared with those at low predicted risk for both, and those at high predicted risk for only CVD or only CKD. This suggests that CVD and CKD risk in HIV-positive persons should be assessed together. These data also suggest that the primary prevention and effective management of these comorbidities, prioritising those interventions that have been repeatedly shown to be effective in the general population, will convey the same if not greater benefits for the population of HIV-positive people. Primary prevention and the effective management of comorbidities should be incorporated into the development of guidelines and defining future research priorities for HIV-positive people.
|
10.1371/journal.pntd.0001278 | The Value of Educational Messages Embedded in a Community-Based Approach to Combat Dengue Fever: A Systematic Review and Meta Regression Analysis | The effects of various dengue control measures have been investigated in previous studies. The aim of this review was to investigate the relative effectiveness (RE) of different educational messages embedded in a community-based approach on the incidence of Aedes aegypti larvae using entomological measures as outcomes.
A systematic electronic search using Medline, Embase, Web of Science and the Cochrane Library was carried out to March 2010. Previous systematic reviews were also assessed. Data concerning interventions, outcomes, effect size and study design were extracted. Basic meta-analyses were done for pooled effect size, heterogeneity and publication bias using Comprehensive Meta-analysis. Further analysis of heterogeneitity was done by multi-level modelling using MLwiN. 21 publications with 22 separate studies were included in this review. Meta-analysis of these 22 pooled studies showed an RE of 0.25 (95% CI 0.17–0.37), but with substantial heterogeneity (Cochran's Q = 1254, df = 21, p = <0.001,). Further analysis of this heterogeneity showed that over 60% of between study variance could be explained by just two variables; whether or not studies used historic or contemporary controls and time from intervention to assessment. When analyses were restricted to those studies using contemporary control, there was a polynomial relationship between effectiveness and time to assessment. Whether or not chemicals or other control measures were used did not appear have any effect on intervention effectiveness.
The results suggest that such measures do appear to be effective at reducing entomological indices. However, those studies that use historical controls almost certainly overestimate the value of interventions. There is evidence that interventions are most effective some 18 to 24 months after the intervention but then subsequently decline.
| Dengue fever is a mosquito-borne viral infection that is widespread in the tropics. Each year there are an estimated 50 million infections worldwide. Preventing infection relies on controlling the mosquitoes that spread disease. Unfortunately it is still not clear what does and does not work in the control of the mosquito vector. There have been several systematic reviews into control of dengue fever but still no consensus has been reached. This lack of consensus reflects the substantial heterogeneity in published effectiveness of studies of dengue control interventions. Prior systematic reviews have not adequately addressed this heterogeneity. We used multi-level modelling meta regression to investigate what variables modify the effectiveness of studies of educational messages embedded in a community-based approach. Most of the between study variation was explained by two variables, study design and time from intervention to assessment. In particular, studies using historic controls substantially overestimated the effectiveness of the intervention compared to those studies using contemporary controls. When the analysis was restricted to just those studies using contemporary controls, effectiveness was highest about 12 to 24 months after intervention but then declined.
| Dengue fever (DF) is an acute viral disease affecting all age groups. It occurs mainly in tropical and subtropical areas, the predominant vectors being the mosquitoes Aedes aegypti and albopictus, which become infected with any of the four dengue viruses and transmit the disease via a bite to humans [1]. Some 2.5 billion people (two-fifths of the world's population) are now at risk from dengue, and the WHO currently estimates that there may be 50 million dengue infections worldwide every year [2]. Depending on the year, tens to hundreds of thousands of cases of the severe and potentially fatal form of the disease, dengue haemorrhagic fever, and dengue shock syndrome (DHF/DSS) occur [3].The incidence of DF has increased dramatically in recent decades. Its proliferation is influenced by many mechanisms – these include population growth with unplanned urbanisation (and consequent overburdening of water and sanitation systems), increases in domestic and international travel, transportation of commodities such as tyres, lack of political will to intervene, and limited financial and human resources to implement effective control measures [4]. The disease has become endemic to more than one hundred countries in Africa, the Americas, the Eastern Mediterranean, South-East Asia and the Western Pacific. Of these, South-East Asia and the Western Pacific are the most seriously affected. There is currently no vaccination for DF, and no medications that can treat DHF or DSS directly, so at present the only way of controlling or preventing the spread of the virus is to combat vector mosquitoes directly.
For many years, spraying with insecticides, such as malathion, has been the main method of control, though this has often had limited success [5]. Other interventions aimed at controlling the mosquito population, have been tested with varying success. For example, the Puerto Rican government, in response to the threat of a DHF epidemic, developed an integrated approach consisting of community-based dengue control programs to complement traditional chemical-based approaches [6]. They encouraged the public to reduce or eliminate containers in and around homes, gardens and villages. These containers, which include discarded plastic packaging, metal cans, and rubber car tyres, are capable of holding water which would then harbour larvae, and allow mosquitoes to proliferate [7]. It has become clear, from the number of projects that have been initiated in recent years that community-based programs are now regarded by both national and international health agencies as the primary long-term solution for prevention and control of DHF/DSS in Asia and the Americas [3].
A recent systematic review carried out by Erlanger and colleagues investigated the effect of different types of dengue vector control interventions, including biological, chemical, environmental and integrated vector management, on well established entomological parameters [4]. Their aim was to compare the effects of these interventions, in order to find the most efficacious. They identified 56 publications with extractable data that had compared the impact of 61 different dengue control interventions with control communities or with the same community prior to the intervention. The authors concluded that dengue interventions are effective in reducing vector populations, particularly when interventions use a community-based integrated approach. An earlier systematic review by Heintze et al. specifically looked at community-based dengue control programmes, and concluded that the evidence that such programmes were effective, either alone or in combination with other programmes was weak [7]. Yet another recent systematic review also concluded that there was little evidence to support the efficacy of mosquito abatement programs due to poor study design and lack of congruent entomological indices [8] An important criticism that can be levelled at these systematic reviews is that there was substantial heterogeneity in study design and in the size of any effect that made it difficult to draw definitive conclusions. In particular some studies used historical control periods whilst others used other contemporary communities as controls. Many of the studies included multiple interventions in combination whilst others studies were of a single intervention. Furthermore the control communities may or may not have had one of more interventions themselves. We argue that these issues make it particularly difficult to disentangle the value of educational messages embedded in a community-based approach, or identify the most successful approach. Although Erlanger and colleagues did undertake subgroup analysis around types of intervention, neither of these studies adequately investigated sources of heterogeneity in effect size (the magnitude of any association between the outcome and predictor) making the drawing of any definitive conclusions problematic.
One of the recent trends in meta-analysis has been the increasing use of methods that aim to investigate causes of heterogeneity in effect size between published studies rather than rely on pooled effects sizes that can often be difficult to interpret [9], [10]. By this way it is hoped that additional insights can be gleaned into how study design and context, such as use of control interventions in control groups, may affect the outcome. Such insights could give some understanding of in what situations these interventions may, or may not, have benefit. This paper reports a deeper analysis of papers that have attempted to determine the impact of educational messages embedded in a community-based approach, which we define as community based intervention that had any element where members of the public were given information or exhortations intended to change their behaviour, on entomological indicators of risk of dengue disease.
The primary outcome of this review was to establish which, if any, of these interventions was most effective in reducing larval indices. Usually entomological effectiveness was measured using one or more of three widely utilised indices: the Breteau index (BI), container index (CI) and house index (HI). The BI specifies the number of containers with Aedes spp. larvae per 100 houses, the CI represents the percentage of water containers positive for Aedes spp. larvae, and the HI gives the percentage of houses with water containers holding immature Aedes spp. [4]. In addition, one study used the average number of positive containers per house (C+/H) [11].
Included studies were required to: firstly refer to control of dengue fever, and secondly have studies investigating an educational intervention alongside a ‘control’ approach or standard management program. Studies also had to look at quantitative outcomes, whether these were the BI, CI, HI or C+/H. Next, these studies had to be community-based, whereby members of the community partook in the interventions or played a major role. Conversely, studies based in laboratory or semi-field settings were excluded, as were purely observational cross-sectional and qualitative studies. Studies were not limited by language of publication.
The primary measure of effect size was relative effectiveness (RE) with 95% confidence intervals. RE is the ratio between the entomological index in the intervention group and in the control group. Consequently the more effective the intervention the lower the RE. An RE of 1.0 would indicate no effect. Where confidence intervals were not given, these were back-calculated from the P value. Where only the entomological index/indices were presented for each group RE and its 95% confidence intervals were estimated using Monte Carlo modelling with @Risk™. The distributions of the indices for the intervention and control groups were taken from the papers. Then values were repeatedly sampled from each distribution and the value sampled from the intervention distribution divided by that sampled from the control sample to give the RE. From the repeat samplings the distribution of the RE was then determined to give mean and 95% confidence intervals.
The review carried out by Erlanger et al. investigated a range of interventions, including entomological and community measures taken in a variety of settings [4]. The objective of this review was to systematically analyse only the publications which included an educational element to their interventions (even if other non-educational interventions were also included). However, we used a rather broad definition of educational intervention to include any community based intervention that had any element where members of the public were given information or exhortations intended to change their behaviour. This was followed by a rigorous up-to-date search strategy, detailed below, which was carried out in order to retrieve references which had been produced since publishing date of the existing review (September 2008).
A structured electronic search of Medline, EMBASE, Web of Science and the Cochrane Database of Systematic Reviews was carried out up to March 2010. This was performed in the format: [dengue or dengue haemorrhagic fever or dengue virus or Aedes aegypti] AND [arthropod vectors] AND [community based] AND [intervention]. Reference lists were checked for additional publications to the ones found in the initial search, which fulfilled the inclusion criteria.
From the initial search results, all titles and abstracts were assessed independently by two reviewers, with disagreements being resolved by discussion. From these, a list of papers to include was made, and full text articles obtained.
Once the publications had been assessed as meeting the prescribed quality and inclusion criteria, and having considered the references used by Erlanger et al., data was extracted systematically, using a standardised form. Data was extracted from the existing systematic review, but also updated with the most recent studies found in the search. Where follow-up occurred over several time points, the longest follow up time point results were included, as this provides the most realistic indicator of long-term effectiveness of the intervention. Data was extracted on the outcome measure, study design, time of follow-up after intervention, what other interventions were used and the nature of the educational component.
Where confidence intervals were not presented in the original paper, these were derived by a process of back-calculation from the presented P value. Initial analyses were done with Comprehensive Meta-analysis (CMA) Version 2.2.050 [12]. All four main entomological indices were included in the analyses. If more than one entomological index were reported in the same study, then a single outcome measure was calculated as the geometric mean of the different entomological index by CMA using the within program option to combine effect sizes from different types. CMA was used to calculate heterogeneity, determine potential effects of publication bias and pooled estimates of effect size. In order to determine whether combining REs using different entomological indices was valid, Pearson correlation coefficients were calculated as was paired t tests between them.
Subsequent analyses of the impact of moderator variables were done using MLwiN [13]. A basic three level model was constructed to account for studies with multiple comparisons [14]. Each of the putative modifier variables were put singly into the model and those with p<0.2 included in a multiple modifier model. In the multiple model, any modifier with p> = 0.2 was then removed and the model rerun until all modifiers in the model had p<0.2. The proportion of the between study variance explained by the final model was derived from τ2 (between-studies variance) in the model with no modifiers and in the final model.
Searches of Medline, EMBASE, Web of Science and the Cochrane Database of Systematic Reviews identified 491 original papers for assessment. Figure 1 shows the flow diagram detailing the search process and inclusion of studies in this review. Of these 491 articles, 456 articles were excluded based on abstract alone because of inappropriateness of subject or study design. A total of 35 papers were obtained in full text. Of these, 14 full text papers were excluded, deemed to be unsuitable with regard to participants, the intervention used, outcomes of the study, or study design. This left 21 papers of which 11 were based in South America, 9 were based in South East Asia, and the remainder were based in Fiji and French Polynesia. The earliest study was published in 1967 [15] and the latest in 2009 [5]. One paper [11] had two study arms that included interventions of interest and each study arm is referred to as separate study were included, giving 22 studies in total.
Studies varied with regard to types of educational component, study design and control groups. The included studies and summary of their characteristics are listed in Table S1. The educational components included the use of print or broadcast media, public lectures, in-home training by public health staff, home visits and targeting school children. The exact mix of interventions varied between studies. Three different approaches were used in the study designs: 6 studies used an historical control period, measuring outcomes in the same village at baseline and at a later time point (‘historical’ control group), 11 studies included a control arm with no additional treatment as well as an intervention arm (‘no treatment’ control group), and 5 studies included a control arm exposed to some anti-mosquito activity, along with an intervention arm (‘some intervention’ control group). The studies also varied in terms of whether or not the intervention communities received other interventions. A total of 9 studies included some form of chemical intervention, as well as the educational component. This varied from the use of malathion spraying both in- and outdoors, to larviciding with the use of abate. Another 8 studies used various additional “other” (i.e. not chemical) measures in the intervention group – these ranged from covering and disposal of containers capable of holding water, to community clean-up campaigns, to the use of other species such as Mesocyclops in order to predate the Aedes spp.
The Pearson correlation coefficients for the REs from the three main entomological indices showed high correlation between them BI-CI 0.68, BI-HI 0.66, CI-HI 0.97. Furthermore, there was no significant difference in the mean RE given by the different entomological indices using a pared t test. We conclude that combining the different entomological indices was valid. The result of the meta-analysis performed on all 22 studies is shown in Figure 2. Using the random effects model, the pooled risk ratio was 0.25 (95% CI 0.17–0.37). However, there was substantial heterogeneity in the effect size (Cochran's Q = 1254, degrees of freedom (df) = 21, p = <0.001). There was no evidence of publication bias (Figure 3).
In order to investigate the sources of heterogeneity further a series of multi level meta-regression analyses were run with potential modifier variables. The results of the initial analyses are shown in table 1. The most significant single modifier variable was whether or not the study used a historical (comparing the same community before and after the intervention) or a contemporary control (comparing the intervention community with another control community). Those studies that used contemporary controls had a much reduced effect size compared to historical controls (Regression coefficient (B) = 2.08, Standard Error (SE) = 0.65). Two other predictor variables, combining a chemical with the community intervention and time at follow-up almost achieved significance. Whether or not other interventions were also used (but not including chemicals) did not achieve statistical significance. Two measures of time at follow-up were tested, the untransformed and log transformed months. The results were very similar between these two time measures and the untransformed used in subsequent analyses as this was marginally more significant and also easier to interpret. In addition the relationship between chemical spraying and RE was further tested as some studies included chemical spraying in both intervention and controls and others in intervention only. Perhaps not surprisingly, chemical spraying where this was applied in both intervention and control arms had almost no effect on RE whilst chemical spraying in the intervention arm but not control arm was associated with a significant improvement in RE (−2.08, SE 0.78).
All variables with p<0.2 in the single modifier variable analyses were included (historic or contemporary controls, time at follow-up and chemical spraying in intervention but not control group) in a final model as shown in table 2. It can be seen that two modifier variables remain historical v contemporary control and study duration. In particular those studies using contemporary controls gave much smaller effect sizes than those using historical controls (B = 2.21, SE = 0.66) and effect sizes improved with longer delays till the follow-up assessments (B = −0.083/month, SE = 0.03). Using chemicals in the intervention group but not in controls was not significant. These three variables were able to explain 64% of the between study variance in the original dataset, though the remaining between study variance was still significant (τ2 = 1.07, SE = 0.39, z = 2.77, p = 0.006). Excluding chemical spraying from the model was still able to explain 61% of the between study variance.
The relationship between RE, choice of control and time to assessment is illustrated in figure 4. Here the difference between RE and choice of control is very clear. It can also be seen that within each category of control the relationship between RE and time to follow-up is more complex. For those studies with historic controls there is a very steep decline with time to assessment. For those studies with contemporary controls there is still a suggestion of a decline over the first 12 months then this levels out and possibly even reverses. This is reflected in the regression equation (table 3) where Log RE is predicted by the time to assessment and time to assessment squared. Indeed, the polynomial equation of time to follow-up was able to explain 44% of the between study variance in the studies with only contemporary controls.
Chemical spraying in the intervention but not control arms of the study were also included (B = 0.54, SE = 0.44, p = 0.254). Although this variable was not significant and subsequently dropped from the final model it is notable that the chemical spraying was if anything associated with reduced effectiveness of the community intervention.
The pooled results of the 22 studies in this meta-analysis suggest an important impact of educational messages embedded in a community-based approach on reducing larval indices. However, there was substantial heterogeneity in effect size between the different studies. This large heterogeneity in effect size reflects the very different study designs in the included studies. As discussed above the studies may or may not have included interventions additional to the community components, they may have used historic or contemporary controls, the controls may or may not have had some form of non-educational intervention. Consequently interpretation of the pooled effect size is difficult. However, the majority of the heterogeneity was explainable by just two variables, the choice of control and the time from intervention to assessment.
The impact of choice of control was particularly marked with studies using historical controls finding much stronger effect sizes than those using contemporary controls (table 2). After adjusting for the time to assessment, anyone basing their judgement of the effectiveness of educational interventions based on historical controls would over-estimate the value of educational interventions by more than 10 fold compared to studies that used contemporary controls (RE = 13.2, 95%CI 4.1–42.5).
If the impact of one aspect of the study design is so great, it begs the question which is the correct study design to use. It could be argued both ways. In favour of the use of historic controls is the argument that at least the populations being compared are geographically the same. The arguments in favour of the contemporary controls include the fact that entomological indices may change from one time to another for reasons totally unrelated to the intervention. Indeed it could be argued that as interventions are usually implemented when the risk of dengue fever is particularly high, it is very likely that entomological indices will improve substantially whatever the intervention. The problems with historical controls are well known to medical researchers [16], [17]. Indeed, the comment has been made that “most historical control groups are compromised for some reason” [18]. In studies with historic controls it has also been argued that the biases are worse the longer the time between the control and intervention. Our analysis would certainly support this suggestion for entomological control measures. Consequently we would argue that studies using historic controls be excluded from any assessment of the effectiveness of dengue vector control programmes.
The polynomial relationship between RE and time to assessment is interesting. From the regression model of only those studies with contemporary controls, the highest effectiveness is seen somewhere around 18 months after the educational intervention. This is consistent with the suggestion that people may need time to learn, but after that their effort and good intentions may slip without reinforcement.
As regards the value of non community interventions in addition to the community interventions, we found little evidence of any effect. In our single modifier analyses additional chemical application did appear of value when the control group did not receive chemical applications. However, in the model with control type and time to assessment, it was not significant. In the model including only studies with contemporary controls there was no evidence that chemical application provided any additional value over that achieved by education alone. However, there were only two studies that included chemical application in the intervention arm and not the control arm [19], [20]. One of these studies supplied sand abate to the villagers, and the other used water treatment with temephos and outdoor spraying from the ground with malathion. Clearly, one cannot draw any definitive conclusions based on two studies in a meta-regression analysis. However, in this regard the study of Espinoza-Gomez and colleauges deserves special mention [11]. In this, well conducted study the authors randomly allocated houses to one of four intervention groups: no intervention, indoor chemical spraying only, education only and combined education and chemical spraying. For our meta-analysis, initialy compared education with no intervention and then compared education plus chemical with chemical only. In their original analysis using a two way ANOVA, the authors found that only education was effective at reducing larval indices and that chemical spraying gave no benefit either alone or in combination with education. In addition, a recent systematic review of the value of the effectiveness of peridomestic space spray also found little benefit of chemical spraying [21]. Why peridomestic spraying has uncertain benefit is unclear.
With regard to comparison of educational interventions against one another, results showed that no single intervention modality (such as the use of print or broadcast media, lectures, training by public health staff, home visits or targeting school children) nor the number of different modalities used together was found to improve the RE significantly (data not shown). However, few if any studies were designed to compare different educational intervention modalities and so we would argue that this issue remains unanswered until specific studies are designed to address the relative effectiveness of different educational modalities.
Although meta-analyses of experimental studies such as randomised controlled trials are usually taken to provide high quality evidence of cause and effect, meta-regression analyses as presented here have the evidential status of observational studies. One of the other issues with meta-analyses of public health interventions is that often it is impossible to adequately blind the study participants or the study investigators. For example this has been raised as a major issue for studies of the effectiveness of household water treatment [10], [22]. Wood et al., in their research into evidence of bias associated with different study designs, found that lack of blinding could be associated with an apparent effect of about 30% for subjective outcomes [23]. For objective outcomes they found no evidence of such bias. Clearly the studies included in the analysis were not blinded. Whether or not the entomological indices are subjective or objective measures are open to debate. We would argue that these indices are semi-subjective and are potentially open to some form or bias due to lack of blinding of the assessors. However, even accounting for this the RE is much greater than could be explained purely by observer bias.
It has been underlined by Erlanger et al., that ‘relative effectiveness’ as numerical evidence of reduction of entomological measures, does not necessarily equate directly to reduction in pathogen transmission' [4]. Other factors such as villages sharing water supplies and garbage disposal may also enable disease transmission even if control within the village was good [11]. Gubler and Clark, in their review of dengue interventions as a whole, state that it must be kept in mind that in the types of community approaches assessed in this review, with these types of community approaches and strategies, it is expected that ordinary members of a community assume responsibility for activities that have historically been conducted by governmental bodies [3]. They suggest that for this reason, it would be very optimistic to expect immediate changes.
In their systematic review of the literature, Erlanger et al. concluded that dengue vector control is effective in reducing vector populations [4]. Since that review three additional systematic reviews have been published none of which came to the same conclusion as Erlanger [7], [8], [21]. These later studies basically came to the conclusion that the quality of the published evidence, something that Erlanger et al. did not adequately address, was too poor to form a definitive conclusion. We would generally agree with the three later studies. In particular we have shown that a major problem is that studies with historical controls strongly overestimate RE compared to those with contemporary controls. Nevertheless, after accounting for the use of historic controls we still found evidence that supports the value of educational messages embedded in a community-based approach in reducing entomological indices of risk. We also showed that there is some evidence that the value of such interventions may decline after 18 to 24 months. With the evidence currently available it is not possible to say what types of educational modalities are most effective. There is a need to reassess whether other interventions add any further value to educational interventions. Finally, the issue also remains whether entomological indices alone is always translated into disease reduction.
|
10.1371/journal.pcbi.1003065 | Interface-Resolved Network of Protein-Protein Interactions | We define an interface-interaction network (IIN) to capture the specificity and competition between protein-protein interactions (PPI). This new type of network represents interactions between individual interfaces used in functional protein binding and thereby contains the detail necessary to describe the competition and cooperation between any pair of binding partners. Here we establish a general framework for the construction of IINs that merges computational structure-based interface assignment with careful curation of available literature. To complement limited structural data, the inclusion of biochemical data is critical for achieving the accuracy and completeness necessary to analyze the specificity and competition between the protein interactions. Firstly, this procedure provides a means to clarify the information content of existing data on purported protein interactions and to remove indirect and spurious interactions. Secondly, the IIN we have constructed here for proteins involved in clathrin-mediated endocytosis (CME) exhibits distinctive topological properties. In contrast to PPI networks with their global and relatively dense connectivity, the fragmentation of the IIN into distinctive network modules suggests that different functional pressures act on the evolution of its topology. Large modules in the IIN are formed by interfaces sharing specificity for certain domain types, such as SH3 domains distributed across different proteins. The shared and distinct specificity of an interface is necessary for effective negative and positive design of highly selective binding targets. Lastly, the organization of detailed structural data in a network format allows one to identify pathways of specific binding interactions and thereby predict effects of mutations at specific surfaces on a protein and of specific binding inhibitors, as we explore in several examples. Overall, the endocytosis IIN is remarkably complex and rich in features masked in the coarser PPI, and collects relevant detail of protein association in a readily interpretable format.
| Much of the work inside the cell is carried out by proteins interacting with other proteins. Each edge in a protein-protein interaction network reflects these functional interactions and each node a separate protein, creating a complex structure that nevertheless follows well-established global and local patterns related to robust protein function. However, this network is not detailed enough to assess whether a particular protein can bind multiple interaction partners simultaneously through distinct interfaces, or whether the partners targeting a specific interface share similar structural or chemical properties. By breaking each protein node into its constituent interface nodes, we generate and assess such a detailed new network. To sample protein binding interactions broadly and accurately beyond those seen in crystal structures, our method combines computational interface assignment with data from biochemical studies. Using this approach we are able to assign interfaces to the majority of known interactions between proteins involved in the clathrin-mediated endocytosis pathway in yeast. Analysis of this interface-interaction network provides novel insights into the functional specificity of protein interactions, and highlights elements of cooperativity and competition among the proteins. By identifying diverse multi-protein complexes, interface-interaction networks also provide a map for targeted drug development.
| Protein-protein interaction (PPI) networks aim to capture the interactions between proteins that mediate many of their molecular functions [1]–[3]. However, with one node per protein and one edge per binary interaction, PPIs provide only a coarse rendering of the nuanced molecular level interactions. With exposed surfaces ranging from tens to hundreds of residues, proteins may present multiple distinct binding interfaces. Each interface can mediate binding to a single partner, or to multiple partners. The cooperative or competitive character of these interactions tunes protein availability in the cell, the formation of higher order complexes, and ultimately many important biological functions. Proteins with multiple binding interfaces can bring together distinct partners to assemble transient or permanent complexes. In contrast, multiple distinct partners competing for a single shared interface may function to connect disparate functional modules in the cell [4], [5], with such competitive binding having arisen, for instance, as a result of gene duplication [6], [7]. Distinguishing the types of binding interfaces a protein uses for each interaction partner is a key step to resolving the cooperativity inherent in functional protein interactions. Moreover, a protein interaction network with resolved interfaces helps to connect gene mutations with disease [8], and to identify possible drug targets, with inhibitors of protein-protein binding receiving increasing attention [9]–[12]. In particular, by targeting interfaces shared in multiple binding interactions, one may be able to shut down entire pathways, whereas targeting more isolated interactions offers a route for a more measured intervention. Assigning interfaces to protein interactions thus has both fundamental and practical relevance.
To refine the coarse protein-protein interaction network and to capture these important structural and chemical aspects of interactions [5] requires the identification of the binding domains or interfaces on each protein. Importantly, one needs to distinguish on the basis of clear rules between binding partners that target overlapping or distinct surface regions. By systematically cataloguing these details it is possible to create not only a map of shared versus distinct binding interactions [5], [13], [14], but an entirely new sub-network of the protein interaction network, as we do here. A PPI network with interfaces overlaid on the proteins highlights the number of interfaces each protein uses to mediate binding and the number of binding partners per interface (see Figure 1b). An interface-interaction network (IIN) is what one gets by visualizing the protein interface connectivity as separated from the underlying PPI network. Unlike in the PPI network representation, in the IIN representation distinct patterns of connectivity between interfaces emerge, and this network topology can be analyzed to yield insight into the specificity and possible cooperation and competition of protein interactions.
Although the importance of structural details in protein interactions has led to increasing efforts to identify protein-binding interfaces in a systematic way [13], [15]–[18], PPI networks with interfaces overlaid on them and detailed IINs have not previously been created. Earlier studies have used protein structures combined with homology modeling [5], genomic data [19], and docking algorithms [20]–[22], to both assign and infer [23] binding interfaces. While it would be possible to construct an IIN from the residue level details collected from some structural data [21], [23], both the accuracy and coverage of the network would be limited by the errors inherent to homology modeling or docking methods, and by the fact that crystallized protein complexes cover only a small percentage of known protein-protein interactions [20]. In particular, using structural homology to infer binding partners provides important guidance but may overestimate the number of binding partners because structural homologs reflect evolution but not necessarily shared functions [24], and small differences in sequences can separate specific from non-specific binding [25]. Proteins with disordered regions and without structural or domain information would be absent from the network, thus sampling only subsets of interaction types within the proteome. The presence of false positives in the IIN would obscure the diverse patterns that emerge in the network and distinguish the network structure from that of the parent PPI. The topology of a network reflects functional pressures acting to connect nodes in specific ways, whether the nodes are proteins, interfaces, or airports. One force shaping the PPI network is the need to transmit information across diverse functional modules, resulting in a giant connected component with few unconnected proteins. Differences in the IIN topology imply different functional and physical forces acting on the component interfaces.
In order to analyze the specificity, cooperativity, and topological properties of an IIN one requires both an accurate assessment of shared and distinct binding interfaces and a dense collection of protein-protein interactions. Therefore, we here combine both structure-based computational approaches with literature-curated biochemical data to build an IIN for the proteins involved in clathrin-mediated endocytosis (CME) in yeast [26]. CME is a central pathway for internalizing cargo such as nutrients and signaling molecules into the cell. Assigning the interfaces mediating each protein interaction would be severely limited if we relied on structural data alone, as several of these interactions are mediated by short peptide motifs on disordered regions. Furthermore, any uncertainty associated with using models of possible domain interactions is completely bypassed by exploiting the wealth of established domain and residue information known from biochemical experiments. Through this process, we are able to carefully evaluate whether a protein's interaction partners bind to the same interface, or to distinct interfaces. Figure 1a illustrates the results of the interface assignment for the yeast actin protein, ACT1, and Figure 1b shows the PPI network with interfaces overlaid for the subset of actin binding proteins. This representation captures both the ability of actin to bind several proteins simultaneously through four distinct interfaces, and the competition between multiple proteins to bind each of these interfaces. The resolution of this network exceeds in two important ways that of networks obtained by only defining PPI network edges as competing. First, because all of actin's binding partners compete with at least a few others to bind actin, all PPI edges from actin would be marked as competing. Therefore it would not be possible to distinguish which of actin's binding partners can bind at the same time. Second, because an edge connects two proteins, a network with edges marked as competing does not clarify which protein surface (say, actin's or its partner's) is actually shared, as sharing can occur on one or both proteins.
In addition to collecting detailed data on protein structures, a particular advantage of our curated approach is to eliminate false positives from the PPI by creating a coherent and consistent picture of the protein interactions. We identify the specific mechanism mediating an observed protein-protein interaction and determine whether the interaction is direct or indirect. Of particular concern are indirect interactions, mediated through intervening proteins, because they are not always distinguishable from direct interactions in high-throughput affinity purification/mass spectrometry (AP/MS) [3], protein-fragment complementation assay (PCA) [27] or, to a lesser extent, yeast two hybrid (Y2H) experiments [3]. Literature sources also document protein pairs tested and found to not bind. Therefore, by curating the literature we do not predict new interactions but we do remove spurious interactions. We also compile the number and types of experiments used to identify the interfaces in each protein-protein interaction, as the interfaces can vary from high-resolution selections of specific residues to low-resolution large regions of the protein. This compilation provides a starting point for improving the resolution of the structural interaction.
The CME network constructed and characterized in this way reveals significant complexity with permanent and dynamic assemblies of few or many proteins, a mixture of binding modes with both shared and distinct binding, and both large and small binding interfaces. This detailed information is necessary for building models of protein-protein interactions where both competitive and cooperative binding reactions contribute to function. The accuracy and coverage of the protein IIN we have generated allows us to draw generalizable insights about the structure of the IIN, the overlap of binding interfaces, the identification of indirect interactions, and the implications towards the biological functions with the parent PPI. Compiling this information for more parts of PPI networks will help prune indirect and spurious interactions, highlight areas of poorly resolved structural and biochemical characterization, and facilitate investigation of the physical and evolutionary origins of the IIN topology and in turn of protein binding.
The aims of the present work are (1) to develop a general framework for the construction of IINs from a combination of structural and biochemical data that measure the support of proposed protein interactions; (2) to characterize the general network properties of the resulting endocytosis IIN as compared to PPI networks and randomized networks, and (3) to demonstrate the applications of the IIN, including as a resource for predicting response to mutation and to specific binding inhibitors. At the network level, we examine whether the IIN retains the complex characteristics of the PPI network, including a high connectivity, hub structures, local clustering, and a scale-free character manifested in power-law distributions of the number of binding partners. We also quantify the fragmentation of the fully connected parent PPI network into separate interface modules at the IIN level. We provide several examples for the use of the IIN in selecting possible drug targets and in predicting the effects of mutations by identifying specific pathways of communication between proteins via their interfaces. For the CME proteins we discuss the central role of SH3 domains and multi-interface proteins. We emphasize that the process of assigning protein interfaces has generated not only a useful map of interactions among these highly-studied proteins but has highlighted the difficulties associated with trying to make automated assignments, including overlapping residues and inconsistencies between sources. Therefore, we discuss the insights derived from our interface procurement process that are relevant for high-throughput methods of interface determination.
As a first step, we constructed a curated PPI network of 56 proteins involved in CME in yeast [26]. Following the approach described in Methods, we first combined 337 edges downloaded from BioGRID [28] via the Saccharomyces Genome Database (SGD) [29] and 49 additional distinct edges collected from IntAct, MINT, DIP, and BIND. 177 edges had interfaces assigned to both proteins in the interaction and nine additional edges were added from literature evidence. We note that for these 56 proteins, we observed significant overlap in the interactions reported in each protein-protein interaction database, as listed in Table 1. Of the assigned protein-protein interactions, sixteen had two binding modes, and two had three binding modes, resulting in a total of 206 assigned interface-interface interactions from 186 assigned protein-protein interactions. We removed 35 edges from the original network because they were suspected to be indirect, shown not to bind in further experiments, or they occurred only in a study of yeast prions, suggesting that the observed binding may not normally be functional. For 28 interactions identified in multiple high-throughput studies no evidence from the literature was found to assign interfaces, and 145 interactions were found only in one reference without sufficient information to assign an interface. Nearly all of the 145 unassigned interactions that were implicated in a single experiment came from high-throughput studies, and because the proteins in the CME subset form the connected clathrin coat and actin patch together, many of these observed interactions could be indirect.
The differing support for the CME protein interactions is represented visually in Figure 2 and collected along with specific details of their assignment in tabulated form in Table S1. The tabulated list contains all the currently known interactions between these 56 CME proteins and the interface assignment status effectively ranks them in terms of their reliability. Interactions with interfaces assigned are further classified in Figure 3 according to the experimental data used to make the assignment. The blue edges in the Figure 2 network are unresolved interactions that have the most evidence (more than one study) supporting their potential functionality in the cell. The red edges are most likely to be artifacts of the experimental probes of their interaction, on the basis of evidence listed in Table S1. Of the red ‘false positive’ edges, most were indirectly interacting through a larger complex. A few had been shown in detailed biochemical characterizations not to bind to one another.
In Figure 3 we also distinguish the amount and type of evidence used to support each interface assignment by coloring each edge. The number of interface interactions assigned directly from crystal structures is shown in black, and represents a minor fraction of the total assignments for this network. For blue edges, both interfaces were resolved using biochemical studies, typically by truncating or mutating the constituent proteins. We note that although these binding interactions have been tested in vitro and in some cases in vivo (thicker blue lines), some of the interfaces encompass large folded domains rather than specific surface binding residues. These domains could therefore be segregated further into more than one interface given additional resolution of the specific residues involved in each interaction. Green and cyan edges had one or both interface inferred. For these assignments we relied on homology to other proteins either through sequence, function or crystal structure. Alternatively, a lack of competition for binding to a surface or a lack of any structural or functional homology was sometimes used to infer distinct vs shared interfaces. Finally, the yellow edges are speculated to be distinct interfaces due to a lack of observed similarity to known partners or domains, and as such have the weakest support.
We determined the degree distributions of both the original and the curated PPI networks, as a statistical measure of the number of interaction partners of each node. Upon going from the full combined-database PPI network (plus the added 9 edges; see Methods) to the curated PPI network, the decrease in total edges results in a less dense network in which the average number of partners per protein dropped significantly from 13.5 to 6.4. Although large PPI networks typically have degree distributions characterized as power law or truncated power laws [30], [31], neither the curated PPI nor the combined-database PPI have degree distributions statistically consistent with a power law density (Figure 4a). The deviations from a power law could be due to the small size of the two networks (only 56 proteins) and the fact that they are all part of a functionally related module. Such deviations would be exacerbated in the original un-curated network, where the distribution is more uniform for N<10, by spurious and likely indirect interactions within the set of proteins absent in the curated PPI network. Lastly, both of the original and curated endocytosis PPI networks had high clustering coefficients (0.56 and 0.46, respectively; see Methods) indicating that proteins that interacted had partners that were likely to interact with one another. Not surprisingly, the clustering coefficient of 0.28 in a full yeast PPI collected from several large-scale studies in yeast [3], [27] is lower, since the CME proteins were specifically chosen to be part of the same functional module.
The majority of proteins in this endocytosis network have multiple interfaces, with an average number of 3.5 distinct binding interfaces per protein. The correlation between protein size and the number of interfaces is quite weak (R = 0.26). One reason for this weak correlation is that several of these proteins have additional binding partners outside the CME network module considered here. Another reason is the size of the interfaces varies broadly, from just a few residues (for NPF motifs [32]) to hundreds of residues (e.g., the clathrin-clathrin leg binding [33]). For example, LAS17 is a medium sized protein at 633 residues that has the most interfaces thanks in part to five short proline-rich domains (PRDs) we assigned as distinct interfaces due to their specificity towards different binding partners [34], [35]. The black edges in Figure 3 connect interfaces determined through crystal structures that therefore are defined by a subset of non-contiguous residues. Interfaces from biochemical studies tend to lack single residue resolution but instead span stretches of residues or complete domains. In the 3D protein structure, only a fragment of these residues would be expected to contact binding partners and as such some of these interfaces could be split or refined further.
It is interesting to compare the change in network properties between the curated PPI network (shown in Figure 3a with interfaces overlaid) and its IIN (Figure 3b). In general, a PPI network and its IIN should have equal numbers of edges, but it is possible for an IIN to have more edges if a pair of interacting proteins has multiple modes of binding to one another. Proteins that act as alternating subunits in a symmetric complex, for example, will contact two copies of the same partner through distinct interfaces. The CME IIN contains several instances of multiple binding modes, resulting in an increase in edges from the PPI. Such distinct modes for the same two proteins to bind one another can act as a regulatory mechanism controlling the accessibility of surfaces on the protein, or as sources of extra stability to the protein-protein interaction. For example, the protein CRN1 contains two distinct actin-binding domains that bind separate regions on the actin surface and are modulated by the nucleotide bound state of actin. Through these multiple binding modes, CRN1 can have opposite roles in either inhibiting or activating the severing of actin filaments [36]. In another example, the SH3 domain of LSB3 binds three distinct PRDS on LAS17. The PRDS on LAS17 follow one after another and the flexibility to bind any one of them to the SH3 domain of LSB3 could help stabilize the binding interaction at different geometries as part of the higher-order actin patch assembly.
The IIN contains more nodes than the PPI, with each node now representing a distinct interface rather than a protein. In general, one would expect such an increase because proteins are known to have evolved multiple domains or interfaces to bind specific partners. The increase in nodes is much greater than the increase in edges from the PPI to the IIN, and therefore the IIN is substantially less densely connected than the PPI, with the average degree dropping from 6.4 to 2.06. In this now sparsely connected network, the clustering coefficient has dropped from 0.46 in the PPI to zero in the IIN. To quantify the significance of this result we generated randomized versions of the IIN that maintain the same number of nodes and the same degree distribution. We find that the randomized networks have distinctly higher clustering coefficients than the IIN (Table 2), suggesting that the structure of the IIN has evolved against having interfaces that bind to one another sharing the same partners. This is in contrast to the PPI network, where the relatively large clustering coefficient reflects the likelihood that two proteins that interact with one another share interaction partners.
To further quantify the significance of the local structural elements in the IIN, we evaluate the relative abundance of all 6 different types of 4-node motifs in the network. As shown in Table 2, their abundance in the IIN differs significantly from randomized networks. In particular, 4-node hubs with one shared interface binding to four non-shared interfaces are abundant in the actual IIN, and 4-node chains with four shared interfaces forming a linear chain of interactions are suppressed. The low abundance of these motifs is expected from network specificity optimization [37]. Interestingly, 4-node squares formed by four shared interfaces binding as in A1-B1, B1-A2, A2-B2, B2-A1 are also enriched in the IIN, forming a motif that has high specificity and can arise from gene duplication. Collectively these abundance shifts suggest that the local structure of the IIN is not random but reflects distinct evolutionary mechanisms acting on its topology.
Moving to the global properties of the IIN, we find that the distribution of binding partners per interface follows a power law quite well, with most interfaces having only a single binding partner (Figure 4b). The degree distribution of the IIN is constrained by the parent PPI degree distribution but not fully determined by it, as a PPI can theoretically give rise to many IINs with distinct numbers of nodes and connectivity (but each IIN uniquely defines its parent PPI). Hence a power-law distribution of the number of partners per interface is not a trivial outcome of having a power-law distribution of the number of partners per protein. We do expect that the number of nodes in the IIN will increase relative to the PPI and therefore the number of partners per node will be split between more nodes (assuming the number of edges stays about the same). How exactly the degree distribution changes from PPI to IIN then depends on whether it is mostly the highly connected hub proteins that are split about equally between multiple interfaces, or whether some interfaces retain large portions of binding partners and several single partner interfaces are created. In the CME network, the maximally connected node in the PPI (actin) is split between interfaces, but not evenly, such that one interface retains the majority (16) of the 23 binding partners. Overall, the IIN contains a significant number of highly connected nodes, just as in the PPI. The biggest change in the degree distribution from the PPI to the IIN was the formation of many single-partner interfaces in the IIN, whereas a protein in the PPI was more likely to have at least 3 partners. We discuss further below whether these trends might be conserved in other IINs.
Another distinguishing feature of the IIN is its fragmentation into modules, unlike the densely connected PPI. Compared to randomized networks, the CME IIN has a diverse distribution of module sizes, with many small fragments, whereas randomized networks all have a single giant connected component alongside many small fragments (Figure 5). In fact, the number of interfaces in each CME fragment again appears to follow a power law distribution with an exponent of about −2 (Figure 5). As a result, isolated small modules dominate, but larger connected networks even at the interface level are not uncommon. One must keep in mind, though, that here we focus on only a limited, functionally defined module. In future studies, it will thus be interesting to examine other IINs resolved at the same level of detail.
The modules in the IIN start to show clustering of interfaces with shared properties, although to varying degrees. In Figures 3a and 3b, we colored the interfaces according to specific domain types that are repeated in the network: PRDs and SH3 domains; EH domains and NPF motifs; phosphorylation sites and kinase domains; clathrin boxes; acidic domains; and subunit-subunit interfaces. As seen in Figure 3a, at the PPI level these interface types are mixed (i.e., distributed across different proteins); by contrast, we find them to be clustered into separate IIN modules. In randomized networks such clustering is not observed. This clustering of interface types reflects the need for binding interfaces to maintain high specificity towards their complementary binding partners and against binding towards unrelated interface sequences [37]. We note that our choice of defining all phosphorylation sites as distinct interfaces places them all in the same module (see Methods), whereas an alternative definition (for example, treating any phosphorylated residues overlapping with other interfaces as forming shared interfaces) would distribute some of them throughout the network. By contrast, the actin ACT1.2 interface is part of a large module with significant heterogeneity in domain types, as discussed further below. Because these binding interfaces do not all contact the same residues of the ACT1.2 interface, they do not all classify according to a single domain type. The convergence of these distinct partners to bind a single protein surface seems more likely a result of functional selection rather than duplication and divergence [38].
The IIN shares some of the scale-free characteristics of PPI networks [30], yet differs markedly in a number of network topological properties, including a lower average degree of the IIN and a more fragmented structure. While strictly applying only to the CME IIN, we expect many of these results to be conserved in IINs derived from larger PPI networks. First, the comparison of the IIN structure with randomized networks suggests evolutionary pressure acting on the IIN to prevent both giant connected components and a high clustering coefficient (where two interacting interfaces have the same partners). Second, interfaces that have only a single partner should be robustly conserved even for larger networks because they frequently mediate inter-subunit contacts (see light green nodes in Figure 3), and can evolve to high specificity [37]. A noticeable increase in singly connected nodes when transitioning from PPI to IIN would contribute to a steep power law-type degree distribution as a general trend. If a few hub interfaces and many single interfaces were maintained in other IINs, their degree distributions would resemble power laws. The degree distribution of interface partners is noteworthy because power-law distributions indicate networks robust against attacks on specific nodes [39], as would occur from mutations to specific binding surfaces or targeting by binding inhibitors.
In a separate study, we will pursue the hypothesis that the structure of the IIN evolved to minimize nonspecific binding, and that therefore the network features of the IIN encode important physical and biological functions of the proteins. Since minimization of nonspecific binding is a physical pressure common to all proteins [37], [40], we would predict that these topological features would then be conserved in all IINs, not just for CME proteins.
A number of distinct patterns emerge in the CME IIN. From the degree distribution of the IIN, we can contrast the properties of single interfaces from hub interfaces. More than a quarter of the single partner interfaces come from interfaces between subunits of a multi-subunit complex like ARP2/3 [41]. Dimerization interfaces also tend to be single partner interfaces. The most highly connected interfaces, or hub interfaces, are a surface on the actin protein with 16 partners, and several SH3 domains. The actin surface is distinct from the SH3 domains in that its binding partners do not all conform to the same binding type. The binding interface ACT1.2 is a relatively large and flat region spanning parts of subunits I, II and III of actin (Figure 1a), where not all binding partners use the same set of residues to stabilize their interactions, but the overlap is still significant. While it is certainly possible that with additional residue information this interface could be refined and split into more than one binding site, the extensive sharing of the ACT1.2 interface is consistent with earlier studies that found flat interfaces to provide a better platform for binding a large variety of partners [42], as geometrical packing need not be as optimized. Furthermore, we note that the nucleotide binding state of actin strongly tunes the affinity for its distinct partners.
The IIN overlaid on the PPI reinforces that many of these endocytic proteins are able to bind multiple partners simultaneously because of the number of distinct interfaces. This directly observable insight would be lost if one only categorized protein-protein interactions (i.e., edges in the PPI) as either competing or not, since many proteins have multiple shared interfaces. The interface assignments also highlight redundancy in the network, where the recruitment of a particular protein during the endocytic pathway could happen via multiple mechanisms, as many of the proteins are chimeras of the most frequently represented domains [43] in this network. Of the endocytic proteins, 16% contain SH3 domains, whereas in the entire yeast proteome <1% of proteins contain SH3 domains.
The designation of distinct domains on each protein allows one to contrast the specific structural elements present in these CME proteins versus CME proteins in other organisms. Much of the CME pathway between yeast and mammals is conserved. However, a major distinction is that CME in yeast requires the actin network to initiate the membrane invagination [44], whereas in mammals the actin network is engaged only in some cases in the later stages of vesicle budding [45]. It is interesting to note that of the 9 CME proteins in yeast without functional homologs in mammals [26], all but one (PAL1) engage in SH3 or PRD interactions (LSB3, LSB4, LSB5, BBC1, AIM21, BSP1, AIM3, APP1). This finding is statistically significant, having only a ∼0.15% probability to occur by chance (as determined by the probability of choosing at random 8 or more proteins out of 9 that have SH3 or PRD interactions, with 22 candidates among the 56 proteins of the CME network). The SH3 domains in the CME network recruit proteins throughout the progression of the vesicle budding process after the initial clathrin coat assembly [35]. The abundance of SH3 domains in yeast CME proteins likely reflects the central role of their interactions in connecting the growing clathrin coated pit to the actin cytoskeletal network of yeast.
Distinguishing interface domains in each protein also enables direct visual identification of multi-interface proteins that act to bring together multiple proteins with different functions, which again is not possible if one only marks edges in the PPI as competing. Both PAN1 and SLA1 have many interfaces that can connect simultaneously to both the scaffolding proteins of the clathrin pit formation (through PAN1's EH domain [32] and SLA1's clathrin box [46]), and to the actin polymerization proteins via PRDs, SH3 domains, and acidic domains. LAS17, on the other hand, does not connect directly to the scaffold proteins of clathrin pit formation but rather has distinct interfaces to bind both SH3 proteins and the ARP2/3 complex. While the role of LAS17 is not fully understood [26] and appears to involve both activation and inhibition of actin branching [47], the designated interface-interface pairs provide a basis for grouping the many functions of this protein along with distinct CME proteins according to domain types (including PRDs, acidic domains, the C-helix and WH2 domains). Lastly, some multi-interface proteins in the network, such as Arc15 and Arc19, contain only subunit-subunit interface domains, indicating that they function as structural components of a multi-subunit complex.
Designing any ligand, and in particular a drug molecule, to bind exclusively to its intended target without cross-reactivity requires not only positive selection for the specific target but also negative selection against related targets [37]. The clustering of interfaces in modules in the IIN provides a tool for predicting which binding partners of an interface are the most selective for its surface and do not bind to related domains. For example, both RVS167.2 and the SH3 domain of YSC84/LSB4 (YSC84.1) bind several of the same PRDs. Obtaining target specificity for only one of those interface sites benefits from knowing which PRDs are specific to only one of these interfaces. The interfaces VRP1.0, BSP1.3, and ABP1.1 that bind RVS167 but not YSC84, and ABP1.5, LAS17.6 and AIM21.0 that bind YSC84 but not RVS167, could be used as templates for targeting only one of the two SH3 domains.
Collectively, the information on interface connectivity and protein connectivity combined in a network format provides important guidance for the selective inhibition or activation of specific pathways, for drug targeting, and for predicting response to surface mutations. The PPI network is essential for identifying which proteins interact in a functional pathway, but the details of the IIN allow one to isolate specific binding sites while conserving the functionality of other sites. The IIN also allows one to predict how drugs designed as roadblocks along a certain pathway could be bypassed by alternate available interface interactions.
For example, one might expect that inhibiting or mutating ‘hub’ interfaces, much like knocking out ‘hub’ proteins, would induce a more severe phenotype. RVS167.2 is found to interact with 12 PRD interfaces as part of the PRD-SH3 IIN sub-network in the top right of Figure 3b. These interactions are not immediately apparent in the PPI network, lacking interface resolution. While early studies [34], [48], [49] already pointed to the prevalence of these interactions, their functional importance and temporal recruitment in endocytosis is emerging only now [35], [50], [51]. Mutations of the SH3 domain of RVS167 that leave its membrane shaping BAR domain intact still significantly alter the endocytic phenotype [50]. A substantial phenotypic response to such a localized mutation would be anticipated from the IIN because removing that particular node removes multiple edges. However, the fragmented and clustered structure of the IIN also provides a more detailed perspective on the response to deleting this node. Although targeting the SH3 domain of RVS167 would inhibit 12 RVS167 binding interactions, one can see in the IIN that most of those interface partners can also bind to alternate SH3 domain containing proteins and all of the interface partners are on proteins with a PRD that can bind an alternate SH3 domain. These alternate pathways may help explain why mutations of the SH3 domain of RVS167 do not eliminate endocytic function in yeast [50].
The inhibition of particular binding sites would have unexpected results if there were nodes or edges missing from the network. For example, truncation of the clathrin N-terminal domain (CHC1.1 in Figure 3b) was accurately predicted to cause a severe endocytic phenotype by preventing recruitment of clathrin to the membrane. However, when the known binding sites on the N-terminal domain were mutated, the expected result was not observed, and this led to the identification of duplicate binding sites in the N-terminal domain [52]. We do note that the CME IIN overlaid on the PPI proposes another mechanism for recruitment of clathrin to the membrane via binding of the clathrin light chain (CLC1) to SLA2, which can then bind the membrane or other membrane bound proteins. This interaction may be too weak to recruit clathrin on its own, or SLA2 may be too small to bridge the large separation from the clathrin light chain to the membrane.
The IIN also indicates sets of opposing or reciprocal mutations, or truncations that should result in the same phenotypic response. The prediction of the identical responses assumes that the binding partners of the targeted interface act independently of other binding partners. The extent to which these assumptions are violated could suggest allostery or cooperativity between the affected partners. Identifying interfaces for reciprocal mutations could then offer a tool for testing cooperativity or dependence between binding interactions or for identifying missing interfaces. In the clathrin CHC1.1 interface example given above, the IIN would predict a reciprocal mutation to all five clathrin boxes to give the same phenotype as the removal of the CHC1.1 interface. If, instead, clathrin were still recruited to the membrane, then one expects other clathrin N-terminal binding sites to be missing from the network.
In another module, the EH domains are shown with their NPF motif binding partners. Based on the specificity of these interfaces for one another only in the IIN, one would expect that mutations to either the NPF motifs (including all copies) or to the EH domains (including all copies) would generate the same phenotype. The extent to which they do not match would first indicate possible missing nodes from the network. Alternatively, the result could indicate that one of these domains acts cooperatively with another domain to affect the global behavior of the protein, not just this specific interaction.
In terms of the anticipated biological response to mutation of either the EH domain or the NPF motifs (assuming independence), this interaction helps stabilize a scaffold of proteins at the membrane that recruits the clathrin trimer. From the IIN combined with the PPI, cutting these edges out of the network would not prevent any of the proteins from connecting to the membrane or the early coat module, as PAN1 could still connect via SLA2 and EDE1 via SYP1. Clathrin and actin would still be recruited normally. What this mutation should affect is crosslinking between these proteins and therefore clustering of these proteins in one place on the membrane. If crosslinking and clustering of proteins is necessary for efficient coat formation then eliminating these interactions could decrease or slow down clathrin-pit formation.
As one of the main challenges in IIN construction, there is more than one way to define whether a binding interface on a protein is shared between multiple binding partners or is completely distinct. The two main criteria we use to characterize shared and distinct interfaces are (1) if the same residues are present in both interfaces, (2) if the binding of one protein partner would interfere with the binding of another partner due to structural overlap or allosteric effects. Both criteria are important to the function of the proteins in the cell. Concerning the first criterion, the sequence makeup of the interface is central to achieving binding specificity, as even proteins with the same domain structures do not necessarily share the same partners [35]. Furthermore, the residues involved in a binding interaction are not only important for binding to their specific partner but also for avoiding the formation of nonfunctional interactions with the other proteins in the cell [37]. This negative selection on an interface can contribute to optimizing the specificity and strength of functional binding interactions [25]. Concerning the second criterion, determining whether two potential binding partners can both bind at the same time to form a trimer is important for modeling the dynamics of protein association, as competition for binding partners will affect concentrations of available protein. The same is true if a protein has repeated copies of the same domain and can therefore bind multiple copies of the same binding partner. However, it may not always be possible to assign distinct interfaces that meet both criteria of sequence specificity without any steric obstruction and therefore in this work we consistently emphasize residue detail where the information is available. Otherwise we did use competition for binding partners as grounds for defining shared versus distinct interfaces. In future work it would be valuable to annotate both the residues involved in each interface as well as whether each pair of distinct interfaces on a protein can bind their partners simultaneously.
The procedure of manually assigning interfaces has also highlighted some important issues for consideration in computerized interface assignment. For one, residue overlap does not necessarily mean that proteins compete for binding to the protein, as demonstrated by multi-subunit complex formation (Table 3). None of the interface interactions within the complex would be considered shared because they are all bound together at the same time. The majority of interfaces do not overlap, but ∼30% of the bound partners share one or more residues in their interactions. Most commonly the overlap was only one or two residues, and the corresponding percentage of the interface varied substantially depending on the size of the protein. Thus it seems reasonable to allow 1–2 residues of overlap before defining interfaces as shared. This policy is also consistent with the assignment of different domains as distinct interfaces, even though the 3D structure of the protein might produce some residue overlap between two distinct domains. We note that here we did not use a strict cutoff in our assignments because through manual curation we treated each interaction on a case-by-case basis, merging residue level detail with experimental data on simultaneous binding partners.
For attempts at homology modeling or docking, it would first be useful to assess how reliable a purported interaction might be. Particularly in the case of interactions involving subunits of multi-protein complexes, many of the interactions are actually indirect. Arp2, for instance, has relatively high homology to actin and shares several binding partners; however, Arp2 acts as part of a multi-subunit complex and binds to these shared partners (such as LAS17) in distinct ways. Also, higher thresholds for sequence similarity could be warranted in particular cases, such as SH3 domains, where small variations in sequence distinguish specific from nonspecific partners [25].
One of the major distinctions between the procedure used here and current automated methods is the inclusion of detailed information on binding interfaces between proteins from biochemical studies, not just from high-resolution protein structures. This information preempts or complements the use of homology or docking models of protein interactions. Unfortunately, the domain or interface details from these studies is not collected in a convenient database, whether it is the specific residues that comprise the interface or more general information on inhibition or competition between binding partners. There are also ambiguities and inconsistencies in existing data that are difficult to resolve without combining multiple literature resources in a coherent analysis. Nevertheless, mining these data would provide a valuable resource for generating more complete networks of interface-interface interactions.
Our protein list is composed of 56 proteins that were selected because they all participate in the yeast clathrin-mediated endocytosis pathway and have been identified as central components [26]. We downloaded the physical interaction partners of the 56 proteins of the endocytic functional module in yeast via the Saccharomyces Genome Database (SGD) [29] interaction list compiled from BioGRID [28] and directly from the IntAct, MINT, DIP, and BIND protein interaction databases. We kept only the interactions between the subset of 56 proteins to define the initial set of experimentally determined protein-protein interactions. We disregarded genetic interactions, as they do not imply that the proteins directly interact with one another, but rather that their expression or phenotype is correlated. The overlap in databases was quite large for these proteins, with BioGRID containing the largest number of interactions and missing interactions coming not from missed references but from missed interactions within the same references.
Given a PPI network, the first step in assigning the binding interfaces was collecting information on the particular proteins from the SGD [29]. The SGD combines information from various databases on each yeast gene. The major data sources we used were the list of referenced physical interactions loaded from the various PPI databases and the available PDB structures. The protein tab also provides a useful guide to the size, sequence, domain structure, and function of each protein.
Crystal structures of complexes were available for a few of the protein interactions, including the ARP2/3 complex and several actin binding interactions (shown as black edges in Figure 3). We ensured that we matched the numbered PDB residues (which sometimes started at zero arbitrarily) to the correct sequence region on the protein of interest. For protein homologs from species other than yeast, the sequence alignment is also provided for positioning the interface on the yeast protein of interest. Compared across species, actin has high (87%) sequence homology and structures from other species were simply used as proxies for the expected interaction in yeast. To assign residues involved in the protein interfaces from a PDB complex we used a 4-Å cutoff between non-hydrogen atoms and required that at least 3 residues contacted one another in each interface. Cofactors such as metal ions and water molecules were not considered in assigning whether two proteins interacted or which residues formed the interfaces. Some of the protein structures had multiple missing residues for crystallization purposes, such that the assigned interface may be smaller than in the complete protein. By using the PDB structures we eliminate all indirect interactions that are often assigned to protein subunits of a large complex in high-throughput AP/MS and PCA. We did not use any predicted models of protein complexes [23] because direct information was generally available through literature studies and because protein homologs (e.g., Arp2 and actin) do not always share the same set of binding interactions.
In most cases, crystal structures were not available and instead the literature references from the PPI databases were used to assign interfaces. Binding to proteins outside the endocytic network, as listed in the SGD, was ignored. Nearly all of the edges to which we assigned interfaces were implicated as binding in more than one experiment. We have collected all the justifications for each assignment into a spreadsheet with references (see Table S1), categorized the support for each interface assignment with edge colors in Figure 3, and below we describe additional criteria we used to define the interfaces for the specific cases of kinase binding and SH3 domains binding to PRDs.
Several of the endocytic proteins have SH3 domains (BZZ1, ABP1, LSB3, LSB4, RVS167, BBC1, MYO3, MYO5, and SLA1) and PRDs to which SH3 domains bind (VRP1, LAS17, MYO5, APP1, AIM21, AIM3, SCD5, BBC1, ABP1, ARK1, PRK1, INP52, SCP1, BSP1, SLA1, SYP1, GTS1). We took advantage of several large-scale studies [34], [35], [48] focused on identifying which PRDs bind to which SH3 domains by compiling all interactions noted for our 56 proteins (including those interactions missing from the PPI databases). Tonikian et al. [35] provide the most recent and comprehensive study to identify PRDs by combining data from three independent experiments. We assigned the PRD and SH3 interfaces if the interactions were observed by Tonikian et al. and at least one other experimental study. As one exception to this criterion, if there is only one supporting experiment, yet that experiment found a different PRD site, then the interface was left unassigned. Lastly, if more than two references reported binding and the PRDs were different, the two PRDs were combined into one binding site. Binding multiple PRDs on the same protein has been experimentally demonstrated [34], but Tonikian et al. only report the most likely PRD, so this does not rule out additional PRDs. We merged the two SH3 domains of BZZ1 to improve the consensus of their binding partner interfaces but kept the two SH3 domains of SLA1 separate. We separated the multiple PRDs of LAS17 into distinct binding sites as multiple lines of evidence implicated specific binding partners for specific regions. These details are collected in Table S1, under tabs 2 and 3.
The endocytic protein subset contains three kinases (ARK1, PRK1, AKL1) and similar to the SH3 domains, the specificity of kinases for their phosphorylation targets has also been studied at large scale [53], [54]. We here again compiled the interactions from Breitkreutz [53], Mok [55], and Ptacek [54] and their collaborators (again including some interactions missing from the PPI databases), and assigned the interactions if the binding was reported in at least two references. Because most of the targets in Mok et. al. [55] were predicted but not verified, we included these sites as references only if they were also experimentally tested or observed in previous mass spectrometry experiments. These details are collected in Table S1, under tab 4.
In some cases the data implicating two proteins as interacting only came from high-throughput studies and these interactions were generally unassigned. Others came from a literature source that did not isolate binding interfaces, with no additional evidence available from homologs or functionally related proteins. Edges that were identified between the ARP2/3 complex subunits and other proteins were considered indirect if PDB structures or biochemical evidence implicated a specific subunit in the direct interaction. For a few interactions, evidence from the literature suggested that such proteins did not bind directly to one another upon further investigation, and as a result these edges were removed. We note these in the interaction table. For example, we were unable to find any evidence for the protein RVS161 forming direct physical interactions with any proteins other than RVS167. Furthermore, there was some biochemical evidence suggesting that proposed edge interactions were mediated via RVS167 rather than directly through RVS161 [56], as they operate as an obligate dimer.
We added 9 new edges to the network to account for literature studies providing evidence for the binding interactions. These were largely actin related interactions that lacked references in the PPI databases but have been well established as functionally important binding partners of actin. One was an SH3-PRD interactions defined in two separate publications that were missing from the database.
Several of these proteins have domains known to bind at the membrane that are important to their function in endocytosis. Therefore we pointed these out on the protein-interface interaction network in Figure 3a to facilitate the prediction of functional responses to mutation.
As the first criterion to assign an interface, we used the residues involved in the binding, if available. Specific residues were available from PDB structures and for several peptide motifs like PRDs [34], [35] and clathrin boxes [32], [46], [57]. If two partners of a protein bind to an interface using some overlapping residues we did not automatically classify the interface as shared. There are two reasons for this decision, the main one being that sharing one or a few residues does not mean those two proteins cannot bind simultaneously. To demonstrate this point we calculated the percent of distinct interface pairs within a multi-subunit protein complex that had overlapping residues. For each of the complexes we considered, there are some pairs of interfaces that have one or more residues in common (Table 3). Even if the interfaces are defined at the atomic rather than residue level, there is still a fraction of atoms within the cutoff distance of both binding partners. The second main reason is that even if the binding partners cannot bind simultaneously, the specificity and stability of their interactions may be mediated through chemically distinct binding sequences. For example, we chose to treat a kinase's phosphorylation binding site as distinct from other protein binders that may interact with the phosphorylated residue because of their distinct binding modes. However, if the residue overlap is substantial, as is the case for many proteins that bind to actin in similar but not identical ways, then the interface is considered shared.
When the specific residues of the folded protein interfaces were not available, the next description of the interface was the domain structure represented by sequential sequence residues (e.g., the SH3 domain contained in residues 1–51). These domains were generally identified in biochemical studies and the sizes of the domains varied from a few residues (e.g., clathrin boxes) to hundreds of residues (e.g., coiled-coil domains). In some cases the assigned interfaces may not represent a known domain but they are designated as unique interfaces because they do not overlap with any of the protein's other binding partners. Lastly, if residue level detail is not available, then the fact that two binding partners are competing with one another is used as justification for listing the interface as shared.
To summarize, we did not use a strict cutoff of overlapping residue numbers for defining shared versus distinct interfaces. All subunits of a multi-subunit complex were assigned distinct interfaces for these inter-subunit contacts because they could clearly bind simultaneously. This is despite the fact that pairs of proteins could have as many as 10 overlapping residues if a long disordered region of a protein sat at the seam of an interaction between two other proteins. For most biochemical studies, stretches of residues were identified and shared interfaces were assigned when proteins bound to overlapping stretches of residues and there was no evidence that they could bind simultaneously. For the distinct surfaces in actin, there was in some cases overlap between residues, but there was also evidence that the proteins could bind simultaneously. For example, several actin binding proteins bind to the actin filaments, and therefore they can bind simultaneously with the actin-actin binding interactions, despite overlapping with some residues.
In representations simpler than the IIN, edges in the PPI network have been marked as shared. To extend the representation to full interface assignments, one must keep track of possible overlap in all pairs of binding interactions for each protein. Given a protein that has k binding partners, there are k(k−1)/2 possible pairs of partners sharing an interface. To keep track of the interface assignments, each protein had its own file with a k-by-k matrix indicating the overlap between the k binding partners (Table 4). The diagonal entries are null and the off-diagonal entries of the symmetric matrix are 0 if the two partners use separate interfaces and 1 if the two partners use the same interface. Some protein-protein interactions are controlled by more than one set of interfaces and would require an additional entry into the matrix. The binding interfaces from each protein can then be consolidated into a network representing a connected set of interface interactions. We note that in a matrix representation it is possible to define a case where one interface overlaps with two others that do not overlap with each other, and this detail cannot be captured in a simple interface network picture. This would be the case, e.g., if two proteins A and B bind to two distinct parts of a protein X and the third protein C binds across those two complete interfaces on protein X. However, this issue can easily be fixed by splitting protein C's interface into two interfaces to bind the two parts of protein X. For example, this splitting was done for LAS17's CA region that binds to ARP3 through both its C interface and its separate A interface [47].
We evaluated clustering coefficients of our networks using the expression [58]where Nclosed(i) counts how many distinct pairs of the k(i) partners of interface i have an edge between them to form closed triangles with node i. Self-loops were ignored in this calculation. We also use a global clustering coefficient Cglobal as the number of distinct closed triangles Ntriangle in the network divided by the total number of distinct triplets,with Nopen the number of open triplets.
We computed degree distributions, p(k), where k counts the number of partners per node (either protein or interface), and p(k) is the probability for finding a node in the network with that degree. For the degree distribution, we note that we treated self-loops as a single partner, rather than the standard method of treating a self-loop as counting as degree of 2, so that the degree would reflect the physical number of binding partners per protein (or interface).
We enumerated the 4-node motifs present in our networks by identifying all distinct sets of 4-node subgraphs that are connected by at least one path (each node in the subgraph can be reached by the others). There are six distinct 4-node subgraph architectures [59] and we note that they are all counted mutually exclusive to one another, i.e., a set of 4 nodes uniquely classifies as one of the six subgraphs. A single node may belong to more than one 4-node subgraph. Hub and chain motifs have 4 nodes connected by 3 edges, flag and square motifs have 4 nodes connected by 4 edges, and the other two 4-node subgraphs contain 5 and 6 edges.
To generate networks that shared the same number of interfaces, edges, and the same degree distribution as the IIN in Figure 3b, we used the Monte Carlo method of Maslov and Sneppen [60]. Specifically, in a trial move two interfaces were selected randomly and a partner from each of these interfaces was randomly selected. The partners were then swapped between interfaces, unless one of these new edges already existed, in which case the move was rejected.
We fit our degree distributions to power laws using the maximum likelihood method, where the discrete data is fit to a power law distribution x−γ/ζ(γ) normalized over the range x≥xmin [31]. We measure the goodness-of-fit using the Kolmogorov-Smirnov metric and calculate the p-value for the data being drawn from a power law density using the method of ref. [31]. For the p-value calculation, our null hypothesis is that the data is drawn from a power-law density. Therefore, a small p-value of <0.05 would reject this null hypothesis and demonstrate that our data is not described by a power law. A large p-value, on the other-hand, indicates that the data is consistent with the hypothesis that it was drawn from a power law distribution.
|
10.1371/journal.pgen.1000832 | U87MG Decoded: The Genomic Sequence of a Cytogenetically Aberrant Human Cancer Cell Line | U87MG is a commonly studied grade IV glioma cell line that has been analyzed in at least 1,700 publications over four decades. In order to comprehensively characterize the genome of this cell line and to serve as a model of broad cancer genome sequencing, we have generated greater than 30× genomic sequence coverage using a novel 50-base mate paired strategy with a 1.4kb mean insert library. A total of 1,014,984,286 mate-end and 120,691,623 single-end two-base encoded reads were generated from five slides. All data were aligned using a custom designed tool called BFAST, allowing optimal color space read alignment and accurate identification of DNA variants. The aligned sequence reads and mate-pair information identified 35 interchromosomal translocation events, 1,315 structural variations (>100 bp), 191,743 small (<21 bp) insertions and deletions (indels), and 2,384,470 single nucleotide variations (SNVs). Among these observations, the known homozygous mutation in PTEN was robustly identified, and genes involved in cell adhesion were overrepresented in the mutated gene list. Data were compared to 219,187 heterozygous single nucleotide polymorphisms assayed by Illumina 1M Duo genotyping array to assess accuracy: 93.83% of all SNPs were reliably detected at filtering thresholds that yield greater than 99.99% sequence accuracy. Protein coding sequences were disrupted predominantly in this cancer cell line due to small indels, large deletions, and translocations. In total, 512 genes were homozygously mutated, including 154 by SNVs, 178 by small indels, 145 by large microdeletions, and 35 by interchromosomal translocations to reveal a highly mutated cell line genome. Of the small homozygously mutated variants, 8 SNVs and 99 indels were novel events not present in dbSNP. These data demonstrate that routine generation of broad cancer genome sequence is possible outside of genome centers. The sequence analysis of U87MG provides an unparalleled level of mutational resolution compared to any cell line to date.
| Glioblastoma has a particularly dismal prognosis with median survival time of less than fifteen months. Here, we describe the broad genome sequencing of U87MG, a commonly used and thus well-studied glioblastoma cell line. One of the major features of the U87MG genome is the large number of chromosomal abnormalities, which can be typical of cancer cell lines and primary cancers. The systematic, thorough, and accurate mutational analysis of the U87MG genome comprehensively identifies different classes of genetic mutations including single-nucleotide variations (SNVs), insertions/deletions (indels), and translocations. We found 2,384,470 SNVs, 191,743 small indels, and 1,314 large structural variations. Known gene models were used to predict the effect of these mutations on protein-coding sequence. Mutational analysis revealed 512 genes homozygously mutated, including 154 by SNVs, 178 by small indels, 145 by large microdeletions, and up to 35 by interchromosomal translocations. The major mutational mechanisms in this brain cancer cell line are small indels and large structural variations. The genomic landscape of U87MG is revealed to be much more complex than previously thought based on lower resolution techniques. This mutational analysis serves as a resource for past and future studies on U87MG, informing them with a thorough description of its mutational state.
| Grade IV glioma, also called glioblastoma multiforme (GBM), is the most common primary malignant brain tumor with about 16,000 new diagnoses each year in the United States. While the number of cases is relatively small, comprising only 1.35% of primary malignant cancers in the US [1], GBMs have a one-year survival rate of only 29.6%, making it one of the most deadly types of cancer [2]. Recent clinical studies demonstrate improved survival with a combination of radiation and Temozolomide chemotherapy, but median survival time for GBM patients who receive therapy is only 15 months [3]. Due to its highly aggressive nature and poor therapeutic options, understanding the genetic etiology of GBM is of great interest and therefore, GBM has been selected as one of the three initial cancer types to be thoroughly studied in the TCGA program [4].
To that end, numerous cell line models of GBM have been established and used in vast numbers of studies over the years. It is well recognized that cell line models of human disorders, especially cancers, are an important resource. While these cell lines are the basis of substantial biological insight, experiments are currently performed in the absence of genome-wide mutational status as no cell line that models a human disease has yet had its genome fully sequenced. Here, we have sequenced the genome of U87MG, a long established cell line derived from a human grade IV glioma used in over 1,700 publications [5]. A wide range of biological information is known about this cell line. The U87MG cell line is known to have a highly aberrant genomic structure based on karyotyping, SKY [6], and FISH [7]. However, these methods neither provide the resolution required to visualize the precise breakpoint of a translocation event, nor are they generally capable of identifying genomic microdeletions (deletions on the order of a megabase or less in size) in a whole genome survey of structural variation. SNP genotyping microarrays can be used to detect regions of structural variation in the forms of loss of heterozygosity (LOH) and copy number (CN) based on probe intensity, but do not reveal chromosomal joins. To assess the genomic stability of U87MG, the genome was genotyped by Illumina Human 1M-Duo BeadChip microarray. In spite of being cultured independently for several years, the regions of LOH and the CN state of our U87MG genome matched exactly with data retrieved from the Sanger COSMIC database for U87MG [8], which had been assayed on an Affymetrix Genome-Wide Human SNP Array 6.0. This suggests that although U87MG bears a large number of large-scale chromosomal aberrations, it has been relatively stable for years and is not rapidly changing. This suggests that prior work on U87MG may be reinterpreted based on the whole genome sequence data presented here.
The first draft of the consensus sequence of the human genome was reported in 2001 [9],[10]. The first individual human diploid sequence was sequenced using capillary-based Sanger sequencing [11]. Since then, a few additional diploid human genomes have been published utilizing a variety of massively parallel sequencing techniques to sequence human genomes to varying degrees of coverage, variant discovery, and quality typically costing well over $200,000 and several machine months of operation [12]–[16]. For the sequencing of U87MG, we utilized ABI SOLiD technology, which uses a ligation-based assay with two-base color-encoded oligonucleotides that has been demonstrated to allow highly accurate single nucleotide variant (SNV) and insertion/deletion (indel) detection [17]. Additionally, long mate-paired genomic libraries with a mean insert size of 1–2kb allowed higher clone coverage of the genome, which improved our ability to identify genomic structural variations such as interchromosomal translocations and large deletions. While longer insert sizes would improve resolution of some structural variants, during genomic shearing the highest density of large fragments occurs at 1.5kb, allowing a sufficiently complex library to be generated from only 10 micrograms of genomic DNA while still being well powered to identify structural variations. Here, we demonstrate that aligning the two-base color-encoding data with BFAST software and decoding during alignment allows for highly sensitive detection of indels, which have in the past been difficult to detect by short read massively parallel sequencing.
For cancer sequencing, it is important to assess not only SNVs, but indels, structural variations and translocations, and it is preferable to extract this information from a common assay platform. A major characteristic of the U87MG cell line that differentiates it from the samples used in other whole genome sequencing projects published thus far is its highly aberrant genomic structure. Due to its heavily rearranged state, we thoroughly and accurately assessed each of these major classes of mutations and demonstrated that small indels, large microdeletions and interchromosomal translocations are actually the major categories of mutations that affect known genes in this cancer cell line. These analyses provide a model for other genome sequencing projects outside major genome centers of how to both thoroughly sequence and assess the mutational state of whole genomes.
From ten micrograms of input genomic DNA, we performed two and a half full sequencing runs on the ABI SOLiD Sequencing System, for a total of five full slides of data [17]. Utilizing the ABI long mate-pair protocol, we produced 1,014,984,286 raw 50bp mate-paired reads (101.5Gb). In some cases the bead was recognized by the imaging software for only one read, thereby producing an additional 120,691,623 single end reads (6.0Gb). In aggregate, we generated a total of 107.5Gb of raw data (Table 1).
We also performed an exon capture approach designed to sequence the exons of 5,253 genes (10.7Mb) annotated in the Wellcome Trust Sanger Institute Catalogue of Somatic Mutations in Cancer (COSMIC) V38 [8], Cancer Gene Census, Cancer Genome Project Planned Studies and The Cancer Genome Atlas (TCGA) [4] GBM gene list using a custom-created Agilent array. This approach used the Illumina GAII sequencing system [18] to sequence captured DNA fragments using a paired end sequencing protocol. This resulted in 9,948,782 raw 76bp paired end reads (1.51Gb), and a mean base coverage of 29.5×. These reads were used to calculate concordance rates with the larger whole genome sequence dataset.
The Blat-like Fast Accurate Search Tool (BFAST) [19] version 0.5.3 was used to align 107.5Gb of raw color space reads to the color space conversion of the human genome assembly hg18 from University of California, Santa Cruz (http://hgdownload.cse.ucsc.edu/goldenPath/hg18/bigZips/, based on the March 2006 NCBI build 36.1). Duplicate reads, typically from the same initial PCR fragment during genomic library construction, were inevitable and accounted for 16.4% of the total aligned data. These were removed using the alignment filtering utility in the DNAA package (http://dnaa.sourceforge.net). A total of 390,604,184 paired end reads (39.06Gb), 266,635,829 (13.33Gb) unpaired reads, and 62,336,824 (3.12Gb) single end reads were successfully mapped to a unique location in the reference genome with high confidence for a total of 55.51Gb of aligned sequence (Table 1). For the exon capture dataset, we uniquely aligned 8,142,874 paired end reads (1.2Gb) and 1,097,000 (83Mb) unpaired reads for a total of 1.32Gb of raw aligned sequence (Table 2). Using the ABI SOLiD reads, we identified small insertions and deletions (indels), single nucleotide variants (SNVs), and structural variants such as large-scale microdeletions and translocation events. The exon capture Solexa reads were used to validate SNVs identified in the SOLiD sequencing.
The overall pattern of base sequence coverage from the shotgun reads changes across the genome, and as expected is highly concordant with the copy number state as determined by Illumina 1M Duo and Affymetrix 6.0 SNP analysis (Figure 1). Regions of two normal copies, such as chromosome 3, showed even base sequence coverage across their entire length (12.4 reads/base, excluding centromeric and telomeric regions which are not represented accurately in hg18). Meanwhile, regions with one-copy state according to the SNP chip, such as the distal q-arm of chr11 and the distal p-arm of chr6, show about half the base sequence coverage (7.2 reads/base) as a predicted two-copy region. Likewise, predicted three-copy state regions, such as the distal q-arm of chr13, show about 1.5 times the base sequence coverage of a predicted two-copy region. A complete deletion spanning the region on chromosome 9 that includes the CDKN2A gene is also seen in both the SNP chip and ABI SOLiD base sequence coverage. These data show at a very large scale that sequence placement is generally correct and supports the copy number state calls from the array based data.
Single nucleotide variants (SNVs) and small insertions and deletions ranging from 1 to 20 bases (indels) were identified from the alignment data using the MAQ consensus model [20] as implemented in the SAMtools software suite [21]. SAMtools produced variant calls, zygosity predictions, and a Phred-scaled probability that the consensus is identical to the reference. To improve the reliability of our variant calls, variants were required to have a Phred score of at least 10 and further needed to be observed greater than or equal to 4 but less than 60 times and at least once on each strand.
In total, we identified 2,384,470 SNVs meeting our filtering criteria. Of these, 2,140,848 (89.8%) were identified as exact matches to entries in dbSNP129 [22]. Exact matches had both the variant and observed alleles in the dbSNP entry, allowing for the discovery of novel alleles at known SNP locations. In total, 243,622 SNVs (10.2%) were identified as novel events not previously recorded in dbSNP 129. This rate of novel variant discovery is consistent with other whole human normal genome sequences of European ancestry relative to dbSNP [12]. These SNVs were further characterized based on zygosity predictions from the MAQ consensus model, separating SNVs into homozygous or heterozygous categories (Table 3). The observed diversity value for SNVs (θSNV, number of heterozygous SNVs/number of base pairs) across autosomal chromosomes was 4.4×10−4, which is generally consistent with the normal human genome variation rate.
For small (<21bp) insertions and deletions, 191,743 events were detected with 116,964 not previously documented in dbSNP 129. The same criteria as used for SNVs was used for determining if an indel was novel and they were further classified as homozygous or heterozygous using the SAMtools variant caller (Table 4). The observed diversity value (θindel, number of heterozygous indels/number of base pairs) across autosomal chromosomes was 0.38×10−4.
A subset of 38 variants meeting genome-wide filtering criteria, including a 20-base deletion, was tested by PCR and Sanger sequencing with 34 being validated. In summary, 85.2% of SNVs (23/27), and 100% of small insertions (3/3), deletions (4/4), translocations (3/3) and microdeletions (1/1) were validated in this manner (Table S1). While this is a small sample, it demonstrates an overall low false positive rate.
The size distribution of indels identified in U87MG is generally consistent with previous studies on coding and non-coding indel sizes in non-cancer samples [23]–[25]. Small deletion sizes ranged from 1 to 20 bases in size and their distribution approximates a power law distribution in concordance with previous findings [23] (Figure 2A). There is a small deviation from the power law distribution with an excess of 4-base indels in U87MG's non-coding regions (Figure 2A, red bars) [11],[26].
A similar trend is seen with insertions in non-coding sequence with the maximum observed insertion size of 17 bases (Figure 2B, red bars). The maximum insertion size observed is less than the maximum deletion size because it is easier to align longer deletions than it is to align insertions. Some small insertions and deletions are likely to be larger than the upper limit of 17 and 20 bases actually observed, but the 50-base read length limits the power to align such reads directly.
In coding regions, there is a bias towards events that are multiples of 3-bases in length that maintain the reading frame despite variant alleles, suggesting that many of these are polymorphisms (Figure 2A-deletions, Figure 2B-insertions, blue bars). In non-coding regions, only 10.8% of indels are a multiple of 3 bases in size, while in coding regions, 27.0% are 3, 6, 9, 12 or 15 bases in size. This trend is expected based on past observations of non-cancer samples [11],[26].
Observed SNV base substitution patterns were consistent with common mutational phenomena in both coding sequences and genome wide. As expected, the predominant nucleotide substitution seen in SNVs is a transition, changing purine for purine (A<->G) or pyrimidine for pyrimidine (C<->T). Previous studies have observed that two out of every three SNPs are transitions as opposed to transversions [27], and we observed that 67.4% of our SNVs were transitions, while 32.6% were transversions, a 2.07∶1 ratio. (Figure 3) However, in coding regions, there appears to be an increase in C->T/G->A transitions and a decrease in T->C/A->G transitions, whereas genome-wide these transitions were approximately equivalent.
To assess the coverage depth of the U87MG genome sequence, we followed Ley at al. [13] and required detection of both alleles at most positions in the genome. We utilized the Illumina 1M-Duo BeadChip to find reliably sequenced positions in the genome with an understanding that this may lead to bias towards more unique regions of the genome. In order to best use the SNP genotyping array data, we included only those regions that are diploid based on normal frequency of heterozygous calls and copy number assessment. This effectively permitted us to use the heterozygous calls for assessing accuracy of the short read data for variant calling (Figure 1). Only SNPs both observed to be heterozygous and that the Illumina genotyping chip called ‘high quality’ were used, which provided a total of 219,187 high quality heterozygous SNPs for comparison. 99.71% of these were sequenced at least once. After applying variant detection filtering criteria (see Materials and Methods) and assessing concordance between the sequence calls and genotyping array calls, 93.71% of the genome was sequenced at sufficient depth to call both alleles of the diploid genome. This is roughly equivalent to the likelihood of sufficient sampling of the whole genome when repeats and segmental duplications are excluded.
Notably, a variant allele was observed at every position called heterozygous by SNP chip, while a reference allele was observed at 201,414 (97.94%) positions. In other words, the SNV detection algorithm uniformly miscalled the homozygous variant allele. Filtering for quality causes a bias toward identifying SNVs at sites that have higher coverage. That said, after SNV quality filtering, diploid coverage of the cytogenetically normal portions of the genome was 10.85× for each allele, which is clearly adequate for calling over 90% of the base variant positions on each allele at high accuracy.
Because the positions of the genome included on SNP arrays is not a random sampling of the genome, we also assessed mapping coverage genome-wide. Of all bases in the haploid genome, 78.9% of the whole reference genome was covered by at least one reliably placed read. Of that portion of the genome, 91.9% of all bases were effectively sequenced based on passing variant calling filters (Phred>10, >4× coverage, <60× coverage). Thus, a total of 72.5% of the whole genome was sequenced, including repeats and duplicated regions, which is typical of short sequence shotgun approaches.
10.9Mb of genomic sequence was targeted consisting of the amino acid encoding exons of 5,235 genes and were sequenced to a mean coverage of 30× using the Illumina GAII sequencer. Given the larger variability of coverage from the capture data, only a subset of these bases (8.5Mb) was evaluable to determine the false positive variant detection rate from the complete genomic sequence data. This region contained 1,621 SNPs present in dbSNP129. Within the 8.5Mb of common and well-covered sequence in the genomic sequence data and the capture sequence, there were 1,780 SNVs called from the genomic sequence. The same non-reference allele was concordantly observed at 1,631 positions within the capture data. At 149 positions, the non-reference allele was not observed in the capture data, but the reference allele was detected. However, the mean coverage at these 149 positions was significantly lower than that of the other 1,631 positions (p = 0.0003), suggesting that the non-reference allele was not adequately covered and is under called in the capture data. Moreover, of the 1,621 dbSNPs in the region, the capture adequately covered only 1,515. In these data there was a bias for the pull down data to under observe the non-reference allele (Figure S1). The 106 dbSNP positions detected in the ABI whole genome sequence dataset were observed to all call the reported alternate allele from dbSNP. In theory, if these were errors, then non-reference base calls should be randomly distributed to the three alternate base calls. Thus, no discrepancies are reliably identified within the dbSNP overlap when a variant was called in the ABI genomic sequence data.
There were a total of 100 novel SNVs detected in the ABI genomic sequence dataset that were also very well evaluated in the Illumina pull down data with at least 20 high quality Illumina reads, such that the ABI sequence could be well validated. Of these, 2 of the 100 discovered variants in the genomic sequence dataset were not observed in the Illumina pull down sequencing dataset. Thus, of the entire 8.5mb interval there are 2 unconfirmed variants for an estimated false positive error rate of about 3×10−7 for the whole interval. Alternatively viewed, there were 100 novel SNVs, with a 2% error rate in those novel positions. Thus, the de novo false discovery rate may be as high as 2%. Extrapolating to the whole set of 243,622 novel SNVs, we expect up to 4,872 false positives SNVs. These observations are roughly concordant with a sampling of 37 novel SNVs (not in dbSNP) in the whole genome set selected for testing by Sanger sequencing. Of these, 34 out of 37 (92%) were validated.
There are now several publicly available complete genomes sequenced on next generation platforms. We compared the SNVs discovered in U87MG to two of these published genomes: the James D. Watson genome [12] and the first Asian genome (YanHuang) [14]. Further, we simultaneously compared each of these to dbSNP version 129 [22]. Compared with dbSNP, 10.2% of U87MG SNVs, 9.5% of Watson SNVs, and 12.0% of YanHuang SNVs were not present within dbSNP (Figure 4). As U87MG was derived from a patient of Caucasian ancestry, which is confirmed by genotyping, it is unsurprising to see a higher overlap with dbSNP for U87MG than for YanHuang. Between the three genomes themselves, 44.7% of U87MG SNVs overlapped with Watson SNVs while 60.0% of SNVs were in common with YanHuang SNVs. Only 8.5% of dbSNP SNVs were shared between Watson and U87MG, while 11.3% of them were shared between YanHuang and U87MG. Thus, there is not a substantially higher amount of SNVs in the U87MG cancer genome relative to normal genomes.
We utilized the predictable insert distance of mate-paired sequence fragments to directly observe structural variations in U87MG. Our target insert size of 1.5kb gave us a normal distribution of paired end insert lengths ranging from 1kb to 2kb with median around 1.25kb and mean around 1.45kb in the actual sequence data (Figure S2). We identified 1,314 large structural variations, including 35 interchromosomal events, 599 complete homozygous deletions (including a large region on chromosome 9 containing CDKN2A/B, which commonly experience homozygous deletions in brain cancer), 361 heterozygous deletion events, and 319 other intrachromosomal events (Table 5). The 599 complete microdeletions summed up to approximately 5.76Mb of total sequence, while the 361 heterozygous microdeletions summed to 5.36Mb of total sequence. Most of the microdeletions were under 2kb in total size. Because of the high sequence coverage and mate pair strategy each event was supported by an average of 138 mate pair reads. Mispairing of the mate pairs did occur occasionally due to molecular chimerism in the library fabrication process, but such reads occur at a low frequency (<1/40 of the reads). Thus, the true rearrangement/deletion events were highly distinct from noise in well-mapped sequences. Interchromosomal events included translocations and large insertion/deletion events where one part of a chromosome was inserted into a different chromosome, sometimes replacing a segment of DNA. All together, these structural variations show a highly complex rearrangement of genomic material in this cancer cell line (Figure 5). All identified structural variants are summarized in Table S2. We note as well that even when breakpoints are within genome-wide common repeats there can be sufficient mapping information to reliably identify the translocation breakpoint (Figure S3).
The thirty-five interchromosomal events often coincided with positions of copy number change based on the average base coverage (Figure 5). Figure 6 shows two interchromosomal events between chromosomes 2 and 16. The events on chromosome 16 are less than 1kb apart while those on chromosome 2 are about 160kb apart. Based on the average base coverage, there appears to be a loss of genomic material between the event boundaries on their respective chromosomes, shifting from two to one copy. Although we are unable to determine the origin of such an event, it appears that there was an interchromosomal translocation between chromosomes 2 and 16 with a loss of the DNA between the identified regions on each chromosome.
A subset of 3 translocations were confirmed by amplifying DNA from the breakpoint-spanning region by polymerase chain reaction and sequencing by dideoxy Sanger sequencing (Table S1). Each confirmed the predicted breakpoint to within 100 nucleotides of the correct position. In a subset of cases, unmapped short read fragments could be identified from the shotgun short read data that span the breakpoint and are concordant at base resolution with Sanger sequencing of PCR amplified product spanning the breakpoint
The SNVs and indels identified in U87MG were assessed for their potential to affect protein-coding sequence. We considered variants predicted to be homozygous and to affect the coding sequence of a gene through a frameshift, early termination, intron splice site, or start/stop codon loss mutation as causing a complete loss of that protein. We chose to focus on homozygous null mutations for two major reasons. First, this is an interesting set of genes that we can predict from the whole genome data are non-functional within this commonly used cell line. Although heterozygous mutations can certainly affect gene products in multiple ways, it is difficult to assess their effect from genomic data alone. Second, by cross-referencing such null mutations with known regions of common mutation in gliomas we can pick out specific candidates that are of interest to the glioma community.
Of the 2,384,470 SNVs and 191,743 small indels in U87MG, a total of 332 genes are predicted to have loss-of-function, homozygous mutations as a consequence of small variants (Table S3). Of these, 225 genes contained variants matching alleles annotated in dbSNP (version 129), while 107 contained novel variants not observed in dbSNP.
We further divided these homozygous mutant genes by variant type. Of genes mutated by SNVs, 146 contained variants present in dbSNP while only 8 were knocked out by variants not in dbSNP. The ratio of known SNPs causing loss-of-function mutations to total known SNPs (146/2,140,848 = 6.82×10−5) was not significantly different from the ratio of novel SNVs causing loss-of-function mutations over total novel SNVs (8/243,622 = 3.28×10−5; p = 0.04). This indicates that many of the possible de novo point mutations may indeed be rare inherited variants made homozygous by chromosomal loss of the normal allele.
In contrast to the trend in SNVs, small indels that homozygously mutated genes were more often novel. There were 79 genes predicted homozygously mutated by indel variants reported in dbSNP while 99 were predicted mutated by novel indels. Despite this trend, however, there was not a significant enrichment of deleterious indels among the novel indels (99/191,743 = 5.16×10−4) compared to the known indels (79/116,964 = 6.75×10−4; p = 0.08) This suggests that the difference in ratios of novel versus documented SNVs (8 vs. 146) and indels (99 vs. 79) is the result of compositional bias in dbSNP129, which contains a far greater number of SNPs compared to indels.
We also assessed the structural variants in U87MG for whether or not they were likely to affect a gene. Two different criteria were used to determine if translocations and microdeletions impacted a coding region, both predicted to produce an aberrant or nonfunctional protein. Using the UCSC known gene database, we identified 35 genes affected by interchromosomal translocations, 145 affected by complete deletions, 91 affected by heterozygous deletions and 166 affected by other intrachromosomal translocations (Table 4).
Interchromosomal translocation events were significantly enriched for occurring at positions where they would affect genes with 32 out of 35 events (91.4%) occurring within 1kb of a gene (p<0.0001), while only 44.1% of the reference genome is within 1kb of a known gene. In total, intrachromosomal events did not display this enrichment with 145/319 (45.5%) falling within 1kb of a gene (p = 0.67). However, we ran a set of simulations to assess whether microdeletions were enriched to overlap exons because we noted that 585 of our 599 complete microdeletions were less than 10kb in length with a mean size of 1.8kb. We ran 100,000 simulations randomly placing 600 microdeletions of 2kb lengths and determined how many times a microdeletion spanned an exon. In this way, we demonstrated that complete (homozygous) microdeletions under 10kb in size spanned exons slightly more often than by chance with a simulated p-value of .046. Similar assessment of microdeletions greater than 10kb in size did not find evidence of enrichment. These findings suggest that small microdeletions may preferentially occur within genes as opposed to being randomly distributed across the genome, but the signal is not strong from the available data. Genes affected by structural variations are summarized in Table S4.
The annotation tool DAVID was used to further examine the biological significance of the list of likely knockout mutations (including genes affected by SNVs, indels, microdeletions and translocation events) using the EASE analysis module. After gene ontology (GO) analysis, 18 GO terms were nominally enriched and associated with the mutated gene with a p-value < = 0.01 (Table S5). These GO enrichments include cell adhesion (GO:0007155 and GO:0022610), membrane (GO:0044425), and protein kinase regulator activity (GO:0019887).
The list of genes was also compared to the list of cancer-associated genes maintained by the Cancer Gene Census project (http://www.sanger.ac.uk/genetics/CGP/Census/). For SNVs and small indels, eight were observed in the census list, but this is not unexpected given the large number of mutations found in this cell line (p = 0.21). Two CGC genes were affected by complete microdeletions (CDKN2A and MLLT3), and one gene each was affected by heterozygous microdeletions (IL21R) and interchromosomal translocations (SET). These included genes previously annotated as mutated in instances of T cell prolymphocytic leukemia (TCRA and MLLT3), glioma (PTEN), endometrial cancer (PTEN), anaplastic large-cell lymphoma (CLTCL1), prostate cancer (ETV1 & PTEN), Ewing sarcoma (FLI1 and ETV1), desmoplastic small round cell tumor (FLI1), acute lymphocytic leukemia (FLI1 and MLLT3), clear cell sarcoma (FLI1), sarcoma (FLI1), myoepithelioma (FLI1), follicular thyroid cancer (PAX8), non-Hodgkin lymphoma (IL21R), acute myelogenous leukemia (SET), fibromyxoid sarcoma (CREB3L2), melanoma (XPC), and multiple other tumor types (PTEN and CDKN2A).
We also explored the overlap of genes with mutations in GBMs according to the Cancer Genome Atlas (TCGA) with those we predicted are homozygously loss-of-function mutated in U87MG (Table S5). Seven genes mutated in U87MG by SNVs or indels were also found mutated within the TCGA sample (PTEN, LTF, KCNJ16, ABCA13, FLI1, MLL4, DSP). This overlap is not statistically significant (p = 0.16). Ten additional genes overlapped, including two genes mutated by interchromosomal translocations (CNTFR, ELAVL2), three genes mutated by intrachromosomal translocations (ANXA8, LRRC4C, ALDH1A3), and five by homozygous microdeletions (CDKN2A, CDKN2C, MTAP, IFNA21, TMBIM4).
Finally, in order to place the homozygous mutations of U87MG in context relative to GBM mutational patterns as a whole, the Genomic Identification of Significant Targets in Cancer (GISTIC) method [28] was applied to 293 glioblastoma samples with genome wide copy number information available from the TCGA. This yielded a list of significant, commonly deleted regions present across glioblastomas as a group and highlights genes commonly mutated in GBMs. These data indicate that all or parts of chromosomes 1, 6, 9, 10, 13, 14, 15, and 22 are commonly deleted within GBMs as a group. In total, these regions comprise 915,306,764 bases, covering roughly 30 percent of the genome. In order to highlight genes homozygously mutated in U87MG that are within the regions of common loss, we cross-referenced these lists and found that 62/332 (19%) are within the GISTIC defined regions. This does not suggest a significant overlap of homozygously mutated genes in U87MG with commonly deleted regions, but those mutated genes that do overlap may be of increased relevance to cancer. Two of the 62 genes are also in the Cancer Gene Census: PTEN and TCRA. We propose that a subset of the genes mutated in U87 within these commonly deleted regions may be the specific targets of mutation and should be assessed on larger sample sets. (Table S5 and Figure S4).
Reported individual human genome sequencing projects using massively parallel shotgun sequencing with alignment to the human reference genome clearly indicate the practicality of individual whole genome sequencing. However, the monetary cost of data generation, data analysis issues, and the time it takes to perform the experiments have remained substantial limitations to general application in many laboratories. Here we demonstrate enormous improvements in the throughput of data generation. Using a mate-pair strategy and only ten micrograms of input genomic DNA, we generated sufficient numbers of short sequence reads in approximately 5 weeks of machine operation with a total reagent cost of under $30,000. We believe this makes U87MG the least expensive published genome sequenced to date signaling that routine generation of whole genomes is feasible in individual laboratories. Further, the two-base encoding strategy employed within the ABI SOLiD system is a powerful approach for comprehensive analysis of genome sequences and, in concert with BFAST alignment software, is able to identify SNVs, indels, structural variants, and translocations.
Of particular interest in whole-genome resequencing studies such as this one is how much raw data must be produced to sequence both alleles using a shotgun strategy. Here, 107.5Gb of raw data was generated. Of this, 55.51Gb was mapped to unique positions in the reference genome. In effect, this results in a mean base coverage of 10.85× per allele within non-repetitive regions of the genome. Repetitive regions are of course undermapped, as their unique locations are more difficult to determine. This level of oversampling is adequate for high stringency variant calling (error rate less than 5×10−6) at 93.71% of heterozygous SNP positions. There may be some biases in library generation resulting in bases that are not successfully covered even if they are relatively unique, but solutions to this may be found in performing multiple sequencing runs with varied library designs, as suggested in other studies [17].
With rapid advances in the generation of massively parallel shotgun short reads, one of the major computational problems faced is the rapid and sensitive alignment of greater than 1 billion paired end reads needed to resequence an individual genome. We demonstrate a practical solution using BFAST, which was able to perform fully gapped local alignment on the two-base encoded data to maximize variant calling in less than 4 days on a 20-node 8-core computer cluster.
Comparing U87MG SNVs with the James Watson [12] and YanHuang [14] genome projects' SNVs displays differences in SNV detection between the three projects. Being derived from a Caucasian individual, U87MG and James Watson are expected to share more SNVs than U87MG and YanHuang. However, when we compared SNVs between U87MG and these two genomes, more SNVs were actually shared between U87MG and YanHuang. Meanwhile, the YanHuang project called significantly more SNVs in total than both our U87MG sequencing project and the James Watson project. These results stress that utilizing different sequencing platforms (U87MG-ABI SOLiD, James Watson-Roche 454, YanHuang-Illumina Solexa), alignment tools (U87MG-BFAST, James Watson-BLAT, YanHuang-SOAP) and analytical approaches results in finding different quantities of SNVs. The higher genomic coverage in our U87MG sequence relative to James Watson and the increased sensitivity of BFAST relative to BLAT and SOAP were counted on to find highly robust variants. This is particularly important when sequencing a cancer genome because of the interest in finding novel cancer mutations as opposed to common polymorphisms.
The genomic sequence demonstrates global differences in variant type across the coding and non-coding portions of the human genome. By increasing the sensitivity of indel detection, we revealed that small indels have mutated genes at a higher rate than SNVs. A larger proportion of the indels identified are predicted to cause a protein coding change compared to SNVs (178/191,743 indels vs. 154/2,384,470 SNVs).
In U87MG, there is a relative increase in 4-base indels genome-wide, which has been observed in other normal genomes [23]–[25] (Figure 2, red bars). However, indels found in coding regions exhibit a bias toward events that are multiples of 3-bases in length (Figure 2, blue bars) presumably selected to maintain reading frame. Thus, many of these events are likely to be polymorphisms and not disease related genomic mutations [25]. Similarly, the nucleotide substitution frequencies demonstrate a bias in coding regions compared to non-coding. Two-thirds of the substitutions were transitions genome-wide, as expected [27], but there was an enrichment of CG->TA transitions in coding regions (Figure 3). It is well established that the most common source of point mutations and SNPs in primates is deamination of methyl-cytosine (meC), causing transition to a thymine (T) [16],[29], and there is circumstantial evidence of that in U87MG's genome as well.
The resolution of genome-wide chromosomal rearrangements is substantially improved by the mate-pair strategy, coupled with sensitive and independent alignment of the short 50-base reads (Figure 5). Based on published SKY data, we anticipated 7 interchromosomal breakpoints [6]. However, whole-genome mate-paired sequence data revealed the precise chromosomal joins of 35 interchromosomal events, which account for previously observed chromosomal abnormalities in U87MG but at additional finer scale resolution (Figure 5, Figure 6, Figure 7). The translocation events were enriched in genic regions with 32/35 (91.4%) occurring within 1kb of genes. A weaker, but still noticeable enrichment over genes occurs with microdeletions as well, which are generally missed by other experimental techniques like DNA microarrays. Thus, within the overall mutational landscape of this cancer cell line, translocations and structural variants preferentially occurred over genes, supporting a model where cancer mutations occur via structural instability rather than novel point mutations.
Delving into the functional effects of the mutations in U87MG through gene ontology and cross-referencing the literature, we found a large number of known and predicted cancer mutations present in the cell line. There is always a concern when dealing with a cancer cell line that mutations will be more related to its status as a cell line than to the cancer it was derived from. While this remains a concern, the large number of predicted and known cancer genes present in U87MG suggests other genes mutated in it have relevance to cancer as well. Using GISTIC to find regions with common deletions in glioma samples, we highlight 60 genes that are mutated in U87MG and are located in regions that are commonly deleted in GBMs that are not included within the Cancer Gene Census list as potential candidate mutational targets in GBMs (Table S5).
Cancer cell lines are commonly used as laboratory resources to study basic molecular and cellular biology. It is clearly preferable to have complete genomic sequence for these valuable resources. U87MG is the most commonly studied brain cancer cell line and is highly cytogenetically aberrant. While this made the sequencing and mutational analysis more challenging, it serves as a model for future cultured cell line genomic sequencing. Through custom analyses, we found that the mutational landscape of the U87MG genome is vastly more complicated than we would have expected based on the variants discovered in previously published genomes. It is our hope that the increased genomic resolution presented here will direct researchers and clinicians in their work with this brain cancer cell line to create more effective experiments and lead to a greater ability to draw meaningful conclusions in the future.
The NCBI reference genome (build 36.1, hg18, March 2006), genome annotations, and dbSNP version 129 were downloaded from the UCSC genome database located at http://genome.ucsc.edu. A local mirror of the UCSC genome database (hg18) was used for the subsequent analysis of variants using included gene models and annotations. The Watson genome variants were downloaded from Cold Spring Harbor Laboratory (http://jimwatsonsequence.cshl.edu) with bulk data files available from ftp://jimwatsonsequence.cshl.edu/jimwatsonsequence/. The YanHuang variants were downloaded from the Beijing Genomics Institute at Shenzhen (http://yh.genomics.org.cn/) with bulk data files available from http://yh.genomics.org.cn/download.jsp.
U87MG cells were ordered from ATCC (HTB-14) and cultured in a standard way. Genomic DNA was isolated from cultured U87MG cells using Qiagen Gentra Puregene reagents. DNA was stored at −20C until library generation.
Long-Mate-Paired Library Construction: The U87MG genomic DNA 2× 50bp long mate-paired library construction was carried out using the reagents and protocol provided by Applied Biosystems (SOLiD 3 System Library Preparation Guide). A similar protocol was reported previously [17]. Briefly, 45ug of genomic DNA was fragmented by HydroShear (Digilab Genomic Solutions Inc) to 1.0–2.5kb. The fragmented DNA was repaired by the End-It DNA End-Repair Kit (Epicentre). Subsequently, the LMP CAP adaptor was ligated to the ends. DNA Fragments between 1.2–1.7kb were selected by 1.0% agarose gel to avoid concatamers and circularized with a biotinylated internal adaptor. Non-circularized DNA fragments were eliminated by Plasmid-Safe ATP-Dependent DNase (Epicentre) and 3ug of circularized DNA was recovered after purification. Original DNA nicks at the LMP CAP oligo/genomic insert border were translated into the target genomic DNA about 100bp by nick translation using E. coli DNA polymerase I. Fragments containing the target genomic DNA and adaptors were cleaved from the circularized DNA by single-strand specific S1 nuclease. P1 and P2 adaptors were ligated to the fragments and the ligated mixture was used to create two separate libraries with 10 cycles of PCR amplification. Finally, 250–300bp fragments were selected to generate mate paired sequencing libraries with average target genomic DNA on each end around 90bp by excision from PAGE gel and use as emulsion PCR template. Templated Beads Preparation: The templated beads preparation was performed using the reagents and protocol from the manufacturer (Applied Biosystems SOLiD 3 Templated Beads Preparation Guide). SOLiD 3 Sequencing: The 2×50b mate-paired sequencing was performed exactly according to the Applied Biosystems SOLiD 3 System Instrument Operation Guide and using the reagents from Applied Biosystems.
We used an array pull-down capture strategy established in our lab [30]. An Agilent custom array for capturing 5,253 “cancer-related” genes was designed through Agilent e-array system (www.agilent.com). Only the amino acid encoding regions were targeted with 60mer oligos spaced center-to-center 20–30bp. The probes were randomly distributed across two separate 244K arrays. The library for cancer gene capture sequencing was generated following the standard Illumina paired-end library preparation protocol. 5ug of genomic DNA was used for the starting material and 250–300bp fragments were size-selected during the gel-extraction step. In the last step, 18 cycles of PCR were performed in multiple tubes to yield 4ug of product and mixed with 50ug of Human Cot-1 DNA (Invitrogen), 52ul of Agilent 10× Blocking Agent, 260ul of Agilent 2× Hybridization Buffer and 10× molar concentration of unpurified Illumina paired-end primer pairs custom made according to the sequences provided by Illumina (Oligonucleotide sequences, 2008, Illumina, Inc: available on request from Illumina). The mix was then diluted with elution buffer for the final volume of 520ul and then incubated at 95°C for 3 min and 37°C for 30min. 490ul of the hybridization mix was added to the array and hybridized in the Agilent hybridization oven (Robins Scientific) for 65 hrs at 65°C, 20rpm. After hybridization, the array was washed according to the Agilent wash procedure A protocol. The second wash was extended to 5 minutes to increase the wash stringency. After washing, the array was stripped by incubating it in the Agilent hybridization oven at 95°C for 10min, 20rpm with 1.09× Titanium Taq PCR Buffer (Clonetech). After the incubation and collection of the solution, 4 tubes of PCR were performed with each tube containing 96ul of the collected solution, 1ul of dNTPs (10mM each), 1ul of Titanium Taq (Clonetech) and Solexa primers, 1ul each. 15 cycles of PCR was performed at the following condition: 30sec at 95°C, (10 sec at 95°C, 30 sec at 65°C, 30 sec at 72°C)×18 cycles, 5 min at 72°C and hold at 4°C. The amplified product was purified using QIAquick PCR Purification Kit and eluted in 30ul of EB. After confirming the size of the amplicon on 2% agarose gel and measuring the concentration, the amplicon was diluted to 10nM, the working concentration for cluster generation. The Illumina flowcell was prepared according to the manufacturer's protocol and the Genome Analyzer was run using standard manufacturer's recommended protocols. The image data produced were converted to intensity files and were processed through the Firecrest and Bustard algorithms (1.3.2) provided by Illumina to call the individual sequence reads.
We used Blat-like Fast Accurate Search Tool version 0.5.3 (BFAST http://bfast.sourceforge.net) [19] to perform sequence alignment of the two-base encoded reads off the ABI SOLiD to the NCBI human reference genome (build 36.1). Utilizing the local alignment algorithm included in BFAST [31], we were able to simultaneously decode the short reads, while searching for color errors (encoding errors), base changes, insertions, and deletions.
We found candidate alignment locations (CALs) for each end independently. We utilized ten indexes to be robust to up to six color errors, equating to a 12% per-read error rate:
1111111111111111111111
111110100111110011111111111
10111111011001100011111000111111
1111111100101111000001100011111011
111111110001111110011111111
11111011010011000011000110011111111
1111111111110011101111111
111011000011111111001111011111
1110110001011010011100101111101111
111111001000110001011100110001100011111
We also set parameters to use only informative keys when looking up reads in each index (BFAST parameter -K 8), and to ignore reads with too many CALs aggregated across all indexes (BFAST parameter -M 384). If reads mapped to greater than 384 locations, then they were categorized as ‘unmapped’. We then performed local alignment for each of the returned CALs, simultaneously decoding the read from color space searching for color errors (encoding errors), base changes, insertions, and deletions [31]. We choose the “best scoring” alignment, accepting an alignment only if it was at least the equivalent edit distance of two color errors away from the next best alignment. This is approximately similar to a ‘mapping quality’ of 20 or better from the MAQ program output, for reference. We removed duplicate reads using the alignment filtering utility found in DNAA (http://dnaa.sourceforge.net). For single-end and mate-paired reads where only one end mapped, we removed duplicates based on reads having identical stat positions. For mate-paired reads, we removed duplicates where both ends had the same start position.
Illumina generated sequence was aligned to the NCBI human reference genome (build 36.1) using BFAST with the following parameters applied. Each end of the fragment library was mapped independently to identify CALs, utilizing ten indexes to be robust to errors and variants in the short (typically 36bp) reads:
1111111111111111111111
1111101110111010100101011011111
1011110101101001011000011010001111111
10111001101001100100111101010001011111
11111011011101111011111111
111111100101001000101111101110111
11110101110010100010101101010111111
111101101011011001100000101101001011101
1111011010001000110101100101100110100111
1111010010110110101110010110111011
We also set parameters to use only informative keys when looking up reads in each index (BFAST parameter -K 8), and to ignore reads with too many CALs aggregated across all indexes (BFAST parameter -M 1280). We then performed a standard local alignment for each CAL. Reads were declared mapped if a single unique best scoring alignment was identified within the genome. Duplicate reads were filtered out in the same manner as for the ABI SOLiD data.
To find SNVs including SNPs and small indels, we assumed the MAQ consensus-calling model [20] utilizing the implementation in SAMtools [21]. We used a value of 0.0000007 for the prior of a difference between two haplotypes (-r parameter). This was chosen based on ROC analysis of a test dataset (data not shown).
Structural variations were detected using custom algorithms designed to comprehensively search for groups of mate-pair reads with aberrant paired-end insert size distributions that are consistently identifying a unique structural variant in the genome. We utilized the “dtranslocations” utility in the DNAA package (http://dnaa.sourceforge.net) for the primary structural variation candidate search. The utility first selected all pairs for which each end is uniquely mapped to a single location in the human genome and for which the mate-pair reads are not positioned in the expected size range relative to the consensus genome. Then we filter out false positives that are not consistent with a chromosomal difference on an allele. Briefly, the genome was divided into 500-base bins sequentially stepped 100-bases apart from their start positions. Each bin was then paired with other bins on the basis of containing similar ‘mismapped’ mate-pair reads. The aberrant mate-paired reads were defined as reads that were mapping less than 1000 or greater than 2000 bases apart within the reference genome sequence, which is selected based on the insert size distribution calculated from the aggregate dataset (Figure S2). These were then rank-ordered based on the number of mate-pairs meeting criteria, and the destination bin with the most reads within it was paired with a given source bin to create a ‘binset’. Binsets containing less than 4 reads were filtered out, removing 98.3% of the candidates based on having too little evidence supporting them. The resulting list of filtered binsets was then scanned for clusters of binsets. Binset clusters are groups of binsets where the source bins occur within 2000 bases of each other and the destination bins occur within 2000 bases of each other. Redundant binsets were combined and those binset clusters that contain too few (less than 9 binsets spanning at least 1000 bases) or too many binsets (greater than 29 binsets spanning at most 3000 bases—higher is impossible given our insert size distribution) were removed as artifacts. The resulting binset clusters represent the reads immediately flanking structural breakpoint events. This detection process is currently being automated as Breakway (http://breakway.sourceforge.net), but was done using custom scripts at the time of analysis.
The structural variations were then separated into interchromosomal and intrachromosomal events. Intrachromosomal events of less than 1Mb are assessed for deletion status by averaging base coverage within the bounds of the event and comparing it to base coverage 200kb outside the event on both sides. Those that have average interior base coverage less than 25% of the average exterior base coverage are classified as “complete” deletions. Those with average interior base coverage between 25% and 75% that of average exterior base coverage are classified as “heterozygous deletions” (deletions of at least one copy of the region, but with at least one copy remaining).
Variant calls from the SAMtools pileup tool were first loaded into a SeqWare QueryEngine database and subsequently filtered to produce BED files. This filtering criteria required that a variant be seen at least 4 times and at most 60 times with an observation occurring on each strand at least once. For SNVs we further enforced the criteria that SNVs should only be called in reads lacking indels and the last 5 bases of the reads were also ignored. This reduced the likelihood that spurious mismappings were used to predict SNVs and eliminated the lowest quality bases from consideration. For small indels (<21bp) we enforced a slightly different filter by requiring that any reads supporting an indel were only allowed to contain one contiguous indel and these reads were not considered if the indel occurred on either the beginning or end of the read. These criteria, like the SNV criteria, were used to reduce the likelihood of using mismapped reads or locally misaligned reads in the variant calling algorithm. The elimination of reads with indels at the beginning or end of the read was intended to remove potential alignment artifacts caused by ambiguous gap introduction due to lack of information at the ends to guide proper alignment. Together, these filtering criteria reduced the likelihood that sequencing errors were identified as SNV or indel variants. We used scripts available in the BFAST toolset and SeqWare Pipeline to filter and annotate the variant calls. Variants passing these filters were further annotated by their overlap with dbSNP version 129. Variants were required to share the same genomic position as a dbSNP entry along with matching the allele present in the database to be considered overlapping. Mapping to dbSNP allowed us to filter out known SNPs from de novo variants.
Filtered SNV and indel variants were then analyzed for their affect within the genome that is annotated with gene models. This analysis used scripts from the SeqWare Pipeline project and gene models downloaded from the UCSC hg18 human genome annotation database. Six different gene model sets from hg18 were considered: UCSC genes (knownGene), RefSeq genes (refGene, http://www.ncbi.nlm.nih.gov/RefSeq), Consensus Coding Sequence genes (ccdsGene, http://www.ncbi.nlm.nih.gov/CCDS), Mammalian Gene Collection genes (mgcGenes, http://mgc.nci.nih.gov), Vertebrate Genome Annotation genes (vegaGene, http://vega.sanger.ac.uk), and Ensembl genes (ensGene, http://www.ensembl.org). Each variant was evaluated for overlap with genes from each of the 6 gene models. If overlap was detected the variant was examined and tagged with one or more of the following terms depending on the nature of the event: “utr-mutation”, “coding-nonsynonymous”, “coding-synonymous”, “abnormal-ref-gene-model-lacking-stop-codon”, “abnormal-ref-gene-model-lacking-start-codon”, “frameshift”, “early-termination”, “inframe-indel”, “intron-splice-site-mutation”, “stop-codon-loss”, and/or “start-codon-loss”. The variant was also tagged with the gene symbol and other accessions to facilitate lookups. This information was loaded into a SeqWare QueryEngine database to allow for querying and filtering of the variants as needed.
Genes affected by structural variations were assessed in two ways depending on the structural variation type. For interchromosomal translocation events, a gene was considered “affected” when either end of an interchromosomal translocation event fell in a genic region (including the entire coding region plus 1kb up- or down-stream of the gene's coding region). The same criteria were used for all intrachromosomal translocation events. For events that were classified as complete or heterozygous deletions, a gene was considered affected if all or part of a coding exon was deleted.
Homozygous SNVs, small indels, large deletions, and translocation events for variants that included predicted coding sequence changes were tallied. This became a reference list of variants with serious homozygous mutations that likely completely disrupted, or “knocked out”, the normal function or synthesis of the target protein.
For the SNVs and small indels, a “knockout” variant was defined as a homozygous call by the SAMtools variant caller where the variant was predicted by the SeqWare Pipeline scripts to change coding sequence with one or more of the following annotations: “early-termination”, “frameshift”, “intron-splice-site-mutation”, “start-codon-loss”, and/or “stop-codon-loss”. The “early-termination” event represented a stop codon introduced upstream of the annotated stop codon. The “frameshift” represented an indel that resulted in a shifting of the reading frame of the gene resulting in, typically, early termination and non-sense coding sequence. The “intron-splice-site-mutation” referred to a mutation in the two consensus splice site intronic bases flanking exons (GT at the 5′ splice site and AG at the 3′ splice site). Finally, “stop-codon-loss” and “start-codon-loss” simply refer to variants that interrupt the stop or start codons. We chose to not include “coding-nonsynonymous” and “inframe-indel” annotations in this list of knocked out variants because, while potentially serious as these mutations are, they are not guaranteed to result in an unexpressed or non-functional protein. However, homozygous frameshift, early termination, splice site, and stop/start codon loss mutations are very likely to interrupt a gene's expression and translation to functional protein.
As described above, large microdeletions that removed all or part of an exon and interchromosomal translocation events that fell within 1kb of a gene's coding region were also classified as mutated genes.
Once suspect knockout variants were identified, a mapping process was used to translate one or more variants to the gene symbol. This mapping allowed us to condense multiple variants affecting multiple gene models to a more abbreviated list of gene symbols likely to be affected by these knockout mutations. The mapping from variants to gene symbols used variants identified with gene models from the refGene and the knownGene tables in the UCSC hg18 database and mapped these variants to gene symbols using queries against the name field of the knownGene table and the alias field of the kgAlias table. The UCSC table browser was used to accomplish these queries and map the knownGene identifiers to gene symbols via the kgXref table. A similar approach was used for homozygous large-scale microdeletions and translocation events.
The list of knockout genes was uploaded to the Database for Annotation, Visualization, and Integrated Discovery (DAVID, version 2008) to identify enriched Gene Ontology (GO) terms [32]–[33]. Overlap with GO terms from the biological process, cellular component, and molecular function ontologies were considered. The default parameters were used and a p-value cutoff of < = 0.01 was considered significant.
The overlap between the Cancer Gene Census genes and those identified as knockouts in U87MG were compared. The Cancer Gene Census project is an ongoing effort to catalog genes with mutations that have been implicated in cancer [34]. It is a highly curated list that includes annotations for each gene including tumor types, class of mutations, and other genetic properties. We used the gene symbol list from the September 30th, 2009 complete working list, which includes 412 gene symbols.
The overlap between mutations in the Cancer Genome Atlas (TCGA) and those identified as knockouts in U87MG was analyzed. TCGA is an ongoing effort to understand the molecular basis of cancer through large-scale copy number analysis, expression profiling, genome sequencing, and methylation studies among other techniques [4]. It provides information on mutations found by Sanger sequencing on many patient samples. For glioblastoma this includes sequence data aberrations detected in 158 patient samples and 1,177 genes.
The Genomic Identification of Significant Targets in Cancer (GISTIC) method was used to find significant areas of deletion in 293 samples from the TCGA [24]. The GISTIC technique was designed to identify and analyze chromosomal aberrations across a set of cancer samples, based on the amplitude of the aberrations as well as the frequency across samples. This approach produced a series of commonly deleted regions across the set of TCGA GBMs. To calculate the areas of deletion, we used 293 Affymetrix SNP 6.0 samples segmented using the GLAD SNP analysis module [35]. Default parameters of GISTIC were used. GISTIC produces peak limits, wide peak limits, and in addition broader region limits. These commonly deleted broader regions were then scanned for predicted knockout genes in U87MG.
The distribution of small indel sizes was examined for both deletions and insertions. Indels classified as affecting coding-sequence by the SeqWare Pipeline (see above) were compared to those outside coding regions. Raw counts were collected, recalculated as percents of total, and compared directly.
Similarly, nucleotide substitution frequency was examined for SNVs from U87MG both genome-wide and only in coding regions. Once binned appropriately, the SNV nucleotide substitutions were counted, tallied in a table, and graphed as percents of total.
Variants from the Watson and Yan Huang genome were downloaded from each respective project from the following URLs: ftp://jimwatsonsequence.cshl.edu/jimwatsonsequence/watson-454-snp-v01.txt.gz and http://yh.genomics.org.cn/do.downServlet?file=data/snps/yhsnp_add.gff. These files contained variant calls for each genome along with annotations describing the variant as novel or occurring in dbSNP. The Watson genome only contained SNV calls so our comparison was limited to just SNVs. The Yan Huang genome also contained calls indicating heterozygous or homozygous. However, a variant was considered to match between genomes regardless of zygosity state. We compared the overlap of the U87MG genome, dbSNP and each of these genomes in turn. SNVs from U87MG that were considered for comparison had to meet our criteria; variants had to be observed at least 4 times, at most 60 times, at least once per strand, and with a minimum phred score of 10. SNVs in the three-way comparison were said to match if the position and allele matched between the genomes. If both variants matched between U87MG and the other genome and one was annotated in dbSNP, then the other was considered in dbSNP as well. If neither contained annotations from dbSNP the variant was considered novel. A similar process was carried out for variants distinct to each genome. The results were recorded as Venn diagrams showing the overlap between dbSNP, U87MG, and the Watson or Yan Huang genome.
Genomic DNA from U87MG was submitted to the Southern California Genotyping Consortium to be run on the Illumina Human 1M-Duo BeadChip, which consists of 1,199,187 probes scattered across the human genome. The Illumina Beadstudio program was used to analyze the resulting intensity data. Loss of heterozygosity was determined by analyzing B-allele frequency as determined by the Beadstudio program. Normal two-copy regions of the genome are represented by long stretches of probes with B-allele frequencies of 0, 0.5 or 1. Regions of LOH, on the other hand, deviate from this pattern significantly. Copy number was determined by looking at probe intensity.
Primers for validation were designed by targeting regions immediately flanking the event predicted by our whole genome sequence analysis using the Primer3 tool (http://frodo.wi.mit.ed/primer3/). Polymerase chain reaction was performed following standard protocols using Finnzymes Phusion Hot-Start High Fidelity polymerase. Products were run on 2% agarose gel electrophoresis and product purity and size was assessed by staining with ethidium bromide. Sanger sequencing was performed at the UCLA Genotyping and Sequencing core facility using an ABI 3730 Capillary DNA Analyzer. Sequence trace files were analyzed using Geospiza FinchTV. Validation status and PCR primers are listed in Table S1.
Intensities, quality scores, and color space sequence for the genomic sequence of U87 SOLiD were uploaded to the Sequence Read Archive under the accession SRA009912.1/Sequence of U87 Glioblastoma Cell-line. Intensities, quality scores, and nucleotide space sequence for the exon capture U87 Illumina sequence were also uploaded to the Short Read Archive under the same accession. For both datasets, alignment files have been uploaded to the Short Read Archive as additional analysis results.
Variant calls for both datasets are available via a SeqWare QueryEngine web service at http://genome.ucla.edu/U87. This tool allows for querying the variants using a variety of search criteria including coverage, mutational consequence, gene symbol, and others. SeqWare QueryEngine produces results in both BED and WIG format making it compatible with the majority of genome browsers such as the UCSC genome and table browsers. Variant data will be uploaded to SRA as metadata along with the raw sequences. For the whole genome SOLiD alignment, small indels (<21bp), SNVs, large deletions, and translocation events can be queried. For the exon capture Illumina alignment, small indels and SNVs can be queried.
Most software used for this project is open-source and freely available. We created two software projects that were instrumental in the analysis of the U87MG data: BFAST and SeqWare. The color- and nucleotide-space alignment tool BFAST can be downloaded from http://bfast.sourceforge.net and many of our alignment filtering as well as the primary step in structural variation detection can be found in the DNAA package at http://dnaa.sourceforge.net. The SeqWare software project was used throughout the analysis of variant calls. We used the SeqWare LIMS tool for sample tracking, the SeqWare Pipeline analysis programs for annotating variants with dbSNP status and mutational consequence predictions, and SeqWare QueryEngine was used to database and query variant calls and annotations. This software and documentation can be downloaded from http://seqware.sourceforge.net.
|
10.1371/journal.pgen.1005600 | Intermediate Levels of Bacillus subtilis CodY Activity Are Required for Derepression of the Branched-Chain Amino Acid Permease, BraB | The global transcriptional regulator, CodY, binds strongly to the regulatory region of the braB gene, which encodes a Bacillus subtilis branched-chain amino acid (BCAA) permease. However, under conditions that maximize CodY activity, braB expression was similar in wild-type and codY null mutant cells. Nonetheless, expression from the braB promoter was significantly elevated in cells containing partially active mutant versions of CodY or in wild-type cells under growth conditions leading to intermediate levels of CodY activity. This novel pattern of regulation was shown to be due to two opposing mechanisms, negative and positive, by which CodY affects braB expression. A strong CodY-binding site located downstream of the transcription start point conferred negative regulation by direct interaction with CodY. Additionally, sequences upstream and downstream of the promoter were required for repression by a second pleiotropic B. subtilis regulator, ScoC, whose own expression is repressed by CodY. ScoC-mediated repression of braB in codY null mutants cells was as efficient as direct, CodY-mediated repression in wild-type cells under conditions of high CodY activity. However, under conditions of reduced CodY activity, CodY-mediated repression was relieved to a greater extent than ScoC-mediated repression was increased, leading to elevated braB expression. We conclude that restricting increased expression of braB to conditions of moderate nutrient limitation is the raison d’être of the feed-forward regulatory loop formed by CodY and ScoC at the braB promoter. The increase in BraB expression only at intermediate activities of CodY may facilitate the uptake of BCAA when they are not in excess but prevent unneeded BraB synthesis when other BCAA transporters are active.
| Expression of Bacillus subtilis BraB, a branched-chain amino acid permease, is under both negative and positive control by a global transcriptional regulator CodY. The negative control is direct and the positive control is indirect and mediated by another B. subtilis pleiotropic transcriptional regulator, ScoC, which, in turn, is repressed by CodY. Thus, CodY and ScoC form a feed-forward regulatory loop at the braB promoter. In a very unusual manner, the interaction of CodY and ScoC results in high braB expression only at intermediate CodY activities; braB expression remains low both at high and low CodY activities. The novel regulation of braB shows that important, novel regulatory phenomena can be missed by analyzing null mutants in regulatory genes but revealed by using mutants with partial activity.
| BraB is one of three permeases demonstrated to be involved in the uptake of branched-chain amino acids (BCAA) in Bacillus subtilis [1]. Given the important role of BCAA in cell metabolism, it is not surprising that the synthesis of the permeases is strictly regulated and coordinated. The most efficient BCAA permease, BcaP, is subject to very strong transcriptional repression by CodY [2], a global regulator in B. subtilis and other Gram-positive bacteria [3, 4]. A second permease, BrnQ, is subject to strong repression by AzlB, a member of the AsnC/Lrp family of transcriptional regulators, in response to an as yet unidentified signal [5]. The regulation of BraB synthesis has not been previously determined.
A fragment containing the regulatory region between the divergently transcribed iscSB (formerly nifZ) and braB genes was found to bind CodY strongly in vivo in a ChIP-to-chip experiment [6]. Moreover, a strong CodY-binding site in the iscSB-braB intergenic region was also detected in vitro during the global characterization of CodY-binding sites by IDAP-Seq [7]. The latter site is well-placed to serve as a potential site of regulation of braB. However, transcription of neither braB nor iscSB was altered >2.0-fold by a null mutation in codY, as detected in DNA microarray or RNA-Seq experiments [6, 8](http://www.genome.jp/kegg/expression/).
CodY controls directly or indirectly the transcription of more than 200 B. subtilis genes [7, 8]. The DNA-binding affinity of CodY from B. subtilis and many other species is increased by interaction with two types of ligands, the BCAA [isoleucine, leucine, and valine (ILV)] [9–11] and GTP [6, 11–14]. Thus, given the presence of CodY-binding sites in the putative braB regulatory region and the influence of BCAA on CodY activity, it was surprising that expression of braB was only minimally affected by a codY null mutation. We describe here a detailed analysis of the mechanisms by which CodY regulates braB.
The braB gene proved to be directly repressed by two proteins, CodY and another pleiotropic regulator, ScoC (formerly known as hpr or catA) [15–18]. Because CodY also represses scoC [19], CodY and ScoC form a feed-forward regulatory loop [20, 21] in which CodY acts an indirect positive regulator of braB. The opposing effects of fully active CodY balance each other; as a result, braB derepression could only be observed at intermediate levels of CodY activity or when both regulators are inactive. These findings emphasize that the phenotypes caused by null mutations in global regulatory protein genes can be misleading.
The unexpected absence of an effect of a codY null mutation on expression of a gene with a strong CodY-binding site in its putative regulatory region led us to analyze braB transcription in more detail. A primer extension experiment established that the 5’ end of the braB mRNA is located 72 bp upstream of the initiation codon. The sequences TTGACT and TATAAT, with one and no mismatches to the –35 and –10 regions of σA-dependent promoters, respectively, and a 16-bp spacer region, can be identified upstream of the 5’ end location, suggesting that this position does in fact correspond to the transcription start point (Fig 1A). (Since B. subtilis σA-dependent promoters rarely have a 16-bp spacer, our assignment of the -10 and -35 regions may be off by 1 or 2 bp.) A mutation, T(-29)C, located immediately downstream of the likely -35 region, reduced expression of a braB-lacZ fusion 6-fold (1.97±0.35 Miller units, see below), consistent with our assignment of the promoter.
DNase I footprinting experiments showed that CodY protected two sites, I and II, within the 194-bp iscSB-braB intergenic region from positions -62 to -47 and +11 to +50 of the template DNA strand with respect to the braB transcription start point, respectively (Figs 1A, 2B and 2C). Site II is much stronger than site I. Binding to an additional very weak site, III, from positions –143 to –124, which is within the upstream iscSB gene, was observed only at high concentrations of CodY (≥200 nM) (Fig 2B and 2C).
The results of the footprinting experiments are consistent with the identification of a strong CodY-binding site in this area by ChIP-to-chip experiments [6]. Moreover, they confirmed and extended the results of the in vitro IDAP-Seq experiments, which identified a strong core binding site from positions +29 to +43, a much weaker core site, which ends at position -45, and an additional, very weak core site, ending at position -116 and detected only at a very high CodY concentration (1 μM) (core sites only include positions that are essential for CodY binding; the beginning positions of the two upstream core sites could not be determined due to limitations of the IDAP-Seq procedure) [7].
The braB regulatory region contains five 15-bp motifs, which resemble the 15-bp CodY-binding consensus sequence, AATTTTCWGAAAATT [22–24] (we use the terms “site” and “motif” to describe an experimentally determined location of CodY binding and a 15-bp sequence that is similar to the consensus motif, respectively). Site I of the braB gene overlaps CodY-binding motif 1, located between positions -64 and -50, that has 4 mismatches with respect to the CodY-binding consensus (Fig 1A and Table 1). The strong site II overlaps two adjoining versions of the 15-bp sequence, motifs 2 and 3, located between positions +14 and +43, each of which has three mismatches with respect to the consensus motif. Another 15-bp sequence, motif 4, with four mismatches is located from positions +40 to +54 and overlaps motif 3 by 4 bp. Site III overlaps CodY-binding motif 5, with 5 mismatches, located from positions -141 to -127 (Fig 1A and Table 1).
Binding of CodY to upstream braB sites occurred independently of the presence of the downstream site and vice versa (Figs 1 and 3; see below for generation of the truncated fragments), similar to the case for other genes containing multiple CodY-binding sites within their regulatory regions [2, 25, 26].
In gel-shift experiments, CodY bound to DNA fragments containing only sites III and I (braB156) or only site II (braB144) with apparent dissociation constants (KD) of ∼75 nM and ∼4 nM, respectively, compared with ∼3 nM for the full-length fragment, braB242 (Fig 4A, 4B and 4D) (KD reflects the CodY concentration needed to shift 50% of DNA fragments under conditions of vast CodY excess over DNA). Complexes with lower mobility were formed at higher concentrations of CodY for all fragments, indicating apparent changes in stoichiometry of CodY binding (Fig 4).
We constructed a transcriptional fusion (braB242-lacZ) containing a 242-bp fragment that includes the entire iscSB-braB intergenic region (Fig 1). Under conditions of maximal CodY activity, in cells grown in TSS glucose–ammonium medium supplemented with ILV and a mixture of 13 other amino acids (referred to here as the 16 aa-containing medium), fusion expression in a codY null mutant strain was very similar (1.3-fold less) to that in the wild-type strain (Table 2, strains BB3076 and BB3079). Consistent with the lacZ fusion results, only very weak, positive regulation (1.6- to 1.9-fold) in amino acid-containing medium was detected in microarray or RNA-Seq experiments by comparing wild-type and codY null mutant strains [6, 8].
The activity of CodY is reduced to intermediate levels when some amino acids are removed from the medium and decreases strongly in the absence of all amino acid supplements [2, 27]. Expression of the braB242-lacZ fusion in the wild-type strain in the absence of any amino acids or in the presence of ILV only was very similar to that in the presence of 16 aa (11.3 to 14.7 MU versus 12.2 MU). Unexpectedly, almost 3-fold higher activity was found in 13 aa-containing medium (i.e., in the absence of ILV), indicating that CodY, at an intermediate level of activity, may serve as a positive regulator of braB (Table 2).
To test whether expression of the braB gene indeed responds differentially to varying levels of CodY activity in vivo we made use of a previously constructed set of mutant forms of CodY that have different levels of residual responsiveness to ILV. Most of these proteins have alterations in amino acids that form the ILV-binding pocket; they are expressed at wild-type levels and have undiminished activity in effector-independent DNA binding [2, 8, 28]. Since the population of CodY molecules in the cell is in equilibrium between the liganded and unliganded forms of the protein, the unliganded fraction of the population of a mutant protein that has lower affinity for ILV will be greater than for the wild-type protein at a given intracellular ILV concentration. That is, a mutant strain containing a form of CodY that has low affinity for ILV behaves functionally equivalently to the wild-type strain that has a low intracellular pool of ILV.
The analysis of Dataset S1 of Ref. (8) indicates that expression of braB determined by RNA-Seq experiments was up to 4.1-fold higher in three strains containing partially active versions of CodY, F71Y, R61K, or R61H (Fig 5A). The results of the RNA-Seq experiments were confirmed and extended by real-time RT-PCR and by analyzing expression of the braB242-lacZ transcriptional fusion in a larger collection of partial codY mutants (Fig 5B and 5C). The up-and-down expression pattern of the braB fusion, in which maximal activity was seen in mutants with intermediate levels of CodY activity, was in drastic contrast to the plateau-reaching expression pattern of the previously characterized CodY-repressed bcaP283-lacZ fusion (Fig 5F) [2] and all other CodY-regulated genes [2].
A braB242-gfp translational fusion was introduced into the wild-type strain, the codY null mutant, and a codY point mutant (R61K) strain with intermediate residual activity. In all cases, the level of braB expression was rather similar across the cell population (Fig 6), eliminating the possibility that a bistable expression pattern could explain our results. As expected, the codY (R61K) mutant strain had elevated expression compared to the wild-type and codY null mutant strains.
We initially hypothesized that the very unusual pattern of braB regulation observed results from CodY binding independently to negative and positive regulatory sites within the braB regulatory region. If so, the positive and negative effects of fully active CodY might balance each other, but, at intermediate levels of CodY activity, positive regulation might dominate. The CodY-binding sites I and II are located upstream and downstream of the braB promoter in positions appropriate for positive and negative regulation, respectively. To determine their independent effects, we created additional lacZ fusions containing truncated versions of the braB regulatory region lacking the upstream CodY-binding site III (braB184-lacZ and braB162-lacZ) or sites III and I (braB144-lacZ) or the downstream site II (braB156-lacZ) (Fig 1B). (Note that the braB242-, braB184-, braB162- and braB144-lacZ fusions have the identical junction with lacZ; their levels of activity can be directly compared. However, other fusions, such as braB156-lacZ, have different junctions; their activities in wild-type cells can only be compared to the activity of the same fusion in mutant strains or other fusions with a similar junction.)
Surprisingly, deletion of the upstream binding sites III and I did not cause any significant decrease in braB expression in wild-type cells (Table 2; compare strains BB3076, BB3719, BB3811, and BB3122), implying that these are not sites of positive regulation. On the other hand, the braB162-lacZ and braB144-lacZ fusions, but not the braB184-lacZ fusion, were derepressed 12-fold when codY was inactivated (Table 2), suggesting that braB expression is subject to negative regulation by CodY bound to the remaining downstream site II. If so, this regulation must be masked in other fusions by the action of a second repressor that binds to the sequence located between the 5’ ends of braB184-lacZ and braB162-lacZ. Interestingly, no up-and-down expression pattern in mutants with different levels of CodY activity was observed for the braB144-lacZ fusion, which lacks the putative binding site for the predicted second regulator (Fig 5E), suggesting that the latter is responsible for the unusual regulation. As expected from this new model, the braB156-lacZ fusion, which lacks the downstream CodY-binding site II, was not subject to regulation by CodY (Table 2).
To confirm that site I is not involved in braB regulation and to quantify more directly the contribution of site II, we changed the very highly conserved G9 and A10 residues of CodY-binding motifs 1, 2, and 3 to CC (the p1, p2, and p3 mutations, respectively) (Fig 1 and Table 1).
The p1 mutation in site I reduced ~10-fold the affinity of CodY for a fragment containing sites III and I, indicating that site I is the major contributor for CodY binding to this fragment (Fig 4B and 4C). However, as expected from our deletion analysis, the p1 mutation did not affect expression of the braB242-lacZ fusion (Table 3, strains BB3731 and BB3076). Thus, as noted previously, many CodY-binding sites have no physiological significance either because they are not positioned appropriately for regulation or because binding is too weak [7].
The p3 mutation reduced the affinity of CodY for site II ≥10-fold (Fig 4E). The p2 mutation did not affect binding of CodY to site II, but further decreased the ability of CodY to interact with this site if it already contained the p3 mutation (Fig 4F and 4G). Footprinting experiments showed that each mutation affected CodY binding to the region of site II, which corresponded to the respective motif (Fig 3B). Taking together the gel-shift and footprinting results, we conclude that interaction of CodY with motif 2 is weaker than with motif 3 and is partly dependent on simultaneous interaction of CodY with motif 3 (see below for the effect of p2 on braB regulation).
The p3 mutation increased expression of the braB242-lacZ fusion 8-fold consistent with relief from CodY-mediated repression (Table 3, strains BB3729 and BB3076). However, expression of the braB242p3-lacZ fusion was substantially reduced in a codY null mutant strain apparently due to repression by the second regulator (Table 3, strain BB3735). This result suggests strongly that the second regulator is active in codY mutant cells, but not in wild-type cells, i.e., its activity or expression is under negative CodY control. Paradoxically, this indicates that our initial hypothesis that braB regulation is subject to simultaneous positive and negative regulation by CodY was likely to be correct, though positive regulation appears to be indirect and mediated through regulation of the second repressor.
CodY is known to regulate the expression of a small number of other regulatory proteins, including ScoC [6–8, 19, 29, 30]. ScoC is a repressor of multiple genes, including those encoding extracellular proteases and oligopeptide permeases, and is also involved in the regulation of sporulation [15–19, 31–33]. Though microarray experiments did not identify braB as a ScoC target [15], we decided to test whether ScoC is the second regulator of braB expression. No effect of a single scoC null mutation on expression of the braB242-lacZ fusion in TSS + 16 aa was detected (Table 2, strain BB3847). However, in a double codY scoC null mutant, expression of the fusion was 11- to 12-fold higher than in the wild-type strain or in scoC or codY single mutants (Table 2, strain BB3835), indicating that both CodY and ScoC contribute to negative regulation of braB but these effects cannot be dissected if either one of the regulators is active.
Expression of the same fusion in a double null mutant in TSS + 13 aa medium was very similar (Table 2), indicating that our original observation of higher braB expression under these growth conditions in a wild-type strain was indeed due to reduced CodY activity and its effect on ScoC expression.
As expected, in the absence of ScoC, the up-and-down expression pattern of the braB242-lacZ fusion in strains with different CodY activity was replaced by a plateau-reaching pattern, resembling that of the bcaP283-lacZ fusion, which is not subject to ScoC-mediated regulation (Fig 5D and 5F).
Expression of the scoC561-lacZ fusion in strains with different CodY activity also followed a plateau-reaching pattern, characteristic for most genes regulated by CodY, and did not correlate with expression from the braB promoter (Fig 5G).
In DNase I footprinting experiments, ScoC protected two sites, I and II, within the iscSB-braB intergenic region from positions -79 to -68 and +43 to +57 of the template DNA strand with respect to the braB transcription start point, respectively (Figs 1A and 7). A short, weakly protected region, site III (possibly a part of site II), was also detected from positions +16 to +20. Binding of ScoC to the downstream sites II and III was independent of the presence of the upstream site I on the same DNA fragment (Fig 7).
The downstream CodY- and ScoC-binding sites partly overlap (Fig 1A). To address the possibility that CodY and ScoC compete for binding at this location, we analyzed interaction of these proteins with a short, 64-bp braB fragment, containing CodY-binding site II and ScoC-binding sites II and III (Fig 1B). In accord with the results described above, ScoC bound this fragment in gel shift experiments less efficiently (KD≈150 nM) than did CodY (KD≈5 nM) (Fig 8A and 8B). Nevertheless, ScoC, in a concentration-dependent manner, was able to replace CodY efficiently in a preformed braB-CodY complex as evidenced by formation of ScoC-specific complexes with higher mobility and the decrease in the amount of braB-CodY complexes with lower mobility (Fig 8C). The CodY-mediated displacement of ScoC from the preformed braB-ScoC complex cannot be recognized confidently because of the low mobility of CodY-specific complexes (complexes containing both proteins would have a similar low mobility). However, by comparing and Fig 8B and 8D, it is clear that CodY bound much less efficiently to preformed braB-ScoC complexes than to free braB DNA, confirming competition between the two proteins for binding. A similar competition between CodY and ScoC was previously detected at the oppA promoter [19].
Another ScoC-binding site, site IV, was detected further upstream within the divergent iscSB gene (Figs 1A and 7). This site was not present in the braB184-lacZ fusion and therefore was not involved in the regulation described. No consensus ScoC-binding motifs, AATAnTATT [18], with ≤2 mismatches were detected within any of the braB binding sites.
The locations of ScoC-binding sites I and II (Figs 1 and 7) correspond well to the binding sites for the predicted second regulator of braB determined by deletion analysis (Table 2). That is, expression of the braB162-lacZ and braB144-lacZ fusions, which lack the upstream ScoC-binding sites, was not affected by a scoC mutation even if the latter was present together with a codY mutation (Table 2). On the other hand, expression of the slightly longer braB184-lacZ fusion, which includes an intact ScoC-binding site I, as well as the downstream site II, was subject to full ScoC repression (as revealed in a double codY scoC mutant) (Table 2).
Expression of the braB156-lacZ and braB181-lacZ fusions, which carry the upstream ScoC binding site but lack the downstream site II, was also not affected by a scoC mutation (Table 2). A requirement for interaction with two (or more) binding sites within the same regulatory region appears to be a common theme for ScoC-mediated repression [18, 19, 34–36].
The lack of both ScoC- and CodY-mediated regulation explains why the braB76-lacZ, braB156-lacZ, and braB181-lacZ fusions are expressed at the same level in wild-type cells and in codY null mutant cells (Table 2). On the other hand, the braB242p3-lacZ fusion, which lost direct CodY-mediated regulation, is still subject to repression by increased levels of ScoC accumulated in a codY null mutant strain (Table 3).
Interestingly, the p2 mutation, designed to reduce binding of CodY to motif 2 of site II, in fact may affect ScoC interaction with the braB regulatory region. Indeed, the p2 mutation did not affect expression of the braB242-lacZ fusion in a wild-type strain, in which scoC is repressed, but did so in codY null mutant cells, in which ScoC is expressed (Table 3, strains BB3730 and BB3736); the p2 mutation is located 1 bp downstream of ScoC-binding site III (Fig 1).
The expression levels of different fusions and locations of the ScoC-binding sites confirmed that ScoC is the predicted second repressor of braB. As noted above, deleting of one of the ScoC-binding sites resulted in a plateau-reaching expression pattern of the braB144-lacZ fusion in strains with different CodY activity (Fig 5E).
Although previous analysis did not detect any significant regulation of braB by CodY, we now know that braB is subject to complex CodY-mediated regulation by which the protein acts both as a direct repressor and as an indirect positive regulator. The positive effect of CodY is mediated by its repression of the gene encoding a second repressor of braB, ScoC. As a result, braB expression only escapes repression under conditions (e.g., during growth in a medium containing multiple amino acids but lacking ILV) in which CodY activity is limited enough to prevent repression of braB, but high enough to maintain sufficient repression of scoC (Fig 9).
Our previously described repression of scoC by CodY, coupled with ScoC autorepression [19], keeps the level of ScoC relatively low when cells are growing rapidly. Thus, CodY and ScoC are never fully active or inactive simultaneously. When CodY is inactive, the ScoC level is high enough to repress its target genes, including braB. When CodY is fully active, the ScoC level is insufficient for repression, but CodY is able to repress braB to the same level as fully active ScoC. Because we observe higher expression of braB under conditions of partial CodY activity, we suspect that as CodY activity declines, its binding to the braB regulatory region decreases more rapidly than does its binding to the scoC regulatory region. Alternatively, the affinity of ScoC for its braB binding site might be low enough that ScoC needs to reach a relatively high concentration in order to be effective; by the time this happens, CodY-mediated repression of braB is already very low. In addition, it is possible that the competition between more strongly binding CodY and more weakly binding ScoC for interaction with the same region of the braB regulatory region (at the downstream sites for each protein) may contribute to the differential response of braB expression to varying levels of CodY activity. As a result, even relatively small losses in activity of CodY, such as in CodY(F71A), allow neither efficient direct repression by CodY nor sufficient derepression of ScoC, which would compensate for the loss of CodY-mediated repression.
The novelty of braB regulation reinforces the view that important mechanisms of gene regulation can be missed by using regulatory protein null mutants as the only means of genetic analysis. A null mutant has no activity in any environment and at any stage of its life cycle, but in wild-type cells regulatory proteins are rarely, if ever, totally inactive. What normally varies is the fraction of the population of the regulator that is in the active state. Furthermore, interpreting the phenotype of a null mutant usually assumes that a regulatory protein is either only a positive regulator or only a negative regulator of its target gene(s). Studying the behavior of genes at different levels of a regulator’s activity has the potential to reveal more complex mechanisms in detail.
It should be noted that, although the complex pattern of braB regulation is very interesting, it is not common. Combined repression by CodY and ScoC has also been observed for the B. subtilis opp operon and scoC gene itself. However, in case of opp, ScoC-mediated repression was more efficient than CodY-mediated repression and was detected, although at a reduced level, even in codY+ cells [19]. The opposite was true for expression of the scoC gene, whose regulation by CodY was detected even in scoC+ cells [19]. Among CodY-regulated genes, only the braB gene has shown the described up-and-down pattern, i.e., expression was maximal at the intermediate levels of CodY activity [8]. It remains unknown whether additional regulatory inputs, e.g., through SalA-mediated regulation of scoC expression [37], affect interaction between ScoC and CodY.
We have recently characterized three permeases, BcaP, BraB, and BrnQ, involved in the BCAA uptake in B. subtilis cells [1]. The roles of different BCAA permeases in amino acid uptake under different growth conditions should reflect their levels of expression. The bcaP (yhdG) gene encodes the most efficient permease for isoleucine and valine and is one of the genes most highly repressed by CodY; expression of the bcaP gene is virtually abolished in amino acid-rich media [2, 6]. It is very likely that higher activity of BraB is not needed during strong nutrient limitation when CodY activity is very low, because BcaP is fully derepressed. It is also likely that when bcaP and braB are repressed by highly active CodY, the residual activity of BraB, together with BrnQ, is sufficient for the uptake of high concentrations of BCAA. However, the increase in BraB expression at partial activities of CodY may facilitate the uptake of intermediate concentrations of BCAA.
It is not uncommon for two regulators to control expression of the same gene in such a way that the lack of one regulator is fully compensated for by the increased activity of the other regulator and, as a result, no regulatory effect is observed in single null mutant strains. However, when such regulators act independently and do not form a feed-forward regulatory loop, the full compensatory effect should also be observed at intermediate activities of the regulator. The peculiarity of braB regulation is that the full compensatory effect of ScoC is seen only when CodY has very low or no activity.
The feed-forward regulatory loop formed by CodY and ScoC at the braB promoter, known as a type-2 incoherent loop, is an arrangement in which two regulatory proteins repress the same target gene and one of the regulators represses expression of the other [20, 21]. This regulatory mechanism may have evolved specifically to achieve higher expression of the braB gene at intermediate activities of CodY. Genes that are regulated by a single repressor are also expressed at a higher level when activity of the repressor is reduced. However, expression of such genes reaches its maximum only when the repressor is completely inactive; the regulatory mechanism of braB avoids this scenario.
The B. subtilis strains constructed and used in this study were all derivatives of strain SMY [38] and are described in Table 4 or in the text. Escherichia coli strain JM107 [39] was used for isolation of plasmids. Bacterial growth in DS nutrient broth or TSS 0.5% (w/v) glucose-0.2% (w/v) NH4Cl minimal medium was as described [2]. The TSS medium was supplemented as indicated with a mixture of 16 amino acids [40]. This mixture contained all amino acids commonly found in proteins (all concentrations in μg/ml) except for glutamine, asparagine, histidine, and tyrosine: glutamate-Na, 800; aspartate-K, 665; serine, 525; alanine, 445; arginine-HCl, 400; glycine, 375; isoleucine, leucine, and valine, 200 each; methionine, 160; tryptophan, 150; proline, threonine, phenylalanine, and lysine, 100 each; cysteine, 40. In some experiments, ILV were omitted from the amino acid-containing medium.
Methods for common DNA manipulations, transformation, primer extension, and sequence analysis were as previously described [24, 41]. All oligonucleotides used in this work are described in Table 5. Chromosomal DNA of B. subtilis strain SMY or plasmids constructed in this work were used as templates for PCR. All cloned PCR-generated fragments were verified by sequencing.
Plasmid pBB1593 (braB242-lacZ) was created by cloning the XbaI- and HindIII-treated PCR product in an integrative plasmid pHK23 (erm) [24]. The 0.24-kb braB PCR product, containing the entire braB regulatory region, was synthesized with oBB417 and oBB418 as primers. Plasmids pBB1596 (braB144-lacZ) or pBB1772 (braB184-lacZ), containing the braB regulatory region truncated from the 5’ end, were constructed in a similar way using oBB422 or oBB645, respectively, instead of oBB417. Plasmids pBB1597 (braB156-lacZ) and pBB1807 (braB181-lacZ), containing the braB regulatory region truncated from the 3’ end, were created as pBB1593 but using oBB423 or oBB688, respectively, instead of oBB418. Plasmids pBB1803 (braB162-lacZ), pBB1804 (braB76-lacZ), and pBB1808 (braB101-lacZ), in which the braB regulatory region was additionally truncated at the 5’ end, were constructed as pBB1593, pBB1597, and pBB1807, respectively, but using the ApoI and HindIII-digested PCR products that were cloned in pHK23, treated with EcoRI and HindIII.
B. subtilis strains carrying various lacZ fusions at the amyE locus (Table 4) were isolated after transforming strain BB2511 (amyE::spc lacA) with the appropriate plasmids, by selecting for resistance to erythromycin, conferred by the plasmids, and screening for loss of the spectinomycin-resistance marker, which indicated a double crossover, homologous recombination event. Strain BB2511 and all its derivatives have very low endogenous β-galactosidase activity due to a null mutation in the lacA gene [42].
Plasmids pBB1773 (braB242p3-lacZ), pBB1774 (braB242p2-lacZ), and pBB1775 (braB242p1-lacZ), containing 2-bp substitution mutations in CodY-binding motifs, were constructed as described for pBB1593 using fragments generated by two-step overlapping PCR.
In the first step, a product containing the 5’ part of the braB regulatory region was synthesized by using oligonucleotide oBB417 as the forward primer and mutagenic oligonucleotide oBB641, or oBB643, or oBB646 as the reverse primer. A product containing the 3’ part of the braB regulatory region was synthesized by using mutagenic oligonucleotides oBB642, or oBB644, or oBB647 as the forward primer and oligonucleotide oBB418 as the reverse primer. The PCR products were used in a second, splicing step of PCR mutagenesis as overlapping templates to generate a modified fragment containing the entire braB regulatory region; oligonucleotides oBB417 and oBB418 served as the forward and reverse PCR primers, respectively.
Plasmid pBB1776 (braB242p2/p3-lacZ), pBB1801 (braB242p1/p3-lacZ), and pBB1802 (braB242p1/p2-lacZ), containing two mutations, each, were constructed in a similar way, but using a plasmid, containing one of the mutations, pBB1773 or pBB1774, as template for PCR.
Truncated plasmids, containing mutations in the braB regulatory region, were constructed in the same way as plasmids without mutations.
A conversion plasmid for replacing the aphA3 marker for the erm marker, originating from Tn917, was constructed by cloning the 1.5-kb SmaI-StuI fragment of pDG782 [43] into the SnaBI site of pJPM8 [44]. In the resulting plasmid, pBB1560, the aphA3 gene of pDG782, conferring resistance to kanamycin or neomycin, is flanked by 5’ and 3’ parts of the erm cassette of pJPM8; the orientation of the aphA gene coincides with that of erm.
The PCR products containing the regulatory region of the braB gene were synthesized using braB-specific oliginucleotides or vector-specific oligonucleotides oBB67 or oBB358 and oBB102, as the forward and reverse primers, respectively. The reverse primer for each PCR reaction (which would prime synthesis of the template strand of the PCR product) was labeled using T4 polynucleotide kinase and [γ-32P]-ATP. oBB67 and oBB358 start 96 bp or 12 bp upstream of the XbaI site (and 112 bp or 28 bp upstream of the EcoRI site) used for cloning, respectively, and oBB102 starts 36 bp downstream of the HindIII site that serves as a junction between the promoters and the lacZ part of the braB fusion.
The procedures for gel shift and DNase I footprinting experiments were as described [19].
A 0.24-kb braB PCR product, containing the entire braB regulatory region, was synthesized with oBB724 and oBB725 as primers. Plasmid pBB1845 (braB242-gfp) was created by cloning the EcoRI- and SalI-treated PCR product between the EcoRI and XhoI sites of an integrative plasmid pMMB759 (tet), containing a gene encoding a monomeric version (A206K) of GFPmut2 [45]. The braB insert within pBB1845 was identical to the insert in pBB1593 (braB242-lacZ). B. subtilis strain BB4082 carrying the braB242-gfp fusion at the lacA locus was isolated after transforming strain BB2263 (lacA::spc) with pBB1845, by selecting for resistance to tetracycline, conferred by the plasmid, and screening for loss of the spectinomycin-resistance marker, which indicated a double crossover, homologous recombination event.
Cells, containing the braB242-gfp fusion, were grown until mid- to late-exponential phase in TSS + 16 aa medium, centrifuged and resuspended in TSS medium at OD600≈3. The images were collected at a 1,500 ms exposure time using the 100x (1.3 N.A.) objective of the Zeiss Axio Observer.Z1 fluorescent microscope (Zeiss) with the Colibri.2 LED light source, and the ORCA-R2 digital charge-coupled device camera (C10600, Hamamatsu). Zen Pro 2012 software (Zeiss) was used to acquire, view, and analyze the images.
CodY-His5 and His6-ScoC were purified to near homogeneity as described previously [19, 24].
β-Galactosidase specific activity was determined as described previously [46].
RNA isolation, DNA depletion, and cDNA synthesis were performed as previously described [14]. Quantitative, real-time RT-PCR was used to measure steady state braB transcript abundance during exponential growth using oligonucleotides oSRB339 and oSRB340 as described [14], except that we used B. subtilis strain SMY chromosomal DNA to generate the standard curve. rpoC transcript was used to normalize mRNA abundance.
|
10.1371/journal.pntd.0002333 | Imported Amoebic Liver Abscess in France | Worldwide, amoebic liver abscess (ALA) can be found in individuals in non-endemic areas, especially in foreign-born travelers.
We performed a retrospective analysis of ALA in patients admitted to French hospitals between 2002 and 2006. We compared imported ALA cases in European and foreign-born patients and assessed the factors associated with abscess size using a logistic regression model.
We investigated 90 ALA cases. Patient median age was 41. The male:female ratio was 3.5∶1. We were able to determine the origin for 75 patients: 38 were European-born and 37 foreign-born. With respect to clinical characteristics, no significant difference was observed between European and foreign-born patients except a longer lag time between the return to France after traveling abroad and the onset of symptoms for foreign-born. Factors associated with an abscess size of more than 69 mm were being male (OR = 11.25, p<0.01), aged more than 41 years old (OR = 3.63, p = 0.02) and being an immigrant (OR = 11.56, p = 0.03). Percutaneous aspiration was not based on initial abscess size but was carried out significantly more often on patients who were admitted to surgical units (OR = 10, p<0.01). The median time to abscess disappearance for 24 ALA was 7.5 months.
In this study on imported ALA was one of the largest worldwide in terms of the number of cases included males, older patients and foreign-born patients presented with larger abscesses, suggesting that hormonal and immunological factors may be involved in ALA physiopathology. The long lag time before developing ALA after returning to a non-endemic area must be highlighted to clinicians so that they will consider Entamoeba histolytica as a possible pathogen of liver abscesses more often.
| Amœbiasis is caused by Entamoeba histolytica (E. histolytica), a protozoan specific to humans which infects humans by ingestion of contaminated food and water. According to some authors, amœbiasis could be the second leading cause of death from parasitic disease worldwide. It is endemic in tropical countries but can also be diagnosed in industrialized countries, essentially in travelers. One of its main clinical manifestations is amoebic liver abscess (ALA). Abscess can be medically treated by metronidazole, but percutaneous aspiration of the abscess is sometimes performed. Few studies have been performed so far regarding the existence of cases of ALA in the industrialized world. The results of our study reported the existence of 90 cases of imported ALA in France which should raise interest among physicians as well as tourists traveling to endemic areas. Male adults and foreign-born patients presented larger abscesses, suggesting that hormonal and immunological factors could be involved in ALA physiopathology. Foreign-born patients had a longer lag time between the return to France after traveling abroad and the onset of symptoms than European-born ones. This must be highlighted to clinicians who should think about this diagnostic even if a recent travel in a tropical area is not notified by patients.
| Amœbiasis is caused by Entamoeba histolytica (E. histolytica), a protozoan specific to humans. Amœbiasis is present throughout the world, but is endemic in tropical countries where the risk of faeco-oral transmission is high [1]. The main clinical manifestations of amœbiasis are colitis and liver abscess (ALA) [1]. Pleuropulmonar and pericarditis forms also exist but ALA is the most common extraintestinal manifestation of the disease and is one of the etiologies of febrile returning travelers. In Europe, ALA is observed in European-born travelers visiting tropical countries, and in foreign-born patients living in European countries. Complications of ALA are rupture of the abscess into the neighborly cavities of the pleura, pericardium, peritoneum and distant embolic dissemination [2]. ALA was historically responsible for a high number of fatal cases, but since the introduction of medical treatment, the mortality rate is dropped to around 1 to 3%.
The management of ALA is still debated, the indications for percutaneous aspiration being a source of controversies. However most authors recognize that liver abscess puncture or drainage is generally indicated in severely ill patients, patients with large abscess, and if there is a immediate risk of rupture [1]. In France, one single study assessed treatment options for imported ALA and proposed an algorithm for drainage procedures [3]. Size of ALA is determined by radiology and is in most cases between 5 and 10 cm [2], [3], [4], [5]. Nevertheless, no factor has been identified as showing a correlation with the size of the abscess up to date.
It is critical than clinicians should be aware of the presentation, evolution and management of imported ALA.
The aim of this study was to describe the presentation of ALA in the early 2000's in Paris area where there is a high proportion of travelers and foreign born (i.e. born outside the European Union) individuals. We compared the presentation, treatment and outcome of ALA in these two populations. In addition we aimed to identify risk factors associated with abscess size at diagnosis and with percutaneous aspiration. We also aimed to describe evolution of the abscesses when follow up allowed it.
We retrospectively investigated the medical records from patients diagnosed with ALA between January 2002 and December 2006 in 13 teaching hospitals in Paris area which agreed to participate to the study. We collected demographic data, clinical data, laboratory and radiology findings and outcome on the basis of a standardized questionnaire.
Diagnosis of ALA relied on three criteria: travel associated exposure, presence of at least one liver abscess detected by abdominal ultrasound and/or CT scan in a symptomatic patient, and a positive serology for amœbiasis.
Positive serology had to be confirmed by two of the following tests: ELISA, indirect immunoflourescence assay (IFA), indirect haemagglutination (IHA), Latex and counterimmunoelectrophoresis (CIE). Positive serology was defined when titres were above 1∶200 for ELISA, 1∶100 for IFA, 1∶320 for IHA and when there was at least one precipitation arc for CIE.
We distinguished two populations and four patient profiles. The first population was represented by European-born patients defined as patients born and living in France who traveled to tropical areas (European travelers) and patients born in France living for more than 6 months in an endemic region (expatriates). The second population consisted of foreign-born patients defined as patients born in an endemic region, having immigrated to France who traveled back to their country of origin to visit friends and relatives (VFRs), and patients born in an endemic region who had not returned to their country of origin since they emigrated to the EU (immigrants).
We estimated time of apyrexia and biological normalization for patients who had assessment monitoring data. Abdominal ultrasounds and/or CT-scans, for those patients who had them, were reviewed to estimate the disappearance of the abscess.
We analyzed the risk of large ALA at diagnosis, defined as a diameter greater than 69 mm, corresponding to the median abscess size in our patients, using a logistic regression model. Factors investigated were: sex of patients, age (divided into two categories defined by the median age), birthplace (France, Africa and other), patient profile (European travelers, European expatriates, Immigrants and VFR), place of contamination (Africa, Asia and other), HIV status (positive/negative), absence or presence of abdominal pain, laboratory findings such as clinically relevant haemotological and biochemical parameters : anemia defined by haemoglobin lower than 12 g/dL, a White Blood Cells count higher or lower than 10 000 per mm3, a level of C Reactive Protein higher or lower than 200, and cytolysis defined by elevated liver transaminase levels up to the normal, location of the abscess (left or right liver or unknown), number of abscesses (1, 2, 3 and 4 or more), lag time between return from an endemic area and date of the onset of symptoms, in categories based on the median (more or less than 128 days) and lag time between onset and treatment in categories based on the median (more or less than 10 days). For all of these factors, missing values were defined as “unknown”. A descriptive analysis of the dependant and independent variables was performed. A univariate analysis was then conducted, and all factors associated with large ALA having a p-value<0.20 were entered into the multivariate model. A backward stepwise selection procedure was then applied to identify significant (p<0.05) independent variables.
Logistic regression was also used to identify factors associated with the use of percutaneous aspiration. The same factors were investigated, as well as the size of the ALA (≤69 or >69 mm), the treatment and the type of care unit where the patient was admitted (i.e. medical or surgical department). Variables “place of birth” and “patient profile” are correlated. In our model we preferred the variable “patient profile”, which was more relevant to us because the consideration of the place of birth and reason of travelling are more appropriate for clinical practice.
Clinical profile, as well as laboratory and radiological findings were also compared between foreign-born patients (i.e., VFRs and immigrants) and European-born patients (i.e., European travelers and expatriates) using a Student t-test, chi-square test or Fischer's exact test, when required.
Using ultrasound and CT scans, we monitored the time to the disappearance of the abscess using Kaplan-Meier estimates. The Kaplan-Meier curves were compared across the different groups using the log-rank test. Factors associated with the disappearance of an abscess were identified using a Cox proportional hazard model. The factors investigated were the same as those used in the logistic regression model, and, for each factor, the proportional hazard assumption was tested based on Schoenfeld's residuals.
All analyses were performed using STATA 10.0 (Stata statistical software, Stata Corporation, College Station, Texas, USA), and a p-value≤0.05 was considered as significant.
The study received the approval of the French National Commission and Informatics and Liberties under the number 165 29 66 and all data were anonymized.
331 positive amœbiasis serologies were reported. 229 positive amœbiasis serologies were without liver abscess 102 Amoebic Liver Abscesses were diagnosed during the period but 12 medical records were unaivailable (Figure 1). A total of 90 patients with ALA were identified during the study period. The male/female ratio was 3.5 and the median (inter quartile range (IQR)) age at diagnosis was 41 (33–53) years. Age distribution was not different between men and women (t-test, p = 0.46). All the patients had traveled in a region where amœbiasis is endemic: Africa (56%), Asia (19%), and South/North America (4%). There was 38 (51%) European-born patients and 37 (49%) foreign-born patients. We could not determine origin for 15 subjects. The median (IQR) time between the return from the endemic area and the first symptoms was 128 days (61–563). The median time between the first symptoms and diagnosis was 10 days (5–20) and was longer for foreign-born than European-born patients (t-test, p<0.01). Of the 79 patients tested for HIV, 4 (5%) were HIV-positive.
Clinical and laboratory findings are summarized in Table 1. Stool sample examination was performed in 58 patients (64%) (one sample for 34 (38%) and two or more for 23 (26%)). Stool samples were performed before treatment in 11 patients and were found to be positive for Entamoeba cysts in five patients.
Diagnosis of liver abscess relied on ultrasound in 30%, CT scan in 25%, and both in 45% of cases. A single abscess was detected in 77% of cases whereas two abscesses were present in 9% of cases and three or more in the remaining 14%. The median (IQR) diameter at diagnosis was 69 mm (50–290). The main abscess location was the right lobe of the liver (78%).
Pleural effusion was found in 12 patients. Reported complications included portal thrombosis, biliary tract rupture and colectomy due to amoeboma.
Metronidazole was the most commonly used anti-amoebic agent (94.5% of the cases) but others, such as tinidazole (n = 3) and ornidazole (n = 2), were also used. The median (IQR) time between first symptoms and treatment, known for 62 subjects, was 9 days (5–17). Treatment duration was longer than 14 days in 39 subjects (43%). Treatment was complemented by a non-absorbed anti-amoebic luminal agent (i.e., tilbroquinol-tiliquinol) in 80% of the patients.
At least one antibiotic was additionally prescribed in 59 patients (72%). Third generation cephalosporin (n = 34), or amoxicillin alone (n = 4) or amoxicillin-clavulanate (n = 22) were used. Percutaneous aspiration was associated with the medical treatment in 27 patients. No one in this population had surgical treatment.
All the patients were hospitalized. The median (IQR) time to apyrexia after treatment initiation was three days (1–4). Median time for normalization of WBC count and CRP was 5.5 days (1–30) and 15 days (9–30), respectively. Two relapses of ALA were observed, one and three years after initial treatment, respectively. One patient had not received any anti-luminal agent during the first episode and another patient received tilbroquinol-tiliquinol. There was no fatal case.
Our comparison (See Table S1 in appendixes) showed that there were no significant differences between the two populations except that a longer time between return from the endemic area and the onset of the first symptoms was observed for foreign-born patients: 17% of foreign-born patients presented with an ALA more than one year after returning from a tropical area where the disease is endemic, while for European-born patients, all cases of ALA were diagnosed within a year of their return (p<0.01). The maximum lag-time between the time of return from the endemic area and the onset of symptoms was 14 years in foreign-born patients and 12 months in their European-born counterparts.
Data were available for 86 abscesses. Table 2 shows the results of the logistic regression analysis of factors associated with abscesses larger than 69 mm at the time of diagnosis. In multivariate analysis, men compared to women and patient older than the median age of 41 were more likely to present with a larger abscess (odds ratio (OR) = 11.25, [95% confidence interval (CI): 2.42–52.29] and OR = 3.63 [95% CI : 1.26–10.48] respectively), as were immigrants when compared with the European-born travelers' group (OR = 11.56, [95% CI : 2.10–63.41]). The estimated specificity and sensibility of the model were 71% and 70% respectively.
The only factor significantly associated with percutaneous aspiration in multivariate analysis was the type of department where the patient was admitted: those initially admitted to a surgical unit were more likely to have percutaneous aspiration than those admitted to a medical unit (OR = 10.0, [95% CI: 2.70–37.03]) (See Table S2 in appendixes). We couldn't identify any statistical interactions which would be clinically relevant in our models.
CT-Scan and/or ultrasound post-treatment monitoring was available in 24 cases. Using Kaplan Meier estimates, the observed median time to abscess disappearance was 7.5 months (Figure 2). Using a Cox proportional hazard model (See Table S3 in appendixes), the factor most strongly associated with abscess disappearance was its initial size. Larger abscesses were less likely to disappear than smaller ones: hazard ratio (HR) = 0.40 [95% CI: 0.15–1.03]; p = 0.06). Treatment (duration of antibiotherapy and percutaneous aspiration) was not associated with the probability of disappearance.
We identified 90 cases of ALA in the Paris area during the five-year study period. We showed that men were more likely to present with larger abscesses than women.
This is the second largest series of imported ALA. In 1984, Laverdant et al. reported in France 152 cases collected over 15 years in two French military hospitals [6]. In this historical and descriptive cohort, the clinical presentation was the same as in our study, with a mean abscess size of 65 mm. Nevertheless the population was different (military personnel) and percutaneous aspiration was less frequently performed than in our study (12% vs. 30%). As in other studies, ALA mainly concerns middle-aged men, presenting with fever and abdominal pain as well as elevated WBC count and CRP rate [3], [7], [8], [9], [10], [11].
ALA was much more frequently diagnosed in men than in women as in several other studies [3], [6], [8], [9], [11], [12], [13], [14]. This is unlikely that the higher proportion of men in foreign-born population living in Western countries is a relevant explanation for this. Indeed the male predominance is also found in the subgroup of European travelers (75.7% of foreign-born and 81.6% of European are male). In addition this gender tendency has already been reported for other infectious diseases [15]. For instance, analysis of pyogenic liver abscesses in Denmark over a 30-year period, showed that men were more likely to be affected [16]. This finding is also consistent with the animal model, as Lotter et al. demonstrated that male mice developed larger ALA than female mice [17], [18]. Hormonal factors have been suggested to explain this difference [15]. A recent study in hamsters highlights the role of sex hormones in ALA development [19]. Male hamsters, who were gonadectomized, developed either no or smaller ALAs, suggesting that testosterone could be a host factor favoring the development of ALAs. Other studies strongly support the hypothesis that immunological factors explain this gender difference. Indeed in humans, Snow et al. showed that female serum was more effective in killing E. histolytica trophozoïtes than a male one, thanks to the complement system [20]. In the animal model, female mice recovered more rapidly than their male counterparts due to a higher production of Interferon-γ [17] secreted by Natural Killer T Cells [18].The absence of association with HIV infection and ALA size in our study, is contradictory to this immunological hypothesis, but can be attributed once again to the low number of subjects. In a study involving more than 2000 ALA cases from an endemic region, Blessmann et al. showed a peak incidence at 40–49 of age as in our cohort [8]. Regarding the positive role of testosterone on susceptibility of ALA, we could expect to observe a decrease of incidence of ALA with age.
In our study, we referred to the abscess size as being the size at diagnosis, as opposed to the susceptibility to amœbiasis. For elderly patients, alteration of immunity could be determined as a factor to develop a bigger abscess than for younger patients. This could be explained by the effects of immunosenescence [21]. We have also shown that laboratory findings, especially regarding leucocytosis, C Reactive Protein and hemoglobin, were identical for large and small abscesses. This fact is to be brought to the attention to clinicians who might underestimate ALA size when expecting higher White Blood Cells count, C Reactive Protein level and more severe anemia for bigger abscess. Besides, since no significant difference was observed between the two abscess size groups regarding pleural effusion -which is one of the main complications of ALA-, it seems that localization is more to be linked to pleural complication than size itself in this case.
By comparing patients of European origin and foreign-born patients, we found a longer time between the patient's return from an endemic area and the onset of symptoms and larger abscesses in foreign-born patients. In contrast with this latter result, a study performed in the US showed that travelers born in the US were more likely to have larger abscesses (and chronic illness) compared with patients born in endemic countries [11]. In that study, immigrants came from South America and Asia whereas our patients mainly came from Africa. This longer time to develop ALA and larger ALA may both be explained by immunological reasons due to previous contact with Entamoeba histolytica, possibly leading to a certain protective immunity providing a less acute evolution of the disease. Although few data support this hypothesis, some studies explaining the low rate of relapse by anti-lectin IgA antibody mediated mucosal immunity [22], [23], [24] have suggested that a possible strong protective immunity develops after an episode of intestinal or liver amœbiasis. Another explanation could be that foreign-born patients have delayed access to care as this has already been described for patients infected with HIV [25], [26], a fact which led to their consulting a medical practitioner at a later disease stage. It is worth noting that among foreign-born subjects, immigrants but not VFR, seemed more likely to present with a larger abscess when compared with European-born travelers. This finding, which has not been reported in any other study may be due to better access to care for VFRs as compared to immigrants that are facing economic problems not encountered by VFRs which remain able to travel.
Nevertheless, the long delay before developing the disease, sometimes years for foreign born individuals and months for their European-born counterparts, is of clinical interest and should be highlighted. A few data are available about incubation because most of studies about ALA are performed in endemic area, where exposition is permanent. Some authors have ever suggested a long incubation, easily evaluated in case of imported ALA [5]. The problem in these studies, as in ours, is that only the last travel declared by patients is taken into account, omitting the fact they may have been contaminated during previous travel.
Clinicians should also remember that, when patients initiate treatment, apyrexia is quickly achieved, a fact which should help them to evaluate the efficiency of their treatment. In addition we were able to estimate to a median time of 7.5 months the disappearance of ALA in a quarter of the patients. This highlights the fact that recovery is a long process and that the persistence of lesions, as highlighted by radiology techniques, should not alarm clinicians during follow-up in patients free of symptoms. In addition, based on the admittedly small number of subjects in our Cox model, we found that treatment duration was not associated with any decrease in abscess size.
Percutaneous aspiration is not as effective as prolonged liver drainage but it is difficult to explain why percutaneous aspiration was not associated with any decrease in abscess size. The small number of subjects who had percutaneous aspiration monitored in our study (8 individuals) could explain these findings. Percutaneous aspiration is usually recommended for ALA bigger than 10 cm, but it is still unclear if it is really beneficial for the patient [4], [27]. In our study, patients admitted to a surgical unit were more likely to receive percutaneous aspiration. We can hypothesize that patients with a larger abscess were more likely to be admitted to a surgical unit, but our data showed that there was no difference between surgical or medical units in terms of the number of large abscesses. The greater susceptibility of surgeons for surgical treatment compared to medical treatment could explain the highest rate of percutaneous aspiration in their units.
Guidelines for the use of percutaneous aspiration do not exist. Some authors suggest that percutaneous aspiration be used for abscesses with a diameter greater than 10 cm [3], while others suggest it be performed depending on clinical evolution. Our study was not designed to answer this question.
The present study has limitations. First, the retrospective design gives rise to missing data, especially for the time between return in Europe and first symptoms. Place of contamination were unknown for 18 subjects. It could miss some autochthonous case of contamination but, comparing other series of ALA in France, this mode of contamination is exceptional. On the same hand, unknown origin (European or foreign-born) involved 15 subjects. Sensitive analyses, taking account or not these missing subjects, about factors associated to initial size of ALA had similar results. Variables with many missing values were not present in our final models.
The small number of subjects in our study is certainly responsible of a low statistical relevance. In the worst situations, an effect could have been masked in our analysis, but any false association has been shown. The sensibility and specificity of our model about abscess size were around 70%. Our study provides trends in presentation of ALA, but has been proved insufficient to bring irrefutable results.
The low number of subjects with biological and radiological data explains the low number of cases observed for the Cox multivariate model, but this original analysis is at our knowledge the first for imported cases of ALA.
Secondly, our definition of ALA was based on a positive amœbiasis serology associated to a non-bacterial liver abscess. Entamoeba histolytica PCR of the pus aspirate of the abscess is a very useful tool to diagnose ALA. According to case reports, serodiagnostic may lacks sensitivity when compared to PCR, as false positive serologies have been reported in endemic area [5], [28]. PCR has been evaluated in a few studies for imported ALA and should become a gold standard to diagnose ALA when invasive explorations are performed.
At least, our study collected data from 13 hospitals with different types of healthcare management. Longitudinal follow-up studies could better describe recovery and also evaluate factors associated with abscess persistence. Similarly, clinical trials are needed to assess percutaneous aspiration and to perform guidelines for its use in the treatment of ALA, something which can often be adequately treated using drugs alone.
The clinical presentation and outcomes for ALA were similar in European and foreign-born patients. However foreign-born patients presented with ALA later than their Europeans-born counterparts, sometimes several years after travel. Moreover, we showed that men were more likely to present with larger abscesses than women, something which has already been observed in animal model and patients from other countries.
These data support the need for further studies on amoebic liver abscess physiopathology, including the impact of specific immunity and sexual hormones. New tools for an easier and more reliable diagnosis of ALA are also expected as in countries of imported ALA as well as in countries endemic for amœbiasis.
|
10.1371/journal.pgen.1006870 | Intestinal stem cell overproliferation resulting from inactivation of the APC tumor suppressor requires the transcription cofactors Earthbound and Erect wing | Wnt/β-catenin signal transduction directs intestinal stem cell (ISC) proliferation during homeostasis. Hyperactivation of Wnt signaling initiates colorectal cancer, which most frequently results from truncation of the tumor suppressor Adenomatous polyposis coli (APC). The β-catenin-TCF transcription complex activates both the physiological expression of Wnt target genes in the normal intestinal epithelium and their aberrantly increased expression in colorectal tumors. Whether mechanistic differences in the Wnt transcription machinery drive these distinct levels of target gene activation in physiological versus pathological states remains uncertain, but is relevant for the design of new therapeutic strategies. Here, using a Drosophila model, we demonstrate that two evolutionarily conserved transcription cofactors, Earthbound (Ebd) and Erect wing (Ewg), are essential for all major consequences of Apc1 inactivation in the intestine: the hyperactivation of Wnt target gene expression, excess number of ISCs, and hyperplasia of the epithelium. In contrast, only Ebd, but not Ewg, mediates the Wnt-dependent regulation of ISC proliferation during homeostasis. Therefore, in the adult intestine, Ebd acts independently of Ewg in physiological Wnt signaling, but cooperates with Ewg to induce the hyperactivation of Wnt target gene expression following Apc1 loss. These findings have relevance for human tumorigenesis, as Jerky (JRK/JH8), the human Ebd homolog, promotes Wnt pathway hyperactivation and is overexpressed in colorectal, breast, and ovarian cancers. Together, our findings reveal distinct requirements for Ebd and Ewg in physiological Wnt pathway activation versus oncogenic Wnt pathway hyperactivation following Apc1 loss. Such differentially utilized transcription cofactors may offer new opportunities for the selective targeting of Wnt-driven cancers.
| The identification of effective therapy for colorectal cancer, which is a leading cause of cancer-related death, is imperative. Wnt pathway components have promise as therapeutic targets, since more than 90% of colon cancers are triggered by mutations that overactivate this pathway, particularly in the tumor suppressor APC. However, as Wnt signaling is also required for normal intestinal homeostasis, the selective therapeutic targeting of oncogenic Wnt signaling remains a major challenge. Through a forward genetic screen, we previously identified two suppressors of Drosophila Apc1, Earthbound (Ebd) and Erect wing (Ewg), as transcription cofactors of the Wnt pathway. Here, we analyze the roles of these two factors in the Wnt-dependent control of intestinal stem cell proliferation. We find that both Ebd and Ewg are essential for the hyperactivation of Wnt signaling and the consequent epithelial hyperplasia resulting from Apc1 inactivation. Moreover, Ebd, but not Ewg, is also required for the Wnt-dependent maintenance of normal intestinal homeostasis. Together, our findings reveal differential requirements for two highly conserved transcriptional cofactors in Wnt pathway activation versus hyperactivation. The identification of such factors may provide potential selectivity for the targeting of Wnt-driven cancers.
| The evolutionarily conserved Wnt/β-catenin signal transduction pathway directs fundamental cellular processes across metazoans, whereas deregulation of this pathway is associated with numerous human congenital disorders and cancers [1,2]. In the absence of Wnt exposure, β-catenin, a key transcription coactivator, is phosphorylated and targeted for proteasomal degradation by a “destruction complex” comprised of the scaffold protein Axin, the tumor suppressor Adenomatous polyposis coli (APC), and two kinases: glycogen synthase kinase 3 (GSK3) and casein kinase 1α (CK1α). Wnt stimulation inactivates the destruction complex and thereby stabilizes β-catenin, which subsequently translocates to the nucleus and interacts with the DNA-binding transcription factor T-cell factor (TCF) to regulate Wnt target genes [3–5].
The adult mammalian intestine is among the many tissues in which Wnt pathway activation is crucial. Wnt signaling is a key determinant of intestinal stem cell (ISC) maintenance and proliferation during homeostasis [6–10]. Conversely, aberrant activation of the Wnt pathway, which occurs primarily through truncating mutations in APC, initiates the development of the vast majority of colorectal cancers [10–19]. As inhibition of Wnt signaling blocks the division and induces the differentiation of these cancer cells, targeting Wnt pathway components is of great interest for colorectal cancer treatment [20–23].
Many of the same Wnt target genes that are transcriptionally activated in the intestinal epithelium during homeostasis are expressed at aberrantly increased levels in colorectal tumors harboring mutations in APC [18,23–31]. The β-catenin-TCF complex is required for the activation of Wnt target genes in both physiological settings and in these pathological states [7,15,23,32]; however, recent studies have suggested that some of the transcription cofactors interacting with β-catenin-TCF to drive Wnt target gene expression in these two contexts are distinct. For example, B-cell CLL/lymphoma 9 (BCL9) and Pygopus (Pygo), which form a complex with β-catenin and TCF [33–38], are essential only in a subset of tissues during mammalian development [39–45], and are dispensable for Wnt-dependent ISC proliferation and maintenance during homeostasis [40,46]. In contrast, BCL9 and its homolog BCL9-2 are crucial for Wnt-driven intestinal tumor progression [46–52], and Pygo is required for the activation of several Wnt target genes in colon cancer cells [37,47]. These studies suggest that distinct transcription cofactors are utilized in physiological versus pathological states, thereby conferring potential selectivity between Wnt-dependent cell proliferation in normal tissues and tumors. The identification of such novel cofactors that specifically transduce oncogenic Wnt signaling may yield new strategies for the targeting of Wnt-driven cancers.
Through a forward genetic modifier screen for suppressors of Apc1 in the Drosophila retina, we identified Earthbound1 (Ebd1) and Erect wing (Ewg) as context-specific transcription cofactors in the Wingless pathway [53,54]. Ebd1, a member of a protein family containing Centromere Binding Protein B (CENPB) DNA binding domains, physically associates with and bridges β-catenin/Armadillo (Arm) and TCF, thereby promoting the formation and stability of the β-catenin-TCF complex and the recruitment of β-catenin to chromatin [53]. Ewg is a DNA binding transcriptional activator that shares DNA binding specificity with its human homolog, Nuclear Respiratory Factor-1 (NRF-1) [54–57]. We found that Ewg is a physical and functional partner of Ebd1 that promotes the recruitment of Ebd1 to specific chromatin sites [54]. We postulated that recruitment of Ebd1 to chromatin by Ewg enhances the transcriptional activity of the β-catenin-TCF complex, thus promoting Wingless signaling.
Herein, we report that these two Wnt pathway transcription cofactors have distinct functions in the Wnt-directed regulation of the adult Drosophila intestine. Similar to the mammalian intestine, the adult Drosophila midgut undergoes rapid turnover and is replenished by intestinal stem cells (ISCs) [58,59]. We find that both Ebd and Ewg are required for all major consequences of Apc1 inactivation in the adult midgut: the hyperactivation of Wingless target genes, excess number of ISCs, and hyperplasia of the epithelium. By contrast, during intestinal homeostasis, Ebd is essential for the Wingless-dependent control of ISC proliferation, whereas Ewg is dispensable. These studies identify transcriptional cofactors that are differentially required for Wnt signaling in physiological conditions versus the pathological states resulting from hyperactivation of the pathway, providing potential selectivity for therapeutic strategies that target Wnt-driven cancers.
Mammalian genomes encode two APC genes: APC and APC2/APCL with partially redundant roles [60,61]. APC is required in the gastrointestinal tract and serves as a gatekeeper that prevents colorectal cancer [11–19,62–66]. The Drosophila genome also encodes two Apc genes: Apc1 and Apc2 [67–71]. Simultaneous inactivation of both Drosophila Apc homologs results in overproliferation of ISCs and hyperplasia of the intestinal epithelium, resembling the mammalian counterpart [72–75]. However, contradictory findings were reported previously regarding the role of Apc1. Two studies indicated that loss of Apc1 alone results in ISC overproliferation [72,73], whereas another study indicated that Apc1 and Apc2 are fully redundant in this context [75]. To address this controversy, we compared the intestinal epithelium of Apc1Q8 null mutants [67] to controls. In the wild-type intestinal epithelium, ISCs divide asymmetrically to give rise to enteroblasts (EBs) or pre-enteroendocrine (pre-EE) cells, which differentiate into absorptive enterocytes (ECs) or secretory enteroendocrine cells (EEs), respectively [58,59,76–78]. As documented previously [72,73], we found that substantially more progenitor cells (ISCs and EBs) were present in Apc1 mutant midguts, as revealed by the progenitor cell-specific marker esg>GFP (esg-gal4 UAS-GFP) (low magnification: S1A and S1B Fig; high magnification: Fig 1A and 1B and quantification: S2A Fig) [58,59]. Furthermore, the number of EBs, as indicated by the expression of GBE-Su(H)-lacZ [58], was also increased (S2B and S2C Fig). Thus, our results confirmed that loss of Apc1 alone results in excess intestinal progenitor cells.
Moreover, we discovered a novel phenotype that results from Apc1 inactivation: disrupted EC morphology in the midgut epithelium. Levels of membrane-associated Arm in ECs were decreased in Apc1 mutants (low magnification: S1C and S1D Fig; high magnification: Fig 1E and 1F). Furthermore, Discs large 1 (Dlg1), which is normally restricted to the septate junctions between ECs, was instead diffusely cytoplasmic (Fig 1I and 1J). These findings indicate that the cell-cell junctions and the apico-basal polarity of ECs [59,75,79] were disrupted by loss of Apc1. Furthermore, in contrast to the monolayer intestinal epithelium of controls [58,59,74], many ECs were detached from the basement membrane in Apc1 mutant midguts, forming an aberrantly multi-layered epithelium (low magnification: S1E and S1F Fig, Fig 1M and 1N; high magnification: Fig 1Q and 1R and quantification: S2E and S2F Fig). Together, these findings reveal that loss of Apc1 alone is sufficient to result in an aberrantly increased number of progenitors, defects in adhesion and epithelial polarity, and disorganization of the intestinal architecture, in a manner analogous to the pathological consequences of APC inactivation in mammals.
As the severe intestinal defects present in Apc1 mutants were readily detected as early as two days of adulthood, we hypothesized that these phenotypes arise during formation of the adult midgut. To test this hypothesis, we examined the Apc1 mutant gut epithelium shortly after eclosion. Strikingly, compared with the age-matched controls, the midguts of recently eclosed Apc1 mutants (0–4 hours after eclosion) exhibited supernumerary esg>GFP marked progenitor cells (low magnification: S3A and S3C Fig; high magnification: S3B and S3D Fig). The cell-cell adherens junctions, however, remained intact at this time (low magnification: S3A’ and S3C’ Fig; high magnification: S3B’ and D’ Fig). Thus, an excess number of progenitor cells are present prior to eclosion, whereas the disruption of both EC structure and epithelial architecture arise after eclosion. To trace the initial requirement for Apc1, we examined midguts earlier in their development. The adult intestine is derived from the larval gut, but undergoes major histolysis and reformation during pupation [80–82]. Therefore, we examined the gut epithelium of Apc1 mutant third instar wandering larvae, the developmental stage that immediately precedes formation of the adult gut. Notably, supernumerary adult midgut progenitors (AMPs) [80] were not detected in the Apc1 mutant larval guts (S4A–S4B” Fig) and the epithelial structure was normal (S4C–S4D” Fig). Thus, loss of Apc1 initiates intestinal defects during formation of the adult gut during pupation, and these defects increase in severity during adulthood.
To further test the temporal and cell-specific requirements for Apc1 in the midgut, we utilized the temperature-sensitive progenitor cell driver (esgts: esg-Gal4 tub-Gal80ts UAS-GFP) for RNAi-mediated Apc1 knockdown in ISCs and EBs either during formation of the adult gut or during adult gut homeostasis. Of note, dramatic increases in the progenitor cell number were observed in both contexts (adult gut formation: S5A–S5D’ Fig, adult gut homeostasis: S5E–S5H’ Fig; quantification: S5I Fig). Furthermore, we used an inducible “escargot flip out” system (esgts>F/O: esg-Gal4 tub-Gal80ts UAS-GFP; UAS-flp Act>CD2>Gal4 [83]) to mark the progenitor cells and their progeny in which Apc1 was knocked down during adulthood. Compared with controls, increased phospho-histone H3 (pH3) positive cells and stem/progenitor cell lineages were observed within Apc1 RNAi “escargot flip out” posterior midguts (S6A–S6B” Fig; quantification: S6C and S6D Fig), providing further evidence that adult-specific Apc1 knockdown resulted in increased ISC proliferation. Together, these results indicate that Apc1 is required in stem/progenitor cells during both adult gut development and homeostasis, supporting our observations in Apc1 null mutants (Fig 1 and S1–S4 Figs) and previous reports [72,73].
Both the initiation and sustained growth of human colon cancers harboring APC mutations rely on the aberrant activation of Wnt target genes [10,15,16,18,20–31]. To examine whether loss of Drosophila Apc1 also induces the aberrant activation of Wingless target genes in the midgut, we tested three transcriptional reporters for direct target genes of Wingless signaling: frizzled3 (fz3), notum, and naked cuticle (nkd) [84–88]. The Drosophila midgut, like its mammalian counterpart, is subdivided into compartments with distinct histology, gene expression, and physiological functions (Fig 2A) [89–91]. The activation of Wingless signaling is graded along the length of each intestinal compartment; Wingless target genes are activated at high levels at intestinal compartment boundaries and at lower levels within compartments as a function of distance from the boundary [89,92]. We found that inactivation of Apc1 resulted in strong ectopic expression of notum-lacZ [93,94] (low magnification view: S7A and S7B Fig; high magnification view: Fig 2B and 2C), fz3-RFP [95] (low magnification view: S7E and S7F Fig; high magnification view: Fig 2F and 2G) and nkd(UpE2)-lacZ [96–98] (low magnification view: S8A and S8B Fig; high magnification view: S8D and S8E Fig), both at compartment boundaries and within compartments of Apc1 null mutant midguts. To quantify this aberrant increase in expression, we examined fz3-RFP in marked clones of Apc1 null mutant cells generated with the MARCM (Mosaic Analysis of a Repressible Cell Marker) technique [99]. Even at compartment boundaries, where Wingless pathway activity is normally at its peak, a further fivefold increase in fz3-RFP levels was present in Apc1 null mutant clones as compared to the surrounding control tissue (Fig 2J–2L; quantification: Fig 2M). These findings indicate that Apc1 is required to prevent the constitutive activation of Wingless target genes at both compartment boundaries and within compartments of the midgut.
We sought to determine the extent to which Apc1 loss deregulates gene expression. Using Affymetrix microarrays, we found that the expression of approximately 1000 genes, which can be grouped in broad categories, was either up- or down-regulated by more than twofold in Apc1 mutant midguts as compared to wild-type controls (GEO database: GSE99071; S9 Fig). The changes in expression of selected up- or down-regulated genes were validated by real-time quantitative PCR (Fig 3 and S10 Fig). Consistent with the Wingless target gene reporter analysis described above (Fig 2 and S7 and S8 Figs), fz3, notum and nkd transcription were activated significantly in Apc1 mutants (Fig 3A and S10A Fig). Furthermore, a previous study identified Janus kinase/signal transducer and activator of transcription (Jak/Stat) and Epidermal growth factor receptor (Egfr) signaling as key mediators of ISC hyperproliferation in Apc1 mutants [73]. Indeed, the expression of both unpaired 3 (upd3), a ligand of the Jak/Stat pathway, and Socs36e, a downstream target gene of this pathway, were induced upon loss of Apc1 (Fig 3A and S10A Fig). Furthermore, the expression of vein (vn) and spitz (spi), two ligands of the Egfr pathway, was also increased (Fig 3A and S10A Fig). Together, these results suggest that loss of Apc1 results in an extensive deregulation of gene expression.
In a forward genetic screen in the Drosophila retina, we previously identified both Ebd1 and Ewg [53,54] as novel suppressors of Apc1. Furthermore, we found that these two proteins function as transcriptional cofactors that physically interact both with each other and with Arm/Tcf to promote the context-specific activation of Wingless signaling during pupal muscle development [53,54]. However, the limited genetic tools available to analyze Wnt signaling in pupal muscle restricted our ability to identify the roles of Ebd and Ewg in physiological versus pathological Wnt signaling. Herein, we sought to overcome this obstacle by utilizing powerful in vivo assays in the Drosophila intestine.
First, we sought to determine whether Ebd and/or Ewg are required for the phenotypic consequences of Apc1 inactivation in the intestine by examining ebd Apc1 and ewg Apc1 double mutants. Strikingly, the three major defects in the midguts of Apc1 mutants were largely suppressed upon inactivation of either ebd or ewg. First, the numbers of progenitor cells in ebd Apc1 or ewg Apc1 double mutants were similar to those in controls (Fig 1A–1H and S2B–S2D Fig; quantification: S2A Fig). Second, the levels of membrane-associated Arm and the subcellular localization of Dlg1 in ECs were indistinguishable from controls (Fig 1E–1L). Third, in ebd Apc1 and ewg Apc1 double mutants, the midguts reverted to a monolayer epithelium (low magnification: Fig 1M–1P; high magnification: Fig 1Q–1T and quantification: S2E and S2F Fig). Thus, both Ebd and Ewg are required for the excess progenitor cells and epithelial hyperplasia resulting from Apc1 inactivation.
To determine whether Ebd or Ewg are required for the aberrantly high expression of Wingless target genes that results from Apc1 inactivation, we examined the expression of notum-lacZ, fz3-RFP and nkd(UpE2)-lacZ in ebd Apc1 and ewg Apc1 double mutants. Strikingly, upon loss of either ebd or ewg, the hyperactivation of all three Wingless pathway reporters in Apc1 mutant midguts was reduced nearly to control levels (Fig 2B–2E, Fig 2F–2I, S7 and S8 Figs). Thus, not only Arm/β-catenin and TCF, but also Ebd and Ewg are required for the aberrantly increased activation of Wingless target genes in Apc1 mutant midguts.
We sought to test whether Ebd and Ewg are required for the hyperactivation of only a subset of direct Wingless target genes or also have broader effects in the extensive deregulation of gene expression that occurs in Apc1 mutants. Therefore, we analyzed the expression of genes that are selectively up- or down-regulated genes by Apc1 inactivation in either ebd Apc1 or ewg Apc1 double mutants by real-time quantitative PCR (Fig 3 and S10 Fig). Of note, the transcriptional deregulation resulting from loss of Apc1 was rescued in ebd1 Apc1 or ewg Apc1 double mutants for all genes analyzed (Fig 3 and S10 Fig). These findings provide further evidence that the aberrant transcriptional response in Apc1 mutant midguts requires both Ebd and Ewg.
Ewg is a known sequence-specific DNA binding protein [54–57]. Therefore, we sought to determine whether consensus Ewg DNA binding sites are present in the enhancers of genes deregulated by Apc1 loss. As the Wingless target gene reporters notum-lacZ, nkd(UpE2)-lacZ, and fz3-RFP are each hyperactivated in an Ewg-dependent manner following Apc1 loss, and the enhancers within these reporters are well-characterized, we searched for potential Ewg and TCF binding sites in these enhancers. The transcriptional enhancers that drive expression of both the notum-lacZ and nkd(UpE2)-lacZ reporters, which are 2.2 kb [93,94] and 0.6 kb [97], respectively, are directly bound and regulated by TCF through distinct pairs of core consensus sites (SSTTTGWWSWW) and Helper sites (GCCGCCR) [5,96–98,100] (S11A–S11C Fig). We identified similar TCF core consensus binding sites and Helper sites in the 2.3 kb enhancer of the fz3-RFP transgene [95] (S11D Fig). In addition, we found that the fz3 enhancer contains an Ewg consensus binding site (GCGCABGY) [54–57] (S11A and S11D Fig), and that this site is conserved among sequenced Drosophila species (S12 Fig) [101]. In contrast, neither the notum nor the nkd enhancer contains an Ewg consensus binding site (S11A–S11C Fig). Therefore, these findings suggest that the hyperactivation of at least some Wingless target genes in Apc1 mutants may not require direct binding of Ewg to DNA, or alternatively, that Ewg may bind non-consensus DNA sites upon Apc1 inactivation.
RNAi-mediated knockdown of Apc1 specifically in progenitor cells phenocopies the supernumerary progenitors observed in Apc1 null mutants (S5 Fig). Therefore, we hypothesized that Ebd and Ewg act in progenitor cells to mediate the phenotypic consequences of Apc1 loss. To test this hypothesis, we used RNAi-mediated knockdown to concomitantly reduce both Apc1 and ebd or ewg in progenitors. The aberrant increase in progenitor cell number resulting from Apc1 knockdown was largely suppressed upon simultaneous knockdown of either ebd or ewg (Fig 4A–4H’; quantification: Fig 4I). Based on these findings, we conclude that Ebd and Ewg are required in progenitors to mediate the gut defects resulting from Apc1 loss.
Our findings indicate that both Ebd1 and Ewg are required for the aberrantly increased expression of the Wingless target genes resulting from Apc1 loss in the adult midgut (Fig 2 and Fig 3; S7, S8 and S10 Figs). To determine whether Ebd1 and Ewg also promote Wingless target gene expression in the adult midgut under physiological conditions, we analyzed the expression of the Wingless target gene reporter notum-lacZ. Under basal conditions, notum-lacZ expression peaks at both the foregut/midgut (Fig 2B) and the midgut/hindgut boundaries (Fig 5). This expression of notum-lacZ is completely dependent on Wingless pathway activation, as revealed by its loss in fz Dfz2 double null mutant [102] or dsh null mutant clones [103] (Fig 5A–5D”). We found that in many, but not all ebd1 null mutant clones, notum-lacZ expression was eliminated (Fig 5E–5F”), providing evidence that Ebd1 promotes the activation of Wingless target genes not only in hyperactivated states, but also during homeostasis in the adult midgut. In contrast, ewg null mutant clones resulted in no detectable reduction in the expression of notum-lacZ (Fig 5G–5H”). In addition, ewg null mutant clones also did not affect the expression of the other two Wingless target gene reporters, fz3-RFP (S13A–S13B” Fig) or nkd-lacZ (S13C–S13D” Fig), suggesting that Ewg is dispensable for Wingless target gene activation in the adult midgut under physiological conditions. Thus, these findings indicate although both Ebd and Ewg are essential for the hyperactivation of Wingless signaling upon Apc1 inactivation, only Ebd is required for Wnt pathway activation during intestinal homeostasis, whereas Ewg is dispensable.
Wingless pathway activation is crucial for the maintenance of adult midgut homeostasis [89,92]. We thus sought to determine whether Ebd and Ewg are required for this process. To test whether Ewg has a role during adult midgut homeostasis, we first analyzed ewgP1, a hypomorphic allele containing an ewg missense mutation ([54]; note that complete inactivation of ewg results in embryonic lethality). In comparison with controls, ewgP1 midguts contained comparable numbers of progenitor cells (marked by esg>GFP) and displayed normal epithelial architecture (Fig 6A–6B’). Thus, although this allele revealed that Ewg is crucial for the hyperactivated Wingless signaling and intestinal hyperplasia that results from Apc1 inactivation in the adult midgut (Figs 1–3 and S1, S2, S7, S8 and S10 Figs), it exhibits no detectable defects under physiological conditions. To further reduce the level of ewg activity, we examined the midguts of flies transheterozygous for the null allele ewg2 and the hypomorphic allele ewg1, which is the most severe viable combination of ewg alleles (all mutant flies exhibit “erect wing” defects) [54,55]. Notably, no excess progenitors were observed in ewg2/ewg1 transheterozygotes (Fig 6C and 6D; quantification: Fig 6E). These results suggested that consistent with our observation that Ewg is dispensable for physiological Wingless target gene activation, Ewg does not have a role in Wingless-dependent adult intestinal homeostasis. Gut cell type specific RNA-seq under homeostatic condition revealed previously [104] that expression of Ewg is very low during intestinal homeostasis, while Ebd1 is expressed in all gut cell types (Fig 6F). Thus, the possibility arose that loss of Apc1 induces overexpression of ewg and this might explain why Ewg is essential for hyperactivated Wingless signaling but dispensable for physiological signaling. However, RT-qPCR of ewg expression revealed a less than 2-fold increase in Apc1 mutants (Fig 6G), suggesting that an increase in Ewg levels is unlikely the mechanism by which Ewg mediates the consequences of Apc1 loss in the adult midgut.
We next analyzed whether Ebd is required during adult intestinal homeostasis. In Drosophila, two Ebd proteins, Ebd1 and Ebd2, possess partially redundant functions [53]. To elucidate the function of Ebd1 in the midgut, we compared the intestinal epithelium of control (ebd1/+) to ebd1240 null mutants [53]. We found that the number of progenitor cells (marked by esg>GFP or combination of Arm/Prospero staining) was significantly increased in ebd1 mutants (S14A–S14F Fig; quantification: S14M). In addition, in contrast to the control midguts, very few of which (8%) displayed chains of progenitors and none of which exhibited clusters of progenitor cells, the majority of ebd1 mutant midguts (60%) contained chains of progenitor cells and 15% exhibited clusters (S14A–S14F Fig, quantification: S14N Fig). To further determine whether this requirement for Ebd1 in the regulation of ISC proliferation occurs during adulthood, we generated marked ebd1 null mutant clones in adults. We found that ebd1 mutant clones were markedly larger than control clones: 44% of ebd1 mutant clones contained more than 4 cells, as compared to 14% of the control clones (S15 Fig). Together, these findings indicate that in contrast to Ewg, Ebd1 is required for intestinal homeostasis during adulthood, resembling other Wingless pathway components [89,92].
We further sought to determine whether the combined inactivation of ebd1 and ebd2 would result in a more severe phenotype than inactivation of ebd1 singly. Staining for esg>GFP revealed that by comparison with ebd1 single mutants, a larger proportion of ebd1ebd2/ebd1 mutant midguts displayed clusters of progenitor cells (42%), and this proportion increased further in midguts homozygous mutant for both ebd1 and ebd2 (58%) (S14G–S14L Fig; quantification: S14N Fig), indicating that ebd2 inactivation exacerbated the severity of ebd1 null mutant phenotype. To further differentiate the subtypes of progenitor cells that are deregulated by loss of the Ebd1 and Ebd2 proteins, we examined the ISC-specific marker Delta (Dl) [105] and the EB-specific marker GBE-Su(H)-lacZ in ebd1 ebd2/ebd1 mutants, and detected a significant increase in the number of both cell types as compared to controls (Fig 7A–7L; quantification: Fig 7M and 7N). Furthermore, 33% of ISCs (esg+, GBE-Su(H)-) were associated with an EB (esg+, GBE-Su(H)+) in controls, but this number increased to 78% in ebd1 ebd2/ebd1 mutants (Fig 7O). Together, both Ebd1 and Ebd2 are required for homeostasis of the Drosophila midgut as their inactivation leads to an aberrant increase in both ISCs and EBs.
To determine the cell types in which Ebd1 is expressed in the midgut, we immunostained intestines with Ebd1 antibody [53] (S16A–S16A” Fig). Using ebd1 null mutant clones to test the specificity of the Ebd1 antibody, we found that Ebd1 is expressed in enterocytes (S16B–S16C”‘ Fig). We also detected Ebd1 in progenitors and EEs, but based on mutant clonal analysis, it was not clear that this staining was above background (S16B–S16C”‘ Fig). Therefore we conclude that there is strong Ebd1 expression in ECs, and any Ebd1 expression in progenitors cells or EEs is lower than the detection limit of the Ebd1 antibody. We also tested ebd1-Gal4 lines [53] to drive reporter expression, which revealed that Ebd1 is strongly expressed in ECs, but also detectable at lower levels in progenitors and EEs (S17 Fig).
The activation of Wingless signaling in ECs inhibits the proliferation of ISCs non-autonomously to regulate adult intestinal homeostasis [89,92]. Similarly, we found that an abnormally large number of progenitor cells were clustered around ebd1240 or ebd15 null mutant clones (Fig 8A–8B’; quantification: Fig 8C). The excess progenitor cells present near ebd1 clones resulted from aberrantly increased proliferation, as revealed by the number of pH3 positive cells (Fig 8D). Since Wingless signaling is required specifically in ECs to regulate the proliferation of surrounding ISCs [92], we tested whether Ebd1 functions similarly by reducing ebd1 expression in ECs using RNAi-mediated knockdown with the EC-specific driver Myo1A-Gal4 [83]. As compared with controls, knockdown of ebd1 in ECs resulted in increased numbers of progenitor cells (marked by esg-lacZ or a combination of Arm and Prospero staining) and ISCs (marked by Dl) that were present in chains or grouped in clusters (Fig 8E–8K and S18A–S18F Fig). Furthermore, the number of pH3-positive cells increased upon ebd1 knockdown in ECs, confirming the overproliferation of ISCs (Fig 8L). Moreover, as reported previously for inactivation of other Wingless pathway components [92], increased ISC proliferation was observed when ebd1 expression was disrupted during adulthood, but not during development (S18G–S18I Fig). These results were obtained with three independently derived transgenic ebd1 RNAi lines to rule out the possibility of off-target effects. Together, our findings suggest that loss of ebd1, like that of Wingless pathway components, non-autonomously promotes the proliferation of neighboring ISCs.
As Wingless signaling controls the proliferation of ISCs through the Jak/Stat pathway [92], we examined whether Ebd1 analogously controls Jak/Stat signaling. RNAi-mediated knockdown of ebd1 expression in ECs led to significant increases in the expression of upd2 and upd3, ligands for the Jak/Stat pathway (Fig 9A). In contrast, little increase was detected in the expression of decapentaplegic (dpp) or keren (krn), EC-expressed ligands for the TGF-β and EGF pathway, respectively [106–109] (Fig 9A). Similarly, expression of puckered (puc) and keap1, target genes for the two major stress response signaling pathways, JNK (c-Jun N-terminal kinase) and Nrf2 (Nuclear factor 2) respectively [110–112], was not affected (Fig 9B). Thus, Ebd1 specifically regulates the expression of Jak/Stat pathway ligands in ECs, and could thereby control the activation of Jak/Stat signaling in adjacent ISCs. In support of this idea, we found that RNAi-mediated knockdown of ebd1 in ECs induced expression of Socs36e, a direct target gene of the Jak/Stat pathway [113] (Fig 9B). To further test whether Jak/Stat signaling is activated in ISCs upon loss of ebd1 in ECs, we analyzed the expression of the Jak/Stat pathway reporter, stat-GFP [114]. We found that stat-GFP expression increased markedly in ISCs near ebd1 mutant clones (Fig 9C–9C”; quantification: Fig 9D), indicating that the Jak/Stat pathway was activated non-autonomously upon ebd1 inactivation. To determine whether the ectopic activation of Jak/Stat signaling mediates the overproliferation of ISCs resulting from loss of ebd1, we concomitantly knocked down both upd and ebd1 in ECs using RNAi. Dual knockdown of ebd1, and either upd2 or upd3, reduced ISC proliferation in posterior midguts, as indicated by Dl and pH3 staining (Fig 9E and 9F). Therefore, Ebd1 activity in ECs, like that of Wingless pathway components, prevents the non-autonomous activation of JAK/STAT signaling in neighboring ISCs, and thereby inhibits their proliferation.
The observation that ebd1 inactivation results in ISC overproliferation in physiological conditions, but prevents ISC overproliferation in Apc1 mutants presented us with a paradox. Analysis of the spatial and temporal requirement for Ebd1 provided an explanation for these unanticipated results. The Apc1 mutant phenotype emerges during formation of the adult gut during pupation (S3–S5 Figs and S19A, S19C and S19E Fig), a stage in which ebd1 knockdown has no effect (S18G–S18I Fig). Indeed, the midguts of newly eclosed ebd1 mutants exhibited a similar number of EBs by comparison with the age-matched controls (S19A and S19B Fig; quantification: S19E Fig). Furthermore, Ebd1 is non-autonomously required in ECs to prevent ISC overproliferation during adult homeostasis (Figs 8 and 9 and S18 Fig), in contrast to its autonomous requirement in progenitor cells for the gut defects resulting from Apc1 loss (Fig 4). Together, these findings indicate that Ebd1 plays qualitatively different roles in transducing physiological and pathological Wingless signaling, which are temporally and spatially distinct.
In summary, our analysis of two transcription cofactors, Ebd and Ewg, in the Drosophila midgut revealed that both Ebd and Ewg are required for all major consequences of Apc1 inactivation: the hyperactivation of Wingless target genes, excess number of progenitor cells, and epithelial hyperplasia (Fig 10A). By contrast, during intestinal homeostasis, only Ebd, but not Ewg, is essential for the Wingless-dependent control of ISC proliferation (Fig 10B). Together, these findings provide evidence that some context-specific transcription cofactors are differentially required for physiological Wnt pathway activation during homeostasis versus the oncogenic hyperactivation of the Wnt pathway resulting from loss of Apc1, and thus may present opportunities for the therapeutic targeting of Wnt-driven diseases.
Our findings indicate that both Ebd and Ewg are necessary for the aberrantly high-level Wnt target gene activation that mediates the consequences of Apc1 loss. These results provide in vivo evidence that the core β-catenin-TCF transcriptional machinery is insufficient for the transformation of intestinal epithelial cells in Apc1 mutants; cooperation of β-catenin-TCF with Ebd and Ewg is also necessary. As Ebd and Ewg are known to physically interact with each other and with β-catenin, we postulate that the Ebd-Ewg complex acts with β-catenin-TCF to activate the high-level transcription of Wnt target genes in ISCs in Apc1 mutants. Moreover, Ebd, but not Ewg, is required for the Wnt-dependent control of ISC proliferation during homeostasis. Together, these studies reveal that transcription cofactors with context-specific roles in Wnt target gene activation under physiological conditions can be co-opted to function with β-catenin-TCF to promote the global hyperactivation of Wnt target genes following APC loss.
In both mammals and Drosophila, two APC paralogs have partially redundant roles that are dependent on cell context. However, as in humans and mice, inactivation of a single Drosophila APC homolog alone is sufficient to induce ISC overproliferation, as well as defects in intestinal epithelial cell adhesion, cell polarity, and intestinal architecture that recapitulate many aspects of human colorectal cancer. Furthermore, similar to inactivation of human and mouse APC, loss of Drosophila Apc1 results in aberrantly high levels of Wnt target gene expression in the intestine. Our analysis of one Wnt target gene reporter reveals a fivefold increase in its expression in Apc1 mutant cells compared to wild-type cells even at intestinal compartment boundaries, which are the sites with the highest levels of Wingless protein and the highest activation of physiological Wingless signaling in the adult gut. Overall, the expression of approximately 1000 genes is significantly deregulated in Apc1 mutant guts. These results provide evidence that inactivation of Drosophila Apc1 singly results in intestinal hyperplasia and Wingless target gene hyperactivation, in a manner analogous to the pathological consequences that result from loss of mammalian APC.
Our findings also reveal that Ebd and Ewg mediate the intestinal epithelial defects and oncogenic levels of Wnt target gene expression that result from loss of Apc1. In addition, we find that although Ewg is a known sequence-specific DNA-binding protein and is required following Apc1 loss for the high level expression of the Wnt target genes fz3, nkd, and notum through their well-characterized enhancers, Ewg consensus DNA binding sites are present in only one of these three enhancers. Therefore, the direct association of Ewg with DNA might not be required for Ewg’s role in the hyperactivation of Wnt signaling, or Ewg might also interact with non-consensus binding sites. Thus, in these contexts, Ebd might access DNA through its own CENPB-type DNA binding domains [53,115]. Alternatively, these findings raise the possibility that Ewg and Ebd access chromatin via protein-protein interactions instead of direct association with DNA. A precedent for this type of mechanism was documented previously for Fushi tarazu, which activates transcription even when its DNA-binding homeodomain is deleted, through interaction with the DNA-binding transcription factor Paired [116,117].
In the Drosophila intestine, activation of Wingless signaling in ECs non-autonomously restricts the proliferation of surrounding ISCs during homeostasis. Our findings herein suggest that this process requires Ebd. We further find that Ebd is also required for the autonomous hyperactivation of Wingless signaling in ISCs that results in their overproliferation following Apc1 loss. This novel finding reveals that Ebd is required for Wnt signaling during both normal homeostasis of the intestine and its aberrant hyperplasia, in addition to Ebd’s previously documented roles in muscles and neurons [53,54]. Similar to that in other tissues, the role of Ebd in the intestine is context-specific, as not all Wnt-mediated processes are dependent on Ebd; Ebd promotes the Wnt-mediated regulation of ISC proliferation during homeostasis, but is dispensable for the Wnt-dependent specification of cell fate near intestinal compartment boundaries [92]. Conversely, Ewg has no observed role in either of these Wnt-dependent processes. Thus, Ebd functions in an Ewg-independent manner in the adult gut under physiological conditions. These results suggest that Ebd and Ewg do not always function in a complex, and that recruitment of Ebd to chromatin by Ewg [54] is context-specific.
Based on our findings, we propose that mechanistic differences in the Wnt transcriptional machinery underlie target gene activation in physiological versus pathological states. These novel distinctions likely underlie the markedly increased expression of Wingless target genes in the hyperactivated state that results from Apc1 inactivation. Analogously, the mammalian transcription cofactors Pygo and BCL9 also form a complex that enhances target gene activation by β-catenin-TCF in the Wnt hyperactivated state. Neither mammalian Pygo nor BCL9 is required for Wnt-mediated ISC proliferation or maintenance during homeostasis, but both promote Wnt target gene expression in colorectal cancer. Most targeted therapies under investigation disrupt Wnt signaling not only in tumors, but also in normal tissues. Thus, the discovery that transcription cofactor complexes, such as Ebd-Ewg in Drosophila or Pygo-BCL9 in mammals, are essential for supraphysiological signaling but dispensable for most Wnt-dependent physiological processes may distinguish tumors from normal tissues and provide selectivity for therapeutic strategies that target Wnt-driven diseases.
Our findings suggest that the human homologs of Ebd and Ewg might provide novel therapeutic targets for the treatment of Wnt-driven cancers. Jerky (also known as JRK or JH8; [118–123]), the human homolog of Ebd, rescues ebd mutant phenotypes when expressed in Drosophila [53] and promotes the aberrant increase of both cell proliferation and β-catenin-TCF mediated transcription in colon cancer cell lines [53,124,125]. Moreover, aberrantly high levels of Jerky are present in several carcinomas, including colon, breast, and ovarian serous cystadenocarcinoma. Elevated Jerky expression is associated with increased β-catenin nuclear localization and the aberrantly increased expression of Wnt target genes in human colorectal tumors [125]. A possible role for Nuclear Respiratory Factor 1 (NRF1), the human homolog of Ewg, in Wnt signaling awaits future investigation. Together, these findings suggest that inhibition of Jerky, NRF1, or their physical interaction may provide promising therapeutic strategies for colorectal cancer.
Reporters: esg>GFP (esg-Gal4 UAS-GFP) [58], GBE-Su(H)-lacZ [58], fz3-RFP [95], notum-lacZ [93,94], nkd(UpE2)-lacZ [97], esg-lacZ [58], and 10x stat-GFP (destabilized) [114].
Mutant alleles: Apc1Q8 [67], ebd1240 [53], ebd15 [53], ebd2136 [53], Df(3L)9698 [53], ewgP1 [54], ewg2 [55], ewg1 [55], fzH51 Dfz2C1 [102], and dsh3 [103].
MARCM lines: MARCM 82B: y w hs-flp UAS-CD8::GFP; tub-Gal4 FRT82B tub-Gal80/TM6B [126], MARCM 2A: y w hs-flp; tub-Gal4 UAS-mCD8::GFPLL5/CyO act-GFPJMR1; FRT2A tub-Gal80LL9 [127], or y w hs-flp; tub-Gal4 UAS-GFP; FRT2A tub-Gal80/TM6B (A kind gift from the Ohlstein lab), or y w hs-flp tub-Gal4 UAS-dsRed; FRT2A tub-Gal80, MARCM 19A: hs-flp tub-Gal80 FRT19A;; tub-Gal4 UAS-mCD8::GFP/SM6^TM6B [128].
RNAi lines and Gal4 drivers: Myo1A-Gal4 [129], ebd1-Gal4 [53] (DGRC#104336), esg-Gal4 tubGal80ts UAS-GFP/CyO, esgts F/O (esg-Gal4 tub-Gal80ts UAS-GFP; UAS-flp Act>CD2>Gal4) [83], UAS-GFP-lacZ (BDSC#6452), w1118 (BDSC#5905), UAS-Apc1 RNAi#1 (VDRC#51469; Construct ID: 1333), UAS-Apc1 RNAi#2 (VDRC51468; Construct ID: 1333), UAS-ewg RNAi#1 (BDSC#31104), UAS-ewg RNAi#2 (BDSC#31225), UAS-ebd1 RNAi#1 (VDRC#26180; Construct ID: 10952), UAS-ebd1 RNAi#2 (BDSC#35765), UAS-ebd1 RNAi#3 (BDSC#28296), UAS-upd2 RNAi#1 (BDSC#33988), UAS-upd2 RNAi#2 (BDSC#33949), UAS-upd3 RNAi#1 (VDRC#27136; Construct ID: 6811), UAS-upd3 RNAi#2 (BDSC#28575).
Canton S flies were used as wild-type controls. Fly crosses were performed at 25°C unless otherwise indicated.
Primary antibodies were chicken anti-GFP (Abcam, Cat. no. ab13970, 1:10000), rabbit anti-GFP (Thermo Fisher Scientific, Cat. no. A-11122, 1:500), mouse anti-Arm [Developmental Studies Hybridoma Bank (DSHB), N2 7A1, 1:20], mouse anti-Discs large (DSHB, 4F3, 1:20), mouse anti-Prospero (DSHB, MR1A, 1:100), mouse anti-Delta (DSHB, C594.9B, 1:100), rabbit anti-dsRed (Clontech/TaKaRa, Cat. no. 632496, 1:500), mouse anti-β-gal (Promega, Cat. no. Z378B, 1:500), rabbit anti-β-gal (MP Biomedicals, Cat. no. 08559762, 1:5000), rabbit anti-phospho-histone H3 (Ser10) (Millipore, Cat. no. 06–570, 1:1000), rabbit anti-phospho-histone H3 (mix 1–1 of Cell signaling, Cat. no. 9701 (Ser10) and Cat. no. 9713 (Ser28), 1:100), guinea pig anti-Ebd1 ([53], 1:1000), Alexa Fluor 555 phalloidin (Thermo Fisher Scientific, Cat. no. A34055, 1:500) and DAPI (Sigma, 1:400). Secondary antibodies were goat or donkey Alexa Fluor 488 or 555 conjugates (Thermo Fisher Scientific, 1:400), and goat or donkey Cy5 conjugates (Thermo Fisher Scientific/Jackson Immunochemicals, 1:200).
Adult fly intestines were dissected in PBS and fixed in 4% paraformaldehyde in PBS for 45 minutes at room temperature. For Delta staining, intestines were fixed in 8% paraformaldehyde, 200mM Na cacodylate, 100mM sucrose, 40 mM KOAc, 10 mM NaOAc, and 10mM EGTA for 20 minutes at room temperature [130]. Tissues were then washed with PBS, 0.1% Triton X-100, followed by incubation in PBS, 0.1% Tween-20 and 10% BSA for 1 hour at room temperature. The samples were then incubated with primary antibodies at 4°C overnight in PBS, 0.5% Triton X-100. Samples were stained with secondary antibodies for 2 hours at room temperature. Specimens were stained with DAPI (2μg/ml) and mounted in Prolong Gold (Invitrogen). To assess the gut layers, specific mounting set-ups were performed according to a protocol from the Micchelli lab [74]. Larval guts were immunostained in the same way except that wandering third instar larvae were fixed in 4% paraformaldehyde in PBS for only 20 minutes and were incubated with primary antibodies in PBS, 0.1% Triton X-100. Fluorescent images were obtained on a Nikon A1RSi confocal microscope except those in (S6 Fig), which were captured on a Zeiss LSM 780 confocal microscope. Images were processed using Adobe Photoshop software.
Mitotic clones were generated using the MARCM system [99]. Developmental clones were induced in third instar larvae by a single 2–3 hour heat shock at 37°C and examined 1 to 2 days after eclosion. To generate clones in the adult gut, flies were heat shocked for 30 minutes in a 37°C water bath four days post-eclosion. After heat shock, flies were maintained at 25°C for five days before analysis. For quantification of clone size, flies were maintained at 25°C for 14 days post heat shock and only clones in the posterior midguts were included in the analysis.
To induce temporal knockdown in ISCs, control or specific RNAi lines were crossed to the esgts (esg-Gal4 tubGal80ts UAS-GFP/CyO) driver. For knock down during development, crosses were set up at 18°C and shifted to 29°C 6 days later (during the second instar larval stage). Progeny of desired genotypes were dissected 2–3 days after eclosion. For knock down during adulthood, crosses were maintained at 18°C until eclosion, and progeny of desired genotypes were then shifted to 29°C for 14 days before analysis.
To induce temporal knockdown in ECs, RNAi experiments were performed using Myo1A-Gal4 with the temperature-sensitive Gal4 repressor, Gal80ts. Crosses were maintained at 22°C and on the day of eclosion, progeny of desired genotypes were shifted to the restrictive temperature (29°C) for 7 days.
To induce temporal knockdown using the “escargot flip out” system (esgts F/O: esg-Gal4 tub-Gal80ts UAS-GFP; UAS-flp Act>CD2>Gal4), crosses were maintained at 18°C and 3-5-day-old progeny of the desired genotypes were shifted to 29°C. The marked esg+ cell lineages were analyzed 14 days later.
For quantification of ISCs, flies were stained with anti-Delta (Dl) and anti-Prospero antibodies. Images of the midgut R5a region [89] were obtained with a 60x lens and the total number of Dl-positive cells in a field of 0.051mm2 was counted. Similarly, progenitor cells inside a defined field were quantified by counting esg>GFP, esg-lacZ, or small cells with strong Arm staining and absence of Prospero staining. EBs inside a defined field were quantified by counting GBE-Su(H)-lacZ positive cells. For quantification of pH3-positive cells, the total number of pH3-positive cells in the posterior midgut of the indicated genotypes was counted. For quantification of pH3-positive cells near MARCM clones, the number of pH3-positive cells in a field of 4000 µm2 around the clone was counted.
For quantification of intestinal structure, 40x Z-stack confocal images of posterior midguts of desired genotypes were acquired. The maximum number of epithelial layers and maximum epithelial height were measured using NIS-elements software (Nikon).
Quantification of GFP intensity in “esgts flip out” guts was performed by measuring overall GFP intensity within two areas per posterior midgut and normalizing that value by the total number of cells in the field.
Quantification of fz3-RFP intensity was performed with ImageJ (NIH). For each clone examined (total of 57 clones derived from more than 20 guts), intensities of three separate areas within the clone and areas of identical size outside the clone were measured. The average intensities of the three areas were compared to the average intensities of their control counterparts.
Quantification of stat-GFP intensity was performed using Imaris software (Bitplane). Stat-GFP-positive cells within a field (40μm × 40μm) surrounding an ebd1 mutant clone, or in an equal field at least 50μm away from the ebd1 mutant clone, were identified and their intensity was measured.
All statistical tests were performed using Prism (GraphPad Software, USA).
Whole midguts from Canton S (control) or Apc1Q8 7-day-old females were dissected in nuclease-free PBS and processed for transcriptomic analysis. Total RNA from 30 adult midguts per sample was extracted using Trizol following manufacturer’s instructions. Triplicate samples of each of the genotypes were prepared. RNA was sent to the Microarray facility at the University of Manchester where it was used to hybridize Drosophila Affymetrix 2.0 chips. The CEL files were subject to RMA normalization and log2 transformation followed by differential gene expression analysis by the Beatson Institute Bioinformatics department. Microarray data were deposited in the GEO database (GSE99071). GO term analysis was performed via “PANTHER GO-slim” [131].
Whole midguts from 15–20 flies of desired genotypes were dissected in PBS and total RNA was extracted using the RNA miniprep kit (Zymo research). The RNA was subsequently treated with RQ1 DNase (Promega). 1 μg of RNA was reverse transcribed using pdT15 primers and M-MLV reverse transcriptase (Invitrogen). Expression level of candidate genes was quantified using the StepOne Real-time PCR system (Life Technologies) with SYBR green (Life Technologies/Biorad). RNA extraction of three biologically independent samples was performed. Two independent repeats are presented in Fig 3 and S10 Fig, respectively, as mean fold change relative to the internal control (rpl32), with standard deviation. The primers used are listed in Table 1.
|
10.1371/journal.pgen.1004454 | Unraveling Genetic Modifiers in the Gria4 Mouse Model of Absence Epilepsy | Absence epilepsy (AE) is a common type of genetic generalized epilepsy (GGE), particularly in children. AE and GGE are complex genetic diseases with few causal variants identified to date. Gria4 deficient mice provide a model of AE, one for which the common laboratory inbred strain C3H/HeJ (HeJ) harbors a natural IAP retrotransposon insertion in Gria4 that reduces its expression 8-fold. Between C3H and non-seizing strains such as C57BL/6, genetic modifiers alter disease severity. Even C3H substrains have surprising variation in the duration and incidence of spike-wave discharges (SWD), the characteristic electroencephalographic feature of absence seizures. Here we discovered extensive IAP retrotransposition in the C3H substrain, and identified a HeJ-private IAP in the Pcnxl2 gene, which encodes a putative multi-transmembrane protein of unknown function, resulting in decreased expression. By creating new Pcnxl2 frameshift alleles using TALEN mutagenesis, we show that Pcnxl2 deficiency is responsible for mitigating the seizure phenotype – making Pcnxl2 the first known modifier gene for absence seizures in any species. This finding gave us a handle on genetic complexity between strains, directing us to use another C3H substrain to map additional modifiers including validation of a Chr 15 locus that profoundly affects the severity of SWD episodes. Together these new findings expand our knowledge of how natural variation modulates seizures, and highlights the feasibility of characterizing and validating modifiers in mouse strains and substrains in the post-genome sequence era.
| Absence seizures - also known as “petit-mal” - define a common form of epilepsy most prevalent in children, but also seen at other ages, and in related diseases such as juvenile myoclonic epilepsy. Absence seizures cause brief periods of unconsciousness, and are accompanied by characteristic abnormal brain waves called “spike-wave discharges” (SWD) due to their appearance in the electroencephalogram (EEG). Although few genes are known for human absence seizures, perhaps because the underlying genetics are complex, several laboratory rodent models exist, including one caused by mutation of a gene called Gria4. While studying Gria4, we noticed that a mouse strain called C3H can suppress or enhance the frequency and severity of Gria4-associated SWD in a perplexing manner; such effects are generally attributed to “modifier” genes. Here we identify a novel modifier – called “pecanex-like 2”, or Pcnxl2 for short – that reduces the severity of SWD in the C3H substrain in which the Gria4 mutation originally arose. This finding directed us to use of related substrains to locate additional modifiers, one of which has an even more profound effect on SWD episodes. Modifier genes, nature's way of controlling seizure severity, are promising targets for better understanding seizure mechanisms and potential new therapies in the future.
| Laboratory mouse strains are well known to vary in their susceptibility to convulsive seizures, including acute experimentally induced seizures [1], [2], [3], [4], [5], [6], and spontaneous seizures induced by genetic mutation [7], [8], [9], [10], [11], [12], [13]. Most of the known strain effects have been for convulsive seizures, where motor manifestations are obvious, including one modifier identified to date (Kcnv2 as a modifier of Scn2a1Q54 induced convulsive seizures [13]. There has been less attention to such effects in non-convulsive phenotypes, such as absence epilepsy, where the seizures lack a convulsive element. Nevertheless, strain differences were noted for three different genes that cause absence seizures when mutated – Scn8a [14]), Gabrg2 [15] and Gria4 [16], and for at least two the C3H strain generally worsens the absence seizure phenotype, compared with the relatively protective strain C57BL/6J (B6J).
For absence seizures caused by Gria4 mutation, the C3H strain has been a paradox. Each of three very closely-related C3H/He substrains has a similar level of spontaneous spike-wave discharges (SWD) - the distinctive electroencephalographic hallmark of absence seizures in human and in animal models - but only one substrain, C3H/HeJ (HeJ), carries a Gria4 mutation. This mutation is caused by an intracisternal A-particle retrotransposon (IAP) insertion in Gria4 [16], [17], [18]. Although genetic and later functional analysis proved that Gria4 is the cause of these seizures in HeJ [19], two other C3H/He substrains that lack this mutation still have appreciable SWD. At least one of these strains was shown to have a polygenic etiology, with no indication of any effect from proximal Chr 9 where Gria4 resides [18]. The further surprise was when HeJ mice were outcrossed to other strains - even to another C3H substrain, C3HeB/FeJ (FeJ), that does not have frequent SWD - about half of the next generation Gria4 deficient progeny had significantly higher SWD incidence than HeJ itself [16]. Together these results suggested the model whereby the HeJ substrain has both the initial seizure-causing Gria4 mutation, and also a mitigating or protective mutation, without which the seizures would be much more severe. The model also suggests that C3H strains in general have a high baseline susceptibility to absence seizures, compared with strains such as C57BL/6.
IAP insertions like the element in Gria4 are known to cause deleterious mutant phenotypes by reducing RNA expression of the target gene, and the vast majority of spontaneous IAP insertion mutations in mice have arisen in the C3H strain family (reviewed in [20]). While searching for genetic markers to distinguish C3H substrains for analysis of the putative HeJ suppressor, we discovered extensive IAP retrotransposition among C3H substrains. One of the HeJ substrain-private IAP insertions resides in the same chromosomal region as an epistatic modifier of Gria4 absence seizures mapped to Chr 8 [17]. Here we report that this Chr 8 IAP is inserted in the previously unstudied Pcnxl2 (pecanex-like 2) gene, affecting its RNA expression. We further use TALEN mutagenesis to create new Pcnxl2 alleles, confirming that Pcnxl2 loss of expression is responsible for mitigating Gria4 seizures of HeJ mice, accounting for the substrain difference. We also mapped modifiers that differ between C3H and B6J, and show that one – G4swdm1 – has a profound effect on seizure severity. The identification of the first absence seizure modifier Pcnxl2 provides significant traction to the complex genetics of absence seizures in the C3H strain family and possible new mechanisms for mitigating disease.
Prior studies of Gria4-deficient mice showed significant genetic background influence on the incidence and duration of spike-wave discharges (SWD) between C57BL/6J (B6J) and C3H strains [16], [17], whether the natural Gria4spkw1 allele (abbreviated hereafter as Gria4IAP) or the engineered knockout Gria4tm1Dgen (Gria4KO). While mapping the main phenotype to Gria4 in a backcross between C3H/HeJ (HeJ) and B6J, about half of the progeny were noted to have significantly more SWD than HeJ itself, and an epistatic modifier putatively due to this effect was mapped to distal Chr 8 [16], [17]. Interestingly, a similar phenomenon from crosses between HeJ and C3HeB/FeJ (FeJ) suggested this was due to a substrain difference [16], [17]. To confirm this, we backcrossed Gria4IAP allele from HeJ to FeJ and examined SWD in this congenic pair and compared to B6J-Gria4KO for reference. FeJ-Gria4IAP had significantly more frequent and longer SWD than HeJ or B6J-Gria4KO (Table 1). These results indicate that HeJ has a suppressor allele(s) not present in FeJ. It also suggest that further modifiers differ between FeJ and B6J strains.
The HeJ suppressor is a challenge to pursue by classical recombination mapping because of the paucity of genetic markers between FeJ and HeJ substrains, which split into separate lines around 1950 [21]. However, C3H mice are known to have frequent spontaneous germline IAP retrotransposition, with de novo insertions accounting for a significant fraction of the spontaneous mutation load in C3H (see review by [20]). Moreover, almost all de novo IAP insertions, including Gria4IAP, are of a particular IAP subtype - IAP-1Δ1 - containing a characteristic in-frame 1.2 kb deletion of the gag-pol fusio gene [22]. Sequence similarity search of the B6J mouse genome assembly with an oligonucleotide sequence spanning the IAP-1Δ1 deletion detected approximately 200 independent genomic sites, whereas the same region from full-length IAP detected over 700 sites (data not shown), suggesting the feasibility of direct hybridization approaches to identify substrain-specific de novo IAP-1Δ1 insertions.
To determine whether C3H substrains vary significantly in IAP-1Δ1 content, we designed an oligonucleotide specific for the IAP-1Δ1 1.2 kb common deletion to examine proviral-host DNA junction fragments by direct dried gel hybridization and to subclone them by inverse PCR (Fig. 1A). Although it is unlikely that any visible ‘band’ in the direct hybridization represents a single IAP insertion, from the differential pattern– best observed in the 2 kb range - there are ample variation between four substrains examined; with HeJ appearing to have the highest load including at least 20 HeJ-specific bands evident (Fig. 1B). Inverse-PCR was used to identify and clone select IAP-1Δ1 – host junction fragments from FeJ and HeJ. A representative experiment to identify lower molecular weight junction fragments revealed a banding pattern reminiscent of the gel hybridization, and recapitulates the corresponding paucity of bands in this region in the FeJ strain (Fig. 1C). From this and other experiments, bands were excised, cloned and junction fragments sequenced, and PCR assays were developed ultimately leading to identification of at least 26 insertion-host junction fragments that are present in the HeJ substrain and absent from FeJ (Table 2). 9 are intergenic, but the majority are in introns of known or unclassified genes (Table 2). Most HeJ insertions we cloned were ultimately identified in the recent retrotransposon-mining of the whole genome sequence of HeJ and other mouse strains from the Sanger Mouse Genomes Project [23], although substrain specificity was not ascertained in that study.
One HeJ-specific IAP insertion resides at 128.3 Mb on Chr 8, in the same general region of the previously mapped epistatic modifier of Gria4 SWD [17]. This insertion was not seen in any of the 17 inbred mouse strains for which genome sequence is available [23], and the PCR assays used to identify it (Table 2) determined that it was also absent from C3HeB/FeJ (FeJ), C3H/HeOuJ and C3H/HeSnJ substrains (data not shown). This IAP is integrated in the 5′-LTR-3-′LTR orientation in intron 19 of Pcnxl2, the gene that encodes pecanex-like 2. Pcnxl2 is one of three mammalian paralogs of Drosophila melanogaster pecanex, a neurogenic gene about which little is known, but was recently suggested to be part of the notch signaling pathway in the endoplasmic reticulum [24], [25]. In adult mouse brain, Pcnxl2 is expressed highest in the hippocampal pyramidal cell layer, it is also expressed prominently in the area of cerebral cortex corresponding to layer V, and sparsely in the reticular thalamic nucleus (Allen Brain Atlas: see http://mouse.brain-map.org/experiment/show/70239051). As the latter two regions are key in abnormal cortico-thalamic oscillations associated with absence seizures generally and specifically in Gria4 mutants [19], Pcnxl2 is a suitable candidate for the suppressor.
To determine whether the IAP insertion affects Pcnxl2 gene expression, we initially examined Pcnxl2 transcript expression in brain from public datasets, comparing HeJ to other inbred strains. From inspecting publically available mouse strain hippocampal microarray (http://www.genenetwork.org) and in the Sanger Center's whole-brain RNAseq (ftp://ftp-mouse.sanger.ac.uk/current_rna_bams), we noted an HeJ-specific drop in relative expression or abundance of exons downstream of exon 19 (data not shown). To examine this between HeJ and FeJ substrains, by using quantitative RT-PCR we confirmed that Pcnxl2 expression across intron 19 is indeed lower in HeJ compared to FeJ (Table 3, compare sample IDs 1–2).
To confirm that Pcnxl2 encodes the Gria4 SWD suppressor, we used TALEN mutagenesis [26] in the FeJ-Gria4IAP congenic strain to create Pcnxl2 frameshift mutations (Fig. 2A). Two Pcnxl2 exons were targeted separately: exon 16, roughly in the middle of the gene, and exon 29 nearer the 3′ end, the latter containing the so-called “pecanex” domain – a conserved domain of unknown function shared among pecanex paralogues. Three in 106 liveborn mice contained germline mutations and each created a frameshift allele that resulted in a premature translational stop codon (Figure 2).
To test the effect on SWD, EEG was examined in FeJ-Gria4IAP homozygotes carrying each of the three Pcnxl2 mutations (A−11, A+1 and B−2, hereafter referred to as FS1 – frameshift 1, FS2 and FS3, respectively). Pcnxl2FS1 and Pcnxl2FS2 mutations significantly lowered SWD incidence (Fig. 2B) and length (Fig. 2C), showing an additive effect across genotypes. The homozygotes were also slightly more suppressed than HeJ with its natural Pcnxl2IAP allele. An examination of Pcnxl2 transcripts by qPCR showed decreased (between 2 and 3.5-fold; Table 3, compare sample ID's 3–4, 5–6, 7–8) but not eliminated expression in each homozygous genotype; this is expected as frameshift mutations in middle exons often cause nonsense-mediated decay. However, the Pcnxl2FS3 allele did not decrease SWD as effectively as the others (Fig. 2B, C). Since commercial antisera to PCNXL2 are not effective in mouse brain (data not shown), we cannot know whether this is because a partial or truncated protein is still made and has some residual function, or whether nonsense-mediated decay was less complete. Regardless, the effect of at least two independent new Pcnxl2 alleles on SWD, combined with transcript reduction shows that Pcnxl2 encodes the suppressor of Gria4 SWD.
To determine whether Pcnxl2 deficiency can affect SWD in other absence seizure mouse models, we examined double mutant genotypes between FeJ-Pcnxl2FS1 and either Scn8a8J [14], or Gabrg2tm1Spet [15], each of which encodes a dominant, SWD-causing missense mutation in the respective ion channel gene. Pcnxl2 deficiency had no significant effect on SWD on either of these mutants (Fig. 2C), suggesting that Pcnxl2 may be specific for mechanisms that involve Gria4. However, if Gria4-specific, the mechanism appears not to be on IAP mutagenesis itself, because there is no increase in Gria4 RNA expression in Gria4IAP, Pcnxl2FS1 double mutants (Table 3, compare sample IDs 9–10). We also note that the Pcnxl2 paralogue, Pcnx, is expressed widely in adult mouse brain, including likely overlap with Pcnxl2 (http://mouse.brain-map.org/experiment/show/73818801), suggesting the possibility of compensation. However, the Pcnx transcript is not altered in any of the three new Pcnxl2 frameshift alleles (Table 3, compare sample IDs 13–14, 15–16, 17–18), suggesting that if there is any compensation by this overlapping pecanex gene, it is not at the transcript level.
The fact that Gria4 mutant associated SWD are more pronounced when placed on the C3HeB/FeJ (FeJ) strain compared with B6J suggests additional modifiers, i.e. separate from Pcnxl2 whose genotype does not differ between B6J and FeJ (Table 1). To pursue these, we backcrossed (FeJ-Gria4IAP×B6J-Gria4KO)F1 animals to B6J- Gria4KO, creating 89 N2 mice segregating B6J vs. FeJ strain variants on a Gria4 mutant background (either homozygous Gria4KO or compound heterozygous Gria4KO/IAP) and scored their EEG for SWD incidence and length. The broad, almost continuous distribution of each raw SWD measure in this population suggests multiple factors (Fig. 3A, 3B insets). A genome-wide scan of the two SWD measures and their principal components revealed two highly significant regions (Fig. 3A, 3B): proximal Chr 8 (peak LOD, 3.4) and mid Chr 15 (peak LOD, 2.9). The Chr 15 region was significant for SWD length but negligible for incidence, and vice versa for Chr 8 (compare Fig. 3A with 3B), suggesting each region is primarily responsible for one of the two measures, at least in this cross. A region in mid Chr 3 was significant for SWD length only, and several suggestive peaks were observed for at least one SWD measure on Chrs 1, 2 and 5. A pairwise scan was done to look for pairwise epistatic interactions, but no significant interactions were observed (data not shown).
Because of the increased SWD length, we focused on validating the Chr 15 locus, now termed G4swdm1, for Gria4 spike-wave discharges modifier 1. We made a congenic strain in which the middle of Chr 15 was selectively bred from FeJ into B6J-Gria4KO, and N9F1 and N10F1 intercross mice were generated and tested. Highly significant effects were observed this time for both SWD length and incidence across the introgressed interval (Figure 4A, 4B). While G4swdm1 was initially detected as a dominant allele, the congenic intercross reveals additivity: FeJ homozygotes experience significantly longer and more frequent SWD (Figure 4C). Although SWD length was still the leading phenotype, in the B6J background the effect of G4swdm1 on SWD incidence was stronger than in the N2 population, likely reflecting additional genetic complexity. Indeed, several congenic individuals had SWD 15 s long (Fig. 4D); further, when the homozygous interval was placed together with the Gria4 mutant genotype on a (B6J×FeJ)F1 hybrid background, several very striking SWD, exceeding 40 s were observed (Figure 4D). The estimated 95% confidence interval (CI) for G4swdm1 is large, with the narrowed SWD length interval covering 27.2 Mb including 191 protein coding and 75 small RNA or unclassified genes (File S1), although the “bumpy” likelihood curve suggests the possibility of multiple modifiers. Among these genes are 27 that have 56 non-synonymous coding or potential splice altering SNPs or indels, between published C3H/HeJ and B6J genomes; further, comparison of FeJ exome sequence to HeJ revealed no similar coding variants (File S1). To gain additional evidence for candidacy, we examined gene expression by RNAseq, using somatosensory cortex and thalamus as tissue source, comparing parent strains B6J and FeJ to each other. In the SWD length 95% CI, of 93 mRNAs expressed, 15 had abundance differences (e.g. q<0.1; File S1). Most were modest (<5% change) but for Ly6a, one of several members of a cell membrane protein-encoding gene family, FeJ had a 21% transcript reduction in thalamus and 32% in cortex. Although Ly6a is best known for expression in lymphocytes, it is also expressed in brain and at least one Ly6a knockout allele has prenatal lethality [27]. Further refining the G4swdm1 critical interval, and more extensive testing of existing mutants or creation of others by mutagenesis is required before the correct G4swdm1 candidate(s) can be identified.
Here we unravel the genetic complexity of Gria4-deficiency absence seizure susceptibility in C3H mice. First, we show genetically that Pcnxl2 deficiency accounts for the unusual substrain difference in spike-wave discharges (SWD) in Gria4 mutants on C3H/HeJ (HeJ) compared to C3HeB/FeJ (FeJ). We determined that two new frameshift alleles, generated directly in FeJ-Gria4IAP by TALEN mutagenesis, confer SWD mitigation even more than the natural, presumably hypomorphic Pcnxl2 IAP insertion allele of HeJ. This insertion is of the same IAP-1Δ1 subtype that caused the original Gria4IAP mutation. Given increased incidence and severity of SWD in FeJ-Gria4IAP compared to HeJ-Gria4IAP, and the absence of the Pcnxl2 IAP from two other C3H/He substrains, we think it is reasonable to speculate that Gria4IAP conferred a selective disadvantage to the progenitors of HeJ mice, one that was later diminished upon fixation of the Pcnxl2IAP insertion. These findings also suggest that a number of apparent IAP element differences between C3H substrains, nominally 20% of the IAP-1Δ1 pool, remain a potentially powerful source of genome plasticity, which naturally would not be restricted to neurological phenotypes. Although it was suggested previously that most functional IAP insertions in mouse strains were lost because of deleterious effects, [23], clearly some remain functional – perhaps in the case of Pcnxl2 due to selective advantage such as the one we hypothesize.
The further phenotypic difference between FeJ and B6J strains highlights the existence of additional genetic influences on Gria4 associated absence seizures. To begin to dissect these, genome scans revealed several potential modifiers, affecting one or the other SWD measure. By using a congenic strain we validated one modifier locus – G4swdm1 on Chr 15 and observed its striking effect on SWD length; in F1 hybrid mutants, for example, we observed several SWD episodes lasting over 40 seconds. It is difficult to imagine that such frequent and contiguous states of neural hypersynchrony do not have a broader impact on behavior. Despite its clear effect, G4swdm1 accounts for only some of the phenotype difference between strains – perhaps 30% by comparing SWD length of B6J to F1 hybrid (e.g. as illustrated in Figure 4C to 4D). But when multiple interactions are likely, as in complex traits, any such estimates of effect are overly simplistic. Approaches to map such modifiers using conventional quantitative trait locus, especially in small crosses, are limited to loci that show significant main effects despite interactions, or to simple, pairwise interactions when they are quite strong. Novel computational approaches such as CAPE, which incorporates relationships between phenotypic features directly into the gene interaction model [28], may be required to parse more complex interactions in conventional cross designs.
Pcnxl2 is the first absence seizure modifier gene to be identified in any species, and as such it represents the first of what is likely to be many genetic interactions beneath the complexity of absence seizures. The predicted peptide structure of PCNXL2 is similar to that of other pecanex orthologues, with 8 putative transmembrane domains followed by a conserved so-called pecanex domain. But no known primary or predicted secondary structures have been identified that would predict further function. D. melanogaster pecanex (pcnx) localizes to the endoplasmic reticulum (ER) and functional genetic studies utilizing the protein unfolded response as a readout, suggest that it shows maternal inheritance and has a role in notch signaling [25]. From conservation among pecanex family members we might expect that Pcnxl2 is also expressed in the ER. Although much further discussion of function is merely speculation, if it is an ER protein one tempting possibility is a role in trafficking or in posttranslational modification of synaptic receptors such as Gria4 or other compensating ion channel receptors. The prominent Pcnxl2 expression in layer V of the cerebral cortex opens the door to the possibility that it is involved in mediating excitatory output from layer V pyramidal neurons to the reticular thalamus and thalamus. Whether any such function is the result of an acute Pcnxl2 role, or instead a role in circuit development, will require further studies, for example, the creation of a conditional allele to be induced at different ages.
Laboratory mouse strains are well known to vary in susceptibility to experimentally-induced partial or generalized tonic-clonic seizures and there are several well-characterized strain differences modifying the penetrance or severity of so-called monogenic seizure mutations (as discussed earlier) although only one such modifier gene has been identified to date, and interestingly this has also be implicated in human epilepsy [13]. The rate of human gene discovery is rapidly accelerating due to efficient high-throughput exome sequencing [29], but the new progress so far is for syndromic pediatric encephalopathies such as Lennox-Gastaut syndrome. The pace remains much slower for genetic generalized epilepsies, including absence epilepsy. With new mutagenesis technologies such as TALEN and CRISPR to more readily validate candidate genes in large intervals such as those defined in modifier or QTL mapping, the pursuit of modifiers still holds promise for unbiased discovery of new genes, pathways and future novel therapies for idiopathic disease.
All mice were housed and procedures performed with approval of Institutional Animal Care and Use Committee (IACUC). All mice were obtained from The Jackson Laboratory, maintained in a room with a 14 h hour light on/10 h light off cycle, and given free access to LabChow meal and water.
C3H/HeJ (HeJ), C3HeB/FeJ (FeJ), C3H/HeOuJ, C3H/HeSnJ and C57BL/6J (B6J) inbred mouse strains were obtained from The Jackson Laboratory production colonies and subsequently maintained by sib-matings. C57BL/6J.129-Gria4KO congenic knockout mice were originally obtained from Deltagen, Inc, as previously described [16]. FeJ.HeJ-Gria4IAP congenic mice were generated by backcrossing the Gria4spkw1 mutation from its original strain, HeJ, to FeJ for 14 generations, bred to homozygosity and maintained by sib-matings. FeJ-Gabrg2tmSpet congenic mice were generated by backcrossing the Gabrg2tmSpet (also known as Gabrg2R43Q) knockout mutation from B6J.129-Gabrg2tmSpet congenic mice (originally obtained from Bionomics, LTD) successively for at least 20 generations to FeJ and maintained in the same way. FeJ.B6J-Scn8a8J congenic mice were generated by successive backcrossing of Scn8a8J (also known as Scn8aV929F) for at least 15 generations to FeJ and maintained in the same way, as described recently [30]. The 89 backcross mice used for genome-wide mapping were generated by mating (B6J.129-Gria4KO/KO×FeJ.HeJ-Gria4IAP/IAP)F1 hybrids to B6J.129-Gria4KO/KO congenic mice. The 84 B6J-FeJ-G4swdm1 congenic intercross mice were created by successively backcrossing the FeJ alleles for genetic markers in the critical interval on Chr15 from an N2 to the B6J.129 strain while also selecting for Gria4KO homozygotes, and then intercrossing at generation N9 or N10. The generation of Pcnxl2 mutants is described below.
We contracted Transposagen Biopharmaceuticals, Inc. to design and construct plasmids for TALENs specific to exon 16 (TALEN A) and exon 29 (TALEN B) of Pcnxl2. Target sequences were as follows; TGAGCCGGCAGAGCAGTG and GTGAGTAGCTGTCCTGTA for TALEN A and TATTTGCTGACATGGAC and TTGTTCCAGCCATCCGAA for TALEN B. TALEN plasmids were linearized by PmeI endonuclease digestion. One microgram of linearized plasmid was used as a template for in vitro transcription using AmpliCap-Max T7 High Yield Message Maker Kit (CELLSCRIPT) according to the manufactures instruction. A poly(A) tail was added to the synthesized RNA with the A-Plus Poly(A) Polymerase Tailing Kit (CELLSCRIPT) according to the manufactures instruction. The poly(A) tailed capped RNA was purified by ammonium acetate precipitation, resuspended in RNase free water and the concentration determined by spectrophotometry. TALEN mRNA was diluted to 10 ng/υl in RNase free 1X TE (10 mM Tris-HCl, 1 mM EDTA, pH 7.5) immediately before microinjection into embryos obtained from superovulated FeJ.HeJ-Gria4IAP/IAP congenic mice. The genomic DNA made from the tail tip of 63 TALENA mutant founders was amplified with primers aFXTN (5′-CATCGTGGCTGTCGTAATTC -3′) and aRXTN (5′-CATAGCGTGGGAGAGAAAGA-3′). The product was purified and sequenced using primer aF2XTN (5′-GCACACACCACTCATTCATC-3′). The genomic DNA made from the tail tip of 43TALENB mutant founders was amplified with primers bFXTN (5′- GCTTTGTAATGTGGGTTCTG-3′) and bRXTN (5′- GGTTCTCTACTTCAGCCTATG-3′). The product was purified and sequenced using primer bF2XTN (5′-GAACTCGGGATCCATGTTTG -3′).
For the genome scan, genomic DNA was prepared from tail tips as previously described and sent to Kbioscience (currently LGC Genomics, LLC. Beverly, MA), using a custom single nucleotide polymorphisms (SNP) panel comprised of 187 roughly evenly spaced SNPs. Prior to interval mapping, the raw traits SWD incidence and SWD length were rank-ordered and normal quantile transformed, and from which principal components were derived using JMP software (SAS Institute); once obtained, both principal components were also rank- and normal-quantile transformed, also using tools in JMP. The computer program J/qtl [31] was then used for genome-wide interval mapping of the initial backcross, and for Chr 15 interval mapping of congenic intercross mice. To control for a modest effect (p<0.04) of Gria4KO/KO homozygous null vs Gria4KO/IAP compound heterozygous null/hypomorph genotypes segregating in the N2 cross, a marker linked to Gria4 on Chr 9 was used as a covariate. Sex-averaged genetic map coordinates for SNP markers were obtained from the Mouse Genome Informatics database at The Jackson Laboratory (http://informatics.jax.org). For interval mapping, the multiple imputation model was used and permutation shuffling employed to determine genome-wide significance thresholds.
C3H/HeJ-Pcnxl2IAP (IAP insertion in intron 19 of C3H/HeJ) was genotyped in standard PCR conditions and agarose gel electrophoresis using one assay for the wild-type allele (primers c8-128.3F 5′-AGCGATGAGGACTGTGGTTT-3′; c8-128.3R 5′-CGAGCCCTTCAGCTACTCAC-3′) and a second assay for the insertion allele (primers c8-128.3F 5′-AGCGATGAGGACTGTGGTTT-3′; IAPLTR5′R 5′-GGCTCATGCGCAGATTATTT-3′) giving a 364 bp product from the insertion allele and 432 bp product from the endogenous allele. The TALEN A+1 or frameshift 1 (FS2) allele was genotyped in standard PCR conditions at an annealing temperature of 64°C and agarose gel electrophoresis using one assay for the wild-type allele (primers Pcnxl2A1WF2 5′-GCAGAGCAGTGATCCTTCAG -3′; Pcnxl2A1WR2 5′-CCATAGCGTGGGAGAGAAAGAA -3′) and a second assay for the mutant allele (primers Pcnxl2A1MF2 5′-GCAGAGCAGTGATCCTTCAC-3′; Pcnxl2A1WR2 5′-CCATAGCGTGGGAGAGAAAGAA -3′) giving a 311 bp product for both alleles. TALEN B-2 or FS3 allele was genotyped in standard PCR conditions with an annealing temperature of 67°C and agarose gel electrophoresis using one assay for the wild-type allele (primers Pcnxl2BwtF 5′-TAGATGCTGGTAGGAGTGAAGA -3′; Pcnxl2BwtR 5′-GGCTGGAACAACAACTTTGTGT-3′) and a second assay for the mutant allele (primers Pcnxl2BwtF 5′-TAGATGCTGGTAGGAGTGAAGA -3′; Pcnxl2BmutR 5′-GGCTGGAACAACAACTTTGTAG-3′) giving a 228 bp product from the mutagenized allele and a 227 bp product from the wildtype allele. TALEN A-11 or FS1 allele was genotyped in standard PCR conditions at an annealing temperature of 55°C and agarose gel electrophoresis using primers Pcnxl2AF2 (5′- CACTGTTCTCGGCCTTCTG-3′) and Pcnxl2AR (5′-AGACATGTGGACATGCGTTTA-3′) giving a 123 bp product from the mutagenized allele and a 134 bp product from the endogenous allele.
For direct detection of IAP-1Δ1 insertions, dried gel hybridization was done essentially as previously described [32] except using an oligonucleotide probe that spans the 1.2 gag-pol common deletion of IAP-1Δ1 elements [22]. Briefly, 8 µg of high quality mouse genomic DNA (obtained from The Jackson Laboratory DNA Resource) was digested with restriction enzyme Bgl II, electrophoresed overnight on 0.8% Tris-borate EDTA agarose gels, EtBr-stained and imaged, denatured in NaOH, neutralized and dried for several hours on a flat slab gel dryer with minimal vacuum. 5′-32P radiolabeled 31-nt oligonucleotide probe (IAPd1oligo1-R; 5′ ATACCTCTTATCAGGTTCAGCAGAATAAGCTC-3′) was hybridized overnight, the dried gel was washed and imaged on x-ray film. For cloning of IAP-1Δ1-host junction fragments by inverse PCR: 1 µg of genomic DNA was digested with BglII, diluted, then 10 ng was ligated for 2 hrs at room-temperature using T4 DNA ligase, heat-inactivated, then circular product was amplified in a polymerase chain reaction (PCR) using an oligonucleotide that spanned the IAP-1Δ1 deletion (IAPd1oligo2F, 5′- GAGCTTATTCTGCTGAACCTGATA-3′) paired with an oligonucleotide specific for the IAP LTR (IAPLTR5′, 5′- GGCTCATGCGCAGATTATTT-′3). Amplified fragments were visualized by agarose gel electrophoresis (e.g. Figure 1C), excised and extracted from agarose using Qiagen minicolumns, cloned into Bluescript plasmid by T/A cloning (Invitrogen), transformed into a suitable E. coli K12 host, miniprepped and subjected to Sanger sequencing using T7 and Sp6 vector primers.
Total RNA was prepared from the whole brain of adult HeJ, FeJ, and FeJ-Gria4IAP and the hippocampus of adult FeJ- Gria4IAP either wildtype or carrying Pcnxl2 TALEN mutations with Trizol (Invitrogen) and treated with DNase I (Promega) under the manufacturer's suggested conditions. RNA (2 µg) was reverse transcribed with AMV reverse transcriptase (Promega). The cDNA was diluted 20-fold, and 2 µl was added to DyNAmo HS SYBR Green qPCR master mix (Thermo Scientific) with pairs of the following primers; beta-actinF (5′- ATGCTCCCCGGGCTGTAT-3′) and beta-actinR (5′- CATAGGAGTCCTTCTGACCCATTC-3′), Pcnxl2exon19F (5′-GGATCTCACATCCTGTGCTC -3′) and Pcnxl2exon20R (5′-CCACACGTAGAGTCTCTCAAAC -3′), Gria4exon15F (5′- GGTGGCTTTGATAGAGTTCTGTTACA-3′) and Gria4exon16R (5′- TCTTATGGCTTCGGAAAAAGTCA -3′), Pcnxexon35F (5′-GAACAGCTGGAAAGACTGGA-3′) and Pcnxexon36R (5′-CGATGTGGGACCTTGTACTT-3′). The PCR reactions were analyzed on an Applied Biosystems 7500 Real-Time PCR System. The PCR amplifications from three mice of each strain and/or genotype were run in triplicate. Amplification of the correct size products was confirmed by agarose gel electrophoresis. The ΔΔΧt method was adopted for the calculation of relative transcript levels.
Somatosensory cortex or thalamus was dissected from wildtype or Scn8a8J/+ B6J and FeJ adult male mice in triplicate, and prepared for high-throughput sequencing on the Illumina HiSeq 2000. The Jackson Laboratory Gene Expression Service prepared mRNA sequencing libraries using the Illumina TruSeq methodology. Tissue was placed in RNALater (Qiagen, Inc, MD), RNA was extracted using TRIzol (Invitrogen, CA). For mRNA-Seq, mRNA was purified from total RNA using biotin tagged poly dT oligonucleotides and streptavidin coated magnetic beads followed by quality control using an Agilent Technologies 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA). The mRNA was then amplified and double-stranded cDNA was generated by random priming. The ends of the fragmented DNA were converted into phosphorylated blunt ends. An ‘A’ base was added to the 3′ ends. Illumina-specific adaptors were ligated to the DNA fragments. Using magnetic bead technology, the ligated fragments were size selected and then a final PCR was performed to enrich the adapter-modified DNA fragments since only the DNA fragments with adaptors at both ends will amplify. The sequencing library was first validated using an Agilent Technologies 2100 Bioanalyzer to characterize DNA fragment sizes and concentration. The concentration of DNA fragments with the correct adapters on both sides was then determined using a quantitative PCR strategy, following the kit manufacturer's protocol (Kapa Biosystem, Cambridge, MA). Following library quantitation, libraries were diluted and pooled as necessary. Using the Illumina cBot, libraries were added to the flow cells and clusters were generated prior to 100 bp paired end sequencing on the Illumina HiSeq 2000 (Illumina, San Diego, CA, USA). During and after the sequencing run, sequence quality was monitored using the real time analysis (RTA) and sequence analysis viewer (SAV) software available by Illumina. Following sequencing, demultiplexed fastQ files were generated using the Illumina CASAVA software.
FastQ files were aligned to the C57BL/6J reference genome on a high performance computing cluster using Tophat (http://tophat.cbcb.umd.edu/) for the alignment and RSEM (http://deweylab.biostat.wisc.edu/rsem/) for isoform assembly and quantitation, except that frequency of reads per kilobase was normalized based on quartile instead of the total number of mapped reads. Further analysis was done in R (http://www.R- project.org) using ANOVA and linear modeling to test expression differences by strain, genotype and tissue, and FDR analysis was done in Microsoft Excel (Microsoft Corp).
Adult mice aged between 6 and 8 weeks were anesthetized with tribromoethanol (400 mg/kg i.p.). Small burr holes were drilled (1 mm anterior to the bregma and 2 mm posterior to the bregma) on both sides of the skull 2 mm lateral to the midline. Four teflon-coated silver wires were soldered onto the pins of a microconnector (Mouser electronics, Texas). The wires were placed between the dura and the brain and a dental cap was then applied. The mice were given a post-operative analgesic of carprofen (5 mg/kg subcutaneous) and allowed a minimum 48 h recovery period before recordings. Differential amplification recordings were recorded between all four electrode pairs, providing 6 channels for each subject. Mice were connected to the EEG Stellate Lamont Pro-36 programmable amplifier (Lamont Medical Instruments, Madison, WI) for a 2-hour period on 2 separate days, between the hours of 9 AM and 4 PM during the lights-on period. EEG data were recorded with Stellate Harmonie software (Stellate Systems, Inc., Montreal, Canada) into a database. SWD consist of adjacent, connected spike-wave (or wave-spike) complexes. Recordings were reviewed using low/hi bandpass filters at 0.3 Hz and 35 Hz respectively, and SWD episodes were scored blinded to genotype using the following criteria: at least 2 connected spike-wave complexes (typically spanning at least 0.5 seconds) with amplitudes at least two fold higher than background and observed concurrently in the majority of the 6 recording channels per mouse.
|
10.1371/journal.pgen.1002904 | Retrovolution: HIV–Driven Evolution of Cellular Genes and Improvement of Anticancer Drug Activation | In evolution strategies aimed at isolating molecules with new functions, screening for the desired phenotype is generally performed in vitro or in bacteria. When the final goal of the strategy is the modification of the human cell, the mutants selected with these preliminary screenings may fail to confer the desired phenotype, due to the complex networks that regulate gene expression in higher eukaryotes. We developed a system where, by mimicking successive infection cycles with HIV-1 derived vectors containing the gene target of the evolution in their genome, libraries of gene mutants are generated in the human cell, where they can be directly screened. As a proof of concept we created a library of mutants of the human deoxycytidine kinase (dCK) gene, involved in the activation of nucleoside analogues used in cancer treatment, with the aim of isolating a variant sensitizing cancer cells to the chemotherapy compound Gemcitabine, to be used in gene therapy for anti-cancer approaches or as a poorly immunogenic negative selection marker for cell transplantation approaches. We describe the isolation of a dCK mutant, G12, inducing a 300-fold sensitization to Gemcitabine in cells originally resistant to the prodrug (Messa 10K), an effect 60 times stronger than the one induced by the wt enzyme. The phenotype is observed in different tumour cell lines irrespective of the insertion site of the transgene and is due to a change in specificity of the mutated kinase in favour of the nucleoside analogue. The mutations characterizing G12 are distant from the active site of the enzyme and are unpredictable on a rational basis, fully validating the pragmatic approach followed. Besides the potential interest of the G12 dCK variant for therapeutic purposes, the methodology developed is of interest for a large panel of applications in biotechnology and basic research.
| We exploited the error-prone replication machinery of HIV-1 and its ability to stably introduce transgenes in human cells to develop a novel system, Retrovolution, to generate libraries of mutants of cellular genes. When libraries are screened to isolate variants that modify the phenotype of the human cell for biomedical applications or basic research, false positives often arise from the classical screening procedures performed in vitro or in bacteria. Retrovolution allows an easy screening of the libraries directly in the human cell, where they are generated. We describe the creation and screening of a library of the hdCK (a human kinase activating several anticancer compounds) gene, to identify variants increasing the sensitivity of cancer cells to treatment with low, poorly toxic doses of the anticancer drug Gemcitabine. We isolated a dCK variant inducing death in tumour cells at doses up to 300 times lower than those required for killing non-engineered cells. The mutant presents mutations unpredictable on structural basis and revealed a change in enzymatic properties that accounts for the observed cellular effect. Besides the intrinsic interest of the mutant identified, these results fully validate Retrovolution as a mutagenesis system with broad applications in applied and basic research.
| Broadening the repertoire of natural molecules and generating variants that confer new phenotypes to human cells are appealing perspectives for the development of biomedical applications, and for understanding fundamental cellular processes. To this end, in classical procedures, libraries of mutants are generated in vitro by degenerated PCR or DNA shuffling, and then screened on biochemical bases or with genetic tests in bacteria [1]. However, when the modification of human cells is sought, the mutants identified in these preliminary screenings often do not confer the desired phenotype due to differences in protein folding, post-translational modifications, and to the complex epistatic network that regulates the expression of the phenotype in the cells of higher eukaryotes. Alternatively, the library can be cloned in eukaryotic expression vectors for the screening step, albeit with the drawback of a considerable loss of complexity. The generation and screening of libraries of mutants directly in human cells would constitute an ideal solution to circumvent these problems. Nature provides organisms that are perfectly exploitable for this purpose: retroviruses. Indeed, after entry into the target cell, the viral polymerase (reverse transcriptase, RT) converts the viral genomic RNA, through an error-prone process that generates genetic diversity, into double-stranded DNA, which is then permanently integrated in the genome of the cell.
The human immunodeficiency virus (HIV-1) is the retrovirus with the strongest mutation rate, and constitutes therefore the ideal candidate for developing such approaches. During the process of Reverse Transcription, point mutations are introduced in the proviral DNA with a rate of 1–3.4×10−5 nt/cycle [2]–[4] and genetic diversity is further amplified by recombination [5], [6]. We report here the development of a new methodology (called “Retrovolution”) aimed at generating and screening libraries of cellular genes directly in human cells. In Retrovolution the error-prone replication machinery of HIV-1 is diverted to drive the evolution of cellular genes: by performing successive infection cycles of cell cultures with HIV-1-derived viral vectors containing the sequence target of the evolution inserted in their RNA, libraries of gene mutants are generated in human cells, where they can be directly screened for the desired phenotype. A system for the evolution of non-viral sequences based on the use of engineered HIV-1 had been previously developed for the optimization of the Tetracycline-regulated expression system [7]. In that system, the isolation of an improved phenotype was strictly linked to the efficiency of replication of the viral particle, whereas in Retrovolution selection is not linked to viral replication and the use of replication-defective, non-cytopathic vectors allows the isolation and expansion of cells with the desired phenotype, which is not possible when replication-competent HIV particles are used.
While relatively simple proteins or sequences, as a modified tag or reporter gene, have traditionally constituted the target for setting up evolution procedures [8], [9], to evaluate the power of the Retrovolution system we aimed at generating and isolating variants of a protein modifying a complex cellular phenotype, which could have a potential therapeutic interest. Namely, we wished to mutagenize the human deoxycytidine kinase (hdCK, NM_000788.2 mRNA NCBI accession number) gene with the goal of isolating an enzyme variant able to increase the sensitivity of cells to a widely used category of anticancer compounds, deoxycytidine (dC) analogues. The hdCK, physiologically involved in the phosphorylation of deoxycytidine, deoxyadenosine and deoxyguanosine in the salvage pathway for the deoxyribonucleotide biosynthesis, is a pivotal enzyme in the activation of deoxycytidine analogues used in chemotherapy (AraC and Gemcitabine, used in the treatment of leukemias and solid tumours). These compounds, administered as nucleosides, must be transformed by cellular enzymes in their triphosphate form to become active and induce apoptosis in the cell by interfering with DNA replication and repair [10]. The first phosphorylation step, catalyzed by the human deoxycytidine kinase, determines the overall efficiency of the process [11]. Accordingly, there is a direct correlation between the level of dCK expression and the sensitivity to deoxycytidine analogues in different types of tumor [12], [13]. The appearance of resistance forms among tumor cells, often due to a loss of dCK activity, and the toxicity of high concentrations of nucleoside analogues for non-tumor cells, constitute the major limits of the treatments with these compounds [14], [15]. The sensitization of cancer cells to low doses of drugs by insertion of a gene coding for a drug-metabolizing enzyme constitutes an appealing approach for gene therapy applications [16]–[18]. A dCK variant that could preferentially phosphorylate nucleoside analogues with respect to the natural substrate deoxycytidine, and therefore induce cell suicide upon exposure to low concentrations of prodrug, would constitute a good candidate for these “suicide gene” therapies in cancer treatment, but also in transplantation medicine. In this case, indeed, it would provide a human (and therefore poorly immunogenic) gene to use as negative selection marker to insert in transplanted cells to counteract their eventual uncontrolled proliferation in vivo.
So far, attempts at constructing a dCK mutant with improved ability in phosphorylating deoxycytidine analogues have been based on the information obtained from the 3D structure of the protein [19], [20]; these dCK variants based on the rational design displayed an overall increased catalytic activity, resulting in a highly improved affinity for the natural substrate and a less improved affinity for the prodrug, which is not predictive of the induction of a sensitization phenotype in vivo. By evolving the dCK gene with Retrovolution procedure we could directly screen for dCK variants inducing cell death in presence of doses of the deoxycytidine analogue Gemcitabine lower than the ones needed to kill cells bearing a wt dCK, and this allowed us to isolate a mutant consistently sensitizing cancer cells to the prodrug.
The Retrovolution system, outlined in Figure 1, is based on the insertion of the gene target of the evolution process in the genomic RNA (gRNA) of conditional replication-defective, HIV-1 derived vectors pseudotyped with the Vesicular Stomatitis Virus envelope (VSV-G) (Figure 1A). Repeated cycles of transduction of producer cells with these vectors, as indicated in Figure 1B, mimic successive infection cycles by HIV-1, during which mutations will be inserted in the gRNA and re-shuffled by recombination [21], [22], an event favored when transductions are performed at a high Multiplicity Of Infection (MOI). During the procedure, HIV-1 Reverse Transcriptase will introduce mutations all along the viral gRNA and, while those falling in functional regions of the vector will be selected (positively or negatively), the others, including those falling in the target gene, will accumulate in an unbiased fashion. This process generates a library of mutants of the target gene directly inserted into the vector tasked with delivering the transgene to the human cell, and requires minimal intervention from the experimenter. During the evolution procedure, cells that did not receive the viral vector and therefore do not contain an integrated proviral DNA are removed by puromycin treatment (the puromycin resistance gene being inserted as a selectable marker in the gRNA of the vectors, see Figure 1A). As an alternative, GFP can be used as a selectable marker instead of puroR and successfully transduced cells can be sorted by FACS at every round: the use of GFP would fasten the evolution procedure (a complete puromycin selection takes three to four days), but would require a much more important manual intervention. The average number of positions that will be mutated in the target gene will depend on the number of “infection cycles” performed, on the size of the target gene itself, and on the mutation rate of the reverse transcription process, as outlined in Figure 2A. Once reached the ideal complexity, the library of mutants will be screened by transducing target cells (the human cells to which the intended phenotype is to be conferred) at a low MOI (≪1) to minimize the possibility that one cell will receive more than one vector, followed by clonal analysis (Figure 1B). As the vectors are pseudotyped with the VSV envelope, which can infect virtually any kind of mammalian cells, the library can be screened in all human cell types.
A library of dCK variants was generated by using the dCK coding sequence (dCK-cs) as the target gene (Figure 1A) of the Retrovolution procedure. By sequencing part of the library after 9 cycles of transduction we could calculate an ongoing mutation rate of 1×10−4 nt/infectious cycle. Based on this, the library was screened upon reaching generation 16, at which a complexity of 1.5 mutated positions per dCK-cs (783 bp) was expected (Figure 2A). A further sequencing of a subset of the library at this generation confirmed the presence of 1.4 mutated positions per dCK gene (Table 1). In the dCK variants sequenced all kinds of mutations were present with a bias toward G>A transitions (62.3%), consistent with the mutational pattern of HIV-1 RT [23]. While the sequencing of the dCK gene at different generations of the Retrovolution system confirmed that mutations were inserted randomly along the gene (Figure 2B), the sequencing of the viral backbone of the corresponding clones revealed a consensus of 12 mutations in the U3 region of each vector (Figure 2C). These consensus mutations appeared in the promoter region of HIV-1 LTR since the F8 generation and were fixed in the population ever since. Moreover, their nature diverged from the typical HIV-1 mutation pattern since the percentage of G>A transitions was consistently lower (33%). No mutations were instead found in the other cis-acting regions essential for genome packaging, reverse transcription, integration and nuclear export. Overall, this strongly suggests that the viral sequences underwent selection for an optimized transcription of the genomic RNA from the U3 promoter in the specific context of the producer cells, while preserving the functionality of the vector, and allowing the production of a library of randomly mutated variants of the target gene.
To screen the library for the presence of dCK variants that confer an increased sensitivity to low concentrations of the deoxycytidine analogue Gemcitabine, viral vectors from generation 16 were used to transduce HEK-293T cells at a low MOI (MOI = 0.03) to ensure that each cell did not contain more than one variant of the transgene, and single clones were isolated by limiting dilution and puromycin selection. We then measured the viability of 76 individual clones in the presence of increasing concentrations of Gemcitabine (10, 35 and 70 nM) and calculated the cell death rate as the number of alive cells at concentration X of Gemcitabine divided by the number of cells at 0 Gemcitabine. The experiment was repeated three times using, as controls, HEK-293T cells either untransduced or transduced with a wt-dCK containing vector. The two higher Gemcitabine concentrations tested resulted to be too strong for an effective screening since all samples, including controls, displayed more than 80% of cell death (data not shown). At 10 nM Gemcitabine, instead, 6 of the 76 clones tested yielded, in at least one of the three independent experiments, a cell death rate higher than both untransduced HEK-293T cells and cells bearing the wt dCK transgene (Figure 3A). These six clones were selected for further analysis.
The 6 clones isolated in the preliminary screening were further characterized for the ability of their transgenes to induce sensitization to Gemcitabine in the Messa10K cell line, uterine sarcoma cells that express an inactive dCK and are, therefore, highly resistant to this drug [13]. These cells provide an ideal background for testing the ability of the library of dCK mutants to increase sensitivity to Gemcitabine, since the phenotype generated by the introduction of a single copy of a dCK variant will not be masked by the activity of the wt dCK protein present in the cells. For each HEK-293T clone selected, viral particles were rescued by transfection with the plasmids encoding HIV-1 Gag-Pol and VSV Env, and used to transduce Messa10K cells at a MOI<1. Since the effect induced by a transgene on the cell phenotype can be strongly influenced by the position of integration of the proviral DNA within the target cell genome, transduction of the Messa10K cells was performed following a procedure, outlined in Figure S1, that generates, for each transgene, a population of cells containing the same transgenic sequence inserted at different genomic locations (“polyclonal population”). For each dCK variant analysed we established 4 to 9 independent Messa 10K populations with this procedure, that allows to average the possible effects of the insertion site and to highlight the phenotype induced by the transgenic sequence itself. From the analysis of cell death ratios of Messa 10K populations in presence of increasing concentrations of Gemcitabine appeared that one of the 6 transgenes identified on HEK-293T cells, G12 (E171K, E247K, L249M), strongly sensitized cells to the drug. In the 9 G12-Messa10K polyclonal populations tested, indeed, 50% of the cells died at 75 nM Gemcitabine, while at the same concentration only 10% of Messa 10K containing a wt dCK were killed (Figure 3B). The analysis of a broader range of Gemcitabine concentrations revealed that in G12-Messa 10K cells the Gemcitabine IC50 was reduced by 60-fold compared to cells transduced with wt-dCK, and by 300-fold relative to untreated cells [13] (Figure 3C).
As shown by Western Blot (Figure 4A), sensitization of Messa10K cells by G12 is not the consequence, as could result from mutations in the EF1-alpha promoter (Figure 1A), of a higher level of protein expression compared to wt-dCK. Nevertheless, sequencing of the entire G12 gRNA revealed the presence of mutations within as well as outside the dCK-cs and these mutations could influence the phenotype observed. To rule out this possibility we inserted the G12 dCK-cs in a wild-type gRNA backbone plasmid (generating “G12/wt-backbone” gRNA), and tested the deriving viral vectors on Messa10K cells. Also in this case a sensitization to the nucleoside analogue (Figure 4B) was detected, indicating that the Gemcitabine sensitization phenotype is rather due to properties of the dCK variant encoded by G12 than to mutations arisen in other portions of the genomic RNA.
Messa10K cells constitute a model of Gemcitabine-resistant cells lacking an active endogenous dCK, as frequently emerge during chemotherapy [15]. To investigate the effect of the insertion of G12 in tumour cells that express an endogenous functional dCK protein, we assessed the phenotype conferred by the insertion of the mutated dCK to the colon carcinoma cell line HT29, and to the pancreatic cancer cell line BxPC3. In HT29, sensitivity to Gemcitabine is strictly linked to the levels of dCK activity, whereas in BxPC3 the sensitivity to the prodrug depends on a different mechanism, the expression levels of the cytidine deaminase CDA [24], [25], an enzyme that inactivates the fully phosphorylated Gemcitabine by deamination. Therefore, while in the former cells the insertion of an improved variant of dCK should lead to a significant sensitization to the drug, in the latter ones the effect should be reduced. We generated two independent G12 polyclonal populations for each cell line and tested them. Consistently with our hypothesis, we observed a considerably stronger effect of G12 in HT29 than in BxPC3 (Figure 4C and 4D, respectively), with a decrease in the Gemcitabine IC50 with respect to cells containing a wt-dCK from 1.6 µM to 80 nM in HT29 and only of less than two folds (from 40 to 22 nM) in BxPC3. These results support the idea that the Gemcitabine sensitization phenotype is due to an improved ability of the G12 mutated kinase to activate the prodrug in cell culture and underscores that the insertion of a single copy of the transgene can increase the sensitivity of cells to Gemcitabine even in the presence of a wt dCK activity.
To further support the existence of a link between the effect of the mutant in cell culture and its enzymatic activity, we expressed in E. coli the G12 mutant and the wt type dCK proteins, purified them and characterized their efficiency of phosphorylation of Gemcitabine and of dC in vitro. While Gemcitabine was phosphorylated with comparable efficiencies by the wt and G12 dCKs (left panel of Figure 5A), the natural substrate dC was phosphorylated at almost undetectable levels by G12, with a dramatic drop with respect to the efficiency of phosphorylation observed with the wt dCK enzyme (right panel of Figure 5A). The G12 dCK mutant therefore displays an altered substrate specificity, constituting a kinase that specifically phosphorylates Gemcitabine, even in the presence of the natural substrate of wt dCK. These observations underscore that, in vivo, an increased sensitivity to the prodrug can be efficiently achieved through a decreased ability of the kinase to phosphorylate its natural substrate, rather than through an improved efficiency of phosphorylation of the prodrug itself. So far, efforts at improving the efficiency of phosphorylation of nucleoside analogues by dCK have relied on the rational design of mutants either of the active site region [19] or of Ser74, the phosphorylation of which has been described to modulate the dCK enzymatic activity in vivo [20], [26]. Although the resulting enzymes displayed an increased overall catalytic activity in vitro, the relative efficiency of phosphorylation (expressed as the ratio Kcat/KmGemcitabine/Kcat/KmdC) of the drug with respect to the natural substrate was decreased compared to the wt kinase, a trend opposed to that observed for the G12 mutant we describe (Figure 5B). Consistently with the view that these mutants should not impact sensitivity of cells to Gemcitabine treatment, when we constructed a vector carrying the sequence of the triple mutant A100V/R104A/D133A [19] and used it to transduce Messa10K cells, no altered sensitivity to Gemcitabine was observed (Figure 5C).
We describe here the development of a new system of mutagenesis of cellular genes aimed at identifying genetic variants of interest for applications in basic and applied research. A crucial feature of the system is the possibility of performing, easily and in a controlled manner, a straightforward screening of the library in the human cell. This pragmatic approach allows to overcome the generation of false positives variants, a frequent problem encountered upon in vitro screening of libraries. The application of the Retrovolution system led to the identification of a dCK mutant that fulfils the long-sought feature of increasing sensitivity of tumour cells to Gemcitabine treatment, based on the presence of mutations unpredictable on a rational basis. The screening step was performed following a protocol that ensures that the isolated mutant sensitizes the cells independently from the integration site of the provirus, therefore constituting a good candidate as “suicide gene” in gene therapy.
The mutagenic potential of Retrovolution relies on the error-prone nature of the replication machinery of HIV-1, and on the frequent occurrence of recombination, that reshuffles the pre-existing mutations contributing to increase the diversity of the library. We show here that these two sources of genetic diversification prompt enough diversity as to lead to the identification of mutants of interest. With a relatively reduced number of cycles the system allowed the creation of a library in which each variant of the target gene was characterized by 1.5 mutated positions on average. A higher proportion of mutations would have increased the chances of producing a high percentage of non-functional proteins. An alternative approach for generating a library of mutants of cellular genes contained in lentiviral vectors would have been constituted first by a mutagenesis step of the cellular gene through conventional methods as error-prone PCR, followed by the insertion of the library in the lentiviral gRNA plasmid. With respect to this approach, the method we describe, while slower in the generation of a complex library, provides the advantage of circumventing the unavoidable loss of complexity inherent to the step of cloning the library in the lentiviral gRNA plasmid. Another way to accelerate the acquisition of mutations would have been the use of error-prone reverse transcriptases or mild mutagens. These procedures were not privileged, though, due to the inherent drawback of reduced yield of vector production [27], [28].
Depending on the nature of the target gene, the experimental settings of Retrovolution can be adapted to introduce selection ongoing during the mutagenesis steps. In the present work this was not possible since we targeted a gene for which a negative selection, as the induction of cell death, needed to be applied. The targeting of the dCK gene was aimed, despite the intrinsic difficulty of the screening procedure, at the isolation of variants of a gene with a relevant biological interest for biomedical applications. The properties of the mutant identified, indeed, constitute the second major point emerging from this work.
Improving the efficiency of the currently clinically employed anticancer drugs and overcoming resistances arising during treatments is a major goal of cancer research. To improve the efficiency of the treatments with nucleoside analogues like Gemcitabine, gene therapy approaches aimed at sensitizing cancer cells through the introduction of a transgene that improves the efficacy of drug activation have been proposed. The hdCK is the best candidate transgene for this purpose as the kinase is responsible for the limiting step in intracellular activation of deoxycytidine analogues currently used in clinical treatments. The competition between the natural substrate of dCK and the prodrug inside the cell constitutes an obstacle to the efficient activation of the prodrug itself. The rational design of mutants in the catalytic site of the kinase attained the goal of improving phosphorylation of the prodrug in the past, but the mutants isolated displayed a concomitant, and stronger, increase in the efficiency of phosphorylation of the natural substrate [19], [20]. As a result, phosphorylation of the prodrug is expected to remain disfavoured with respect to the natural substrate and these mutants fail, as we confirmed (Figure 5C), to confer a drug sensitization phenotype to the cell. The G12 mutant, on the contrary, has an increased specificity for the phosphorylation of Gemcitabine due to the fact that it completely lost the ability to phosphorylate its natural substrate. This results in a reduction of the competition between the two substrates in vivo and in a sensitization of the cells to the prodrug. The biochemical mechanism through which an increased sensitization to the drug is obtained is therefore opposed to what generally tried to obtain by engineering the dCK on a rational basis.
The screening for the isolation of the G12 mutant was performed on Messa 10K cells, where the most striking sensitization was observed. These cells constitute a model of tumoral cells acquiring resistance to deoxycytidine analogues due to a loss of dCK activity, a fundamental problem that has to be bypassed for an improvement of cancer treatment. Nevertheless, the G12 mutant has an effect also on tumour cells presenting an endogenous dCK activity, like HT29 and HEK-293T. Therefore, besides a potential application in cancer treatment, G12 also constitutes a promising suicide gene to use as negative selection marker in cells used in transplantation medicine. Suicide genes are inserted in cells transplanted for therapeutical purposes to specifically ablate them in the event they would undergo uncontrolled proliferation in vivo. To this end exogenous genes are generally used, the most frequently employed being the Herpes Simplex Virus thymidine kinase gene in association with gancyclovir treatment. A problem of increasing relevance in clinical gene therapy, though, is constituted by the immune response raised by the patient against the protein encoded by the suicide gene [29], [30]. A suicide gene of human origin, as the mutant described here, is expected to have a faint possibility of being highly immunogenic and therefore represents an ideal candidate for this application.
Besides conferring an increased sensitivity to Gemcitabine through an unexpected mechanism, the G12 mutant is characterized by mutations localized in regions unpredictable on a rational basis. The human dCK is a globular, dimeric protein in which each monomer is formed by a core of 5 beta-sheets surrounded by 10 alpha-helices [19] (Figure 6A). G12 dCK carries the mutations E171K, E247K, and L249M, the first two of which induce a charge change in conserved residues (Figure 6B). Amino acids 247 and 249 are located in the “base-sensing loop” (alpha-helix 10), which influences the folding of the protein upon binding of ATP or UTP as phosphate donor [26], [31] and, consequently, also affects substrate binding. Indeed, a variant containing only these mutations, which was identified during the screening of the library (mutant E8, Figure 3B), slightly increased Messa10K cells sensitivity to Gemcitabine treatment, although with a cell death ratio that did not significantly differ from that of the wt-dCK (p>0.05). The marked sensitization observed with G12 thus requires the presence of the additional mutation, E171K. Residue 171 is located at the base of alpha-helix 7 that, with alpha-helix 4, is involved in the generation of the interface of the dCK dimer [19], and its mutation in a residue of opposite charge could potentially influence the efficiency of formation of the active dimer itself. Gel filtration analyses, though, revealed that the extent of dimerization of G12 does not differ from that of the wt-dCK (Figure S2). It is therefore likely that, as observed for other proteins [32], [33], the mutation rather triggers long-distance changes in the quaternary arrangement of the protein, possibly involving the region of the active site.
In conclusion, Retrovolution allowed the identification of a promising suicide gene to use in cancer treatment but also as negative selection marker in transplantation medicine by allowing isolating a variant of the dCK that would not have been predictable on a rational basis. The property of producing the library and allowing its screening directly in human cells, ensuring that each mutant emerging from the procedure will be relevant for modifying the phenotype of the cells, is central to these findings. Retrovolution thus opens new avenues for the modification of genes conferring complex phenotypes of interest for a broad field of applications in basic and applied research. With the exploitation of its combination of genetic flexibility and ability to deliver transgenes to human cells, the lifestyle of one of the most important pathogens of our recent history appears far from being fully exploited.
HEK-293T cells were obtained from the American Type Culture Collection (ATCC) and grown in Dulbecco's Modified Eagle's Medium (Gibco) supplemented with 10% FBS and 100 U/ml pennicillin-100 mg/ml streptomycin. Messa10K cells were kindly provided by LP. Jordheim and grown in RPMI medium supplemented with 10% FBS and penicillin-streptomycin. HT29 cells were kindly given by JN. Freund and grown in DMEM+ 10% FBS. BxPC3 cells were kindly provided by M. Dufresne and grown in RPMI+ 10% FBS. All cells were incubated at 37°C+5% CO2.
For the creation of the first generation of viral vectors, 8 plates containing 5×106 HEK-293T cells were transfected using the calcium phosphate protocol with 10 µg of pCMVΔR8.91 plasmid [34], 5 µg of pHCMV-G plasmid [34]–[36] and 10 µg of genomic plasmid sdy-dCK (encoding the RNA outlined in Figure 1A). 48 h after transfection, the virus-containing supernatant was collected, filtered through a 0.45 µm filter and concentrated 40-fold in Vivaspin 20 columns (MWCO 50 KDa, Sartorius Stedim Biotech). Transfections for the creation of subsequent generations were performed with 10 µg of pCMVΔR8.91 and 5 µg of pHCMV-G plasmid. Transductions during the evolution procedure were performed on 5 million HEK-293T cells with 1 ml of 40× viral vectors (MOI>100). Transductions for the isolation of single clones were performed with 1 ml of 1∶500 dilution of viral vectors (MOI<1) on 3.5×106 HEK-293T cells; after transduction, cells were seeded at 100 cells/well in 96-well plates in DMEM+ puromycin. For the generation of polyclonal populations of target cells, transduction was performed on 1×106 cells with 1 ml of 1∶5 diluted viral vectors rescued from the isolated clones. Transduced cells were selected by adding puromycin 24 h after transduction using 0.6 µg/ml to HEK-293T cells, 0.5 µg/ml to Messa10K cells, 0.8 µg/ml to HT29 cells and 0.4 µg/ml to BxPC3.
Clones of the F8 and F16 generation of the dCK library that had been isolated by limiting dilution were expanded in 10 cm plates and pelleted. Cells were lysed with 250 µl of Cell Direct PCR (VIAGEN). The dCK transgene was amplified from 1 µl of the lysate with oligonucleotides on the vector, EF1 (5′-gatgtcgtgtactggctccg-3′) and PGK (5-gatgtggaatgtgtgcgagg-3′), flanking the transgene. Sequencing of the PCR fragment was performed by the GATC sequencing service.
The curves for the calculation of the number of Retrovolution cycles needed to have one mutation per target gene (Figure 2A) were drawn based on the formula p/m×n where p = n°of mutations wanted per target gene, m = mutation rate/nt, n = size of the target gene in nt.
Cells were seeded at 5000 cells per well in a 96-well plate and grown overnight at 37°C. Two rows were used for each population. 12 h after seeding, increasing concentrations of Gemcitabine (0–400 nM or 0–100 µM) were added to each well and incubation was continued for an additional 72 h. Cell viability was measured by MTT test (CellTiter 96 Non-Radioactive Cell Proliferation Assay, Promega) and the number of living cells in each well was evaluated by measuring the OD at 570 nm. For each population, the fraction of living cells was calculated as OD570 concentration X Gem/OD570 concentration 0 Gem, to estimate the sensitivity of the population to the prodrug.
Messa10K cells transduced with the different dCK variants were lysed in 1X RIPA buffer, and 6, 12, 24 µg of total protein (evaluated by Bradford) for each cell type were loaded on a 12% bis-tricine gel (Invitrogen). After transfer on a PVDF membrane, the dCK proteins were analysed by western Blot with 1∶4000 dilution of a polyclonal anti-dCK antibody (rabbit, Sigma-Aldrich) and 1∶3000 anti-rabbit HRP (BioRad) conjugated secondary antibody and detected by autoradiography.
To evaluate whether the average amounts of cell death occurring in the Messa10K populations containing different dCK variants were significantly different from the value shown by Messa10K wt-dCK, a two sample t-test was applied for the different concentrations of Gemcitabine.
The wt-dCK and G12 sequences were cloned in the pET14b plasmid and expressed in E. coli cells BL21 DE3 pLysE. Protein expression was induced by adding 0.1 mM IPTG, and cells were collected after 4 h of growth at 37°C. His-tagged proteins were eluted with 250 mM Imidazole from His-Trap TM FF Columns (GE HealthCare), the Histidine tag was removed using the S-Tag Thrombin Purification Kit (Novagen), and dCK and G12 were further purified by gel filtration on S-200 Sephacryl columns (GE Healthcare). Purified proteins were used for the dCK activity assay or stored at −80°C in 30% glycerol.
The efficiency of phosphorylation of the natural substrate deoxycytidine and of the prodrug Gemcitabine were measured for purified wt-dCK and G12 in a NADH-based assay as previously described [37]. All reagents were purchased from Sigma (France) except Gemcitabine (Lilly France SAS). Enzymes were assayed at RT at a concentration of 0.3 µM with Gemcitabine or 0.9 µM with dC. Gemcitabine was used at concentrations between 10 µM and 1 mM and dC at concentrations between 5 and 50 µM. ATP was 4 mM. All experiments were performed in triplicate.
|
10.1371/journal.pgen.1005760 | Arabidopsis ERF1 Mediates Cross-Talk between Ethylene and Auxin Biosynthesis during Primary Root Elongation by Regulating ASA1 Expression | The gaseous phytohormone ethylene participates in the regulation of root growth and development in Arabidopsis. It is known that root growth inhibition by ethylene involves auxin, which is partially mediated by the action of the WEAK ETHYLENE INSENSITIVE2/ANTHRANILATE SYNTHASE α1 (WEI2/ASA1), encoding a rate-limiting enzyme in tryptophan (Trp) biosynthesis, from which auxin is derived. However, the molecular mechanism by which ethylene decreases root growth via ASA1 is not understood. Here we report that the ethylene-responsive AP2 transcription factor, ETHYLENE RESPONSE FACTOR1 (ERF1), plays an important role in primary root elongation of Arabidopsis. Using loss- and gain-of-function transgenic lines as well as biochemical analysis, we demonstrate that ERF1 can directly up-regulate ASA1 by binding to its promoter, leading to auxin accumulation and ethylene-induced inhibition of root growth. This discloses one mechanism linking ethylene signaling and auxin biosynthesis in Arabidopsis roots.
| Ethylene is a gaseous phytohormone that plays critical roles in plant development and defense. It is well known that ethylene inhibits primary root elongation through effects on auxin. However, it is not clear how ethylene signal is translated into auxin. In this report, the highly ethylene-responsive transcription factor ETHYLENE RESPONSE FACTOR1 (ERF1) is demonstrated to positively regulate ASA1, encoding ANTHRANILATE SYNTHASE α1, a rate-limiting enzyme in Trp biosynthesis where auxin is derived, by directly binding to its promoter and activating ASA1. Consequently, auxin biosynthesis is promoted, leading to increased auxin accumulation and thus inhibition of primary root elongation. This study unravels a molecular mechanism that bridges ethylene signaling and auxin biosynthesis in primary root elongation.
| Phytohormones are central regulators of plant root growth and development. Each root development process is determined by a network of interacting signals to give the final architecture of the root[1]. Ethylene and auxin have been shown to regulate some of the same developmental processes, including primary root elongation[2–4]. Although the crosstalk between ethylene and auxin in regulating primary root elongation is well characterized[3,5–9], there is still a significant lack of understanding of the molecular mechanism that links ethylene signaling and auxin biosynthesis.
Ethylene, a gaseous plant hormone, acts as a key regulatory signal during the plant life cycle[10,11]. The biosynthesis of ethylene begins from methionine, which is converted to S-adenosyl-methionine (SAM) by SAM synthetase. Then a family of 1-aminocyclopropane-1-carboxylic acid (ACC) synthases (ACS) converts SAM to ACC. This reaction is a rate-limiting step and a key regulatory point in ethylene biosynthesis[12]. Finally, ACC is converted to ethylene by ACC oxidase (ACO)[13,14]. In the Arabidopsis genome, there are 12 members in ACS gene family, which display overlapping temporal and spatial expression patterns and are responsive to a variety of biotic and abiotic stresses and hormones, such as auxin[15,16]. Apparently all major components of ethylene signal transduction have been identified by the successful isolation of a series of ethylene response mutants and a precise ethylene signaling pathway has been established[17–21]. Once ethylene is synthesized, it is perceived by any of five membrane bound protein receptors ETHYLENE RESPONSE1 (ETR1), ETR2, ETHYLENERESPONSE SENSOR1 (ERS1), ERS2, and ETHYLENEINSENSITIVE4 (EIN4), which possess sequence similarity to bacterial two-component His kinases[22–24]. The binding of ethylene to its receptor results in inhibition of a Raf-like Ser/Thr protein kinase CONSTITUTIVE TRIPLE RESPONSE1(CTR1)[25]. Inhibited CTR1 loses its ability to phosphorylate and repress a positive component of the ethylene signal pathway, the membrane protein ETHYLENE INSENSITIVE2 (EIN2)[26]. The active form of EIN2 stabilises the transcription factors of the EIN3 family located in the nucleus. The EIN3 proteins subsequently bind to the promoters of the ERF genes and activate their transcription[27,28]. Thus a transcriptional cascade commencing with the sensing of ethylene is triggered to produce the ethylene response.
ERFs, which contain an AP2 DNA-binding domain, form a plant-specific superfamily of 122 transcriptional factors in Arabidopsis[29]. ERFs influence a variety of functions involved in plant development and also play important roles in response to biotic and abiotic stresses[30–34], through specifically binding to sequences containing GCCGCC motifs (GCC-box) in the regulatory region of downstream genes[35]. It was reported that GCC-box is not well conserved in ethylene responsive genes, suggesting that other types of transcription factors may also be activated by EIN3 and involve in transcriptional cascade caused by ethylene[7]. ERF1 (AT3G23240) is a downstream component of the ethylene signaling pathway and is directly regulated by EIN3 at the transcriptional level[27]. It is well known that ERF1 is a key integrator of the jasmonic acid (JA) and ethylene signaling pathways involved in the regulation of defence response genes such as b-CHI and PDF1.2[36]. Ethylene signaling is also involved in plant responses to both salt and water stress, as ethylene insensitive mutants were more salt sensitive[37–39]. ERF1 also plays a positive role in abiotic stress responses such as salt, drought, and heat stress[40]. In addition to responding to biotic and abiotic stress, ERF1 further mediates ethylene responses in developmental processes, such as the inhibition of primary root growth and hypocotyl elongation in the dark. This has been confirmed by the production of transgenic plants with constitutively activated ERF1, which displayed phenotypes similar to that are observed in ctr1 mutant, EIN3-overexpressing plants, and wild-type plants treated with ethylene[25,27,41]. Recently, ERF109 was shown to mediate crosstalk between JA signaling and auxin biosynthesis[42].
Root growth relies on two essential developmental processes: cell division in the root meristem and elongation of cells produced by the root meristem[43]. Root cell elongation can be affected by diverse endogenous and exogenous factors such as ethylene[3], auxin[44], and calcium[45]. Ethylene, and its precursor ACC, reduces root elongation in a concentration-dependent manner by inhibition of the cell elongation process[4,6]. The crosstalk between ethylene and auxin has been well investigated[3,6]. The most interesting discovery for auxin/ethylene crosstalk in recent years is that Arabidopsis pyridoxal-phosphate -dependent aminotransferase, VAS1, uses the ethylene biosynthetic intermediate methionine as an amino donor and the auxin biosynthetic intermediate indole-3-pyruvic acid as an amino acceptor to produce L-tryptophan and 2-oxo-4-methylthiobutyric acid[46]. Many mutants that affect auxin synthesis, distribution, or signaling also result in abnormal responses to ethylene[8,47–50], such as, mutants of AUX1 and EIR1/AGR/PIN2 involved in auxin transport, AXR2/IAA7 and AXR3/IAA17 in the auxin signal pathway, or the auxin receptor TIR1, which all exhibit ethylene-insensitive root growth[5,20,49,51–53]. Auxin biosynthetic genes encoding enzymes such as WEI2/ ANTHRANILATE SYNTHASE α1 (ASA1), WEI7/ASB1, TAA1, and TAR1, which are regulated by ethylene, also exhibit ethylene-insensitive root growth[8,9]. YUC genes also play an important role in root responses to ethylene[54]. These studies suggest that the inhibition of primary root growth caused by ethylene requires auxin biosynthesis, transport, or signaling. WEI2 encodes the α-subunit of the enzyme anthranilate synthase in Trp-dependent auxin biosynthesis. Its expression in roots can be induced by ethylene and wei2 mutations cause ethylene-insensitive root growth phenotypes[8]. In Arabidopsis roots, ethylene promotes auxin biosynthesis in a ASA1-dependent manner[8], this is an important molecular mechanism by which ethylene exerts its effect on promoting auxin biosynthesis. However, the molecular mechanism for regulation of ASA1 by ethylene is not well understood.
Here, we report that ERF1, a downstream AP2 transcription factor in the ethylene signaling pathway, positively regulates auxin biosynthesis during inhibition of ethylene-mediated primary root growth. Transgenic plants with constitutive expression or knockdown of ERF1 displayed similar root development phenotype to mutants of ethylene signaling. ERF1 affected auxin accumulation through directly binding to the ASA1 promoter and positively regulating ASA1 expression. Our results indicate that ERF1 plays a pivotal role in the inhibition of ethylene-induced primary root growth in Arabidopsis and acts as the crosstalk node between ethylene and auxin in primary root elongation.
To study the role of ERF1 in ethylene response, we examined the ethylene-induced expression of ERF1 in 5-day-old wildtype seedlings (Col-0) grown on Murashige and Skoog (MS) medium with 10 μM ACC for 0–12 h. The ERF1 transcript level in whole seedlings was measured by quantitative real-time PCR (qRT-PCR). Consistent with previous studies[36,40], the ERF1 transcript level was rapidly induced by ACC treatment. ERF1 increased in a biphasic manner showing a 4-fold increase after 0.5 h, and a second plateau increasing by approximately 8-fold at 3 h. Thereafter, the ERF1 transcript levels remained high (Fig 1A).
The spatial expression pattern of ERF1, was determined by assaying ß-glucuronidase (GUS) activity of transgenic plants carrying an ERF1pro:GUS construct, in which a GUS reporter gene was under the control of the ERF1 promoter (ERF1pro, a 3.0kb promoter fragment). ERF1 was mainly expressed in the maturation zone of the primary root of seedlings (Fig 1B). To examine the response of ERF1 to ethylene, we treated 5-day-old ERF1pro:GUS seedlings with 10 μM ACC for 0–6 h and found that the expression of ERF1 was induced in the maturation zone and cotyledons as the ACC treatment time increased. Stronger induction was observed in the upper maturation zone (Fig 1B).
To determine the long-term response of ERF1 to ethylene, we grew wildtype seedlings on MS medium with either 0 or 0.8 μM ACC for 5 d and examined the relative transcript levels. The abundance of ERF1 increased about 21 times by ACC treatment compared with the 0 μM ACC control (Fig 1C). Transgenic seeds carrying ERF1pro:GUS were directly germinated on MS supplemented with or without 0.8 μM ACC for 5 d before GUS analysis. Strikingly in ACC-supplemented seedlings, GUS staining increased along the entire root and with weaker staining was also visible in the primary root tip compared to the 0 μM ACC control (Fig 1D).
GUS staining of ERF1pro:GUS transgenic lines showed that ERF1 expression was mainly present in roots and old leaves (S1A–S1E Fig). ERF1 nuclear-specific subcellular localization was demonstrated in transgenic plants carrying a CaMV 35S-driven ERF1 construct fused to GFP (S1F Fig). These results suggest that ERF1 may play a role in root development in response to ethylene signaling.
To confirm ethylene-dependent induction of ERF1 expression, we analysed the ERF1 transcript level in the ethylene signaling mutants ein2-5, ein3-1, ein3-1eil1 and compared them to wildtype seedlings. ERF1 transcription was also measured in lines of constitutive or β-estradiol inducible expression, ctr1-1 and EIN3-FLAG (iE/qm) (EIN3ox), respectively. Expression of ERF1 could not be induced by ACC in the ein2-5, ein3-1 and ein3-1eil1 mutants, while it was constitutively expressed in the ctr1-1 and EIN3ox seedlings (Fig 1E). These results suggested that ethylene-induced expression of ERF1 is dependent on the ethylene signaling pathway.
To investigate the function of ERF1 in Arabidopsis root growth, we generated transgenic knockdown and overexpression lines. The phenotype of these plants was confirmed with the analysis of ERF1 expression (S2 Fig). From three overexpression lines and two RNAi knockdown lines (Fig 2A), we found that the primary roots of the lines overexpressing ERF1 (ERF1ox) were stunted. In contrast, the RNAi lines showed longer primary roots compared to the wildtype controls (Fig 2B and 2C). Similar results of root elongation were observed in etiolated seedlings (S3 Fig). Root elongation in ERF1ox lines was similar to mutants with an activated ethylene signal pathway, while root elongation in ERF1 RNAi lines was similar to mutants with defects in ethylene signal pathway (S4A Fig).
Ethylene signaling mutants ein2-5, ein3-1, ein3-1eil1, ctr1-1, and EIN3ox also display abnormal primary root development with or without ACC treatment[25,26,55–58]. The roots of ein2-5, ein3-1 and ein3-1eil1 seedlings, in which the corresponding genes positively regulate the ethylene signal pathway, were longer and more insensitive to ACC compared to wildtype (S4A Fig). However, the roots of ctr1-1 and EIN3ox, in which ethylene signaling was significantly active, were shorter than wildtype roots.
To explore the primary root elongation of the ERF1-related transgenic lines, we investigated the primary root in three developmental zones, the differentiation zone (DZ, also known as maturation zone), the elongation zone (EZ), and the meristem zone (MZ). Root meristem size was measured as the cell number from the quiescent centre (QC) to the first elongated cell in the cortex[59]. No significant change in meristem cell number was observed among the transgenic lines compared to the wildtype (Fig 2D and 2E). However, cell length had dramatically decreased in the DZ of ERF1ox lines, whereas it was greater in the DZ of the RNAi lines compared to wildtype (Fig 2D and 2F). These results suggest that ERF1 contributes to root length via altering cell elongation but not cell division (Fig 2D–2F), which is a similar effect to that produced by exogenously applied auxin[60]. This agrees with the root inhibition mechanism resulting from ACC treatment and the ethylene signal pathway mutants, ein2-5 and ctr1-1[3,6,61].
Since ERF1 is downstream of the ethylene signal pathway, we proposed that the inhibition of root elongation by ethylene might be ERF1-dependent. To analyze this, wildtype, ERF1ox, and RNAi lines were grown on MS plates for 5 d, then transferred to plates with or without ACC for 3 d, and primary root lengths were measured. We found that the primary root lengths of the RNAi lines were greater than controls, but those of ERF1ox lines were shorter than controls (Fig 2G). Furthermore, the difference between ERF1 expression levels in ERF1 RNAi lines and wildtype was augmented by ACC treatment (S4B Fig). Taken together, these results suggest that ethylene inhibits primary root elongation mainly through ERF1. ERF1, as a positive regulator of ethylene signaling, appears to play a negative role in regulating root cell elongation, leading to dramatically shortened primary roots under conditions of hyperactive ethylene signaling.
The reduced primary root elongation phenotype in ERF1 overexpression lines was similar to wildtype plants grown on the medium containing auxin[6], consistent with the proposal that ethylene enhances auxin biosynthesis to inhibits root elongation [3,6]. To study whether ERF1 enhances auxin accumulation in Arabidopsis roots, Firstly, we introduced a DR5:GUS reporter, an auxin reporter responding to endogenous auxin[62], into ERF1 knockdown and overexpression lines (Fig 3A and S5A Fig). Expression of DR5:GUS in primary roots, including the MZ and DZ was significantly increased in the ERF1 overexpression background. GUS expression occurred even in the absence of added ACC, in the root as well as in the cotyledons and hypocotyl (Fig 3A and 3B).The distribution of GUS staining in the primary root tip extended to other tissues compared to wildtype, i.e., also occurring in the epidermal cells of the MZ (Fig 3B, #12). In addition, DR5:GUS expression in DZ occurred not only in the xylem as for wildtype, but also in epidermis, cortex and endodermal tissues (Fig 3B, #2, #6, and #12). In contrast, the GUS activity decreased in the primary root tip and stele in ERF1 knockdown background compared with wildtype (Fig 3B, RNAi-1 and -2 compared with Col-0, respectively). The difference of DR5:GUS expression in these lines were more evident when treated with ACC for 24 h (Fig 3B). Moreover, the IAA content was higher in the ERF1 overexpression lines and lower in the RNAi lines compared with wildtype (Fig 3C). In accordance with the increased auxin, the expression of IAA1 and IAA2, which are auxin-responsive marker genes, was activated in overexpression lines of ERF1 and down regulated in knockdown lines (Fig 3D and 3E). Our results indicate that, ERF1 restrains primary root elongation by increasing auxin.
Ethylene positively regulates the transcription of ASA1 to increase auxin and inhibit root elongation[8]. This finding prompted us to examine whether ERF1 functions as a direct regulator of ASA1. Previous studies showed that ERF1 binds to a specific cis-element (a GCC-box related sequence) upstream of its target genes, to regulate downstream gene expression[27,40].We analysed the promoter sequence of ASA1 and found one GCC-box at 27 bp upstream of the translational start codon. We sought to test whether this GCC-box could provide a handle for ERF1 to regulate ASA1 directly.
To determine if there was a direct physical interaction of the ERF1 protein with the ASA1 promoter sequence, we conducted an electrophoretic mobility shift assay (EMSA) with the full-length ERF1 protein fused to a maltose binding protein (MBP-ERF1) that was expressed in E. coli and purified through affinity chromatography. As shown in Fig 4A, the ERF1-MBP fusion protein was able to specifically bind digoxigenin-labelled DNA probes that contained the GCC-box motif of the ASA1 promoter. Moreover, the binding specificity was confirmed by competition with native DNA probes, or probes carrying a mutated GCC-box. Unlabelled native DNA probe was used as a competitor and unlabelled promoter fragment containing the mutant form of the GCC-box motif as a non-competitor. The EMSA results showed that ERF1 specifically binds to the promoter sequence containing a GCC-box of ASA1 promoter in vitro, but not to the DNA probe containing the mutant GCC-box sequence (Fig 4A). In addition, we carried out a yeast-one-hybrid assay that showed ERF1 was able to bind to the GCC-box sequence in the ASA1 promoter in yeast cells (Fig 4B).
To confirm whether this specific binding occurs in planta, we generated transgenic Arabidopsis plants expressing the 35S promoter-driven HA tagged ERF1 construct (S6 Fig). Transgenic lines showing expected phenotypes were used for chromatin immunoprecipitation (ChIP) with anti-HA antibodies (Roche, USA). As shown in Fig 4C, the chromatin immunoprecipitated with anti-HA antibodies was significantly enriched for the ASA1 promoter containing GCC-box fragments in the ChIP-PCR assay. This was further confirmed by quantitative real-time PCR performed using the same ChIP products and PCR primers flanking GCC-boxes in ASA1 promoter (Fig 4D). These results suggest a specific binding of ERF1 to the promoter of ASA1 in vivo.
To investigate the consequence of ERF1 binding to the GCC-box of the ASA1 promoter, we measured the ASA1 expression level in ERF1 transgenic lines by qRT-PCR. As predicted, the transcript level of ASA1 in ERF1 overexpression lines was significantly increased compared to wildtype without additional treatment. A small but statistically significant reduction of ASA1 transcripts in ERF1 knockdown lines was observed (Fig 5A). The difference in ASA1 expression levels between knockdown lines of ERF1 and wildtype was further increased by ACC treatment (S4C Fig). In addition, the increased expression level of ASA1 in 35Spro:ERF1 plants agreed with the microarray results in a previous report that ASA1 was upregulated in 35S:ERF1 transgenic plants[36]. These results suggest that ASA1 is indeed a downstream target of ERF1.
To confirm further whether the expression level of ASA1 was affected by ERF1, we introduced the ASA1pro:GUS reporter into ERF1 knockdown and overexpression background and examined the primary root phenotype (Fig 5B and S5B Fig). Consistent with the ASA1 expression level in ERF1 transgenic lines (Fig 5A), we found that the intensity of GUS staining in ASA1pro:GUS significantly increased in an ERF1 overexpression background and slightly reduced in the ERF1 knockdown lines compared to wildtype (Fig 5C). Moreover, this difference was further exaggerated in response to ACC treatment (Fig 5C). Furthermore, the GUS staining patterns produced by ASA1pro:GUS in the ERF1 knockdown lines was weaker than that of wildtype with ACC treatment. Taken together, these results suggest that ERF1 positively regulates ASA1 expression in response to ethylene.
Given that ethylene-inhibited root elongation involved auxin biosynthesis which is enhanced by ERF1 through regulating ASA1 expression, we predicted that loss of ASA1 would decrease the sensitivity to ethylene. To test this, we grew wildtype and asa1 mutants (asa1-1 and asa1-2) on MS medium with or without ACC. The asa1 mutants grown on media containing different concentrations of ACC displayed longer primary roots than wildtype seedlings (Fig 6A and 6B). Our results support the notion that inhibition of root elongation by ethylene is ASA1-dependent. This observation is consistent with a previous finding that asa1 mutants are insensitive to ethylene under dark conditions[8]. Furthermore, as shown in S7 Fig, ERF1-RNAi lines, asa1, and ein2-5 showed reduced response to ACC with respect to the effect of the ethylene on primary root growth compared to Col-0. For the reduced sensitivity of root to ACC, asa1 and ein2 were more evident than ERF1-RNAi lines, which might be due to the limitation of ERF1-RNAi materials, and/or some other factors participating in this process.
The primary root elongation of mutants (ein2-5, ein3-1eil1 and ein3-1) was less sensitive to ACC compared to wildtype. The mutants with enhanced ethylene signals (ctr1-1, EIN3ox) showed dramatically shortened primary roots under normal conditions, mimicking wildtype seedlings treated with ACC[25,26] (S4A Fig). Since ASA1 functions downstream of CTR1[8], we measured the ASA1 transcript levels in these mutants in normal conditions and found that the expression level of ASA1 in ein2-5, ein3-1eil1 and ein3-1 was lower than that of wildtype and conversely higher in ctr1-1 and EIN3ox lines (Fig 7A). To confirm this further, we introduced ASA1pro:GUS reporter into ein2-5, ein3-1, ctr1-1, and EIN3ox backgrounds and found that the GUS staining pattern closely correlated with the qRT-PCR results (Fig 7B and 7C). Strong GUS activity in the root and cotyledons was observed in ctr1-1 and EIN3ox backgrounds (Fig 7B). These results indicated that ASA1 is downstream of these ethylene signal pathway components.
To confirm this genetically, we introduced an estradiol-inducible ERF1 overexpression in asa1 mutant background (ERF1ox asa1-1). When ERF1 is overexpressed, the primary root of ERF1ox asa1-1 is significantly longer than that of ERF1ox and a little shorter than Col-0 (Fig 8, S8 Fig). The result demonstrates that ASA1 is downstream of ERF1, and suggests that ethylene-inducible ERF1 controls root elongation through regulating ASA1 expression. There may be other targets of ERF1 which also participate this process since some difference in primary root length was observed between ERF1ox asa1-1 and asa1-1 (Fig 8D).
One of the best studied effects of ethylene on roots is the inhibition of root elongation[5,20]. A number of studies have indicated that ethylene inhibits root development through interaction with auxin. Ethylene has been shown to increase auxin synthesis, auxin transport to the elongation zone, and auxin signaling at the root tip[3,5–9,47,52]. ERF1, a downstream transcription factor in the ethylene signal pathway was reported to reduce primary root growth in the dark when constitutively expressed[27]. Until now, no detailed and explicit mechanism has been provided for its role in primary root elongation. In this study, we demonstrated that ERF1 directly regulates the expression of ASA1, a key enzyme in Trp biosynthesis where auxin is derived and known to play an important role in ethylene-regulated root development[8]. This work elaborates the mechanism by which the transcription factor ERF1 participates in primary root development and directly mediates crosstalk between ethylene and auxin biosynthesis during root elongation.
Through analyses of root response to ethylene with overexpression and knockdown lines of ERF1 (Fig 2A–2C), we found that the length of the primary root was closely correlated to ERF1 expression. These results imply that ERF1 is involved in ethylene-mediated root elongation.
ERFs belong to a large gene family. Only a few ERF mutants show obvious phenotypes, probably due to functional redundancy. However, ERF1 knockdown lines displayed longer root under normal and added ACC conditions (Fig 2B, 2C and 2G), indicating that ERF1 plays an important role in ethylene-inhibited root elongation. Furthermore, ERF1 controls primary root elongation by reducing cell elongation, but not cell division (Fig 2D–2F), which is consistent with ethylene signal pathway mutants[6].
Ethylene upregulates auxin biosynthesis in Arabidopsis seedlings to enhance inhibition of root elongation[6]. High auxin levels are known to reduce root growth[3]. Some auxin biosynthesis genes are ethylene responsive, and if mutated, cause some defects in root growth in the presence of ACC[8,9]. To understand how ERF1 mediates ethylene signaling in primary root elongation and particularly auxin biosynthesis, we analysed the promoters of all genes which participate in auxin biosynthesis, and found that two genes including ASA1 and YUCCA2 contained a GCC-box which can be specifically bound by ERF1. A recent study showed that ERF109, another member of the ERF family that is highly responsive to JA signaling, directly regulates both ASA1 and YUC2 and mediates crosstalk between JA and auxin biosynthesis[42]. The primary root elongation of the yucca2 mutant, in the presence of added ACC, did not differ from wildtype. Considering that ASA1 is ethylene responsive, asa1 is ethylene insensitive in root-elongation[8], and there is a GCC-box in the promoter of ASA1 which could be bound by ERF1, we hypothesized that ASA1 might be a direct target of ERF1.
To confirm our hypothesis, we conducted in vitro binding (EMSA), yeast-one-hybrid, and chromatin immunoprecipitation (ChIP) experiments and confirmed that ERF1 could directly bind a conserved GCC-box element in the promoter of ASA1 in vitro and in vivo (Fig 4). We further confirmed our hypothesis by analyzing ASA1pro:GUS in ERF1 knockdown and overexpression background. As we expected, the expression of ASA1 was remarkably increased in ERF1 overexpression background but reduced in the knockdown lines (Fig 5). Upon ACC treatment, the staining of ASA1pro:GUS in ERF1 knockdown background becomes darker but still relatively weaker than that in Col-0 background at the root tip and DZ (Fig 5C), indicating that the ethylene responsiveness was not completely removed. In the ERF1 knockdown lines, the induction of ASA1 by ethylene was reduced but some induction was still retained compared to the wildtype (S4C Fig), which is consistent with the primary root phenotype (Fig 6). This may be due to incomplete suppression of ERF1 by RNAi technique, alternatively, there may be additional components involved in this process. For instance, EIN3 was suggested to directly regulate ASA1 based on the data of EIN3 ChIP-Seq experiments [63].
Meanwhile, analyses of ASA1pro:GUS reporter in ethylene signal pathway mutation background (ein2-5, ein3-1, ctr1-1, EIN3ox) showed that when ethylene signal pathway was enhanced, the expression of ASA1 was also enhanced. Conversely, if ethylene signal pathway was blocked, the expression of ASA1 was reduced and the induction by ACC was also impaired (Fig 7). These results explicated that ASA1 is downstream of these ethylene signal pathway components.
Taken together, our results support a model in which ethylene stimulates auxin biosynthesis in roots through ethylene-responsive transcription factor ERF1 that positively regulates ASA1. As a consequence of activating ASA1 expression, ERF1 increases the accumulation of auxin, which in turn decreases root elongation and alters root architecture.
Surface-sterilised Arabidopsis seeds were treated for 10 min in 10% bleach and cold-treated for 3–4 d at 4°C, and grown on Murashige and Skoog (MS, 1 x salts) medium with 1% sucrose for the indicated days (see Figure legends). Seedlings were then transferred to plates supplemented with or without ACC for the indicated days, and the plates were placed vertically in a growth room. ACC (Sigma-Aldrich) was dissolved in water and prepared as a 2 mM stock solution. Arabidopsis thaliana 7-d-old seedlings were transferred to soil and grown to maturity in a growth room. All plants were grown under long-day conditions (16-h light / 8-h dark) at 22–24°C. Every experiment was repeated at least three times.
Arabidopsis thaliana ecotype Columbia-0 (Col-0) was used. Some plant materials used in this study were previously described: ein2-5[26], ein3-1[41], ein3-1 eil1-1[52], ctr1-1[25], EIN3-FLAG (iE/qm) (EIN3ox) [64], ASA1pro:GUS [8], DR5:GUS[65].
Plants were obtained from transformation of Col-0 plants with the ERF1pro:GUS construct, which contained a 3.0-kb promoter sequence with primers (S1 Table). ERF1 over-expressing plants (ERF1ox #2, #6, and #12) were obtained from transformation of Col-0 with the 35Spro:ERF1 construct, which contains full-length ERF1 (At3g23240). ERF1 knockdown plants were obtained from transformation of Col-0 with the 35Spro:ERF1:RNAi construct. 35S:HA-ERF1 were obtained from transformation of Col-0 plants with a vector comprised of a 35S promoter and HA sequence in front of the full-length ERF1 cDNA. 35Spro:ERF1-GFP transgenic lines were constructed in Col-0 Arabidopsis. ERF1ox transgenic plants used in genetic analysis of crossing ERF1ox with asa1 were obtained from transformation of Col-0 plants with the vector comprised of an estradiol-inducible promoter in front of the full-length ERF1.
To prepare the ERF1pro:GUS construct, the primer set ERF1pro-F and ERF1pro-R (S1 Table), was used to amplify a 3.0 kb sequence from the ERF1 promoter region. This was cloned into the binary vector pCB308R. To prepare the ERF1 over-expression construct, the primer set, ERF1-F and ERF1-R and full-length ERF1 fragments were amplified and cloned into the binary vector pCB2004[66]. To prepare the ERF1 knock-down construct, the primer set ERF1RNAiF and ERF1RNAiR was used to amplify a 0.18 kb sequence from the ERF1 coding region, and cloned into the binary vector pCB2004B. To prepare the HA:ERF1 over-expression construct, the primer set ERF1HAF and ERF1HAR was used to amplify the HA-ERF1 coding sequence, and cloned into the binary vector pCB2004. To prepare the 35Spro:ERF1-GFP construct, the product amplified by ERF1GFPF and ERF1GFPR was cloned into the binary vector pGWB5[67]. All of these constructs utilised the gateway system technology. To generate the ERF1ox transgenic plants with estradiol-inducible promoter, the PCR product amplified by pER8-ERF1 F and pER8-ERF1 R primers was cloned into the vector pER8 for transformation[68].These constructs were then individually transformed into Agrobacterium tumefaciens Strain (C58C1), and introduced into Arabidopsis plants by the floral dip method[69]. More than 30 transgenic lines were obtained for each construct.
Histochemical staining for GUS activity in transgenic plants was conducted as previous described[70]. Seedlings were grown on MS medium as above, and stained in fresh GUS staining solution (1 mg/mL X-glucuronide in 0.1 M potassium phosphate, pH 7.2, 0.5 mM ferrocyanide, 0.5 mM ferricyanide, and 0.1% Triton X-100) at 37°C in the dark for the indicated time. After incubation, seedlings were cleared with a series of ethanol solutions (100%, 50% and 30%) and photographed. Images were captured using an OLYMPUS IX81 microscope and HiROX (Japan) MX5040RZ. For in planta GFP analysis, seedlings were stained in 10 mg/mL propidium iodide for 8 min and washed twice in water. Propidium iodide fluorescence and GFP were imaged under a ZEISS710 confocal laser scanning microscope: 488-nm and 543-nm lines of the laser were used for excitation, and emission was detected at 510 nm and 620 nm, respectively. More than 15 seedlings were examined for each transgenic line, and at least three independent experiments were displayed.
Total RNA was extracted with TRIzol reagent (Invitrogen) from seedlings or root samples. cDNA was prepared by TransScript RT kit (Invitrogen) for RNA reverse transcription, and was used for real-time quantitative RT-PCR. All quantitative RT-PCR assays were performed with a SYBR Premix Ex Taq II kit on StepOne real-time PCR system (Applied Biosystems) according to the manufacturer’s instructions. Expression levels of target genes were normalized to Arabidopsis UBQ5. All qRT-PCR experiments were performed at least three biological replicates. The primers used in this study have been listed in S1 Table.
The free total IAA content was measured by ELISA as described[71].
The procedures of the yeast one-hybrid assay were displayed as described previously with minor modifications[72]. A DNA fragment encoding ERF1 was amplified with the primers ERF1Y1H F and ERF1Y1H R, and cloned into pAD-GAL4-2.1 (AD vector) to produce the pAD/ERF1 plasmid. Two complementary core 30-bp DNA single strands including GCC-box in the promoter of ASA1, named ASA1Y1H F and ASA1Y1H R, were annealed and cloned into the reporter plasmid via SacI and MluI restriction sites, pHIS2 (BD vector), which contained the nutritional reporter gene, HIS3.Two complementary DNA strands with three copies of the DNA sequence GCC-box, named GCCY1H F and GCCY1H R were annealed and cloned into plasmid pHIS2 through SacI and MluI restriction sites as a positive control. The pAD/ERF1 construct containing the ERF1 cDNA sequence and the pHIS2 reporter construct containing the GCC-box cis-element were co-transformed into Y187 yeast cells. For the negative control, the pAD/ERF1 plasmid and the pHIS2 empty plasmid were co-transformed into Y187 yeast cells. Yeast was cultured in SD/–Trp–Leu medium and then transferred to SD/–Trp–Leu–His medium containing10 mM 3-aminotriazole (Sigma) with dilutions as indicated. The plates were incubated at 30°Cfor 4 d and the results were observed. Growth of yeast cells on the SD/–Trp–Leu–His medium in the presence of 10 mM 3-aminotriazole indicated that the transcription factor could bind this cis-element, and activated relevant gene expression.
One gram each of six-day-old 35S:HA-ERF1seedlings and Col-0 plants were harvested for ChIP experiments. The procedures were conducted essentially as previously described[73]. The enrichment of DNA fragments in the ASA1 promoter was amplified by PCR using the following primer pairs: ChIP-ASA1pro F, ChIP-ASA1pro R, and β-tubulin8 was used as negative control. The resultant PCR-products were resolved by electrophoresis on 2% agarose gels. The results displayed above represent at least three independent repeats.
To analyse the binding quantitatively, a qRT-PCR assay was performed on the basis of the procedure described previously[74]. The relative quantity value is presented as the DNA binding ratio. The same primers for the above PCR analysis were used for qRT-PCR.
The fragment of the full-length ERF1 coding sequence was amplified by PCR with primers ERF1EMSA F and ERF1EMSA R, and cloned into the pMAL-C2 vector via EcoRI and XbaI restriction sites, to construct a plasmid for the expression of recombinant MYC2 protein in Escherichia coli.
Five individual synthetic 30-bp single-stranded DNA molecules containing the GCC-box were used, namely ASA1EMSA F-DIG, ASA1EMSA R, ASA1EMSA F, ASA1EMSA-Mu F and ASA1EMSA-Mu R. These DNA fragments were annealed with their complementary oligonucleotides, ASA1EMSA F-DIG with ASA1EMSA R for EMSA labelled probe, ASA1EMSA R with ASA1EMSA F for EMSA competitive probe, ASA1EMSA-Mu F with ASA1EMSA-Mu R for EMSA non-competitive probe. EMSA was performed according to a DIG Gel Shift kit, 2nd Generation (Roche).
|
10.1371/journal.pgen.1001263 | Signaling Role of Fructose Mediated by FINS1/FBP in Arabidopsis thaliana | Sugars are evolutionarily conserved signaling molecules that regulate the growth and development of both unicellular and multicellular organisms. As sugar-producing photosynthetic organisms, plants utilize glucose as one of their major signaling molecules. However, the details of other sugar signaling molecules and their regulatory factors have remained elusive, due to the complexity of the metabolite and hormone interactions that control physiological and developmental programs in plants. We combined information from a gain-of-function cell-based screen and a loss-of-function reverse-genetic analysis to demonstrate that fructose acts as a signaling molecule in Arabidopsis thaliana. Fructose signaling induced seedling developmental arrest and interacted with plant stress hormone signaling in a manner similar to that of glucose. For fructose signaling responses, the plant glucose sensor HEXOKINASE1 (HXK1) was dispensable, while FRUCTOSE INSENSITIVE1 (FINS1), a putative FRUCTOSE-1,6-BISPHOSPHATASE, played a crucial role. Interestingly, FINS1 function in fructose signaling appeared to be independent of its catalytic activity in sugar metabolism. Genetic analysis further indicated that FINS1–dependent fructose signaling may act downstream of the abscisic acid pathway, in spite of the fact that HXK1–dependent glucose signaling works upstream of hormone synthesis. Our findings revealed that multiple layers of controls by fructose, glucose, and abscisic acid finely tune the plant autotrophic transition and modulate early seedling establishment after seed germination.
| Among the many plant sugar metabolites, glucose signaling has received the most attention. Although fructose is also an abundant hexose, its signaling role in plant growth and development has not been addressed clearly and systematically to date. We found that fructose functions as a regulatory sugar metabolite and interacts with signaling by the plant hormones abscisic acid (ABA) and ethylene in A. thaliana. The fructose-dependent growth response is mediated by FRUCTOSE INSENSITIVE1 (FINS1), which encodes an ancient metabolic enzyme, putative fructose-1,6-bisphosphatase. Interestingly, the catalytic function of FINS1 in sucrose biosynthesis is dispensable for its regulatory role in fructose signaling. FINS1 appears to act downstream of GLUCOSE-INSENSITIVE1, which is involved in ABA synthesis. Overall, it is evident that although fructose and glucose have unique regulatory pathways, they also share some signaling interactions with plant stress and defense hormones and coordinate early seedling establishment of A. thaliana. Fructose affects cell signaling in mammals and causes various metabolic syndromes. However, a direct relationship between fructose and physiological diseases has not been established yet. Because FINS1 is evolutionarily conserved, our genetic evaluation of its signaling function may provide useful information about fructose signaling in animals as well as plants.
| Myriad metabolic pathways enable cells to sustain life with basic carbon and nitrogenous compounds. Thus, the integration of metabolite status, which reflects external and internal living conditions, into cellular activities (e.g., gene expression) is a pivotal process that equips organisms with the ability to survive and proliferate. For example, cellular metabolites often serve regulatory roles in modulating organism growth and development, from unicellular bacteria and yeasts to multicellular animals and plants [1]–[6]. To sense and transduce such metabolite signals, organisms have developed sophisticated biochemical and cellular mechanisms.
Glucose is an evolutionarily conserved regulatory sugar molecule in many different organisms [1]–[6]. It has multiple roles as an energy source, building block, and osmotic regulator, and also acts as a potent signaling molecule that regulates gene expression and controls organism growth and development. For example, in yeast, glucose is sensed by at least four different types of sensors, Hxk2, Snf3, Rgt2 and Gpr1, and regulates gene expression and cell growth [4]. In mammalian pancreatic islet β cells, glucose signaling may be a function of the total amount of ATP generated via catabolism [6].
In plants, glucose [7]–[9], sucrose [10]–[12], trehalose-6-phosphate [13], and low energy/high AMP concentrations [14], [15] function as cellular signaling molecules in specific regulatory pathways that modulate plant growth and development. Of these signaling metabolites, glucose has been studied the most comprehensively in plants. Glucose signaling modulates the gene expression of enzymes in the glyoxylate cycle [16] and the photosynthesis pathway [17], and is also involved in the developmental decision of whether to progress to normal seedling establishment after seed germination [18].
Glucose-mediated developmental repression is largely dependent on HEXOKINASE1 (HXK1) [7]–[9]. HXK1's function in glucose-mediated developmental repression is mostly independent of its catalytic activity and integrates glucose signaling with other plant hormone such as auxin and cytokinin. HXK1-independent glucose signaling has also been reported in plants. For instance, expression of the genes encoding chalcone synthase, phenylalanine ammonia-lyase, and asparagine synthase responds to glucose signaling, but their regulation is independent of HXK1 activity [3], [19]. A recent study further demonstrated that a refined low-glucose condition can uncouple HXK1-dependent and -independent glucose signaling responses during early A. thaliana seedling establishment [9], [20].
In both animals and plants, the developmental roles and regulatory functions of hexoses other than glucose have remained largely unknown. However, within the last few years, dietary fructose was implicated in mammalian cell signaling perturbation and metabolic syndromes such as insulin resistance, obesity, type 2 diabetes, and high blood pressure [21], [22].
Plant triose phosphates synthesized by photosynthetic activity are stored as transitory starch in chloroplasts or converted into sucrose in the cytoplasm through a series of enzymatic reactions carried out by fructose-1,6-bisphosphatase (FBP), UDP-glucose pyrophosphorylase, sucrose phosphate synthase, and sucrose phosphatase [2]. Sucrose is then stored in vacuoles or cleaved into glucose and fructose by invertases or UDP-glucose and fructose by sucrose synthases [23]. Thus, following sucrose hydrolysis, fructose becomes one of the prevalent hexoses in plants and has long been proposed as a possible signaling molecule [24]. Nevertheless, fructose signaling in plants has remained largely unexplored. Recently, Kato-Naguchi et al. [25] showed that the fructose analog psicose induced root growth inhibition in lettuce. Fructokinase (FRK), which performs the same catalytic function as HXK, but with fructose as the substrate rather than glucose, was the first fructose enzyme to be studied for a putative regulatory role in fructose signaling [24], [26], [27]. Although FRK is involved in modulation of plant growth, a regulatory role in fructose signaling was ruled out [28]; hence, little is known about fructose signaling and its regulatory pathways.
In this study, we used a cell-based functional screen and a reverse genetics assay to investigate the signaling role of fructose in A. thaliana. We identified FRUCTOSE INSENSITIVE1 as an indispensable regulatory factor in the signaling pathway. Here, we report the molecular and genetic characterization of fins1 in a fructose signaling context, and its close interactions with ABA signaling during early seedling development.
To evaluate the regulatory role of fructose signaling in plant developmental modulation, we examined A. thaliana seedling growth on 6% (w/v) fructose agar medium with full-strength Murashige and Skoog (MS) salts. Wild-type (WT; Ler and Col accession) seedlings grown on high-fructose medium exhibited a typical early developmental arrest, which was manifested by inhibition of hypocotyl and root growth and repression of cotyledon expansion and chlorophyll accumulation (Figure 1A and 1B). Although the seedling development repression pattern caused by high fructose was similar to that caused by high glucose (6%) [7], [20], fructose caused slightly more root growth inhibition than glucose (Figure S1). Mannitol, an osmotic control, did not induce the same seedling repression, suggesting that the observed phenotype was a developmental response to fructose signaling.
Recently, we refined glucose assay conditions for growing A. thaliana seedlings and showed that the high glucose requirement is due to the high nitrogen content in MS media [9], [20]. When MS salts were omitted, 2% glucose induced equivalent seedling growth repression to 6% glucose media including MS. However, decreasing the concentration of fructose in the absence of MS salts had little effect on seedling growth (Figure S2); this suggested that nitrogen had a different effect on fructose and glucose signaling.
Indeed, further experiments indicated that fructose and glucose signaling does rely on distinct sensors. The glucose-insensitive HXK1-null mutant gin2-1 (gin2) exhibited normal fructose sensitivity, as did transgenic gin2-expressing WT HXK1 and its catalytically inactive mutants HXK1S177A or HXK1G104D (Figure 1A). These data confirmed that the glucose sensor HXK1 was dispensable in fructose signaling. Although HXK1 carries out metabolic activities for both glucose and fructose, it does not appear to be involved in fructose signaling. This may reflect the fact that HXK1 has an approximately 100-fold higher affinity for glucose compared to fructose [28]–[30]. In a previous study, root growth inhibition in lettuce was reported in the presence of either the fructose analog psicose or the glucose analog mannose [25]. However, the HXK inhibitor mannoheptulose restored root growth in the presence of mannose, but not psicose. These results are further evidence that psicose/fructose signaling is independent of HXK function.
Plant sugar signaling, mainly glucose and sucrose, interacts with stress and defense hormone signaling pathways and coordinates seedling growth and development [1]–[3], [23], [31]. For glucose signaling, gin1, gin5, and gin6 were respectively identified as alleles of aba-deficient2 (aba2), aba3, and aba-insensitive4 (abi4) in the ABA pathway, and gin4 was found to be a new allele of constitutive triple response1 (ctr1) in the ethylene pathway [31]–[36]. These mutants have been selected repeatedly from various independent screens for sugar responses, further confirming that sugar signaling interacts with ABA and ethylene response pathways during early seedling development [37]–[40].
To test whether fructose signaling interacts with plant stress/defense hormones, we observed the early developmental response of ABA and ethylene mutants on a 6% fructose agar medium with MS salts. Unlike WT and gin2, both gin1-3 (gin1) and ctr1-1 (ctr1) seedlings were not only insensitive to high glucose, but also overcame fructose repression and developed green cotyledons (Figure 1B). GIN1/ABA2 encodes a short-chain dehydrogenase/reductase in ABA synthesis, and CTR1/GIN4 encodes a putative mitogen-activated protein kinase kinase kinase that functions as a negative regulator of ethylene signaling [31], [33]. Therefore, fructose signaling appears to interact positively with ABA signaling via hormone biosynthesis, whereas it is likely antagonized by ethylene signaling. Interestingly, in ABA-deficient gin1 mutants, cotyledon repression was de-repressed by fructose, but root repression was not (Figure 1B); however, glucose relieved both cotyledon and root growth repression in gin1 mutants (Figure S1) [20]. This indicated that fructose repression of root growth was independent of ABA biosynthesis, unlike cotyledon greening. This observation revealed differential seedling responses to fructose and glucose in an organ-specific manner.
We further monitored marker gene expression using real-time PCR with cDNA templates generated from mRNA of five-d-old seedlings grown on MS agar medium containing 6% glucose, fructose, or mannitol. Expression of the photosynthesis-related CHLOROPHYLL A/B BINDING PROTEIN2 (CAB2/AT1G29920) gene was markedly repressed in WT by both glucose and fructose (Figure 1C and 1D). Gene expression was similarly repressed in gin2 seedlings by fructose, but not by glucose (Figure 1C). However, CAB2 expression was de-repressed in both gin1 and ctr1 seedlings (Figure 1D). The CAB2 gene expression patterns in the mutants reflected the fructose resistance revealed by their phenotypes (Figure 1B). Taken together, these data indicated that fructose signaling was mediated through a unique/unknown sensor, but shared a downstream pathway with glucose signaling, which interacted with the plant stress and defense hormones ABA and ethylene to modulate early seedling development in A. thaliana.
Although the application of high sugar to A. thaliana growth media has been criticized because it is not a normal physiological condition, it is unclear how much sugar is actually taken up by roots, how fast it is metabolized or fluxed, and in which suborganelles the sugar is partitioned. These factors could affect developmental responses to high sugar levels. Glucose and sucrose nanosensors, which detect cytoplasmic levels of sugar content, have demonstrated that plant roots take up sugars supplied in growth media rather efficiently [41]. To comprehensively understand sugar uptake and allocation in plants, apoplasmic sugar levels, sugar distribution in subcellular organelles, and fluxes for specific sugars need to be monitored more closely. Further development of sugar nanosensors will hopefully lead to a better understanding of sugar sensing and signaling [42].
To learn more about the specific regulatory components involved in fructose signaling, we took advantage of a cell-based functional screen using transient expression of the A. thaliana mesophyll protoplast system [43]. Because fructose caused deficient chlorophyll accumulation in A. thaliana (Figure 1A and 1B), we reasoned that fructose signaling may affect photosynthetic gene expression in a manner similar to that of glucose signaling [7], [17]. To monitor the fructose signaling response in leaf mesophyll protoplasts, we generated a reporter construct with an approximately 0.5 kb CAB2 promoter fused to the firefly luciferase gene (CAB2-fLUC). In leaf mesophyll protoplasts, CAB2-fLUC activity was downregulated by fructose, but not by the osmotic control mannitol (Figure S3). We then screened several enzymes involved in fructose metabolism, including putative cytoplasmic FBP (AT1G43670), FRK1 (AT5G51830), and PFK1 (AT4G29220) for their potential roles in fructose signaling (Figure 2A). Of these enzymes, putative FBP (we tested two independent constructs, FBP_3 and FBP_4) had the greatest suppressive effect on CAB2-fLUC activity (Figure 2B). CAB2 promoter activity seemed to be suppressed even without high-fructose treatment, possibly because plant cells became hypersensitized to endogenous fructose when putative FBP was overexpressed. To test if putative FBP enzyme activity is required for CAB2 gene repression, we generated a catalytically inactive form, FBPS126AS127A (SSM), based on domain conservation in plant and animal FBPs (Figure S4). The dual mutation of S126A and S127A in FBP caused a loss of FBP enzymatic activity in protoplasts (Figure S5). This mutation probably distorted the local structure and prevented FBP121D from associating with a divalent ion that is necessary for the enzyme activity [45]. Interestingly, the catalytically inactive form FBPS126AS127A suppressed CAB2-fLUC activity in the same manner as the wild-type FBP (Figure 2B). This result indicates that the regulatory function of putative FBP in fructose signaling may be independent of its catalytic activity in sugar metabolism, similar to how HXK1 functions in glucose signaling [9], [19].
Surprisingly, we observed putative FBP in both the cytoplasm and nucleus (Figure 2C). We were not able to determine whether the nuclear localization of putative FBP depends on cellular fructose signaling (Figure S6), since it is almost impossible to generate zero-fructose conditions in plant cells. However, the nuclear localization of putative FBP certainly suggests that it could be directly involved in fructose-dependent gene regulation. Based on the initial functional screen and localization test in plant cells, we hypothesized that putative FBP was a regulatory factor in fructose signaling.
To study the role of putative FBP in fructose signaling in whole plants, we first obtained a T-DNA insertion mutant that did not accumulate full-length FBP transcript and genetically characterized FBP's function in fructose signaling (Figure 3A). The fins1 seedlings exhibited fructose-insensitive growth responses with progressive cotyledon greening with chlorophyll accumulation (Figure 3B) that was independent of osmotic effects (Figure S7), but displayed glucose-sensitive developmental arrest phenotypes. Since “fructose insensitivity” was the first phenotype that we encountered with this fbp mutant, we designated the allele fructose insensitive1 (fins1).
In fins1 protoplasts, FINS1 expression (using the two independent constructs FBP/FINS1_3 and FBP/FINS1_4) clearly suppressed CAB2-fLUC activity (Figure 3C), which was similar to the effect of HXK1 [7]. To investigate FINS1 function in fructose-mediated gene regulation, we examined marker gene expression in WT and fins1 seedlings grown on 6% fructose agar media with MS salts. Consistent with their growth phenotypes (Figure 3B), CAB2 expression was markedly repressed by fructose in WT, but not in fins1 seedlings (Figure 3D). A key transcription factor in ABA signaling, ABI4 (also known as GIN6/AT2G40220) [44] was induced by fructose in WT but not in fins1 seedlings. In contrast, ETHYLENE RESPONSE FACTOR1 (ERF1/AT3G23240), an ethylene response transcription factor [46], was repressed by fructose in WT, but de-repressed in fins1 seedlings. However, the change in ERF1 expression levels was relatively weak compared to other marker gene responses. These data showed that FINS1 had a central role in fructose-inducible gene regulation.
To verify that the fructose insensitivity exhibited by fins1 was due to the loss of FINS1, we complemented the fins1 mutant with FINS1 cDNA using an Agrobacterium system. Transgenic lines with FINS1 expression levels similar to that of WT were selected by gene transcript and protein levels using reverse transcriptase–dependent PCR and protein blot analysis, respectively (Figure 3E). The selected complementation lines had restored sensitivity to fructose and exhibited seedling developmental arrest similar to that of WT Col seedling (Figure 3F); this confirmed that loss of FINS1 function in fins1 seedling was responsible for fructose insensitivity. Furthermore, a fins1 mutant expressing catalytically inactive FBPS126AS127A also restored fructose sensitivity WT levels (Figure 3G and 3H and Figure S8A). The seedling response was specific to fructose, and did not occur in the presence of mannitol (Figure S8B). This response verified that the function of FINS1/FBP in fructose signaling was independent of its catalytic activity in sugar metabolism, as shown by the results of the cellular assay (Figure 2B).
As stated previously, unlike in the glucose assay, in which the high nitrogen levels of MS salts necessitated a high concentration of glucose, 2% fructose without MS salts did not cause the same phenotypic effect as 6% fructose with MS salts (Figure S2). Consequently, it was not clear whether fructose signaling was related to nitrogen signaling. To address this, we tested the effect of different concentrations of fructose on fructose-mediated seedling developmental responses without osmotic pressure, as well as the sugar-antagonistic effect of nitrate (Figure S9). At 3% fructose, fins1, FINS1-complemented fins1, gin2, HXK1-complemented gin2, and WT seedlings did not exhibit any obvious developmental phenotype (Figure S9), as was the case for 2% fructose (Figure S2). However, all of these seedlings exhibited severe growth repression at 5% fructose. Strikingly, at 4% fructose, fins1 showed a clear insensitivity, and FINS1-complemented fins1 restored seedling developmental arrest to a WT-like phenotype (Figure S9). The glucose-insensitive gin2 seedlings displayed consistent fructose-mediated developmental arrest phenotypes. Some of the extreme sensitivity of gin2 could have been due to its accession, because Ler was hypersensitive compared to Col at the same fructose concentration. These results confirmed that nitrogen affects fructose and glucose signaling in different ways [20]. Together with the initial cell-based functional screen, the reverse genetics analysis revealed the regulatory role of FINS1 in fructose signaling during early A. thaliana seedling establishment.
FBP isozymes have multiple roles in plant sugar metabolic pathways at different subcellular locales [47]. Chloroplast-localized FBP (AT3G54050) has 50% sequence homology to cytoplasmic FBP in A. thaliana and is mainly involved in starch biosynthesis [48]. Cytoplasmic FBP is involved in sucrose metabolism and is inactivated under dark conditions, mainly due to the increase in fructose-2,6-bisphosphate in some species [47], [49]. Consistent with these previous findings, etiolated WT, fins1, and FINS1-complemented fins1 seedlings did not show any striking phenotypic differences when they were grown on MS agar medium containing 6% glucose, fructose, or mannitol in completely dark conditions (Figure 4A–4C). This result suggested that FINS1 mainly mediated fructose signaling under light conditions (Figure 3B, 3F, and 3H).
The genetic repression of FINS1 results in shifting sugar metabolism in favor of starch over sucrose synthesis, but does not affect A. thaliana growth [47]. To physiologically compensate for the decrease in sucrose content during the day, starch breakdown and sugar export are enhanced at night in A. thaliana [47] and tobacco [50], but not in rice [49]. Because FBP/FINS1 plays a central role in sucrose synthesis, we tested whether low sucrose in fins1 was a direct cause of its fructose insensitivity [47], [50], [51]. When we observed seedling growth phenotypes on MS agar media containing 6%, 10%, or 12% sucrose in the presence of light, fins1 seedlings were resistant to developmental arrest at high concentrations of sucrose. However, gin2 was resistant only up to 10% sucrose (Figure 4D–4F), indicating that sucrose levels were irrelevant to the fructose insensitivity of fins1 seedlings.
Sucrose is converted to fructose and glucose or UDP-glucose and fructose in plant cells and then is likely integrated into FINS1-dependent or HXK1-dependent signaling. Thus, the strong sucrose resistance of fins1 seedlings (Figure 4F) indicated that fructose became a predominant hexose after sucrose hydrolysis [2], [23], [24]. This finding was supported by a previous observation using a fluorescence resonance energy transfer–based nanosensor, which showed that a measurable cytoplasmic glucose level was induced within 10–20 s of sucrose application to A. thaliana roots [41]. To obtain further molecular insights into the interconnected nature of sugar signaling, we have currently performing a comprehensive analysis of transcriptome changes.
In summary, the fructose insensitivity of fins1 seedlings was most likely not caused by the loss of FBP catalytic activity or by lower sucrose in the mutant [47], because (1) the fructose-responsive CAB2 promoter activity was modulated by FINS1/FBP, but not by FRK1, which is also involved in the sucrose synthesis (Figure 2B); (2) the fructose signaling response was modulated similarly by catalytically active or inactive forms of FBP (Figure 2B and Figure 3H); and (3) high sucrose did not induce fins1 seedling developmental arrest (Figure 4F).
Upon fructose treatment, we noted a slightly more inhibition of root growth (Figure 3B, Figure S1) and a marked ABA-dependent gene response (Figure 3D). These results led us to examine the interaction between fructose and ABA signaling. To do so, we generated transgenic gin1 seedlings that overexpress FINS1. We then analyzed the epistatic relationship between FINS1 in fructose signaling and GIN1 in the ABA pathway. FINS1-overexpressing gin1 seedlings exhibited a seedling developmental arrest phenotype like that observed in WT seedlings on 6% fructose agar medium with MS salts (Figure 5A). The fructose-dependent seedling response was not due to high osmotic effects, because seedlings grew similarly on 6% mannitol agar medium with MS salts (Figure S10). Thus, fructose signaling appears to be integrated into FINS1 downstream of GIN1, which is involved in ABA synthesis. Interestingly, gin1 seedlings that overexpress the plant glucose sensor AtHXK1 display glucose insensitivity, suggesting that glucose sensing by AtHXK1 occurs upstream of ABA synthesis [32]. Taken together, these findings indicate that although both fructose and glucose signaling crosstalk with ABA signaling during early seedling establishment, FINS1 and HXK1 function downstream and upstream of the ABA pathway, respectively.
To further test whether FINS1 has a critical role in the ABA pathway, WT, gin1, and FINS1-overexpressing gin1 seedlings were grown on MS agar media containing different concentrations of ABA (Figure 5B–5D). All of the seedlings displayed characteristic developmental arrest phenotypes at a saturated level of 1 µM ABA (Figure 5B). Notably, FINS1-overexpressing gin1 seedlings, but not WT or gin1 seedlings, displayed similar growth inhibition at a sub-potent level of 0.5 µM ABA (Figure 5C). This result supports the notion that FINS1-dependent fructose signaling worked downstream of ABA synthesis (Figure 5A). Because these transgenic lines did not show any growth inhibition in the absence ABA (Figure 5D), it is unlikely that the growth response of the FINS1-overexpressing gin1 was caused by accelerated ABA synthesis rather than increased sensitivity to ABA.
Based on the results shown in Figure 5, we decided to investigate the definitive role of FINS1 in ABA signaling. When seedling growth was observed on MS agar media containing 1 µM ABA, fins1 and the constitutive ethylene signaling mutant ctr1 exhibited ABA insensitivity compared to WT, gin1, and gin2 (Figure 6A). Nevertheless, the fins1 phenotype clearly differed from that of ctr1 seedlings, suggesting that the ABA insensitivity of fins1 may not be directly related to an alteration in ethylene sensitivity.
To elucidate the function of FINS1 in ABA-mediated gene regulation, we monitored the gene expression of ABI1 (an ABA negative regulator) and ABI3, ABI4, and ABI5 (ABA positive regulators) in fructose-insensitive fins1, fructose/glucose-insensitive gin1 and ctr1, and glucose-insensitive gin2 seedlings, and as well as in WT seedlings. ABI1 expression was higher in fins1, ctr1, and gin2 compared to its expression in WT and gin1 (Figure 6B). In contrast, expression of the ABA positive regulators (ABI3, ABI4, and ABI5) was suppressed in fins1, and suppressed to an even greater extent in ctr1 (Figure 6C–6E). The higher level of gene suppression in ctr1 correlated with its stronger ABA-insensitive response (Figure 6A). The ABA-dependent seedling phenotypes and gene expression patterns of fins1 further supported the idea that fructose signaling closely interacted with ABA signaling through FINS1. Unlike HXK1 in glucose signaling, FINS1 may not acts as a fructose sensor, because FINS1 binds more readily to fructose-1,6-bisphosphate than to fructose for its catalytic activity (Figure S5). However, it remains to be determined if fructose directly binds to putative FBP and acts as an allosteric regulator of the protein. Further elucidation of the biochemical and cellular processes underlying the interactions between GIN1 and FINS1 will provide a better mechanistic understanding of how fructose signaling controls early seedling establishment.
We have identified fructose as a novel hexose signal that modulates early establishment of A. thaliana seedlings via a pathway that is distinct from glucose signaling (Figure 1). Genetic analyses revealed that fructose signaling interacted positively with ABA and negatively with ethylene, similar to high glucose signaling. Using a cell-based functional screen and reverse genetic analysis, we uncovered a regulatory role for FINS1/FBP in fructose signaling that is independent of its catalytic activity (Figure 2 and Figure 3). fins1 seedlings also showed sucrose insensitivity, indicating that alteration of sucrose content by loss of FINS1 is irrelevant to the fructose insensitivity of fins1 (Figure 4).
The growth response of transgenic gin1 seedlings expressing FINS1 to fructose and ABA indicated that fructose signaling was acting downstream of ABA synthesis (Figure 5). The ABA response was consistently compromised in fins1 seedlings (Figure 6). Further explorations of the biochemical connections among GIN1/ABA2, GIN2/HXK1, and FINS1/FBP within a sugar-signaling context will provide a better mechanistic understanding of hexose signaling processes during early seedling establishment (Figure 7). However, it is apparent that multiple layers of interactions/cross-talk among glucose, fructose, and ABA signaling pathways tightly modulate plant growth promotion and inhibition, and provide developmental plasticity during the plant autotrophic transition following seed germination.
Approximately 0.5 kb of the CAB2 promoter was amplified by PCR and fused to LUC to create the CAB2-fLUC reporter construct [38]. All of the effector constructs were generated by inserting the cDNA between the 35SC4PPDK promoter and the NOS terminator in a plant expression vector for protoplast transient assays and then verifying by DNA sequencing.
Plants were grown in soil at 23°C for 20–22 d under 60 µmol/m2/s with a 13 h photoperiod. Protoplast isolation and transient expression assays were carried out as described previously [38]. All of the protoplasts transient assays were performed with UBQ10-renillaLUC (UBQ10-rLUC) as an internal control. The reporter activities were calculated based on the fLUC/rLUC ratio and normalized to the values obtained without treatment or effector expression.
Plasmid constructs for transgenic plants were generated by inserting the cDNA of FINS1 between the 35SC4PPDK promoter and the NOS terminator in a mini-binary vector pCB302 [8] and expressing it in fins1 or gin1 mutant plants. The transgenic lines expressing transgenes at levels similar to those of WT were selected and used for further analyses. We analyzed the phenotypes of transgenic plants/seedlings from at least two independent lines at the T2 or T3 generation, except for catalytically inactive FINS1_ssm-complemented fins1 (cSSM), which was used at the T1 generation. FINS1/FBP protein expression was analyzed using a cytoplasmic fructose-1,6-bisphosphatase–specific antibody (Agrisera, #AS04043) or HA antibody (Roche).
For sugar repression assays, seedlings were grown on MS (Caisson Laboratories) agar medium containing 6% glucose (Sigma), fructose (Sigma), or mannitol (Sigma) for 5 d under constant light (60 µmol/m2/s). A germination test was performed to determine the ABA sensitivity of each genotype grown on half-strength MS agar medium containing 1% sucrose and a designated amount of ABA under a photoperiod of 16 h light/8 h dark. For the sucrose assay, seedlings were grown on 6, 10, or 12% sucrose MS agar medium with a photoperiod of 16 h light/8 h dark until they showed a clear phenotype. For each experiment, seeds were stratified at 4°C for 4 d before plating. The results were confirmed through several replications.
For gene expression analysis, total RNA was isolated by the Trizol method (Invitrogen) and 1 µg of total RNA was used for cDNA synthesis [15]. We investigated glucose- and fructose-mediated gene regulation and their interactions with ABA and ethylene signaling by monitoring marker gene expression in WT and hormone mutants. Gene expression was quantitatively measured using real-time PCR with cDNA templates generated from the RNA of 5-d-old seedlings grown on MS media containing 6% glucose, fructose, or mannitol. Gene expression values in seedlings grown on mannitol served as osmotic controls. Real-time PCR was carried out with iQ SYBR Green dye-added PCR mix (Bio-Rad). Tubulin4 (AT1G04820) or elongation initiation factor4a (ELF4a, AT3G13920) transcript was used as a real-time PCR control with gene-specific primers. Detailed primer sequences are listed in Table S1. Each primer set was pretested by PCR for a single gene product. Experiments were repeated three times with consistent results.
|
10.1371/journal.pcbi.1000171 | Sizing Up Allometric Scaling Theory | Metabolic rate, heart rate, lifespan, and many other physiological properties vary with body mass in systematic and interrelated ways. Present empirical data suggest that these scaling relationships take the form of power laws with exponents that are simple multiples of one quarter. A compelling explanation of this observation was put forward a decade ago by West, Brown, and Enquist (WBE). Their framework elucidates the link between metabolic rate and body mass by focusing on the dynamics and structure of resource distribution networks—the cardiovascular system in the case of mammals. Within this framework the WBE model is based on eight assumptions from which it derives the well-known observed scaling exponent of 3/4. In this paper we clarify that this result only holds in the limit of infinite network size (body mass) and that the actual exponent predicted by the model depends on the sizes of the organisms being studied. Failure to clarify and to explore the nature of this approximation has led to debates about the WBE model that were at cross purposes. We compute analytical expressions for the finite-size corrections to the 3/4 exponent, resulting in a spectrum of scaling exponents as a function of absolute network size. When accounting for these corrections over a size range spanning the eight orders of magnitude observed in mammals, the WBE model predicts a scaling exponent of 0.81, seemingly at odds with data. We then proceed to study the sensitivity of the scaling exponent with respect to variations in several assumptions that underlie the WBE model, always in the context of finite-size corrections. Here too, the trends we derive from the model seem at odds with trends detectable in empirical data. Our work illustrates the utility of the WBE framework in reasoning about allometric scaling, while at the same time suggesting that the current canonical model may need amendments to bring its predictions fully in line with available datasets.
| The rate at which an organism produces energy to live increases with body mass to the 3/4 power. Ten years ago West, Brown, and Enquist posited that this empirical relationship arises from the structure and dynamics of resource distribution networks such as the cardiovascular system. Using assumptions that capture physical and biological constraints, they defined a vascular network model that predicts a 3/4 scaling exponent. In our paper we clarify that this model generates the 3/4 exponent only in the limit of infinitely large organisms. Our calculations indicate that in the finite-size version of the model metabolic rate and body mass are not related by a pure power law, which we show is consistent with available data. We also show that this causes the model to produce scaling exponents significantly larger than the observed 3/4. We investigate how changes in certain assumptions about network structure affect the scaling exponent, leading us to identify discrepancies between available data and the predictions of the finite-size model. This suggests that the model, the data, or both, need reassessment. The challenge lies in pinpointing the physiological and evolutionary factors that constrain the shape of networks driving metabolic scaling.
| Whole-organism metabolic rate, B, scales with body mass, M, across species as [1](1)where B0 is a normalization constant and α is the allometric scaling exponent, typically measured to be very close to 3/4 [2]. The empirical regularity expressed in Equation 1 with α = 3/4 is known as Kleiber's Law [3],[4].
Many other biological rates and times scale with simple multiples of 1/4. For example, cellular or mass-specific metabolic rates, heart and respiratory rates, and ontogenetic growth rates scale as M−1/4, whereas blood circulation time, development time, and lifespan scale close to M1/4 [5]–[9]. Quarter-power scaling is also observed in ecology (e.g., population growth rates) and evolution (e.g., mutation rates) [2],[10],[11]. The occurrence of quarter-power scaling at such diverse levels of biological organization suggests that all these rates are closely linked. Metabolic rate seems to be the most fundamental because it is the rate at which energy and materials are taken up from the environment, transformed in biochemical reactions, and allocated to maintenance, growth, and reproduction.
In a series of papers starting in 1997, West, Brown, and Enquist (WBE) published a model to account for the 3/4-power scaling of metabolic rate with body mass across species [1], [12]–[14]. The broad theory of biological allometry developed by WBE and collaborators attributes such quarter-power scaling to near-optimal fractal-like designs of resource distribution networks and exchange surfaces. There is some evidence that such designs are realized at molecular, organelle, cellular, and organismal levels for a wide variety of plants and animals [2],[14].
Intensifying controversy has surrounded the WBE model since its original publication, even extending to a debate about the quality and analysis of the data [15]–[28]. One of the most frequently raised objections is that the WBE model cannot predict scaling exponents for metabolic rate that deviate from 3/4 [16],[29], even though the potential for such deviations was appreciated by WBE themselves [1]. If this criticism were true, WBE could not in principle explain data for taxa whose scaling exponents have been reported to be above or below 3/4 [29]–[34], or deviations from 3/4 that have been observed for small mammals [35]. Likewise, the WBE model would be unable to account for the scaling of maximal metabolic rate with body mass, which appears to have an exponent of 0.88 [36]. It is important to note that the actual nature of maximal metabolic rate scaling is, however, not without its own controversy; see [37] for an argument that maximal metabolic rate scales closer to 3/4 when body temperature is taken into consideration.
Much of the work aimed at answering these criticisms has relied on alteration of the WBE model itself. Enquist and collaborators account for different scaling exponents among taxonomic groups by emphasizing differences in the normalization constant B0 of Equation 1 and deviations from the WBE assumptions regarding network geometry [26], [38]–[40]. While these results are suggestive, it remains unclear whether or not WBE can predict exponents significantly different from 3/4 and measurable deviations from a pure power law even in the absence of any variation in B0 and with networks following exactly the geometry required by the theory. Although WBE has been frequently tested and applied [40]–[53], it is remarkable that no theoretical work has been published that provides more detailed predictions from the original theory. Also, work aimed at extending WBE by relaxing or modifying some of its assumptions has hardly been complete; many variations in network structure might have important and far-reaching consequences once properly analyzed. This is what we set out to do in the present contribution. We show that a misunderstanding of the original model has led to the claim that WBE can only predict a 3/4 exponent. This is because many of the predictions and tests of the original model are derived from leading-order approximations. In this paper we derive more precise predictions and tests.
For the purpose of stating our conclusions succinctly, we refer to the “WBE framework” as an approach to explaining allometric scaling phenomena in terms of resource distribution networks (such as the vascular system) and to the “WBE model” as an instance of the WBE framework that employs particular parameters specifying geometry and (hydro)dynamics of these networks [1],[14]. (We shall detail these assumptions and define terminology more accurately in section “Assumptions of the WBE model”.)
Our main findings are: 1. The 3/4 exponent only holds exactly in the limit of organisms of infinite size. 2. For finite-sized organisms we show that the WBE model does not predict a pure power-law but rather a curvilinear relationship between the logarithm of metabolic rate and the logarithm of body mass. 3. Although WBE recognized that finite size effects would produce deviations from pure 3/4 power scaling for small mammals and that the infinite size limit constitutes an idealization [1], the magnitude and importance of finite-size effects were unclear. We show that, when emulating current practice by calculating the scaling exponent of a straight line regressed on this curvilinear relationship over the entire range of body masses, the exponent predicted by the WBE model can differ significantly from 3/4 without any modifications to its assumptions or framework. 4. When realistic parameter values are employed to construct the network, we find that the exponent resulting from finite-size corrections comes in at 0.81, significantly higher than the 3/4 figure based on current data analysis. 5. Our data analysis indeed detects a curvilinearity in the relationship between the logarithm of metabolic rate and the logarithm of body mass. However, that curvilinearity is opposite to what we observe in the WBE model. This implies that the WBE model needs amendment and/or the data analysis needs reassessment.
Beyond finite-size corrections we examine the original assumptions of WBE in two ways. First, we vary the predicted switch-over point above which the vascular network architecture preserves the total cross-sectional area of vessels at branchings and below which it increases the total cross-sectional area at branchings. These two regimes translate into different ratios of daughter to parent radii at vessel branch points. Second, we allow network branching ratios (i.e., the number of daughter vessels branching off a parent vessel) to differ for large and small vessels. We analyze the sensitivity of the scaling exponent with respect to each of these changes in the context of networks of finite size. This approach is similar in spirit to Price et al. [40], who relaxed network geometry and other assumptions of WBE in the context of plants. In the supplementary online material Text S1, we also argue that data analysis should account for the log-normal distribution of body mass abundance, thus correcting for the fact that there are more small mammals than large ones. Despite differences in the structure and hydrodynamics of the vascular systems of plants and animals [1],[13], detailed models of each yield a scaling exponent of 3/4 to leading-order. In the present paper, we focus on the WBE model of the cardiovascular system of mammals. All of our assumptions, derivations, and calculations should be interpreted within that context. Finite-size corrections and departures from the basic WBE assumptions are important in the context of plants as well, as shown in recent studies by Enquist and collaborators [26], [38]–[40].
In final analysis, we are led to the seemingly incongruent conclusions that (1) many of the critiques of the WBE framework are misguided and (2) the exact (i.e., finite-size corrected) predictions of the WBE model are not fully supported by empirical data. The former means that the WBE framework remains, once properly understood, a powerful perspective for elucidating allometric scaling principles. The latter means that the WBE model must become more respectful of biological detail whereupon it may yield predictions that more closely match empirical data. Our work explores how such details can be added to the model and what effects they can have.
The paper is organized as follows. For the sake of a self-contained presentation, we start with a systematic overview of the assumptions, both explicit and implicit, underlying the WBE theory (section “Assumptions of the WBE model”). In Text S1, we provide a detailed exposition of the hydrodynamic derivations that the model rests upon. These calculations are not original, but they have not appeared to a full extent before in the literature. While nothing in section “Assumptions of the WBE model” is novel, there seems to be no single “go to” place in the WBE literature that lays out all components of the WBE theory. Our paper then proceeds with a brief derivation of the exact, rather than approximate, relationship between metabolic rate and body mass (section “Derivation of the 3/4 scaling exponent”). We then calculate the exact predictions for scaling exponents for networks of finite size (section “Finite-size corrections to 3/4 allometric scaling”) and revisit certain assumptions of the theory (section “Making the WBE model more biologically realistic”). In section “Comparison to empirical data” we compare our results to trends detectable in empirical data. We put forward our conclusions in the Discussion section.
The WBE model rests on eight assumptions. Some of these assumptions posit the homogeneity of certain parameters throughout the resource distribution network. Any actual instance of such a network in a particular organism will presumably exhibit some heterogeneity in these parameters. The object of the theory is a network whose parameters are considered to be averages over the variation that might occur in any given biological instance. For the sake of brevity, we refer to such a network as an “averaged network”. The impact of parameter heterogenity on the scaling exponent is very difficult to determine analytically. (Section “Changing branching ratio across levels” addresses a modest version of this issue numerically.)
Using the above assumptions, we can derive how metabolic rate, B, varies with body mass, M, which is the fundamental result of WBE. The key insight is that body mass is proportional to blood volume (following from Assumption 6) and that blood volume is the sum of the volumes of the vessels over all the levels of the network. Using Assumptions 2–6, this sum can be expressed in terms of properties of capillaries, providing a direct link to metabolic rate (owing to Assumption 8). Upon expressing blood volume in terms of capillary properties, we can separate terms of the sum that are invariant (by Assumption 7) from others that vary with the total number of capillaries. The total number of capillaries is directly proportional to the whole-oganism metabolic rate, because each capillary supplies resources at the same rate regardless of organism size (Assumption 7). This ties body mass to metabolic rate. We now provide the formal derivation.
Using Assumptions 2–4 the total blood volume or total network volume (assuming the network is completely filled with blood and ignoring the factor of 2 that may arise from blood in the venous system, which returns blood to the heart) can be expressed as the sum(5)where the volume of a vessel is that of a cylinder. Next, we use the scale-free ratios γ, β<, and β>, defined by Equations 2–4 resulting from Assumptions 5 and 6, to connect level k to successively higher levels and all the way to the capillary level N:(6)where is the volume of a capillary and Ncap = NN = nN the number of capillaries. The first sum ranges over the area-preserving regime and the second sum is over the area-increasing regime. The first sum is a standard geometric series. Observing that then yields(7)where N̅ is the fixed number of levels from the capillaries to the level where the transition from area-increasing to area-preserving branching occurs. Since , we have , and the second sum in Equation 6 is simply(8)Combining these results we have(9)This equation can be re-expressed as(10)where(11)are both constant with respect to body mass. Equation 10 will play a fundamental role in the following sections.
Given this simple relation between total blood volume (or network volume) and the number of capillaries, it is straightforward to relate metabolic rate, B, to body mass, M. Using Assumption 8, the whole-body metabolic rate is just the sum total of the metabolic rates enabled by the resources delivered through each capillary. Let the contribution to total metabolic rate enabled by a capillary be Bcap. By Assumption 7 Bcap is constant across organisms. Thus, B = NcapBcap, or simply B∝Ncap. Inserting this into Equation 10, invoking Assumption 7 that Vcap is independent of body mass, and using Assumption 6 to recognize that M∝Vblood yields(12)with C2 a constant, , , and are new constants.
Letting the number of levels in the cardiovascular system, N, tend to infinity—which necessarily means that body mass, M, and metabolic rate, B∝Ncap = nN, become infinitely large—we conclude that(13)This is the celebrated result that has been empirically observed for nearly a century. Equation 13 is approximately true as long as(14)
It is essential to realize that the prediction of a 3/4 scaling relationship only holds in the infinite M-limit. The approximation becomes less accurate as organisms become smaller, corresponding to smaller metabolic rate B. The exact relationship is Equation 12, or 10, which does not predict a pure power law but a curvilinear graph of ln B versus ln M. Forcing such a curve to fit a straight line will therefore not produce an exact value of 3/4, except when the magnitude of the correction term is small compared with 1 (see Equation 14), the measurement error, or the residual variation in the empirical data. Given that the importance of these deviations will be larger for smaller organisms, it would in principle be interesting to look more carefully at finite-size effects for small fish or plants, because the smallest mammals are considerably larger than the smallest fish or plants [25],[26],[29],[33], although we do not perform such an analysis here. Different taxa often span different ranges of body size and exhibit a particular relative proportion of small to large organisms. These characteristics will likely lead to different measured scaling exponents.
We conclude that the WBE model actually predicts variation in scaling exponents due to finite-size terms whose magnitude depends on the absolute range of body masses for a given taxonomic group. These predictions can be tested against the allometric exponents reported in the empirical literature.
To quantify finite-size corrections, we focus on Equation 10 because the blood volume, Vblood (∝M), and the number of capillaries, Ncap (∝B), are really the fundamental parameters of the theory. Proceeding in this way, we avoid the additional constants C2 and Bcap. By inspecting Equation 10, we see that finite-size effects can become manifest in two different ways. First, even in the absence of network regions with area-increasing branching (N̅ = 0), there are corrections to the 3/4 scaling exponent. Second, the switch-over point N̅ from area-preserving branching to area-increasing branching determines the relative contributions of these two network regimes, and has the potential to considerably influence the scaling exponent. To quantify these effects, we consider three cases: (i) a network with only area-preserving branching (section “Networks with only area-preserving branching”), (ii) a network with only area-increasing branching (section “Networks with only area-increasing branching”), and (iii) a mixture of the two with a transition level (section “Networks with a transition from area-preserving to area-increasing branching”).
The discrepancies between the WBE model and data might be addressed in several ways: (i) by correcting for biases in the empirical distributions of species masses; (ii) by adding more detail to any of the WBE assumptions; (iii) by relaxing the assumptions. In Text S1 (Figure S4), we exemplify case (i) by accounting for the fact that most mammals, in particular those that have been measured, are of small mass. The body-size distribution across species is approximately log-normal. By sampling body sizes according to such a distribution and using the same numerical methods as in section “Finite-size corrections to 3/4 allometric scaling” above, we determined that the overall effect on the scaling exponent is essentially negligible (the exponent is slightly lowered). In section “Modifying the transition level between area-preserving and area-increasing regimes”, we illustrate approach (ii) by altering the level at which the transition from area-preserving to area-increasing branching occurs, as well as the width of the region over which it extends, as motivated by complexities in the hydrodynamics of blood flow. These considerations affect the scaling exponent, but the change is too small to restore the 3/4 figure. In section “Changing branching ratio across levels”, we illustrate approach (iii) by relaxing the assumption of a constant branching ratio (Assumption 4). We show that systematic changes in the branching ratio can significantly lower the measured scaling exponent and lead to intriguing non-linear effects that depend on where the transition from one branching ratio to another occurs.
Savage et al. [23] published an extensive compilation of empirical data for basal metabolic rate and body mass of 626 mammals. In this section we compare the dependency of scaling exponents on body mass as obtained from this dataset to our predictions for scaling exponents with finite-size corrections. We sorted organisms according to body mass and grouped them, starting with the smallest exemplar, into disjoint bins spanning one order of magnitude each. We then analyzed this data compilation in three ways. First, we determined the scaling exponents for successive cumulations of bins. At each addition of a bin, we computed a linear regression on the entire cumulated data, plotting the resultant scaling exponent against the range of sizes. In other words, the first scaling exponent is determined for the first order of magnitude in body mass, the second exponent is determined for the first two orders of magnitude, and so on. This is similar in spirit to the procedure used for analyzing and presenting the numerical data in section “Finite-size corrections to 3/4 allometric scaling”. The result is shown in Figure 10A. In a second approach we proceeded similarly, but starting with the largest order of magnitude in body mass, then successively adding bins of smaller orders (Figure 10B). Lastly, we computed the scaling exponent for each bin separately (Figure 10C).
The panels of Figure 10 show the results with error bars based on the 95% confidence intervals obtained from ordinary least squares (OLS). In panels 10A and 10B, the exponents exhibit an increasing trend with body mass. Panel 10C shows a similar trend for bins that correspond to intermediate mass ranges. These are the bins that contain most of the data points. There is too much scatter at either end of the body mass distribution to make a statement about the entire range for panel 10C. We find that for those ranges and aggregations with smallest scatter (as determined from error bars), the scaling exponent approaches the 3/4 figure from below. Although these data are suggestive, it would be incautious at this point to assert that the data flatten out at 3/4 for some maximum mammalian size. Given the current dataset, however, an “asymptotic” 3/4 scaling seems a reasonable guide.
The concave increase of the scaling exponent with body mass is most consistent with a finite-size WBE model based on pure area-preserving branching throughout the network, see section “Networks with only area-preserving branching”. (The concave increase of the scaling exponent, Figure 3B, corresponds to a convex relationship between metabolic rate and body mass, see the schematic in Figure 2.) Recall that in our numerical studies of section “Networks with only area-preserving branching” the scaling exponent approached 3/4 in a concave fashion from below, while networks built entirely with area-increasing branching (section “Networks with only area-increasing branching”) have scaling exponents that always lie above 3/4, converging to an accumulation point at 1. Networks built with a mixture of these branchings (section “Networks with a transition from area-preserving to area-increasing branching”), approach 3/4 scaling in a convex fashion from above, opposite to the trends seen in Figure 10. (The convex decrease of the scaling exponent, Figure 6A, corresponds to a concave relationship between metabolic rate and body mass, see the schematic in Figure 5.)
A similar analysis of a more limited dataset for heart rate (26 data points) and respiratory rate (22 data points) [23] also shows a trend that is not easily reconciled with our finite-size corrections for networks with a mixture of area-preserving and area-increasing branching. In WBE, heart rate ω and respiratory rate R are both predicted to scale as ω∝R∝M−α/3 (see Table S1 and related text in section “Impact of finite-size corrections on additional WBE predictions” of Text S1). Since our calculations in section “Networks with a transition from area-preserving to area-increasing branching” yield scaling exponents, α, that approach 3/4 from above as body mass increases, we expect the scaling of heart and respiratory rates to both have exponents that are bounded by the maximum value of −1/4. The WBE model with finite-size corrections predicts α≈0.81. Hence, heart and respiratory rates should scale as M−0.27 and asymptote to −1/4 with increasing mass. That is, there should be very little change in the scaling exponent when analyzing data for either small or large mammals. This does not match empirical heart rate data well. Regressing on the first three, four, and six orders of magnitude in body mass yields exponents of −0.33, −0.27, and −0.25, respectively. The match is worse for respiratory rate data. Regressing on the first two, three, five, and seven orders of magnitude in body mass gives exponents of −0.64, −0.44, −0.34, and −0.26, respectively. We observe a convergence to −1/4, but over a much larger range of scaling exponents than expected.
While the WBE model has been predominantly interpreted in the context of interspecific scaling [9],[23], metabolic rate also varies with body mass during development. Such intraspecific data [29],[69] sometimes exhibits a concave curvature across growth stages ranging from young to adult mammals. Indeed, our finite-size corrections for the canonical WBE model predict a concave curvature of ln B versus ln M. However, they also predict an asymptotic approach to a slope of 3/4 for large mammals, and thus a fitted OLS slope for the entire body mass range that is greater than 3/4, as schematically shown in Figure 5. In his Table 5, Glazier [29] reports slopes from 29 intraspecific regressions for 14 species of mammals. From these, we compute an average slope α = 0.70; in this dataset, 20 of the slopes are smaller than 3/4 and only 9 of the slopes are larger than 3/4. This is inconsistent with our predictions. Moreover, the average body mass range of mammals, for which Glazier reports intraspecific regressions, spans only half an order of magnitude. Yet, our calculations show that several orders of magnitude in body mass are required to detect curvilinearity from finite-size effects, as seen in Figure 6A. We thus conclude that the curvature revealed by these intraspecific datasets is either unrelated to finite-size effects or fails to support the finite-size corrected canonical WBE model.
It is important to note that empirical data for the inter- and intraspecific case (especially for restricted size classes) are rather limited. We therefore do not wish to overstate the strength of our conclusions. We merely report discrepancies between the predictions of the canonical WBE model and limited sets of data. We anticipate that further data acquisition, statistical analysis, and model refinement will bring theory and data into agreement.
Over the past decade, the WBE model has initiated a paradigm shift in allometric scaling that has led to new applications (e.g., [2],[70],[71]), new measurements and the refinement of data analysis (e.g., [41]–[53]), and the recognition of connections between several variables that describe organismic physiology [1],[23]. However, WBE has also drawn intense criticism and sparked a heated debate [15]–[28].
In section “Assumptions of the WBE model”, we provide a detailed presentation of the complete set of assumptions and calculations defining the WBE model. While none of these originated with us, the literature lacked, surprisingly, an exhaustive exposition. (In particular, the consequences of Assumption 6 are a distillation of hydrodynamical calculations that we summarize in Text S1.) In section “Derivation of the 3/4 scaling exponent”, we connect each step in the derivation of the main WBE result to the assumptions it invokes. In this way, we provide a self-contained platform for motivating, deriving, and interpreting our results.
One of our main objectives is to clarify that the WBE model predicts (and thus “explains”) the 3/4 exponent of the scaling law relating whole-organism metabolic rate to body mass only as the limit of infinite network size, body mass, and metabolic rate is approached. Although this fact was appreciated by WBE in their original work [1] the nature of this approximation has been broadly misunderstood in the subsequent literature, e.g., [16],[29]. In this work, we conduct a systematic exploration of finite-size effects in the WBE framework and find that these effects yield measurable deviations from the canonical 3/4 scaling exponent, shifting the actual prediction to a value closer to 0.81 when published parameters are employed [1],[14]. This finding has major implications and immediately clarifies some contentious issues. On the one hand, the common criticism that the WBE model can only predict a scaling exponent of 3/4 is incorrect. As we show in section “Finite-size corrections to 3/4 allometric scaling”, a continuum of exponents can be obtained as a function of body-mass. On the other hand, the 0.81 figure (obtained for N̅ = 24 and n = 2) shifts the predicted exponent for mammals away from the canonical figure of 3/4 that reflects current data analysis. In section “Impact of finite-size corrections on additional WBE predictions” of Text S1 we report the finite-size corrections for several variables related to vascular physiology that were documented in the original WBE paper [1].
A major consequence of the curvilinear relationship between ln B and ln M predicted by the model is the fact that the scaling exponent, as measured by a simple power law regression, will show a dependence on the absolute masses of the organisms in question. Notably, our numerical calculations for area-increasing branching in Figure 4 are consistent with the linear scaling of metabolic rate versus body mass that has been observed for small fish [32]. Indeed, with minor modifications, our Equations 19 and 20 could be used to test the form of isometric scaling observed in young and small fish. It should be noted, however, that the magnitude of these finite-size corrections depends strongly on certain network properties, such as N̅.
Furthermore, we find evidence for size-dependent relationships in the available empirical data for mammals (section “Comparison to empirical data”). Specifically, we find that the measured scaling exponent tends to increase with body mass, indicating that the empirical data (of log metabolic rate versus log body mass, or, equivalently, ln Ncap versus ln Vblood) exhibits convex curvature (i.e., the type of relationship dramatized in Figure 2). However, networks constructed with a mixture of area-increasing and area-preserving branching can never produce scaling relationships with exponents less than 3/4 and, although 3/4 scaling is approached in the limit of networks of infinite size, the exponents always approach 3/4 from above (unlike in Figure 10). Mixed networks of this type display inherently concave curvature of the log metabolic rate versus log body mass relationship (i.e., the type of relationship dramatized in Figure 5). That is, a group of organisms of larger sizes will yield smaller fitted exponents than a group of organisms of smaller sizes. Yet, empirical data are best fit by a power law with an exponent less than 3/4 and demonstrate convex curvature in several datasets of log metabolic rate versus log body mass. Thus, assuming that this represents the actual curvature in nature, either (i) a transition between radial scaling regimes does not occur, potentially contradicting Assumption 6 of the WBE model, or (ii) at least one assumption of the WBE model must be modified.
The case for pure area-increasing branching (hypothesis (i) above) within the WBE model is somewhat problematic. The only way for such a network to be consistent with Assumption 6 would be to posit that the transition from area-preserving to area-increasing regimes occurs at a vessel radius smaller than a capillary; in this case, this transition would in principle exist but would simply never actually be observed in nature. A number of facts contradict this explanation. For one, estimates place the transition at vessel radii of about 1 mm. Despite the fact that predictions of where the transition might occur are problematic (see section “Modifying the transition level between area-preserving and area-increasing regimes”), the estimate is unlikely to be 3 orders of magnitude larger than the actual value (since capillary radii are on the order of 1 µm in radius). A further complication is that a pure area-preserving network would theoretically not be able to “slow down” blood flow due to the conservation of volume flow rate for an incompressible fluid. The fact that blood flows much more quickly in the aorta than it does in the capillaries would tend to argue that area-increasing branching must occur somewhere in the network. Finally, there is the simple fact of Murray's Law; empirical findings squarely place β for small vessels in the neighborhood of n−1/3, strongly implying that area-increasing branching is in fact dominant when vessel radii are small [54],[57].
In our hands, empirical data seem most consistent with networks built with purely area-preserving branching, although the lack of very high-quality data for both metabolic rate and body mass makes it difficult to be absolutely certain of this trend. The reasoning outlined above makes hypothesis (i) appear somewhat unlikely. This leaves us with a riddle: cardiovascular networks with architectures that support the scaling trends observed for real organisms would seem to violate Assumption 6 of the WBE model. We are thus led to believe that some modification of assumptions 2–8 is necessary to explain the concavity in the data and an empirical scaling exponent less than 3/4. While a model that aligns with the empirical evidence might differ from the canonical WBE model (assumptions 2–8 plus specific values for the parameters N̅ and n), we believe such a model will squarely remain within the WBE framework (assumption 1, that is, the exploration of allometric scaling in the context of resource distribution networks).
Resolving this paradox will likely require intensive further data analysis and extension of the canonical WBE model. It is clear that work in this area would benefit from a more detailed empirical understanding of cardiovascular networks themselves. Although data for the coronary artery in humans, rats, and pigs exist [51]–[54],[72],[73], along with measurements for the vascular system in the lungs of armadillos [74], stringent tests of the core WBE assumptions require measurements throughout the body, in a larger variety of species, and for vessels farther away from the heart. Measurements are needed especially for the number of levels from the heart to the capillaries for different species, the scaling ratios of vessel radii (β = rk+1/rk) and vessel lengths (γ = lk+1/lk), vessel blood flow rates, and branching ratios (n = Nk+1/Nk). Such data will help to assess the extent to which mammalian vascular systems are space filling (Assumption 5), the scope of area-preserving and area-increasing branching (Assumption 6), the value(s) of the branching ratio throughout the network (Assumption 4), and the degree of symmetry or asymmetry in branchings and scaling ratios (Assumption 3). Analyzing intraspecific variation in network geometry may also enable a quantification of selection pressures for optimality with respect to energy loss, as implied by Assumption 6 (see Figures S1 and S2 in Text S1). Advances in fluorescent microspheres [74], plasticene casting, imaging, and image analysis all hold promise for a careful gauging of the vascular system.
In this paper we have begun the process of relaxing some assumptions of the canonical model. Although these modifications produce interesting results, they do not fully address the riddles discussed above. Addition of further biological realism, such as asymmetric branching or the flow characteristics of the slurry of blood cells at small vessel sizes, may generalize the WBE model from an asymptotic predictor of metabolic scaling into a universal theory that provides an understanding of which properties of resource distribution networks are most relevant for metabolic scaling in any given biological context. This will enable testing the very soundness of the WBE framework (Assumption 1) and the extent to which the cardiovascular system shapes one of the most wide ranging regularities across animal diversity.
|
10.1371/journal.pntd.0004248 | Prevalence of and Factors Associated with Human Cysticercosis in 60 Villages in Three Provinces of Burkina Faso | Taenia solium, a zoonotic infection transmitted between humans and pigs, is considered an emerging infection in Sub-Saharan Africa, yet individual and community-level factors associated with the human infection with the larval stages (cysticercosis) are not well understood. This study aims to estimate the magnitude of association of individual-level and village-level factors with current human cysticercosis in 60 villages located in three Provinces of Burkina Faso.
Baseline cross-sectional data collected between February 2011 and January 2012 from a large community randomized-control trial were used. A total of 3609 individuals provided serum samples to assess current infection with cysticercosis. The association between individual and village-level factors and the prevalence of current infection with cysticercosis was estimated using Bayesian hierarchical logistic models. Diffuse priors were used for all regression coefficients. The prevalence of current cysticercosis varied across provinces and villages ranging from 0% to 11.5%. The results obtained suggest that increased age, being male and consuming pork as well as a larger proportion of roaming pigs and percentage of sand in the soil measured at the village level were associated with higher prevalences of infection. Furthermore, consuming pork at another village market had the highest increased prevalence odds of current infection. Having access to a latrine, living in a household with higher wealth quintiles and a higher soil pH measured at the village level decreased the prevalence odds of cysticercosis.
This is the first large-scale study to examine the association between variables measured at the individual-, household-, and village-level and the prevalence odds of cysticercosis in humans. Factors linked to people, pigs, and the environment were of importance, which further supports the need for a One Health approach to control cysticercosis infection.
| Taenia solium is an infection that is transmitted between pigs and humans. Humans may get infected with the larvae of Taenia solium, which results in cysticercosis, an infection common in pig farming communities where there is poor sanitation and free roaming pigs. Most published studies on this infection have included less participants covering a restricted geographic area, thereby resulting in a limited understanding of the important risk factors for infection. Our study aimed to examine important individual-, household- and village-level characteristics associated with current infection using baseline data from 3609 participants living in 60 villages across three provinces in Burkina Faso. Blood samples from village participants were taken to assess whether they were infected with cysticercosis. We found that eating pork, especially in other village’s markets, being older and male, living in a poorer household, not having access to a latrine, and living in a village where a larger percentage of pigs are left roaming were associated with infection. Soil pH and composition may also play a role in infection. Our results suggest that interventions that include human and veterinary health as well as environmental components should be considered to effectively control cysticercosis in such settings.
| Taenia solium cysticercosis, a zoonotic infection transmitted between humans and pigs, is considered as an emerging infection in Sub-Saharan Africa. In the Sahel and West Africa region, the pig population has increased by 23% between 1985 and 2005, the largest increase in all animal populations during that period [1]. The pig population more than doubled in Burkina Faso between 2000 and 2008 [2]. Nearly 98% of pigs in the country are raised in a traditional manner in small holder farming communities, and left free roaming to fetch their own food [2]. The Joint Monitoring Programme (UNICEF/WHO) estimated the percentage of improved sanitation in rural Burkina Faso to be 8% at the end of 2008, far below the target of 54% set for 2015 by the Burkina Faso National Program for Drinking Water Supply and Sanitation [3]. The increase in primarily traditionally raised pig population and the lack of improvement in sanitation in rural areas are ideal conditions for the spread and maintenance of T. solium infection in humans and pigs. This is likely to have great consequences on public health since humans, when infected with the larval stages of the infection, may develop neurocysticercosis (NCC). NCC is a preventable cause of multiple neurological manifestations including epilepsy, severe chronic headaches, and focal deficits, to name but a few [4–6].
Serological studies using tests to detect antigens or antibodies have demonstrated the presence of the infection in humans [7–9, reviewed by 10] and pigs (reviewed by [11]) in several countries of West Africa, but with some large variation in estimates from country to country and from village to village within countries. Part of the variation could be explained by the use of antibody-detecting tests such as the EITB [12], which measure past exposure to and current infection with living metacestodes, and antigen-detecting tests [13,14], which measure current infection with living metacestodes. For example, in Burkina Faso, no human was found to have a strong positive AgELISA result in a village where very few pigs were present, while the prevalence of strong positives in humans was between 1.4% and 10.3% and in pigs between 32.5% and 39.6% in two villages where pigs were raised [15,16].
Previous studies conducted in West Africa have included a small number of participants or of villages, limiting the ability to detect associations, and especially weaker ones, between village-level as well as individual-level variables and infection. A better understanding of individual and community-level factors associated with infection is essential to developing effective control strategies to reduce the burden of this devastating disease.
The aim of this study is to estimate the magnitude of association of individual-, household- and village-level factors with current human cysticercosis infection in 60 villages located in three Provinces of Burkina Faso.
This study reports on the baseline cross-sectional component of a large community randomized-control trial aimed at estimating the effectiveness of an educational package on reducing the cumulative incidence of current infection of human and porcine cysticercosis. The baseline cross-sectional study took place between February 2011 and January 2012.
The provinces of Nayala (Region of Boucle du Mouhon), Boulkiemdé and Sanguié (Region of Centre-Ouest) were selected for inclusion in the study. Boulkiemdé and Sanguié are among the provinces with the largest number of pigs in the country (191,438 and 145,923 heads respectively in 2010 [17]. Nayala has an estimated pig population of 41,521 and was selected because of local reports that humans tend to defecate in pigsties where most pigs are kept during the rainy season and because it was neighbouring the other two provinces. In each province, all departments where pigs were raised (30 of 31 departments) were selected. Within each department, two villages meeting the eligibility criteria were selected at random. The exclusion criteria were: being located on a National or Provincial road; being the Chef-Lieu (Capital) of the Region or of the Province; being located within 20 km of Koudougou or Ouagadougou. For the purpose of the parent community-based trial, the inclusion criteria were: having a population of at least 1000 people at the 2006 census; being present on the map from the Institut Géographique du Burkina (from the year 2000); being separated from another participating village by at least 5 kilometers. A third village meeting all the eligibility criteria was also selected at random in each department as a replacement if the village initially sampled were to be found not eligible during the field visit (for example, too few households raising pigs, refusal from the village leaders which happened in one instance). Only one village met the eligibility criteria in the Department of Zamo (Province of Sanguié). A village located in the Department of Pouni but right on the border with Zamo was selected as the second village for that Department. The location of all participating villages is illustrated on Fig 1.
A concession (compound) is defined here as a group of households living in a residential development, often fenced, where the authority of a concession chief is recognized. A household is defined as a socio-economic unit where members live together, share resources and satisfy the food and other essential needs of all members. In general, a household includes a man, his wife or wives, their children, and any other person who shares resources with other members of the household.
In each village, the research team of five or six interviewers started by visiting the concession of the village Chief. The field team moved counter clockwise to enumerate each concession. In each concession, starting on the right from the entrance, the field team enumerated each household, as well as the number of members and if someone raised piglets less than 12 months old or reproductive sows in each household.
A stratified random sampling approach was adopted to select concessions. The strata were the presence of reproductive sows, piglets, or no pigs. First, the list of all concessions where there was at least one household raising sows was made and corresponding concession numbers were placed in a bag. A village member was asked to sample 10 numbers from the bag. If 10 or less concessions were raising sows, all of those concessions were invited to participate. Next, the numbers of all concessions where at least one household raised piglets aged 12 months or younger were placed in a bag, including concessions where sows were also raised but that were not selected in the previous step. A village member was asked to sample 30 additional numbers from the bag. If 30 or less concessions raised piglets less than 12 months of age, all of those concessions were invited to participate. Lastly, the numbers of all concessions not yet sampled were placed in a bag. A village member was asked to sample an additional 40 numbers. This resulted in the sampling of a total of 80 concessions in each village, with at least 10 concessions with sows and at least 30 concessions with piglets aged less than 12 months, except in a few villages in Nayala where there were less than 40 concessions raising pigs in the whole village. Fig 2 summarizes the sampling strategy for the concessions. Only one Chief of a concession refused participation in a village of Boulkiemdé and was replaced by another concession in the village.
Once the Chief of a sampled concession gave his/her consent to participate in the study, he was asked to enumerate all households in the concession with the names of their Heads. Numbers corresponding to the households were placed in a bag and the chief of the concession was asked to sample one number from the bag, or, if there was only one household in the concession, that household was selected. The Chief of the selected household was asked for his consent to participate in the study and to enumerate all members of his household. Only one Chief of household refused participation in Boulkiemdé and asked that his father’s household be sampled instead of his, as a mark of respect for his elder. All household Chiefs consented to participate.
Each eligible household member was allocated a number. The inclusion criteria for participants were that they were at least 5 years old, had lived in the village for at least 12 months and were not planning to move in the next three years. The head of the household was asked to pick a number from the bag. The person with the corresponding number was asked for his/her consent to participate in the study which included a screening questionnaire on epilepsy, progressively worsening severe chronic headaches as well as a blood sample to test for the presence of antigens to the larval stages of cysticercosis. Participation in the study involved answering the screening questionnaire and having blood collected three times in a three-year period for the purpose of the parent community-based randomized controlled trial. The sampling process (starting with the concession sampling) was repeated until 60 people consented to being screened and having blood collected three times over three years duration of the randomized trial. If the sampled person refused to consent to the serological component of the study, s/he was asked for his/her consent to participate in the screening-only study for a maximum of 20 screening-only participants per village. If the sampled individual did not consent to either the serological or screening study, the head of the household was asked to sample another number from the bag for that household until a consenting individual was found. There were a few refusals to the serological participation, and all cases were replaced as described above. No one refused to participate to the neurological component of the study.
Each consenting participant was asked to answer an individual questionnaire including questions about socio-demographic factors, pork consumption behavior, knowledge of T. solium infection and life cycle, and screening questions on seizures, epilepsy, and severe chronic headaches. Since the focus of the current study is to estimate the magnitude of association between potential risk factors and current cysticercus infection, results from screening for epilepsy and progressively worsening severe chronic headaches are not reported. In addition, the Chief of each participating household was asked to list assets owned by members of the household (ie bicycles, carts, livestock, etc). The mother of each household was asked questions about preparation of pork and access and use of latrines by members of the household. The building material of the house’s roof, floor and walls was also measured at that time. In each village, the owners of at most 10 sows and 30 piglets aged less than 12 months were interviewed regarding their pig management practices (at most 40 pig owners per village). All questionnaires were conducted and recorded on Personal Digital Assistants (PDAs). The PDAs recorded the geographical coordinates of each concession.
We used indicators of wealth as suggested in Gwatkin et al. (2000) [18] to estimate the wealth quintile of each household. Several wealth indicators were missing and therefore imputed using the missMDA package in R [19,20]. Principle Component Analysis was run in R using the PCA command of the FactoMineR package [21]. The imputed and known were exported back to Stata where PCA analysis was conducted and the fitted values were used to obtain the percentiles that were used to classify the households into 10 wealth groups.
Between 18 March and 25 November 2014, the field team returned to each village to take soil samples to measure soil composition and pH using the LaMotte Soil Texture Unit test (code 1067) and LaMotte pH Test kit (code 5024), respectively. The soil composition test estimates the percentage of sand, silt and clay in the soil. In each village, a soil sample was taken in each of the four cardinal directions (within 2 km) using a plastic pot filled to 125 ml, having scraped the soil surface to a depth of about 5 to 10 cm. The four samples were then thoroughly mixed, dried and cleaned of coarse elements such as stones. A subsample of this mixture (about 125 ml) was then taken and stored for analysis of soil composition and pH, according to LaMotte procedure.
After an average of 7 weeks (range of 0 to 140 days), a physician went to each village to take blood samples from the 60 participants who consented to the serological component of the study and to examine all participants who had screened positive for seizures, epilepsy or severe chronic headache. Any participant confirmed as having epileptic seizures, epilepsy or severe chronic headache who had not initially consented to the serological follow-up were asked to provide a blood sample for (clinical) diagnosis purposes. However, to avoid an over-selection of people with neurological symptoms in our analysis, only those who initially consented to the serological analysis component of the study are included in the analysis. Blood samples of 8 ml were drawn by venipuncture using a syringe, preferably at the antebrachium vein, using 10 ml Venosafe serum gel tubes. The tubes were placed in a cooler, left to decant, and the sera were collected and put in two pre-labelled tubes at the end of each day or the following day. The sera were placed in freezers (-20°C) at most three days after the blood draw. The sera were brought to the IRSS (Institut de Recherche en Sciences de la Santé) in Bobo-Dioulasso every 4 to 8 weeks and kept thereon in a freezer at -20°C until analyses took place.
The serum samples were tested for presence of excretory secretory circulating antigens of the metacestode of T. solium using the B158/B60 enzyme-linked immunosorbent assay (ELISA) [13]. The test was found to have a sensitivity of 90% (95%BCI: 80–99%) and a specificity of 98% (95%BCI: 90–99%) to detect current infection in a study conducted in Ecuador [22].
Descriptive analyses on the study population were first conducted, followed by assessing the association between each potential risk factor and the prevalence of current cysticercus infection at the individual-level and at the village-level. Household characteristics measured through the mother and chief questionnaires were attributed to each individual since only one individual was sampled per household (and concession). Consequently household-level variables were included at the individual-level in all analyses. All descriptive analyses were conducted in Stata 13.1.
Data obtained from pig owners were considered as village-level variables. These included the percentage of pig owners letting their animals roam all seasons, practicing home slaughtering of pigs and asking for inspection at the time of slaughter. Soil composition, soil pH and the season when human samples were collected (dry vs wet) were also considered as village-level variables. The effect of the coverage of the 2012 filiariasis mass drug delivery campaign, which provided albendazole (400 mg) with ivermectin, obtained from the Ministry of Health, was also explored at the village-level.
The effects of variables which may impact the contamination of the environment with human feces were explored at the individual-level and at the village-level. These included the use of latrines to defecate reported by interviewees, the access to a latrine as reported by the mother, the household wealth quintile (as an indicator of general hygiene and sanitation), the knowledge about taeniasis including report of self-infection, and the reporting of pork consumption at home and outside the home.
Individual-level variables explored that were not likely to directly influence environmental contamination included age, gender, education and occupation, although the effect of age was modelled separately for each province. Age was categorized because it was not linear in the logits.
The only concession-level variable explored was the type of concession sampled (i.e. sow, piglet or other).
Given the sampling strategy, concession, household and individual characteristics were all attributed to the individual-level in the analysis. To take the stratified nature of the sampling into account, models including the type of concession sampled were run but did not modify the estimated medians and 95%BCI of the estimates. Hence, results from the simpler model are being reported here.
Bayesian hierarchical logistic models were fitted to estimate the prevalence odds ratios between each variable of interest and the prevalence of current cysticercus infection. Some models were also run with Bayesian hierarchical log-binomial models and resulted in similar estimates when convergence was achieved. At the first level, current cysticercus infection was assumed to follow a Bernoulli distribution. The logit of this distribution was modeled using the individual-level variables and village-level random-effect intercepts. At the second level, each village-level intercept was assumed to follow a normal distribution. The mean of the random-effect intercepts were modelled as a linear regression using the village-level variables, including the province in some models. The effect of the provinces with a random-effect on villages was not important, and therefore all presented models exclude province effects. Diffuse priors were used for all regression coefficients. Missing independent variable values were imputed using the mean value of non-missing data, assuming an ignorable missingness mechanism. When the mean values varied by province, province-level means were used for imputation. Fit was measured by comparing deviances. Convergence was assessed by looking at the history and b Rubin Gelman plots. Some of the more complex models required large numbers of iterations and thinning of 100 to obtain stable estimates. All models were run in WinBUGS [23].
The protocol and consent forms were approved by the University of Oklahoma Health Sciences Center Institutional Review Board and by the Centre MURAZ ethical review panel (Burkina Faso). All participants were read the consent form and any questions they had were answered to the best knowledge of the field staff. Each participant was given a bar of soap to thank them for their time. Information about the study provided in the consent forms were read and explained to each potential participant (mother of the household, chief of the household, participant, pig owners) by the field workers who took the time to answer all questions. Consenting participants signed the consent forms when able or put a cross when not. All consents were witnessed by a local villager. Children aged more than 10 were asked for their assent, but parents consented for all children aged less than 18 years old.
A total of 4795 villagers consented to being screened for epilepsy and severe chronic headaches three times during the course of the parent community-based randomized trial. All three mother, chief of the household and individual questionnaires were missing for three individuals who were excluded from the analyses. Of the remaining 4792 participants, 4788, 4772 and 4778 had information from the chief questionnaire, mother questionnaire or individual questionnaire available, respectively. A total of 3609 participants consented to participate in the serological component of the parent randomized trial and provided sufficient blood at baseline to be analyzed. The characteristics of individuals participating in the serological and screening-only component of the study are described in Table 1. Since individuals living in concessions where pigs were being raised were first asked to participate in the serological component of the study, there was a larger proportion of participants who provided blood who either raised pigs or consumed pork. The participation proportion was similar according to other characteristics although females and more educated people tended to be more likely to consent to the serological follow-up component of the study.
A total of 120 individuals tested positive for current infection with cysticerci. The prevalence of current cysticercosis varied considerably across provinces and villages (Figs 1 and 3) ranging from 0% to 11.5%, although the 95%CI were wide. In the Province of Sanguié, no individual tested positive in seven (35%) of the 20 villages studied. In contrast, this was the case in only three (10%) and one (10%) villages in Boulkiemdé and Nayala, respectively.
Table 2 provides estimates of the prevalence according to different individual-level characteristics of the participants as well as their associated prevalence proportion ratios. The univariate analyses suggested that older males, people living in a household with lower wealth quintiles and those consuming pork had higher prevalences of infection. In addition, access to a latrine as reported by the mother of the household was associated with a reduced prevalence of infection.
Table 3 shows the linear regression coefficients between the prevalence of infection in each village and the village-level variables. The percentage of participants reporting eating pork, and particularly those eating pork in someone’s or their household, and the percentage of participants reporting having had a tapeworm infection were associated with a slight increase in the village-level prevalence of infection. The percentages of silt in the soil and of sand in the soil were associated with a decrease and increase in the prevalence, respectively.
Table 4 shows the results of three candidate models with the lowest deviances. Models 2 and 3 include soil indicators which were measured up to 36 months after the baseline visit while model 1 does not include these variables. The individual-level variables were common to all models and all resulted in similar magnitudes of association with the prevalence odds of infection. Being aged more than 50 years old had a stronger effect on the prevalence odds in Boulkiemdé than in the other two provinces. Males had cysticercosis prevalence odds of nearly 2.6 when compared to females. Eating pork at another village market had the strongest effect, while eating pork at the village market or at home also increased the prevalence odds of infection when compared to never having eaten pork. Living in a household with higher wealth quintiles and having access to a latrine both decreased the prevalence odds of infection, with those not having access to a latrine having a prevalence odds of about 2.8 times higher than those having access to a latrine. The effect of a self-history of taeniosis and knowledge of taeniosis became negligible in models adjusted for age (ie age was a strong confounder of taeniosis history and knowledge).
The major differences in the three models come from the inclusion of the type of soil and soil pH. When these variables are excluded, the percentage of pigs not penned during the rainy season (i.e. left roaming or tethered) led to a weaker association with the prevalence odds of cysticercosis. An increase in the alkalinity of the soil was associated with a lower prevalence odds of cysticercosis. This effect was only noted when the percentage of silt or sand in the soil was also included in the model. We present here the effect of the percentage of sand, but the percentage of silt had an opposite effect of similar magnitude to that of sand.
This cross-sectional study is the most widespread ever conducted, estimating factors associated with the prevalence of current infection of human cysticercosis. Our study is unique in its inclusion of 60 villages located in three provinces and the evaluation of over 3600 people living in these villages. The inclusion of only one individual per household and concession reduced the dependence among observations, thus maximizing the power to detect individual and household-level factors associated with the prevalence. Our hierarchical model also includes potential risk factors measured at the individual- and village-level, thus respecting the sample size of each unit. Finally, by including the village as random-effects, and having explored the impact of incorporating the type of concession which was part of the sampling scheme, we are effectively using a model-dependent approach to adjusting for the sampling scheme, thus reducing the potential biases which may be introduced by clustered sampling [24]. To our knowledge, only two studies conducted in Sub-Saharan Africa had adjusted for the cluster nature of the sampling or the infection [7,25].
We found that the prevalence of current infection with cysticercosis varied from 0% to 11.5% in the sampled 60 villages. These villages were sampled with the goal of conducting a community-based randomized controlled trial and participants were selected based on the presence of pig raising in their household. Therefore, the overall prevalence cannot be generalized to the three provinces nor to the country as a whole. Nonetheless, the village-level prevalences of current infection with cysticerci measured with the AgELISA are within the range of those reported by others using the same test in community-based studies conducted in three rural communities of West Cameroon (from 0.4% (0.2%;0.7%) to 3% (0.3%;11.2%) depending on the locality) [7], one small village in Sénégal (with 7.7% (5.3%-10.7%)) [8], and 20 villages in Zambia (with 5.8% (4.1%-7.5%) [26]. Other community-based studies have reported higher overall prevalences of current cysticercus infection (ie not village-specific) in one village in the Democratic Republic of Congo (with 21.6% (18.2%-25.0%) [25] and 13 villages in Tanzania where very high porcine cysticercosis prevalence levels had been reported (with 16.7% (14.2%;19.2%)) [27].
A unique characteristic of our study is its ability to explore between-villages variation in prevalence. Although the villages were sampled with a set of inclusion criteria necessary for the randomized trial, considerable variation in the prevalence was observed among them as well as among the three sampled provinces, although the credible intervals were wide. Such variation in areas was also observed in a study conducted in three areas of West Cameroon [7], different districts of a village in the Democratic Republic of Congo [25], and six departments of Bénin [9], although the latter study used an EITB to detect exposure to infection [12] instead of current infection. This confirms observations by others that current cysticercus infections in humans occur in clusters [26, 28–30], often around taeniosis carriers. The very clustered nature of cysticercosis calls for great care in attempting to generalize results from studies conducted in a small number of villages or communities to larger areas or to a country.
The individual-level factors found to be associated with the prevalence odds of current infection are similar to those reported in other community-based studies conducted in Sub-Saharan Africa. An increased prevalence odds of current cysticercus infection in adults aged 30 or more as compared to individuals aged 7–30 years old was observed in all three provinces. This confirms what was observed in a study of 720 participants living in 20 villages of Zambia where the prevalence odds increased after the age of 30 years old when compared to those aged 0–9 years old [26] and a study in Tanzania where the prevalence odds was increased in individuals aged 36 or more when compared to those aged 15–25 years old [27]. Moreover, in the province of Boulkiemdé, a further increase in the prevalence odds was observed in people aged 50 years old or more. Such increase in prevalence in older people has been observed in studies conducted in the Democratic Republic of Congo (POR of 2.8 95%CI: 1.14; 3.81 for those aged 70 or more compared to those aged 0–9 years old) [25] and in West Cameroon (seroprevalence of 2% in those aged 46 years old or more and 0.1% in those aged 15 or less) [7]. A study conducted in 1989 in Bénin found an age-pattern of sero-prevalence of exposure to cysticercosis which is very similar to that observed in Boulkiemdé, with an initial increase at 30 years of age, followed by a further increase after 50 years old [9]. The Zambian study did seem to show a tendency for higher prevalence of current cysticercosis after the age of 50, but the small sample size in older age groups may have reduced the power to detect such increase [26]. A study conducted in Ecuador suggested that the increase in current infection in older age could be linked to reduced immunity in older age groups [31]. This is further supported by findings from a cohort study conducted in Zambia which showed that while sero-reversion rates were higher than seroconversion rates in people aged less than 60, such difference disappeared in people aged 60 years old or more [32]. The reason why an additional increase in older ages was only observed in one province is difficult to explain and may require future studies, but may be linked to differential at risk behavior in this group of older men not captured elsewhere.
The increase of prevalence odds of current infection with or exposure to cysticercosis in males has been reported by some community-based studies conducted in Sub-Saharan Africa [9,15,25,33], but not all [7,8,26]. The study conducted in the Mbozi district of Tanzania reported an impact of gender on exposure to cysticercosis measured with the EITB, but the reporting is problematic because the gender among whom the prevalence odds is reported to be increased is inconsistent (males in the results section and females in the discussion section and Table 3) [27]. In that same study, the effect of washing hands by dipping in multiusers buckets was associated with a decrease prevalence odds of exposure to infection and an increased prevalence odds of current infection, casting doubts about the reported results. The difference in results between studies could be real, or could be attributable to the reporting of crude associations in some studies [9] and that resulting from multivariable analyses in others [7,8,15,25, 26,33]. In our study, the association between being male and the prevalence odds persisted after adjusting for pork consumption, wealth quintile of the household, and access to a latrine. This confirms that factors which bring males to get exposed to T. solium eggs more than females, such as poor hand hygiene or consumption of fresh produce such as fruits and vegetable that have not been cleaned properly, or eating meals outside the home that could be prepared by foodhandlers infected with taeniasis may be playing a role. Behavioral studies comparing male and female food consumption and hand hygiene behavior are warranted to develop future interventions.
Where pork is consumed was found to have an impact on current infection with cysticerci. Since cysticercosis can be acquired by either self-infection or through food or hands contaminated with the eggs of T. solium, and since the effect of a self-reported history of taeniosis became negligible once gender and age were included in the model, contamination of food with T. solium eggs may play an important role in this population. To our knowledge, this is the first time that the effect of where the pork is consumed is reported. In Sub-Saharan Africa, only one study in Zambia found that not boiling pork before its consumption played a role among older females based on a classification tree model [33]. Other studies conducted in Latin America, although in univariate analyses, had reported an association between pork consumption and exposure to cysticercosis [34,35].
Having access to a latrine was associated with reduced prevalence odds of current infection with cysticerci. This had been mentioned in the small scale study of Secka et al. conducted in Sénégal [8], but was not confirmed in a multivariable analysis. It was also observed in relation to the prevalence of exposure to infection in univariate analyses in Colombia [35,36] and Honduras [37]. The effect of having access to a latrine by family members had a stronger effect than that of participants declaring that they used a latrine to defecate. This is in agreement with a recent qualitative research conducted in Eastern Zambia which showed that latrines are usually considered as public among neighbors [38]. Our results support that access to latrines is an indicator of environmental contamination with taeniid eggs around the household.
Poor living conditions, as indicated by the three lowest wealth quintiles in our study, have also been reported to be associated with increased prevalence odds in previous studies, in these cases using univariate analyses. In an urban area of Honduras, Sanchez et al. [37] reported that several indicators of poor living conditions such as raising pigs, lack of potable water, lack of sanitary toilets and earthen floor were associated with an increased prevalence odds of exposure to cysticercosis. Poor hygienic conditions of the household was associated with an increased prevalence odds of exposure to infection in the area of Morelos, Mexico [34].
The inclusion of 60 villages in our study allowed us to assess the village-level effect of the type of pig management on the prevalence odds of current cysticercus infection. To our knowledge, this is the first time that a study has a number of sampled villages large enough to explore and measure village-level factors. Previous studies conducted in Latin America had found that ownership of pigs was associated with the prevalence odds of exposure to infection [28,35–37]; we found that how pigs are managed at the village-level has an impact. No previous studies conducted in Sub-Saharan Africa had found such effect. Indeed, ownership of pigs by the respondent did not yield an association with the prevalence of cysticercosis while pig ownership at the household level did in univariate analyses. This should be considered in future studies. In the study areas, almost all pigs were left to roam during the dry season. However, villages where a larger percentage of pigs were not penned during the rainy season had higher prevalence odds of current cysticercosis.
This design also allowed us to find an association between the type of soil and the soil pH and the prevalence odds of current infection. An increase in the pH of the soil was associated with lower village-level prevalence odds of current infection. In an experimental study of inactivation of T. solium eggs in different temperature, dryness and pH conditions, an increase in alkalinity in an alkaline environment (pH from 12.1 to 12.7) was linked to an increase in inactivation rates of the eggs while the opposite was true for in an acidic environment (pH from 5.1 to 5.5) [39]. The soil pH in our study villages was at an average of 6.8 with a range from 5.4 to 8.2. The soil was therefore nearly neutral on average, and it is difficult to say if the results are consistent to these found in [39]. However, it could be hypothesize that the eggs are more tolerant to the generally more acidic environment of the gastro-intestinal tract, which could favor better egg survival in slightly more acidic soil. An increase in the percentage of sand in the soil was associated with an increased village-level prevalence odds of current infection. Perhaps taeniid eggs are more easily disseminated from sandy soil onto vegetables and water through wind. The fact that the soil was sampled nearly three years after the baseline study may also have an impact, although it is unlikely for the soil to change its pH extensively through time.
A self-report of a history of infection with taeniosis was associated with current infection with cysticerci only in univariate analyses. This association was confounded by age, and became non-significant in the hierarchical model also including pork management variable. Others had found associations between the self-report of taeniosis and exposure to cysticercosis in univariate analyses [34]. This underlines the importance of conducting multivariate analyses to identify factors with the strongest impact on infection.
This study had some limitations. The most important one is that the soil samples were collected nearly three years after the baseline study, therefore, these associations should be interpreted with great care. This is why results were presented with and without including the soil analyses. Future studies using follow-up data of the randomized trial among the control group should be able to confirm (or not) this association. Only one person per household was sampled and therefore, factors affecting clusters within households could not be evaluated.
In conclusion, this study is the first to assess the association between several individual-, household, and village-level variables and the prevalence odds of cysticercosis in humans at a large scale. We found that factors linked to people, pigs, and the environment were of importance. This further supports the need for a One Health approach to control this infection.
|
10.1371/journal.pcbi.1000298 | Steady-State Kinetic Modeling Constrains Cellular Resting States and Dynamic Behavior | A defining characteristic of living cells is the ability to respond dynamically to external stimuli while maintaining homeostasis under resting conditions. Capturing both of these features in a single kinetic model is difficult because the model must be able to reproduce both behaviors using the same set of molecular components. Here, we show how combining small, well-defined steady-state networks provides an efficient means of constructing large-scale kinetic models that exhibit realistic resting and dynamic behaviors. By requiring each kinetic module to be homeostatic (at steady state under resting conditions), the method proceeds by (i) computing steady-state solutions to a system of ordinary differential equations for each module, (ii) applying principal component analysis to each set of solutions to capture the steady-state solution space of each module network, and (iii) combining optimal search directions from all modules to form a global steady-state space that is searched for accurate simulation of the time-dependent behavior of the whole system upon perturbation. Importantly, this stepwise approach retains the nonlinear rate expressions that govern each reaction in the system and enforces constraints on the range of allowable concentration states for the full-scale model. These constraints not only reduce the computational cost of fitting experimental time-series data but can also provide insight into limitations on system concentrations and architecture. To demonstrate application of the method, we show how small kinetic perturbations in a modular model of platelet P2Y1 signaling can cause widespread compensatory effects on cellular resting states.
| Cells respond to extracellular signals through a complex coordination of interacting molecular components. Computational models can serve as powerful tools for prediction and analysis of signaling systems, but constructing large models typically requires extensive experimental datasets and computation. To facilitate the construction of complex signaling models, we present a strategy in which the models are built in a stepwise fashion, beginning with small “resting” networks that are combined to form larger models with complex time-dependent behaviors. Interestingly, we found that only a minor fraction of potential model configurations were compatible with resting behavior in an example signaling system. These reduced sets of configurations were used to limit the search for more complicated solutions that also captured the dynamic behavior of the system. Using an example model constructed by this approach, we show how a cell's resting behavior adjusts to changes in the kinetic rate processes of the system. This strategy offers a general and biologically intuitive framework for building large-scale kinetic models of steady-state cellular systems and their dynamics.
| Computational models help quantify the reaction dynamics and regulatory modes in complex biochemical systems [1]–[5], particularly when a system is so intricate that its behavior cannot be predicted by intuition alone. The building blocks for constructing large reaction networks are often available in numerous databases [6]–[9] and journal archives. Here, one can obtain many of the experimentally-derived elementary reaction steps, kinetic constants, or rate laws for individual steps in a given biochemical system or pathway. Despite this wealth of information, however, compiling these data to construct models with accurate system-wide behavior represents a significant challenge in systems biology [10],[11]. Comprehensive models of metabolism have been successfully developed for microbial systems [5],[12],[13] and certain eukaryotic cell types [14]–[16]. These constraint-based models [17] are often represented by stoichiometric networks that lack an explicit description of substrate concentrations, reaction mechanisms, or the transient behavior of the system. Although various strategies have been proposed to incorporate these features into large-scale models [18],[19], the task of assembling complex kinetic models with nonlinear dynamics remains a difficult problem. One of the major obstacles to building accurate kinetic models is the number of unknown parameters in the model that must be estimated using experimental datasets [19], which themselves are often massive, incomplete, noisy, and/or imperfect [20]. A number of parameter estimation methods, such as genetic programming, simulated annealing, and various gradient-based routines [21],[22], have been proposed to infer unknown quantities in biochemical models. Most of these methods address the problem of estimation in purely abstract terms and do not take into account the unique mathematical features of biochemical systems, such as a well-characterized kinetic subsystem (e.g., the dynamics properties of an ion channel [23]). Estimated parameters must still meet constraints imposed by the other experimentally measured parameters in the model.
To address these challenges, we propose a strategy for assembling large kinetic networks that retain the nonlinear dynamics governing individual reactions in the system. The key features of the method are: (i) restriction of steady-state values by subsystem kinetics, (ii) reduction of the steady-state solution space by principal component analysis (PCA), and (iii) combination of independently constructed submodels (modules). The first feature is a Monte Carlo sampling over unknown concentrations with fixed kinetic parameters derived from the literature. The opposite strategy has been used in microbial systems to restrict kinetic parameters based on species concentrations [12]. The second feature, reduction of the steady-state space by PCA, has been applied previously for metabolic systems described by a stoichiometry matrix [5],[13], but not, to our knowledge, for nonlinear systems. In the last step, a full model representation is assembled by combining PCA-reduced, steady-state solutions from each module to form a combined steady-state solution space for the entire system. This global space may then be searched for solutions with accurate time-dependent behavior using any number of established routines 22,24.
The method exploits three properties common to many biological systems: modularity, homeostasis, and known quantitative kinetic relationships among interacting molecular components. Interestingly, this physiology-inspired approach enforces natural constraints on the range of allowable system states and allows one to monitor shifts in steady states due to kinetic perturbations. To illustrate the method with an example, we show how 77 reactions from 17 primary data sources were integrated to construct an accurate model of intracellular calcium and phosphoinositide metabolism in the resting and activated human platelet. Finally, we extend our analysis of this modeling approach by examining the steady-state characteristics of a system that is affected by changes in kinetic rate constants.
Our method builds upon a common representation of biochemical reaction networks [25] consisting of a system of ordinary differential equations (ODEs). In this paradigm, the concentration of each molecule in the system changes with time as a function of the instantaneous values of other concentrations and fixed kinetic parameters in the model. We separate this model description into two parts: The concentration vector (CV) of the model refers to the set of all molecule concentrations at a given instant in time and is denoted by the vector c:(1)The model topology refers to the entire set of kinetic parameters and rate equations that determine how these concentrations evolve with time. Mathematically, this is represented by the vector function f, which defines the rate of change of c with time as a function of the model concentrations and rate parameters:(2)The functional form of each is a sum of rate equations for each reaction that consumes or produces and will generally vary for each molecule. Typical functional forms for f may include, for example, a series of Michaelis-Menten or nonlinear rate expressions. A simple reaction topology is shown in Figure 1A with corresponding ODEs in Figure 1B. It is useful to separate a large model into two or more modules with subset CVs that overlap at reaction edges, as shown in Figure 1A.
Often, the topology of a biological system is better characterized than its CV [17]. For example, the major protein-protein interactions in a signaling pathway may be deduced from mutation or knock-out studies, providing a molecular wiring diagram that links together the various components in the network. For each of these interactions, purified enzymes may be used to measure the strength of the interaction in vitro or to measure the rate of some enzyme-catalyzed reaction in the system. An important caveat is that the kinetic rate constants within the cellular milieu (the cell context) may be different from those obtained in an in vitro experiment with purified components. In contrast, it is generally more difficult to accurately measure the absolute abundance of intracellular enzymes or metabolites in vivo, although progress is being made in this area [26]. Our method thus assumes that the topology of a given system is known and that the unknown set of concentrations exists in a linear space of dimension n in which each species comprises a separate dimension (Figure 1C). The ultimate goal of the method is to efficiently search this concentration space to find a set of values that, when combined with the fixed topology, renders the full model consistent with known resting states and experimental time-series data obtained by perturbation of the cell.
A special situation arises when in equation (2). Under these conditions, the model is said to be at steady state, and the vector is a steady-state solution to the system of ODEs. If f contains nonlinear terms, there may be an infinite number of steady-state solutions for the system of ODEs [25]. This set of solutions occupies some nonlinear subspace of the concentration space exemplified in Figure 1C. To guarantee that nonzero steady-state solutions may be found, the method requires the model topology (and all module topologies) to be balanced, meaning that the production and consumption of each molecule must be equal so that the total mass of the system is conserved. This steady-state assumption [17] is a common constraint in stoichiometric modeling and metabolic flux analysis and is conceptually related to the biological phenomenon of homeostasis [27], in which opposing processes are coordinated to maintain the stability of a cell or organism. For example, a nerve cell may maintain a constant electrochemical gradient by continually transporting ions across a lipid membrane.
The first phase of the method involves generating a compact representation of the steady-state solutions for each module. The steps for module reduction are outlined in Figure 2A. First, conservative bounds are chosen for c based on physiological and practical considerations. For example, a regulatory enzyme is expected be present in at least one copy per cell and not to exceed an intracellular concentration of one molar. Knowledge about the physical size of the system is useful in this step to convert a raw copy number to a concentration. For small systems, this information can provide a rigid lower bound on unknown concentrations [28]. For example, a single molecule in a 6 fL platelet has a concentration of 4 nM. Also, because molecular concentrations can span several orders of magnitude, it is often more efficient to delineate this range of values on a logarithmic scale rather than a linear scale.
Once the sampling distribution for c has been defined, steady-state solutions () for each module are calculated using fixed kinetic parameters for each reaction in the module obtained from the literature [6],[8],[9], novel kinetic experiments, or estimation. For this step, each initial guess is sampled from the distribution for c and combined with the predetermined topology. The combination of fixed rate equations, fixed parameters, and forms a well-posed initial value problem,(3)that may be computed using a numerical solver [29]. For non-oscillating systems, steady-state solutions may be obtained by simulating the system until equilibrium is reached (i.e., until ). Alternatively, one may use any number of multidimensional root-finding routines, such as those available in the GNU Scientific Library [30], to find the closest n-dimensional root to the vector function f using starting guess .
In the third step, a large collection of steady-state solutions for each module is subjected to principal component analysis (PCA). A sample size of 1000 points per unknown concentration is generally sufficient to minimize error due to over-fitting [31]. PCA is then used to transform these points to a new coordinate set that optimally covers the space of steady-state solutions using the fewest number of dimensions. For example, if two molecule concentrations in the steady-state space are highly correlated due to participation in the same reaction, PCA will locate a single dimension to represent each pair of points in the transformed space. Ultimately, these new dimensions will be combined across all modules to search for global solutions that lie in the steady-state space for the fully combined network. Since PCA is a linear method, a steady-state solution space that is highly nonlinear may require more principal component vectors to accurately estimate the solutions. Nonlinear methods of dimensionality reduction, such as kernel PCA [32] or local linear embedding [33], may provide a more compact representation of steady-state solutions spaces in future iterations of the method.
The reduction procedure is illustrated with an example of a human platelet model comprising 4 interlinked signaling modules (Figure 2B). For each module, we used published reaction mechanisms and kinetic parameters to construct the module topologies [28]. Each topology was held fixed while the unknown CVs were sampled from empirically-defined distributions. For this step, we generated more than 109 sets of initial guesses () for each module, computed the initial value problem for each until a steady state was reached ( ), and selected only those steady-state CVs ( ) that were consistent with known concentrations. For example, the concentration of intracellular Ca2+ ([Ca2+]i, Figure 2B) in platelets is known to be ∼100 nM. Thus, only those with [Ca2+]i≈100 nM were kept as part of the steady-state solution space for the Ca2+ balance module. This procedure was used to generate 10,000 steady state solutions for each module for subsequent reduction by PCA. A minimal set of principal component (PC) vectors (those capturing 90% or more of the variance in the solution set) were used as search directions in the final estimation step, in which the transient behavior of the perturbed steady-state was compared to experimental time-series data.
Interestingly, only a small fraction of initial guesses produce steady-state solutions that are also consistent with known concentration values. For example, it was previously shown that only 50,000 of 109 initial guesses (0.005%) in the Ca2+ balance module (Figure 2B) met both requirements and were suitable for further analysis [28]. Among this set of CVs, marginal distributions for individual molecules were often confined to a narrow range of values. As an example, 80% of steady-state solutions for the calcium module contained <1000 IP3 molecules/cell, although initial guesses were sampled uniformly between 1 and 106 molecules/cell. This observation shows that the kinetic topology of these molecular networks places very strong constraints on the range of concentrations that can exist at steady state. In biological terms, this suggests that fixed kinetic properties at the molecular level (e.g., IP3R and SERCA kinetics) can affect not only the dynamical features of a biochemical system but can also determine the abundance of chemical species and the compartmental structures that contain them.
In the final step of the method, the full model is assembled by combining PCA-reduced, steady-state solution spaces from each module into a combined steady-state solution space for the entire system (Figure 3A). This global space is searched for full-length, steady-state solution vectors that satisfy both the individual steady-state requirements of each module and the desired time-dependent properties when the steady-state is perturbed (for example, by increasing the initial concentration of a signaling molecule). For the platelet signaling model, consisting of 77 reactions, 132 fixed kinetic parameters, and 70 species [28], a set of 16 PC vectors representing all 72 unknown variables (70 molecule concentrations, 1 compartment size, and 1 rate constant) in the model were used as search directions in a global optimization routine. The global solution space was searched for models with accurate dynamic behavior using experimental time-series data for ADP-stimulated Ca2+ release (Figure 3A). Equality constraints are imposed during optimization to maintain consistent concentrations of molecules that are present in more than one module. Specifically, for a steady-state space A represented by m PC vectors and a steady-state space B represented by n PC vectors, the projections of each space onto must be equal,(4)where is the unit vector for the shared molecule, . This condition forms a linearly-constrained optimization problem for which a number of efficient routines exist [22]. We used the Asynchronous Parallel Pattern Search (APPSPACK) to perform a derivative-free optimization of the platelet signaling model [24]. A least-squares objective function was used to score the difference between simulated (after perturbation of steady state) and experimental time-series data points. One of the high-scoring steady-state solution vectors for the full model is shown in Figure 3B, along with individual steady-state vectors for each of the four modules. This 72-dimensional vector (i) satisfies the homeostasis constraint in that it is a steady-state solution, (ii) is consistent with the known steady-state levels for 8 of the molecules in the 72-dimensional space, and (iii) predicts the entire dynamic Ca2+ and IP3 response of platelets exposed to ADP (0–100 µM). Additionally, rigid and flexible nodes (steady-state concentrations) in this 72-dimensional space were readily identified when a set of allowable steady-state solution vectors are compared [28].
Resting systems remain in a steady state by the coordinated action of opposing but balanced kinetic processes. Thus, in general, altering one ore more of these rate processes (e.g., increasing the catalytic rate of a reaction) should upset the balance of the system and cause it to adopt a new steady state. Various cell types have been shown to have altered steady-state properties because of mutations that affect the constitutive rates of reactions. For example, patients with type 1 diabetes harbor more Ca2+ ATPase activity in their platelets than healthy volunteers and experience high resting levels of intracellular Ca2+ [34]. In a separate case, a mutation within the tyrosine kinase domain of epidermal growth factor receptor causes significantly higher basal (growth factor-independent) tyrosine phosphorylation levels than the wild-type receptor [35]. Therefore, to examine the changes in steady-state properties caused by kinetic perturbations in our example model, we altered the rates of 3 important regulatory reactions and observed the system response to each perturbation. Each perturbation cause a brief adjustment phase lasting ∼200 s followed by a more gradual phase characterized by a new steady-state profile (Figure 4, left). After 1 hr of simulated time, steady-state concentrations and reaction fluxes were quantified relative to their original steady-state levels (Figure 4, right).
As expected, increasing the rate of Ca2+ release from intracellular stores resulted in higher cytosolic Ca2+ levels (7-fold increase) and 10-fold greater pumping activity by plasma membrane Ca2+ pumps (PMCA), although the new steady-state Ca2+ release flux remained relatively unchanged (Figure 4A). This perturbation also had little effect on the metabolism of phosphoinositides, as indicated by a predominantly green color. In a second perturbation, the inhibition of phospholipase C-β (PLC-β) activity by protein kinase C (PKC) was reduced 10-fold. Since PKC has a negative-feedback role in suppressing the platelet-stimulating activity of PLC-β, this perturbation caused a 2-fold increase in steady-state phosphatidylinositol 4,5-bisphosphate (PIP2) hydrolysis, elevated (inositol 1,4,5-trisphosphate) IP3 concentration, and accelerated Ca2+ release. Interestingly, the same reaction that was initially perturbed with a 10-fold decrease experienced a 10-fold increase in steady-state flux. This was a compensatory effect caused by the negative feedback loop involving Ca2+-regulated activity of PKC, a resulting new hypothesis that can be probed experimentally. In a third example, increasing the hydrolytic activity of PLC-β for the substrate PIP2 by 10-fold caused an expected stimulatory effect, raising intracellular calcium and steady-state levels of cytosolic inositol phosphates (IP3, IP2, and IP) between 2- and 3-fold. Interestingly, reaction fluxes for phosphoinositide hydrolysis were diminished, possibly due to substrate depletion. Taken together, these examples illustrate the system-wide effects of perturbations in the kinetic rate processes. The procedure could easily be extended to examine multiple simultaneous perturbations in both reaction rates and steady-state concentrations.
We have presented a novel strategy for enumerating permissible steady-state solutions to fixed kinetic topologies and combining these solutions spaces to form large kinetic models. This is a practical strategy because kinetic parameters are commonly reported whereas absolute concentrations are not (see, for example, [6],[8],[9]). The method extends the capability to build “genome-scale” models [5],[10],[11],[36] to include nonlinear kinetic features. Through application of the method, we have also explored the implicit restrictions on steady-state solutions that can be imposed by the underlying kinetic structures within a system [4]. This is useful from a physiological standpoint since the regulation and distribution of molecular species in living systems is largely regulated by the coordinated action of synthetic, degrading, and transporting enzymes.
The proposed method requires the model to fulfill a steady-state assumption (i.e., the model must contain nontrivial steady states) even if the system is typically characterized by transient behavior. It is precisely this requirement that allows the model to have the dual functional behavior observed in many biological contexts, such as in cellular signaling responses. At very low levels of activating signal, the model remains at rest by quenching the low level of activating signal through feedback mechanisms or futile cycling. When activating signals are increased, the system responds with the appropriate transient signaling behavior. As an example, a human platelet must remain quiescent under normal circulating conditions, tolerating a number of fluctuations in its surrounding chemical and physical environment. In the presence of the appropriate stimulus, however, it must be able to respond rapidly to bleeding conditions and trigger a precise program of molecular signaling events. Developing a mathematical model that is consistent with two or more biological behaviors is analogous to writing a set of equations that has multiple solutions, each dependent on a given set of initial conditions and parameter values.
Our approach differs critically from metabolic flux analysis and previous genome-scale metabolic network reconstructions [5],[16] because it accommodates nonlinear terms that describe the dynamic behavior of each reaction in the system. Previous large-scale network reconstructions typically use a stoichiometry matrix to represent the gross flux of metabolites in the system [17]. Here, we have preserved the mathematical form of each kinetic rate equations as reported in the literature, allowing models to be built from existing data in a “bottom-up” fashion [10] while still allowing calibration to whole-system experimental data. This feature will substantially improve the accuracy of dynamical system simulation and parameter estimation.
Additional computational savings are provided through modularization. When estimating modules of modest size (5 or less unknown concentrations), we use a brute-force Monte Carlo approach to densely sample the feasible space of initial conditions. Larger networks (20 or more unknowns) cannot be efficiently searched in this brute-force manner, but can be built piecewise by combining subspaces of smaller size that have been densely sampled. Using the naïve Monte Carlo approach, estimating n free parameters is exponential in n. By dividing these parameters into k independent networks, each with n/k free parameters, the estimation procedure becomes exponential in n/k and thus more tractable. By assembling the entire system from smaller, more manageable kinetic modules, data may be used to test the functionality of individual modules before incorporating them into the entire system. In several cases, this approach was shown to offer a substantial computational benefit (e.g., reducing the global search space by over 10,000-fold) by simply requiring a steady-state solution with known subcomponent values. The search space can be reduced further by principal component analysis if there is correlation between free parameters within a module. This was found to be the case for enzymes that have opposing regulatory roles; increasing the levels in one enzyme required a similar increase in the other in order to preserve homeostasis. Lastly, modules sharing common components must hold the same value for that component, which imposes an additional constraint on the steady-state solutions (equation (4)).
As presented, the method exploits known kinetic parameters to restrict unknown concentrations due to kinetic interactions. However, the method is equally valid for estimating unknown kinetic parameters and/or utilizing known concentrations. Both concentrations and kinetic parameters appear indistinguishably as nonlinear terms in the ordinary differential equations that describe the system (Figure 1B). Hence, it does not matter which types of values are known and which are estimated; the procedure is valid for mixed or incomplete sets of unknown values. The use of qualitative data may also be exploited by the method. For example, beginning with a large set of steady-state solutions for a given module, the size of the set may be reduced by determining which solutions in the set contain some qualitative behavior or function. In a previous application of the method [17], a set of 109 steady-state solutions representing calcium balance in a resting platelet were divided into 3 groups, according to their qualitative response to increased IP3 concentration (low, mild, and high response). Using this technique, the functional testing of steady-state modules may be used to eliminate a large subset of the original steady-state solution set. As another example, one may use data from a Western blot to establish the relative abundance between two proteins in the model. This qualitative information may be used to filter the steady-state solutions to a reduced set that is consistent with experimental results. This kinetically-driven, constraint-based approach, which combines a homeostasis requirement with known kinetic parameters and cellular concentrations, naturally enforces numerical limits on unknown system quantities.
|
10.1371/journal.pntd.0003152 | Detection of Rickettsia felis, Rickettsia typhi, Bartonella Species and Yersinia pestis in Fleas (Siphonaptera) from Africa | Little is known about the presence/absence and prevalence of Rickettsia spp, Bartonella spp. and Yersinia pestis in domestic and urban flea populations in tropical and subtropical African countries.
Fleas collected in Benin, the United Republic of Tanzania and the Democratic Republic of the Congo were investigated for the presence and identity of Rickettsia spp., Bartonella spp. and Yersinia pestis using two qPCR systems or qPCR and standard PCR. In Xenopsylla cheopis fleas collected from Cotonou (Benin), Rickettsia typhi was detected in 1% (2/199), and an uncultured Bartonella sp. was detected in 34.7% (69/199). In the Lushoto district (United Republic of Tanzania), R. typhi DNA was detected in 10% (2/20) of Xenopsylla brasiliensis, and Rickettsia felis was detected in 65% (13/20) of Ctenocephalides felis strongylus, 71.4% (5/7) of Ctenocephalides canis and 25% (5/20) of Ctenophthalmus calceatus calceatus. In the Democratic Republic of the Congo, R. felis was detected in 56.5% (13/23) of Ct. f. felis from Kinshasa, in 26.3% (10/38) of Ct. f. felis and 9% (1/11) of Leptopsylla aethiopica aethiopica from Ituri district and in 19.2% (5/26) of Ct. f. strongylus and 4.7% (1/21) of Echidnophaga gallinacea. Bartonella sp. was also detected in 36.3% (4/11) of L. a. aethiopica. Finally, in Ituri, Y. pestis DNA was detected in 3.8% (1/26) of Ct. f. strongylus and 10% (3/30) of Pulex irritans from the villages of Wanyale and Zaa.
Most flea-borne infections are neglected diseases which should be monitored systematically in domestic rural and urban human populations to assess their epidemiological and clinical relevance. Finally, the presence of Y. pestis DNA in fleas captured in households was unexpected and raises a series of questions regarding the role of free fleas in the transmission of plague in rural Africa, especially in remote areas where the flea density in houses is high.
| Fleas are associated with many bacterial diseases such as rickettsioses, bartonelloses and plague. These diseases may be severe, and little is known about their prevalence. Accordingly, we believe that our data shed light on the problem of unexplained fevers in tropical and subtropical African areas. Using molecular tools, we surveyed and studied selected flea-borne agents, namely Rickettsia spp. (R. felis and R. typhi), Bartonella spp. and Y. pestis, in fleas collected in Ituri (Linga and Rethy health zone) and Kinshasa in the Democratic Republic of the Congo, the Lushoto district in the United Republic of Tanzania and in Cotonou in Benin. We found that these bacteria are present in the studied regions. R. typhi and an unidentified Bartonella sp. were detected in fleas from Cotonou (Benin). R. felis and R. typhi were also detected in the Lushoto district (United Republic of Tanzania). Finally, we detected R. felis, Bartonella sp. and Y. pestis in the Democratic Republic of the Congo. As these emerging zoonotic diseases have a global distribution and affect public health, the implementation of vector control strategies is urgent.
| The importance of fleas in human and animal health is largely related to their ability to transmit agents of infectious diseases [1]. The transmission of these zoonotic agents to human occurs mainly through bites or inoculation of feces into pruritic bite lesions [2], [3]. Plague is the most notorious flea-borne disease known to man and is a reemerging public health issue mainly in Africa and South America [3]. The etiological agent of plague, Yersinia pestis, is a facultative gram-negative bacterium restricted nowadays to well defined endemic foci [4], [5]. In the last decade, plague reemerged in old quiescent foci of Algeria [6], the United Republic of Tanzania [7] and Libya [8] and caused remarkable bubonic and pneumonic outbreaks in known endemic foci in Madagascar [9] and in the Democratic Republic of the Congo [10]. Fleas are also associated with other bacterial diseases such as bartonelloses and rickettsioses. Rickettsia spp., the etiological agents of rickettsioses, are intracellular gram-negative bacteria that represent an emergent global threat, particularly in the tropics [11]. R. felis, an emerging pathogen, and R. typhi, the agent of murine typhus (MT), are the main rickettsial pathogens associated with fleas [1], belonging to the spotted fever group (SFG) [12] and typhus group rickettsiae, respectively [13]. Although these two flea-borne rickettsiae are distributed worldwide, R. typhi appears to be more endemic in tropical regions, coastal areas and ports, where its transmission cycles between rats (Rattus spp.) and oriental rat fleas (X. cheopis) [14]. Also, several closely related rickettsiae, referred as Rickettsia felis–like organisms (RFLO), identified in fleas and other arthropods around the world [15]. Likewise, bartonelloses are diseases caused by the fastidious, hemotropic bacteria of the genus Bartonella, especially in debilitated and immunocompromised individuals [16]. Importantly, the list of host species harboring Bartonella spp. includes a large number of mammals, mostly rodents, some of which are kept as pets [17].
An increasing number of papers have reported the occurrence of fleas and human flea-borne infections, especially in relation to wildlife and zoonotic risk. However, the identity and distribution of flea-borne agents in urban, domestic or peridomestic settings have been poorly documented in Sub-Saharan African countries such as the Democratic Republic of the Congo, the United Republic of Tanzania and Benin. Historical data about human infection with Rickettsia and Bartonella species are fragmentary, and virtually nothing is known about the current distribution of these flea-borne zoonotic agents in potential vectors and reservoir hosts in these countries.
In the Democratic Republic of the Congo, recent small-scale surveys have reported serological evidence for Bartonella infection in human patients [18] and molecular data in rodents [19] and fleas [20], suggesting a global underreporting at the country scale. Rickettsioses in humans are mentioned in historical reports; however, their notification remains anecdotal, and the species identification is likely erroneous. Recently though, among febrile military patients in Kisangani, Democratic Republic of the Congo, one patient tested positive in 1999, for the R. typhi antigen using serological tools. In addition, R. felis has been found to circulate in arthropod vectors in Kinshasa [21]. As a general trend, flea-borne agents in fleas are underreported, whereas in the United Republic of Tanzania, a growing number of publications confirm their presence and wide distribution in humans [22] exposed to their bites and in infested rodents [19].
In recent years, our laboratory (Unité de Recherche sur les Maladies Infectieuses et Tropicales, the WHO Collaborative Centre for Rickettsial Diseases and Other Arthropod-Borne Bacterial Diseases in Marseille, France) initiated collaboration with correspondents and universities in the United Republic of Tanzania, the Democratic Republic of the Congo and Benin.
The present survey pursued the objectives of detecting the presence and identity of Rickettsia spp., Bartonella spp. and Y. pestis in flea specimens collected from domestic and peridomestic areas in the Democratic Republic of the Congo, the United Republic of Tanzania and Benin within the context of entomological studies.
Risk assessment was submitted to and approved by the ethical committee and decision board of each institution involved in small mammals trappings, and involved informed consent of the domestic animal owners; ethical approval are available from original publications on mammal hosts on which flea were collected [19], [23], [24]. The Ethical commitee of the University of Antwerp, Belgium and the Sokoine University of Agriculture Morogoro under the project RATZOOMAN granted by the European Commission Framework 5 Programme on International Cooperation, project contract number ICA4 CT 2002 10056, approved the experiment in the South-eastern Africa.
See here technical annex: http://projects.nri.org/ratzooman/docs/technical%20annex.pdf.
The material analyzed consisted of fleas (Siphonaptera) collected in domestic and peridomestic areas in Benin, the United Republic of Tanzania and the Democratic Republic of the Congo (Figure 1). A portion of the collected fleas was used for the present study. A convenient sample was selected according to a good representation of species, host and localities.
In 37 sites in the capital city of Benin, Cotonou (6°21′36″N; 2°26′24″E), rodent fleas were collected from rodents trapped monthly inside human residences and peridomestic areas between November 2009 to July 2010, as described previously [24]. In the United Republic of Tanzania, 17 sites in the Lushoto district (04°40′00″S 38°19′00″E) located in the Tanga Region were surveyed [23], [25]. Lushoto district is a mountainous area where plague was reported from the first time in 1981; this endemic plague focus has however been quiescent since 2004. Between May 2005 and November 2008, fleas were collected – as in Benin – from small mammals in domestic and peridomestic habitats during the dry and rainy seasons. Further details on the rodent measurements and flea collection have been published elsewhere [23], [25].
Finally, in March and April 2007, rodent fleas and free domestic fleas were collected from 4 villages (15 capture sites) in the Linga and Rethy health zones, Ituri district, Orientale Province, the Democratic Republic of the Congo; off-host fleas were collected in 4 villages during an investigation following a plague outbreak that occurred in the third trimester of 2006 [26]. Our investigation area was limited to Djalusene (2°12′10″5 N 30°88′02″7 E) and Kpandruma (2°05′90″1 N 30°88′70″4 E), which had confirmed plague patients, and Wanyale (2°10′11″8 N 30°80′60″5 E) and Zaa (2°14′03″2 N 30°85′65″9 E), which had several suspect cases but were considered control areas at the time of the study. We collected fleas in 40 houses (bedroom) in each village, for 3 nights in a row, using a kerosene lamp hung above a 45-cm diameter tray containing water as described in [27]. In April 2010 and July 2012, additional flea samples were collected from the Ituri district in Rethy village (1°50′N 29°30′E) and in Kinshasa (4°19′19″S 15°19′16″E) by means of light traps in human residences (bedroom) and rodent burrows, and flat tweezers on dogs.
All fleas collected in Benin, the United Republic of Tanzania and the Democratic Republic of the Congo were stored in 70% ethanol and identified morphologically using classical entomologic taxonomic keys. The samples were later processed in the WHO Collaborative Center for Rickettsial Diseases and Other Arthropod-Borne Bacterial Diseases, in Marseille, France.
Fleas were rinsed twice in distilled water for 10 minutes and dried on sterile filter paper; the handling was performed in a laminar flow biosafety hood. The fleas were individually crushed in sterile Eppendorf tubes, as described [28]. A total of 50 µl of DNA was extracted from one half of each flea using the QIAamp Tissue Kit (Qiagen, Hilden, Germany) by QUIAGEN-BioRobot EZ1, according to the manufacturer's instructions. The genomic DNA was stored at −20°C under sterile conditions until used as the template in PCR assays. The remaining portion of each flea was kept at −80°C for an additional control.
All samples were screened by quantitative real-time PCR (qPCR) targeting the biotin synthase (bioB) gene, as previously described [29]. Positive results were confirmed by another qPCR targeting a membrane phosphatase gene with primers (Rfel_phosp_MBF: 5′-GCAACCATCGGTGAAATTGA-3′ and Rfel_phosp_MBR: 5′-GCCACTGTGCTTCACAAACA-3′) and a probe (Rfel_phosp_MBP: 6FAM-CCGCTTCGTTATCCGTGGGACC-TAMRA) designed in our laboratory. The final mixture of the qPCR reaction was composed of 15 µL of mix that contained 10 µL of master mix QuantiTect Probe PCR Kit (QIAGEN, Hilden, Germany), 0.5 µL (20 pmol) of each primer, 0.5 µL (62.5 nmol) of probe, 3.5 µL of RNase DNase-free water and 5 µL of DNA extracted from fleas. qPCR was performed as follows: 15 min at 95°C, followed by 40 cycles of 1 s at 95°C, 40 s at 60°C and 40 s at 45°C, as described [29]. The negative control consisted of DNA extracted from uninfected fleas from our laboratory colony and was used for all the PCR assays in this work. The positive control was DNA extracted from a diluted strain of R. felis from our laboratory in Marseille. Positive results were recorded if the cycle threshold (Ct) value obtained was lower than 36 using the 2 PCR systems.
Samples were screened by qPCR targeting a fragment of the Rpr 274P gene coding for a hypothetical protein, as described previously [30]. Positive results were confirmed by qPCR targeting the glycosyltransferase gene using a previously described Rpr 331 system [31]. qPCR was conducted using the same method as described for R. felis detection. The positive control was DNA extracted from a diluted strain of R. typhi Wilmington (ATCC VR-144) cultured in our laboratory in Marseille.
DNA samples were screened by quantitative real-time PCR targeting the ITS region [32]. Positive samples with ITS primers were then confirmed by standard PCR performed with Bartonella-specific primers for the citrate synthase (gltA) gene, amplifying an approximately 334-bp fragment [33]. The positive control was B. alsatica strain IBS 382 (CIP 105477) DNA extracted from a strain and previously diluted to 10−6.
The success of PCR amplification was verified by 2% agarose gel migration. The products were purified using NucleoFast 96 PCR plates (Machery-Nagel EURL, France) as recommended by the manufacturer. The purified PCR products were sequenced with gltA primers using the BigDye version 1.1 cycle sequencing ready reaction mix (Applied Biosystems, Foster City, CA) with the ABI 31000 automated sequencer (Applied Biosystems). The sequences were assembled and analyzed with the ChromasPro program (version 1.5).
DNA samples were screened by qPCR targeting the plasminogen activator gene (Pla) [6] using primers Yper_PLA_F (5′-ATG-GAG-CTT-ATA-CCG-GAA-AC-3′) and Yper_PLA_R (5′-GCG-ATA-CTG-GCC-TGC-AAG-3′) and probe Yper_PLA _P (6- FAM-TCC-CGA–AAG-GAG-TGC-GGG-TAA-TAGG-TAMRA). Positive results were confirmed with standard PCR targeting the glpD gene, as described [34], and then sequenced using the same method used for Bartonella spp. sequencing. The positive control was Y. pestis DNA extracted from the CSUR P 100 strain, and diluted to 10−6.
In Benin, 886 fleas were collected from 199 sexually mature small mammals of four species, namely, Crocidura olivieri (17/199, 8.5%), Mastomys natalensis (36/199, 18%), Rattus norvegicus (40/199, 20.1%) and Rattus rattus (109/199, 54.7%). Three flea species were collected from rodents, with the oriental rat flea X. cheopis being the most abundant (861/886, 97.1%), followed by X. brasiliensis (24/886, 2.7%) and Ct. felis strongylus (1/886, 0.1%). In the present study, a convenient sample of 199 X. cheopis (picked off Rattus rattus) individuals – 55.78% females and 44.2% males – were selected for an initial molecular screening (the remaining fleas were preserved for subsequent studies).
All fleas tested negative for R. felis and Y. pestis. qPCR performed for the detection of R. typhi revealed 2 positive X. cheopis (2/199, 1%), with a Ct of 32.6 and 34.5, from 2 sites (Bokossi Tokpa and Dédokpo). Bartonella spp. were detected in 69/199 (34.6%) of the fleas (Ct, 31.81, +/−2.97) (24≤Ct≤35) collected from all studied sites (Table 1). DNA sequence analyses of the PCR products of the gltA gene of 8 representative samples (with high Ct values) showed 100% similarity with the Uncultured Bartonella sp. clone Pd5700t (GenBank no. FJ851115.1, 334/334 bp) detected in Praomys delectorum rodents in Mbulu district, northern Tanzania [19]. More information about the Ct value and localization of each positive flea is reported in Supplementary data S1.
A total of 3821 fleas (rodent fleas and free-roaming fleas present in the environment) were collected from localities of the Lushoto district (United Republic of Tanzania) and were distributed into 23 species. A total of 94 fleas belonging to six common species were screened (Supplementary data S2) (20 Ct. f. strongylus, 7 Ct. canis, 20 Ctenophthalmus calceatus calceatus, 20 X. brasiliensis, 20 Pulex irritans and 7 Nosopsyllus incisus. All tested fleas were negative for Y. pestis and Bartonella spp. DNA. However, R. typhi DNA was detected in 10% (2/20) of X. brasiliensis collected from 2 villages (Magamba and Manolo). R. felis DNA was also detected in 20.2% (23/94) of analyzed fleas, including 65% (13/20) of Ct. f. strongylus, 71.4% (5/7) of Ct. canis and 25% (5/20) of Ct. ca. calceatus.
In 2007, in the Linga and Rethy health zones, Ituri district, 1190 fleas captured in households, belonging to 6 species (394 P. irritans, 153 Tunga penetrans, 280 Ct. f. strongylus, 89 Echidnophaga gallinacea, 88 L. a. aethiopica and 186 X. brasiliensis). A total of 123 fleas were conveniently selected for this work (Supplementary data S3). qPCR for R. typhi and Bartonella spp. was negative for all 123 fleas; however, 4.8% (6/123), namely 19.2% (5/26) of Ct. f. strongylus and 4.7% (1/21) of E. gallinacea, contained R. felis DNA (Table 1).
Y. pestis DNA was detected in 3.8% (1/26) of Ct. f. strongylus and 10% (3/30) of P. irritans from 2 villages (Wanyale and Zaa). DNA sequence analyses of the PCR products targeting the glpD gene showed 100% similarity with Yersinia pestis Angola isolated from Angola (GenBank accession no. CP000901.1, 321/333 bp).
In 2010, 111 fleas, belonging to 3 species, were collected in the same district, namely, X. cheopis (62/111, 55.8%), Ct. f. felis (38/111, 34.2%) and L. a. aethiopica (11/111, 9.9%) (Supplementary data S4). qPCR for R. typhi and Y. pestis detection was negative for all fleas (Table 1); however, 9.9% (11/111) of two flea species (Ct. f. felis and L. a. aethiopica) were positive for R. felis. A total of 10 Ct. f. felis from 38 tested (26.3%) and one of 11 L. a. aethiopica (9%) contained R. felis. Bartonella spp DNA was detected in 3.6% (4/111) of fleas, with 36.36% (4/11) from only L. a. aethiopica. Sequencing of the gltA gene fragment from these four Bartonella-positive samples showed 100% similarity with Bartonella sp. MN-ga6 (GenBank no. AJ583126.1, 320/334 bp) detected in fleas collected in South Africa.
Finally, in 2012, from the fleas collected in Kinshasa (Table 1), 56.5% (13/23) of Ct. f. felis collected from 3 dogs was positive for R. felis but negative for R. typhi, Bartonella spp. and Y. pestis by qPCR.
We report the first direct evidence of R. typhi and Bartonella sp. in X. cheopis fleas in Benin (Cotonou). In Lushoto (United Republic of Tanzania), we detected for the first time the presence of R. typhi DNA in X. brasiliensis and R. felis DNA in Ct. f. strongylus, Ct. canis and Ct. ca. calceatus. Finally, in the Democratic Republic of the Congo, we confirmed the presence of R. felis in Ct. felis in Kinshasa and for the first time report the presence of R. felis and Bartonella DNA in L. a. aethiopica and, most importantly Y. pestis DNA in P. irritans and Ct. felis from Wanyale and Zaa villages in the Rethy health zone.
The robustness of our results and the detection of these pathogens in fleas on rodents are supported by the use of a validated method of real-time PCR and subsequent sequencing. The validity of the data that we report is based on strict laboratory procedures and controls that are commonly used in the WHO Center for Rickettsial Diseases, including rigorous positive and negative controls to validate the test. Each positive qPCR result was confirmed by another specific qPCR or confirmed with a successful DNA amplification and sequencing.
R. typhi was detected in X. cheopis collected from Rattus rattus in Bokossi Tokpa and Dédokpo sites (Cotonou, Benin) and in X. brasiliensis from the United Republic of Tanzania. X. cheopis is the primary vector of R. typhi, the etiological agent of murine typhus (MT), in most locations around the world, and X. brasiliensis appears to be an effective vector under experimental conditions [3]. MT is most often a relatively mild disease; yet R. typhi can cause acute febrile illness and death [35]. The diagnosis of MT may be missed or underreported due to its non-specific symptoms or the absence of epidemiological criteria [36], [37] because laboratory tests and validated methods of diagnosis must be performed to confirm the diagnosis [30]. Before our study, R. typhi was never detected in Benin, and it is rarely directly reported in vectors and patients in Africa, specifically in sub-Saharan Africa. R. typhi in African fleas was only detected in X. cheopis fleas collected in Algeria [38]. Additionally, R. typhi has been reported in patients using serological methods in African countries [30]. Cases have been documented in international travelers returning from Tunisia, Morocco, Ivory Coast, Central African Republic, Madagascar, Reunion and Chad [30]. In the United Republic of Tanzania, a seroprevalence study among pregnant women from the port city of Dar es Salaam found a prevalence of 28% [39] and 0.5 to 9.3% in the town of Moshi and the Mbeya region, respectively [22], [40].
R. felis is an emergent agent of infectious disease in humans, and this agent of spotted fever is known to be maintained in cat fleas (Ct. felis) [41], [42]. To date, 12 species of fleas, 8 species of ticks and 3 species of mites have been found to be infected with R. felis [42]. This Rickettsiae has also recently been detected in several mosquito species in sub-Saharan Africa [29], [43], [44]. Interestingly, the R. felis genogroup seems large with recent organisms or genotypes related as R. felis like organisms (RFLO). Our 2 qPCR were specifically designed to amplify R. felis type strain (URRWXCal2). However, the biotin synthase and membrane phosphatase gene sequences of many RFLO are not known. We however know that at least our qPCR system targeting the biotin synthase (bioB) gene do not amplify some RFLO such as Rickettsia sp. RF2125 and Rickettsia sp. SGL01. Recently, a new qPCR assay has been proposed to address this issue by providing new qPCR primers and probe to specifically amplify R. felis OmpB gene fragments [15]. The clinical features of R. felis may include fever, fatigue, headache, generalized maculopapular rash and inoculation eschar(s) [42]. R. felis seems to be a frequent agent of unknown fever in Sub-Saharan Africa [44]. We detected R. felis in 5 species of fleas (Ct. f. strongylus, Ct. canis, Ct. ca. calceatus, L. a. aethiopica and E. gallinacea); some from the United Republic of Tanzania (Lushoto district), and other from the Democratic Republic of the Congo (Ituri District). R. felis had already been detected in the Ituri district [25], but not in E. gallinacea, the fowl flea, and has been previously shown to circulate in arthropod vectors (Ctenocephalides felis) in Kinshasa, the capital city of the country [21]. E. gallinacea is usually found on poultry, and can occurs on rodents (Rattus spp.) foraging in fowl shelters around houses [45]. While chicken DNA has been found in blood meal of fleas collected on rodents in the same area [46] other Rickettsia spp. antibodies have been found in poultry in Brazil [47], whether or not R. felis and R. typhi infects poultry or if poultry can act as a source of infection to human is unknown. Furthermore, no data on the potential vertical transmission of R. felis in E. gallinacea, or on the vectorial transmission of R. felis by E. gallinacea males (females are semi-sessile) between rodents and birds, are available. The questions raised by the findings of the present study in relation to Rickettsia in fleas are of real epidemiological significance and should be further investigated.
Molecular evidence of Bartonella sp. in fleas from the Democratic Republic of the Congo is supported by a recent serological survey in human patients in the Ituri who tested seropositive for B. henselae, B. quintana or B. clarridgeiae [18]. Gundi and collaborators also found that local rodents harbor Bartonella spp. closely related to B. elizabethae or B. tribocorum which shows that a wide variety of Bartonella species is present in the country, and differ according to host [19]. Bitam and collaborators [48] report that B. elizabethae, which causes endocarditis, and B. tribocorum are usually known to be transmitted by X. cheopis fleas. However, while in our study, we detected an Uncultured Bartonella sp., clone Pd5700t (GenBank no. FJ851115.1) in X. cheopis of Benin, we also detected Bartonella sp. MN-ga6 (GenBank no. AJ583126.1) in L. a. aethiopica, from Ituri. This Bartonella sp. had been previously found in the Democratic Republic of the Congo and the United Republic of Tanzania in rodents [19].
The detection of Y. pestis DNA in fleas collected in villages and houses where no current human plague cases had been reported for the last 6 months is puzzling. About 80 species and subspecies of Siphonaptera are known to be carriers and potentially vector of Y. pestis [49], via various transmission mechanisms [50]; in particular in fleas from the genus Xenopsylla (X. cheopis), which played a major role in historical plague pandemics [9]. In the present survey, DNA of Y. pestis was detected in the human flea, P. irritans, and the cat flea Ct. felis in a well known endemic focus of the Democratic Republic of the Congo [51]. In 2006, in the Rethy and Linga health zone more than 600 human cases were reported [52], which triggered the entomological investigation reported previously [25] and the collection of fleas analyzed herein. This survey occurred 6 months after the end of the epidemics, and at the time of the flea sampling, no confirmed human plague cases were reported to the Health centre of the villages (Zaa and Wanyale) or Rethy general Hospital. Several hypotheses can be proposed to explain this finding. A first hypothesis is that infected fleas from rodents, dogs or cats could have been imported in the infested houses, did not bite people and as such no human cases occurred, at the time of collection. A second hypothesis is that infected fleas containing Y. pestis DNA remained infected and alive without biting any potential host or that no human cases were reported to the health authorities which are unlikely due to the recent outbreak and constant surveillance. Other options are that Y. pestis DNA is reminiscent in the flea but the bacterium is either dead (degraded DNA) but the targeted sequences (gene fragment and gene flanking regions are still complete) or alive but in a quiescent form or VBNC state, possibly controlled by epigenetic mechanisms causing virulence gene repression. The human flea (P. irritans) may play an important role in spreading plague via human-to-human transmission as suggested in Lushoto district [27] and could possibly harbor Y. pestis without transmission for several months. Unfortunately no fleas were cultured in the field and the viability of the strain detected cannot be proven, but this finding calls for more research at times post outbreaks in order to answer this question. Similarly, cat fleas could play such a role both in northwest Uganda [53] and in Democratic Republic of the Congo (Laudisoit and al 2014, unpublished data), where C. felis spp. is the most common flea species collected in the domestic environment above a given altitude threshold.
In conclusion, we widened knowledge of the repertoire of flea-borne bacteria present in Sub-Saharan Africa. In our study, we also illustrate the role of fleas in the entomological survey of vector -borne disease, which allow clinicians to confirm the etiological cause for some of the unknown cause of fever in African patients. Future studies on rickettsioses, bartonelloses and other vector-borne diseases should be performed to assess their epidemiological and clinical relevance in tropical and subtropical areas, to estimate the real prevalence and to allow the establishment of antivectorial control plans.
|
10.1371/journal.ppat.1004229 | The Frustrated Host Response to Legionella pneumophila Is Bypassed by MyD88-Dependent Translation of Pro-inflammatory Cytokines | Many pathogens, particularly those that require their host for survival, have devised mechanisms to subvert the host immune response in order to survive and replicate intracellularly. Legionella pneumophila, the causative agent of Legionnaires' disease, promotes intracellular growth by translocating proteins into its host cytosol through its type IV protein secretion machinery. At least 5 of the bacterial translocated effectors interfere with the function of host cell elongation factors, blocking translation and causing the induction of a unique host cell transcriptional profile. In addition, L. pneumophila also interferes with translation initiation, by preventing cap-dependent translation in host cells. We demonstrate here that protein translation inhibition by L. pneumophila leads to a frustrated host MAP kinase response, where genes involved in the pathway are transcribed but fail to be translated due to the bacterium-induced protein synthesis inhibition. Surprisingly, few pro-inflammatory cytokines, such as IL-1α and IL-1β, bypass this inhibition and get synthesized in the presence of Legionella effectors. We show that the selective synthesis of these genes requires MyD88 signaling and takes place in both infected cells that harbor bacteria and neighboring bystander cells. Our findings offer a perspective of how host cells are able to cope with pathogen-encoded activities that disrupt normal cellular process and initiate a successful inflammatory response.
| Translation inhibition is a common virulence mechanism used by a number of pathogens (e.g. Diphtheria Toxin, Shiga Toxin and Pseudomonas Exotoxin A). It has been a mystery how host cells mount a pathogen-specific response and clear infection under conditions where protein synthesis is blocked by pathogens. Using Legionella pneumophila as a model, a bacterium that efficiently blocks the host protein translation machinery, we show here that the innate immune system has devised a mechanism to cope with translation inhibition by selectively synthesizing proteins that are required for inflammation.
| The pathogen-associated molecular pattern (PAMP) hypothesis has been developed to explain how the innate immune system recognizes foreign microbial invaders. By this model, germline-encoded receptors recognize conserved foreign ligands associated with microbes, such as nucleic acids, lipopolysaccharide (LPS), peptidoglycan or flagellin to generate a response directed at clearing the microorganism [1], [2]. More recently, it has become clear that pattern recognition alone does not explain how multicellular organisms are able to differentiate virulent pathogens from harmless commensals and mount a response. It has been proposed that the host immune system can sense the presence of danger and respond to pathogen-encoded enzymatic activities that disrupt normal cellular processes. This mode of recognition, referred to as “effector triggered immunity” has been shown to play a significant role in pathogen clearance both in plants and mammalian cells [3], [4], [5], [6], [7], [8], [9], [10], [11], [12]. Such recognition may be sufficient to activate a host response, but because it occurs simultaneously with PAMP recognition, host cell detection of pathogens likely results from integrating the recognition of microbial patterns together with pathogen-specific activities.
Legionella pneumophila, the causative agent of Legionnaires' disease, promotes intracellular growth by translocating proteins into its host cytosol through its type IV (Dot/Icm) protein secretion machinery [13], [14], [15]. These translocated effectors serve various purposes, including recruitment of ER-derived membrane to the Legionella containing vacuole, inhibition of cell death pathways and manipulation of host lipid metabolism and regulatory pathways [16], [17], [18], [19], [20], [21]. Most importantly for the innate immune response, after contact with macrophages, the bacterium stimulates a pathogen-specific response that is the consequence of simultaneous recognition of PAMPs and pathogen-translocated proteins that results in a unique response to this microorganism [10].
Legionella pneumophila is a pathogen for a broad range of fresh water amoebae, which provide the natural environmental niche for the microorganism and the source of exposure for humans [22], [23]. After aspiration by a susceptible mammalian host, the bacterium is engulfed by alveolar macrophages in the lungs [24]. In cultured macrophages, L. pneumophila provokes signaling through various pattern-recognition receptors (PRRs), such as Toll-like receptors (TLRs) and cytosolic NOD-like receptors (NLRs) [9], [25], [26], [27], [28], [29], [30], [31]. This response is critical for clearance of the microorganism, because mouse mutants defective in these two responses succumb to lethal pneumonia [27].
Interestingly, macrophage challenge with wild type L. pneumophila (Dot/Icm+) triggers a unique transcriptional response in host cells compared to mutants that lack a functional type IV secretion system, supporting the model that there is a pathogen-specific response involved in innate immune recognition [9], [10], [11],[17],[32]. Microarray studies have identified many of these transcriptional targets as being genes controlled by the NF-κB and mitogen-associated protein kinases (MAPKs) transcriptional regulators [9], [17], including downstream dual specificity phosphatases (Dusp1 and Dusp2), stress response genes (Hsp70, Gadd45a, Egr1) and pro-inflammatory cytokines and chemokines (Il1α, Il1β, Tnfα, Il23a, Csf1, Csf2) [9], [10], [11], .
It was recently demonstrated that the pathogen-specific response to Legionella is triggered by the action of L. pneumophila translocated effectors that interfere with host protein translation [10], [11]. Disruption of the host translation machinery serves as a second signal (in concert with signaling from PRRs) to constitute the full innate immune response against Legionella pneumophila [10]. The elimination of five of these effectors is sufficient to block this response, even though it is clear that they are part of a much larger pool of translocated substrates that impinge on host protein synthesis [10], [31]. These inhibitors (the products of the lgt1, lgt2, lgt3, sidI, and sidL genes) modify eukaryotic elongation factor eEF1A and eEF1Bγ of mammalian cells and block protein synthesis both in vitro and in vivo [10], [33], [34], [35].
In addition to blocking elongation, there is evidence that suggests wild type Legionella can also inhibit cap-dependent translation initiation [36]. Recognition of pathogenic Legionella leads to ubiquitination of the mTOR pathway, which in turn suppresses the eukaryotic initiation factor 4E (eIF4E) and prevents the synthesis of various genes [36]. This mode of translation inhibition was shown to induce translational biasing of host cells towards a more pro-inflammatory state [36]. However, it is currently not clear how host cells would be able to mount an inflammatory response when protein translation is blocked by L. pneumophila both at the initiation and elongation stages [9], [17]. It is likely that pattern-recognition would play a role under conditions of intoxication, but the mechanism by which this is regulated is also unclear.
A strong pro-inflammatory cytokine response is crucial for clearance of Legionella pneumophila [9], [31]. The importance of cytokines can be seen in IL-1α, IL-12, IFN-γ and TNF knockout mouse strains that show increased susceptibility to L. pneumophila infection [31], [37]. Moreover, patients treated with TNF-α blockers are at high risk of developing severe Legionnaires' disease [38]. Given the key role that this innate immune response plays in clearance of L. pneumophila, we examined how cytokines and other immune mediators are synthesized under conditions in which the bacterium effectively blocks the host protein translation machinery. We find that although protein synthesis inhibitors induce the transcriptional response and block the translation of most genes, pro-inflammatory cytokine genes can bypass this blockade in a fashion that requires the TLR-adaptor protein MyD88.
We hypothesized that Legionella-induced inflammatory gene transcription will be largely inconsequential, as the transcribed genes cannot be translated to proteins due to the bacterium-derived translation inhibitors. To test this hypothesis, we examined mammalian host protein synthesis predicted to be downstream of MAPK activation following exposure to L. pneumophila.
Bone marrow-derived macrophages challenged with wild type L. pneumophila showed phosphorylation of MAPK members shortly after exposure to the bacterium, with the activation kinetics being almost identical to previous observations ([9], [39], [40]; Figure S1). In the first hour after L. pneumophila challenge, activation of JNK and P38 was independent of the Icm/Dot system, consistent with phosphorylation being the result of Tlr engagement [9]. A second wave of activation was observed beginning at two hours after challenge. This was dependent on the presence of the L. pneumophila type IV secretion system, as the levels of MAPK phosphorylation decayed when macrophages were challenged with the dotA3 mutant that lacks the Icm/Dot system (Figure S1). This two-wave activation, reflecting an early Tlr-dependent and a later L. pneumophila-specific response, mirrors our previous observations with NF-KB activation [41].
Cultured macrophages challenged with wild type L. pneumophila transcriptionally activate a number of dual specificity phosphatase (DUSP) genes, dependent on an intact Icm/Dot system [17], [40]. Bone marrow-derived macrophages challenged with wild type Legionella for 4 hrs showed significant induction of the Dusp1 transcript compared to dotA3 infection (Figure 1A). The transcriptional induction was much higher (over 60 fold) in U937 human monocytes, the cell lines in which dusp1 induction in response to wild type Legionella was first characterized (Figure S2A) [17]. This transcriptional response, however, was not accompanied by translation either in bone marrow-derived macrophages (Figure 1B) or U937 cells (Figure S2B) as DUSP-1 protein levels remained unchanged over the course of infection. Our inability to observe enhanced protein synthesis was not due to limitations with our detection system, because we observed a robust increase in DUSP-1 protein levels in response to the addition of LPS (Figure 1C). Presumably the multiple translocated substrates that inhibit translation elongation [10] frustrate the transcriptional response, preventing translation of induced genes. To test this hypothesis, we challenged cells with an L. pneumophila mutant missing five of the translation inhibitors (Δ5) to determine if the absence of these proteins could allow translation to proceed. Instead, we observed no transcriptional activation of the Dusp1 gene in response to this mutant (Figure S2A). Therefore, induction of genes in the MAPK pathway and frustration of the response are tightly coupled.
We then asked if the response leading to transcriptional upregulation of pro-inflammatory cytokine genes was similarly affected by translational inhibition. To understand how cytokines are regulated in response to L. pneumophila, we measured the transcription and translation of selected pro-inflammatory cytokines in C57/Bl6 (B6) macrophages following L. pneumophila challenge. Flagellin deficient (ΔflaA) mutants were used in these experiments to avoid Caspase 1-dependent cell death downstream from NAIP5/NLRC4 recognition of flagellin by B6 macrophages [42], [43], [44].
Infection of bone-marrow macrophages with virulent (Dot+) L. pneumophila induced Il1α, Il1β and Tnfα transcripts by 6 hrs post infection (Figure 2A). As previously reported, the cytokine response to Dot+ was comprised of MyD88-dependent signaling that is layered on top of MyD88-independent, effector mediated signaling (Figure 2A, black bars) [10]. The response to the avirulent, dotA3 mutants on the other hand, was mostly dependent on MyD88 signaling (Figure 2A, white bars) [9].
To determine if the transcribed cytokine mRNAs were efficiently translated and secreted during infection, WT and MyD88−/− macrophages were challenged with L. pneumophila and cytokine protein levels were measured by western blot and ELISA. Contrary to what we saw for DUSP-1, challenge with L. pneumophila Dot+ led to a significant increase in cell-associated pro-IL-1β levels after 4 hrs of infection (Figure 2B). IL-1α and IL-1β mature forms could also be detected in culture supernatants after 24 hrs (Figure 2C & 2D). Interestingly, challenge of macrophages with L. pneumophila dotA3 mutants accumulated pro-IL-1β transiently, with steady state levels reduced by 12 hr post infection (Figure 2B densitometry), but this was not sufficient to induce the release of mature IL-1β. This is consistent with the hypothesis that in addition to TLR signaling, Icm/Dot translocated substrates are required for persistent pro-inflammatory cytokine activation and secretion [32]. Surprisingly, cytokine translation was severely diminished in MyD88−/− macrophages in response to wild type L. pneumophila (Figure 2B) despite the presence of large amounts of transcripts (Figure 2A). Therefore, a MyD88-dependent signal appears necessary to bypass the translation block induced in response to wild type L. pneumophila.
There is no clear model for why the presence of MyD88 allowed bypass of the translation block. Engagement of MyD88 on the host cell surface may lead to selective bypass of translation inhibition on a subset of transcripts. Alternatively, translation of cytokine transcripts could largely occur in neighboring uninfected cells that have not been directly injected with L. pneumophila translocated proteins, but which have been activated by bacterial fragments liberated by infected cells. We therefore asked if the observed cytokine translation was derived from neighboring bystander cells.
B6 macrophages were challenged with L. pneumophila-GFP strains and macrophages harboring bacteria were sorted from uninfected bystanders. Cytokine transcripts (Figure 3A) and protein levels (Figure 3B) were measured in each population by qRT-PCR and Western blots. To ensure accumulation of TNF-α protein, cells were treated with GolgiPlug (Brefeldin A; Materials and Methods) to prevent secretion of this cytokine. No such treatment was necessary for IL-1α and IL-1β, which accumulate as precursors via an alternate secretion pathway [45].
Relative to cells that had never been exposed to bacteria, challenge of macrophages with L. pneumophila resulted in high levels of I1α, Il1β, Tnfα and Dusp2 transcripts in both infected (GFP+) and neighboring uninfected (GFP−) populations by 4 hrs post infection, (Figure 3A). Depending on the cytokine, the amount of transcription in the bystander cells varied from 10% −30% of that observed in the cells harboring bacteria. Dusp1 on the other hand, was mainly transcribed in GFP+ cells (Figure 3A). More importantly, despite the presence of the translocated protein synthesis inhibitors, macrophages harboring bacteria were able to produce high levels of pro-IL-1α and pro-IL-1β (Figure 3B, GFP+ cells). The kinetics of pro-IL-1α and pro-IL-1β production in infected cells indicated that there was enhanced accumulation of these proteins between 4–6 hrs post-infection (Figure 3B, bottom panel). We will show that during this time window, L. pneumophila translation inhibitors effectively block most protein synthesis in infected cells (below, Figure 4B).
The presence of persistent cytokine synthesis in infected cells was confirmed by intracellular cytokine staining. B6 WT and MyD88−/− macrophages were challenged with L. pneumophila-GFP strains and intracellular cytokine levels were measured by flow cytometry. TNF-α was produced by both macrophages bearing bacteria (GFP+) and bystander cells (GFP−) after challenge with L. pneumophila Δfla, while the dotAΔfla strain induced much lower levels of this cytokine (Figure 3C, top panel 2nd and 3rd boxes). Cells harboring bacteria (GFP+) were a significant source of IL-1α (Figure 3C, bottom panel, 2nd box). Approximately 50% of the cells harboring bacteria showed detectable accumulation of IL-1α (Figure 3C, bottom panels, 2nd box), while bacteria were associated with approximately 33% of the IL-1α-producing cells. Consistent with Figure 2B, translation of IL-1α and TNF-α were both dependent on MyD88 signaling, and accumulation of IL-1a in infected cells was dependent on the presence of the Icm/Dot translocator (Figure 3C, two rightmost boxes in each panel).
Time course analysis of intracellular IL-1α levels using flow cytometry confirmed that the highest level of IL-1α accumulation occurred between 2–6 hrs post infection in the GFP+ population (Figure 3D and 3E). During this time period, the number of cells bearing bacteria that accumulated IL-1α increased from 1% of this population to approximately 50% (Figure 3D). DUSP-1 protein levels on the other hand, remain unchanged between 2–6 hrs (Figure 3E).
The two signals that are received by mammalian cells during L. pneumophila infection (1st signal from TLR activation and 2nd signal from protein translation inhibition) synergize to induce the full cytokine response against the bacterium. It was previously reported that pharmacological inhibitors of host protein translation induce transcription of various stress response genes and cytokines such as IL-6, IL-23, IL-α and IL-1β [10], [31]. We wanted to confirm that translation and secretion of these cytokines could always bypass translation inhibition using the protein synthesis inhibitor cycloheximide (CHX). CHX interferes with protein translation elongation by binding to the E-site of the 60S ribosomal subunit and preventing tRNA translocation [46].
Macrophages were treated with heat-killed Yersinia (HKY) to induce TLR signaling, together with 10 µg/mL of cycloheximide. Addition of the chemical inhibitor at the same time as HKY led to a complete inhibition of TNF-α and IL-1α production in bone-marrow macrophages (Figure 4A). Contrary to what we observed during L. pneumophila infection, addition of CHX dampened the signal received from TLR stimulation (Figure 4 A,B). Surprisingly, even at low concentrations of CHX that permit significant levels of protein translation (Figure S3), CHX was still able to inhibit IL-1β translation (Fig. 4B).
To rule out the possibility that the reduction in IL-1β levels during CHX treatment was due to cell death, we lowered the CHX dose to 0.5 µg/mL (inhibits less than 50% of total host protein synthesis) (Figure S3) and also incubated the cells with the apoptotic inhibitor Z-VAD-FMK (pan-caspase inhibitor). CHX was still able to inhibit IL-1β under these conditions, although it was clear that increased survival of cells was accompanied by higher accumulation of HKY-induced cytokine (Figure 4C).
MyD88-dependent stimulation of mouse macrophages in response to L. pneumophila primarily occurs via Toll-like receptor 2 (TLR2) [29]. A recent report was able to reconstruct the cytokine induction seen during L. pneumophila infection by using the synthetic TLR-2 ligand Pam3CSK4 in combination with Exotoxin A (Exo A), a toxin from Pseudomonas aeruginosa that interferes with translation elongation [31]. Based on this observation, we wanted to determine if specific activation of TLR2 is what leads to cytokine translation in the presence of protein synthesis inhibitors. Macrophages were treated with the TLR2 agonist Pam3CSK4 and pro-IL-1β levels were measured in the presence or absence of the protein synthesis inhibitor cycloheximide (CHX). Drug treatment after addition of the TLR2 agonist led to a large reduction in pro-IL-1β levels (Figure 4D). We also observed a failure to hyperstimulate pro-IL-1b in the presence of another protein elongation inhibitor, puromycin (Figure S4). Similar results were obtained when macrophages were stimulated with another TLR2 agonist lipoteichoic acid (LTA), or a TLR4 agonist LPS, followed by addition of CHX (data not shown). This indicates that the selective synthesis of cytokines may result from host cells sensing a specific mode of protein synthesis inhibition. It is also possible that the selective synthesis of pro-inflammatory cytokines is triggered by a block in translation initiation [36] instead of translation elongation, which would explain why the elongation inhibitors cycloheximide or puromycin were not able to induce the response.
We considered two possible explanations for how cytokines were translated in the presence of the bacterium-derived protein synthesis inhibitors: (1) translation inhibition by L. pneumophila is not efficient, allowing most of the cytokine to be synthesized prior to a complete block; or (2) the host preferentially translates a subset of genes after protein synthesis is shutdown by pathogens.
To distinguish between these possibilities, relative levels of protein synthesis were measured in bone-marrow macrophages at various times following L. pneumophila challenge, using an immunofluorescence readout. B6 macrophages were challenged with L. pneumophila flaA−-GFP at MOI = 10 for 2 hrs and translation of proteins was measured by incorporation of a methionine analog (L-azidohomoalanine, AHA) for an additional 4 hrs. Incorporated AHA was detected by reaction with a fluorescent-labeled phosphine reagent (phosphine-APC), which covalently links to the azido-functional group on AHA (Staudinger ligation reaction) [47].
Protein translation was significantly inhibited in macrophages harboring L. pneumophila (ΔflaA) compared to bacteria lacking the Icm/Dot system (Figure 5A). Furthermore, in macrophage cultures incubated with L. pneumophila, the macrophages harboring L. pneumophila showed selective protein synthesis interference, while the majority of the uninfected cells showed efficient incorporation of the amino acid analog (Figure 5B; compare GFP+ to GFP− population). To determine the time point at which the protein synthesis inhibitors fully shut down global protein translation, pulse-chase experiments were performed in which the methionine analog (AHA) was added for 1 hr intervals starting at 2 hrs post infection (Figure 5C). Between 2–3 hrs post-infection, approximately 40% of the cells harboring L. pneumophila were found in the population that has high levels of protein synthesis. Between 3–4 hrs post infection, we observed a major shift where almost 90% of cells harboring ΔflaA were found in the population having highly depressed protein synthesis. Later time points showed no further blockage in translation, perhaps reflecting the fact that there is a small fraction of wild type bacteria that fail to form replication compartments [48]. This population is predicted to show no significant translocation via the Icm/Dot system and should fail to inhibit protein synthesis (seen in ∼10% of the macrophage population).
To determine if translation-blocked cells could still produce cytokines, macrophage monolayers were challenged with L. pneumophila-GFP+ and protein synthesis was monitored by addition of AHA. The cells were then probed for IL-1α accumulation by immunofluorescence and flow analysis. In the infected GFP+ population of macrophages, the majority of cells that accumulated IL-1α show evidence of an almost complete shutdown of protein translation (Figure 5E; red box). In the absence of MyD88 signaling, both the infected and uninfected populations showed little IL-1α accumulation (Figure 5F). In contrast to IL-1α, there was no DUSP-1 accumulation after L. pneumophila challenge of macrophages (Figure 5D). This confirmed our main hypothesis that there is translation of selected cytokine genes when protein synthesis is inhibited by L. pneumophila and this bypass requires MyD88 signaling.
We have shown that pro-IL-1α and pro-IL-1β accumulate in cells that harbor L. pneumophila between 4–6 hrs post infection (Figure 3B), despite a significant block in protein translation (Figure 5C & E). To confirm that the accumulation we observe in infected cells was due to newly synthesized cytokines over the course of infection, we took advantage of another protein translation assay, SunSET. This assay uses puromycin incorporation into growing polypeptide chains to monitor active protein synthesis [49]. We modified this assay to measure the amount of puromycin incorporated into our protein of interest during a 1-hour pulse period. Accordingly, macrophages were challenged with L. pneumophila for increasing lengths of time, cells were labeled for one hour with 10 µg/mL of puromycin, lysed, and individual proteins were immobilized in assay wells using specific antibodies (Materials and Methods). An ELISA was then used to determine the amount of puromycin incorporated in the immobilized proteins (Figure 5G). Consistent with our intracellular cytokine staining and Western blot data (Figure 3), the highest levels of IL-1α and IL-1b synthesis were detected between 5–6 hrs post-infection. On the other hand, no significant puromycin incorporation was detected for DUSP-1 and RhoGDI proteins, confirming that there is selective synthesis of few genes after L. pneumophila challenge.
To determine the role that L. pneumophila translation inhibitors play in modulating host cytokine synthesis, we used a mutant that lacks the 5 Icm/Dot translocated substrates known to block host protein synthesis (Δ5 mutant) [10]. The level of protein synthesis was first measured using AHA incorporation (Figure 6A, B) and puromycin incorporation (Figure S5) in macrophages infected with Δ5ΔflaA-GFP+. Compared to Dot+ strain (Figure 6B & Figure S5), there was an increase in active protein translation in cells that were infected with the Δ5 mutant, although the cells showed lower levels of protein synthesis than Dot− infected cells (Figure 6B) or the uninfected population (Figure 6A; compare GFP+ to GFP− populations). This is consistent with the hypothesis that in addition to Icm/Dot translocated substrates that act on translation elongation, infection with virulent L. pneumophila also blocks translation initiation [36].
It had been previously reported that in the absence of most known pathways of pattern recognition (MyD88−/− Nod1−/− Nod2−/− macrophages), the cytokine transcriptional response to L. pneumophila was primarily due to the presence of the translocated protein synthesis inhibitors [10], [11]. Using macrophages that are only defective for MyD88, this dependence on the translation inhibitors could be clearly observed for the Il1α, Il1β and Tnfα transcripts (Figure 6C, MyD88−/−, gray vs. black bars). Consistent with our previous data, the transcriptional response in MyD88 knockout macrophages was unproductive, with no evidence that these highly induced transcripts are translated (Figure 6E).
In the case of wild type macrophages, Dot+ and Δ5 infections induced comparable levels of MyD88-dependent cytokine transcription and translation (Figure 6C and E), and this could be observed in macrophages that were sorted by flow cytometry, as well (Figure 6F). This result is in contrast with macrophages lacking MyD88 signaling, in which it is clear that there is protein synthesis inhibitor-dependent induction of cytokine transcripts (Figure 6C), but this induction produces no apparent cytokine translation products. Interestingly, unlike what we see for dotA mutants, infection with Δ5 was still able to induce secretion of mature IL-1α at 24 hrs after infection [31], even though it was to a lesser extent than wild type (Figure 6D).
Therefore, although the protein synthesis inhibitors are responsible for the transcriptional response that occurs in the absence of pattern recognition receptors, the production of cytokine proteins associated with infections by fully virulent strains is not dependent on these translocated substrates.
Cytokine expression is regulated at various stages, including transcription, post-transcriptional processing, translation and secretion. One of the main regulatory steps for IL-1 and TNF production is their transcript stability, which is controlled by their AU-rich elements (ARE) in their 3′-noncoding regions and by various ARE-binding proteins [50], [51]. To determine if lack of cytokine translation in MyD88−/− macrophages was due to mRNA instability, the half-life of Il1β and Tnfα transcripts were compared in WT and MyD88−/− macrophages after L. pneumophila challenge. Macrophages were first infected with L. pneumophila for 2.5 hrs, actinomycin D was added to the medium to block further transcription, and the amount of transcript remaining was measured by qRT-PCR at various time points after the addition of the drug. Tnfα mRNA was highly unstable in the absence of MyD88 when compared to MyD88+/+ macrophages (Figure 7A). However, Il1β transcripts were relatively stable in the absence of MyD88, and the amount of mRNA remaining after 2 hrs was similar to control macrophages (Figure 7A). This indicates that the translation inhibition bypass of Il1β was independent of mRNA stability (Figure 7). In the case of tnfα, however, transcript stabilization via a MyD88-dependent signal may play a role in bypassing translation inhibition.
Another possible explanation for why few pro-inflammatory cytokines, such as IL-1α and IL-1β, bypass translation inhibition could be mRNA abundance. A recent study has shown that these cytokines are the most abundant transcripts within macrophages after challenge with L. pneumophila [36]. Although there was significant transcription of these genes in MyD88−/− macrophages, the total mRNA abundance of IL-1α, IL-1β and TNF was still significantly lower compared to wild type macrophages (Figure 2). To address if mRNA abundance plays a role in bypass of translation inhibition during Legionella infection, WT and MyD88−/− macrophages were pre-activated with the TLR3 agonist poly(I∶C) for 2 hrs to induce NF-kB signaling. We also used heat-killed Yersinia pseudotuberculosis (HKY) to activate NF-kB via TLR4.
2 hr pre-stimulation with 50 µg/mL poly(I∶C) was sufficient to trigger IL-1β transcription (Figure 7B, left graph, white bars) both in wild type and MyD88−/− macrophages. Pre-stimulation with poly(I∶C) followed by L. pneumophila infection increased Il1β transcription initially (after 2 hrs of infection) in wild type macrophages compared to cells that were untreated (Figure 7B, compare black bars with grey bars). Surprisingly, this was reversed by 6 hrs post infection and cells that were pre-treated with poly(I∶C) down regulated their Il1β transcription (Figure 7B, right graph). This was more pronounced in MyD88−/− macrophages (Figure 7B, compare black and grey bars). We observed a very similar phenomenon when cells were pre-treated with heat-killed Yersinia followed by Legionella infection (Figure 7D).
As expected, pro-IL-1β translation was robust in wild type macrophages that were pre-treated with poly(I∶C) or with HKY followed by Legionella infection (Figure 7C and 7E). Interestingly, pro-IL-1β was detected by Western blots in MyD88−/− macrophages after pre-stimulation with poly(I∶C) but the protein level was reduced when the pre-activated cells were challenged with L. pneumophila (Figure 7C). MyD88−/− macrophages that were infected with Legionella alone on the other hand, showed little detectable translation (Figure 7C) despite having higher levels of Il1β transcripts (Figure 7B, right graph, compare black and grey bars). This phenotype was more obvious during pre-treatment with HKY. 2 hr HKY pre-stimulation led to a robust pro-IL-1β translation initially, which was significantly reduced when cells were challenged with L. pneumophila (Figure 7E). Therefore, in the absence of MyD88 signaling, macrophages were unable to overcome the translation inhibition induced by Legionella pneumophila even in the presence of external stimuli.
Inhibition of protein translation is a common virulence mechanism used by many viruses and bacteria. In this study, we showed that host cells have evolved mechanisms to cope with translation inhibition by selectively translating a subset of cytokine genes, including pro-inflammatory cytokines such as IL-1α and IL-1β in response to L. pneumophila challenge. The ability to bypass L. pneumophila translational inhibition is an important determinant of host protection, as mice defective in the IL-1α/IL-1β response and humans exposed to TNF-α inhibitors are highly susceptible to L. pneumophila infection [31], [38].
L. pneumophila challenge of bone-marrow macrophages leads to a dramatic reduction in global protein translation (Figure 5A). The bacterium interferes with protein translation both at the initiation step [36] and elongation step [10]. It has been shown previously that this inhibition triggers the transcription of various stress response genes including NF-kB- and MAPK-regulated genes, heat shock proteins and pro-inflammatory cytokines and chemokines [9], [10], [11]. We show here that L. pneumophila translocated effectors prevent the translation of these genes, resulting in a “frustrated response,” in which there is accumulation of transcripts but no increase in protein levels. A subset of cytokine genes, and potentially other genes that have not been identified yet, are insensitive to this inhibition and get translated in cells that show the highest level of protein synthesis inhibition.
We observed bypass of translational inhibition as an orderly series of events resulting in the accumulation of IL-1α in cells harboring bacteria as well as in bystanders, followed by release of the cytokine into culture supernatants. Initial transcriptional induction and translation of IL-1a occurred independently of the Icm/Dot system, and was associated with TLR-signaling, consistent with the TLR-dependent activation of the NF-κB response known to occur at early time points after L. pneumophila challenge [41]. This was followed by persistent accumulation of pro-IL-1α protein in a process that required both the presence of the Icm/Dot system and MyD88-dependent signaling, indicating collaborative signaling between the two pathways. Surprisingly, accumulation of pro-IL-1α was equally robust in both cells harboring bacteria and in bystander cells, in spite of the translocated protein synthesis inhibitors that are deposited by L. pneumophila. This observation is particularly striking, in that time points showing the strongest inhibition of protein synthesis also resulted in the fastest rate of pro-IL-1α accumulation, arguing that there is selective ribosomal loading of cytokine transcripts in intoxicated cells. In the absence of MyD88 signaling, no such bypass could be observed in either infected or uninfected cells, either because translation requires the extremely high levels of transcription that occur in the presence of MyD88, or there is pattern recognition-dependent bypass of translational inhibition. Accumulation of pro-IL-1α was then followed by its release, which required both the Icm/Dot system and MyD88.
It is conceivable that the L. pneumophila translational inhibitors could be responsible for the induction of pro-IL-1α. Arguing against this model is the fact that a strain that lacks 5 of the known translation inhibitors (Δ5) still induced considerable pro-IL-α accumulation (Figure 6E, F) [31]. The accumulation of cytokine in response to this strain could have resulted from residual translation inhibition that was observed, but it should be noted that MAPK and NF-κB activation resulting from macrophage challenge by this strain is due to a pattern recognition response and is not due to Icm/Dot signals [11]. The initial MyD88-dependent activation that occurs after contact with L. pneumophila may be amplified by unknown Icm/Dot-signals or due to inhibition of translation initiation [36]. A recent study proposed that infection of macrophages with virulent L. pneumophila strains (both Dot+ and Δ5) leads to downregulation of mTOR activity, which is sufficient to suppress cap-dependent protein translation initiation [36]. The second signal that is required for amplifying pattern-recognition could be generated from such translational suppression, and could be the trigger for induction of pro-IL-1α.
Challenge of macrophages that lack MyD88 with L. pneumophila induces cytokine gene transcription, but the transcribed genes fail to be translated. In the absence of MyD88, therefore, the protein translation inhibition takes on global dimensions. The MyD88-dependent bypass of the translation inhibition was independent of transcript stability in the case of Il1β transcript, which is a known strategy for post-transcriptional regulation of cytokines [52], [53]. This surprising result indicates that there may be a previously unrecognized MyD88-dependent signaling pathway that mediates post-transcriptional regulation of cytokine transcripts. It has previously been reported that separate transcriptional and translational signals are required for IL-1β expression [54]. Although it is not clear what these translational signals could be, it is possible that MyD88-dependent signals could result in either enhanced ribosome loading, or could regulate translation via action at the 3′ or 5′ untranslated regions. Alternatively, the role of MyD88 could be totally passive, and merely a consequence of enhancing expression of cytokine gene transcripts. Although L. pneumophila infection causes a large induction of cytokine transcription in the absence of MyD88, these levels are still lower than what is seen when pattern recognition is intact (Figure 2A). This added boost in cytokine gene transcription by MyD88 may be sufficient to push the concentrations of these transcripts above the minimum threshold necessary to support selective translation of these genes under conditions of intoxication. Consistent with this model are previous results from nanostring analysis of macrophage transcripts that are induced in response to L. pneumophila challenge [36]. In this work, it is argued that the primary determinant of translation in cells challenged with L. pneumophila is the relative abundance of a particular transcript. Translation was most likely to occur from transcripts that were the most abundantly expressed after bacterial challenge [36].
A similar phenomenon to that reported here has been observed in the model organism C. elegans upon infection with Pseudomonas aeruginosa. C. elegans intestinal cells endocytose P. aeruginosa Exotoxin A, which shuts down protein translation by inhibiting elongation factor 2 (EF2) [8]. This inhibition leads to the selective translation of ZIP-2, which is required for activation of defense pathways and pathogen clearance [7]. It was proposed that the 5′ UTR of zip-2, which contains several untranslated ORFs (uORFs), was required for the selective translation. Even so, there is no explanation for how ribosome loading and translation can selectively occur in this transcript. Interestingly, there are a few other examples from mammalian cells and yeast, where protein translation inhibition leads to selective translation of few genes that have uORFs at their 5′ UTR. The mammalian stress response transcription factor ATF4 and the yeast transcription factor GCN4 respond similarly to amino acid starvation and protein synthesis inhibition [55], [56]. The 5′ UTR or 3′ UTR of cytokines could potentially be functioning the same way to allow selective protein translation when initiation and/or elongation is blocked by Legionella pneumophila.
Interestingly, pharmacological inhibitors of host protein translation induce transcription of various stress response genes, including pro-inflammatory cytokines such as Il6, Il23, Il1α and Il1β [10], [11], [31]. Secretion of these genes can take place when the highly conserved host elongation machinery is targeted by toxins such as P. aeruginosa Exotoxin A or Corynebacterium diphtheriae encoded diptheria toxins [10], [31]. Diphtheria toxin and Pseudomonas ExoA inhibit eukaryotic elongation factor similar to the mechanism used by L. pneumophila effectors. They modify elongation factor 2 (EF2) of eukaryotic cells by ADP-ribosylation, which has been shown to trigger a strong host immune response [31], [57]. This suggests the presence of a conserved surveillance mechanism the host uses to detect and respond to inhibition of the translation elongation machinery.
A previous study has shown that the cytokine induction seen during L. pneumophila can be mimicked by the addition of P. aeruginosa Exo A in combination with the synthetic PAMP ligand Pam3Csk4 [31]. It seems likely that macrophages are able to selectively bypass the translation block of this toxin in a fashion that is similar to that described in C. elegans. Interestingly, we cannot reproduce this result using a variety of concentrations of the protein synthesis inhibitor cycloheximide, which instead reduces pro-inflammatory cytokine production in response to PAMP challenge. The mechanism by which cycloheximide inhibits protein synthesis is sufficiently different from these toxins to explain why we see differences in the innate immune response against CHX. CHX binds to the E-site of the 60S ribosomal subunit and freezes all translating ribosomes [46].The RNA/ribosome complex remains stabilized and does not dissociate, a phenomenon that may not be perceived as danger by eukaryotic cells. In contrast, both L. pneumophila and ExoA interfere with elongation factor function. There may be a set of modified elongation factors in the host cell that resist action of these effectors, or there may be a population that is sequestered from modification, allowing them to selectively act on cytokine transcripts. In either case, there must be some special property to the cytokine transcript that allows this selective utilization of these active elongation factors. Future work will focus on the nature of these transcripts that allows bypass of translation inhibition.
This study was carried out in accordance with the recommendation in the Guide for Care and Use of Laboratory Animals of the National Institutes of Health. The Institutional Animal Care and Use Committee of Tufts University approved all animal procedures. Our approved protocol number is B2013-18. The animal work, which is limited to the isolation of macrophages, does not involve any procedures of infections of live animals.
L. pneumophila strains Lp02 (referred to as WT) and Lp03 (referred to as dotA3) are streptomycin-resistant restriction-defective thymidine auxotrophs derived from L. pneumophila Philadelphia-1 (Lp01) (Table 1; [58]). The Δ5 and Δ5.ΔflaA strains were kindly provided by Zhao-Qing Luo (Purdue University) [10], [11], [43]. ΔflaA-GFP+, dotAΔflaA-GFP+, Δ5.ΔflaA-GFP+ carry GFP on an isopropyl-β-D-thiogalactopyranoside (IPTG)–inducible, Cm resistant plasmid (Table 1; [17], [59]). Solid medium containing buffered charcoal yeast extract (BCYE) and ACES-buffered yeast extract (AYE) broth culture medium supplemented with 100 µg/mL thymidine were used to maintain L. pneumophila strains [60], [61]. Strains containing the pGFP plasmid were maintained on BCYE plates containing 100 µg/mL thymidine and 5 µg/mL chloramphenicol and grown in AYE containing 100 µg/mL thymidine, 5 µg/mL chloramphenicol and 1 mM IPTG [17].
Bone marrow-derived macrophages (BMDMs) were isolated from the femurs of mice and allowed to proliferate as described [60], [61]. C57BL/6 myd88−/− femurs were kindly provided by Tanja Petnicki-Ocwieja in the laboratory of Linden Hu (Tufts Medical Center). BMDMs were differentiated for 7 days in RPMI containing 30% L-cell supernatant, 10% FBS, 2 mM L-glutamine and 1× Pen Strep (100 U/mL penicillin, 100 µg/mL streptomycin). Cells were lifted and either re-plated for experiments or quick-frozen for later use in FBS and 10% DMSO.
U937 cells (ATCC) were grown in RPMI supplemented with 10% FBS and 1 mM L- glutamine. For differentiation, cells were treated with 10 ng/ml 12-tetradecanoyl phorbol 13-acetate (TPA) for 48 hrs. For L. pneumophila infections, U937 cells were plated in fresh media without TPA and infections were carried out 12–16 hours after plating.
To evaluate protein expression in host cells, C57BL/6 bone marrow-derived macrophages were plated in medium supplemented with 200 µg/mL of thymidine. Cells were challenged with L. pneumophila at the desired MOI, subjected to centrifugation at 1000×g for 5 min and incubated at 37° for the noted time periods. Lysates were collected using 2× SDS Laemmli sample buffer (0.125 M Tric-Cl pH 6.8, 4% SDS, 20% glycerol, 10% beta-mercaptoethanol, 0.01% bromophenol blue). Proteins were electroblotted to PVDF membranes, blocked in milk and analyzed by immunoprobing.
For phospho-specific antibodies, blots were washed of all milk and incubated overnight with 1∶1000 phospho-p38 or phospho-JNK (Cell Signaling) in 5% BSA in phosphate buffered saline (PBS). Rabbit anti-DUSP1 (MKP1 V-15, Santa Cruz) and mouse anti-tubulin (sigma) antibodies were diluted to 1∶200 and 1∶7,500, respectively, in 5% milk in TBST. Goat anti-IL-1α and Goat anti-IL-1β antibodies (R&D systems, AF-400-NA & AF-401-NA) were diluted to 1∶500 in 5% milk in TBST.
For preparation of Heat-Killed Yersinia, wild type Y. pseudotuberculosis strains were grown overnight at 26° in Luria-Bertani (LB) broth. Overnight cultures were heat killed at 60° for 30–60 min and aliquots were frozen at −80° until use.
Macrophages were stimulated with HKY (MOI = 50), LPS (Sigma, 0.1 µg/mL or 1 µg/mL) and Pam3CSK4 (Invitrogen, 2 µg/mL), poly(I∶C) (InvivoGen, 50 µg/mL) for the desired time points. Cells were washed 3× with PBS and lysed in 2× SDS Laemmli sample buffer.
RNA was extracted from mammalian cells using RNAeasy kit (Qiagen). To determine the amount of a particular transcript, a one step, RNA-to-Ct kit (Applied Biosystems) was used according to manufacturer's instructions. Primers used for transcript analysis were as follows: human Dusp1 (5′ TTTGAGGGTCACTACCAG and 3′ CCGCTTCGTAGTAGAG), mouse Dusp1 (5′GGATATGAAGCGTTTTCGGCT and 3′ ACGGACTGTCACGTCTTAGG), mouse Il1α (5′GCACCTTACACCTACCAGAGT and 3′ TGCAGGTCATTTAACCAAGTGG), mouse Il1β (5′ GCAACTGTTCCTGAACTCAACT and 3′ ATCTTTTGGGGTCCGTCAACT), mouse Tnf-α (5′ GCACCACCATCAAGGACTCAA and 3′ GCTTAAGTGACCTCGGAGCT), mouse 18S ribosomal RNA (5′ CGCCGCTAGAGGTGAAATTCT and 3′ GCTTTCGTAAACGGTTCTTCA).
Macrophages were plated in 24 well tissue culture plates (2.5×105 per well) and challenged with L. pneumophila for either 6 hrs or 24 hrs. Supernatants were collected from each sample and 50 µL was used for ELISA. Mouse IL-1α and IL-1β Platinum ELISA (eBiosciences) was used to measure cytokine levels according to the manufacturers manual.
Intracellular cytokine staining was performed as described before [62] with modifications. Differentiated macrophages (∼1×107cells/plate) were washed and challenged with L. pneumophila GFP+ strains in RPMI containing 200 µg/mL thymidine, 5 µg/mL Chloramphenicol and 1 mM IPTG. For measuring TNF production, infections were carried out at MOI-3 and MOI-10 for 9 hrs followed by Golgiplug incubation (BD Bioscience, 1 µL/mL) for additional 5 hrs. For IL-1α and IL-1β, infections were carried out for 6 hrs. Cells were harvested with 10 mL cold PBS, washed twice with FACS buffer (PBS+0.5%BSA+0.05%NaN), incubated with Fc Block (clone 2.4G2) for 20 minutes and fixed with 2% paraformaldehyde overnight at 4°. Macrophage were permeabilized with Perm/Wash buffer (BD Bioscience) on ice for 20 minutes and stained with Alexa Fluor 647-conjugated anti-mouse TNF (BioLegend, clone ALF 161), phycoerythrin(PE)-conjugated anti-mouse IL-1α (BioLegend, clone MP6-XT22), PE-conjugated anti-mouse IL-1β (eBioscience, clone NJTEN3), Rabbit anti-DUSP1 (MKP1 V-15, Santa Cruz) and Goat anti-Rabbit Cy5 (Invitrogen) for 40 minutes. Stained cells were analyzed by BD LSR II flow cytometer.
To detect whole cell translation during defined timepoints, C57BL/6 bone marrow-derived macrophages were challenged with L. pneumophila-GFP+ at MOI-10 for 2 hrs, followed by labeling cells at various timepoints with 50 µM of L-azidohomoalanine (AHA) (Invitrogen) added to the culture medium. Cells were incubated for 1 hr (time course experiments) or for 4 hrs (IL-1α co-staining experiments) to allow incorporation of AHA into growing polypeptide chains. At the end of each incubation period, cells were washed with PBS, fixed with 4% paraformaldehyde for 15–20 min and left overnight in PBS. Incorporation of AHA was monitored by a Biotin- or APC-conjugated phosphine reagent (Pierce) and nuclei were stained with Hoechst 33342 (Molecular Probes). For flow cytometry, fixed cells were blocked with 1× BSA/PBS for 30 min at RT and 100 µM of APC-phosphine was added. Cells were incubated at 37° for 2–3 hrs and excess dye was removed by washing with 0.5% Tween-20/PBS. For IL-1α co-staining, PE conjugated anti-mouse IL-1α (BioLegend, clone MP6-XT22) was added to cells on ice for 30 min. Cells were analyzed by BD LSR II flow cytometer or by BD FACScalibur.
SunSET assay was used to determine the kinetics of protein translation over time [49] with modifications. Macrophages were plated in 6 well plates and infected with ΔflaA for the desired time points. 10 µg/mL of puromycin (Sigma) was added to cells for either 15 min or 1 hr. Macrophages were washed and lysed with IP Wash/Lysis buffer (Pierce) in the presence of protease inhibitors (Roche). Lysates were incubated on ice for 20 min, centrifuged at 13,000 rpm for 5 min and the supernatants were used for ELISA. The protein concentration in each sample was determined by Bradford assay.
ELISA plates were prepared by coating 96 well Nunc MaxiSorp plates with the desired antibody. Polyclonal Goat anti-IL-1α and IL-1β (R&D systems), Rabbit anti-DUSP-1 (Santa Cruz) and Rabbit anti-RhoGDI (Santa Cruz) antibodies were diluted in Carbonate/Bicarbonate buffer (PH = 9.6) to 10 µg/mL and 100 µL was used per well. Plates were incubated overnight at 4°. The following day, plates were brought to room temperature and blocked with 0.5% BSA/PBS for 1 hr. Cell lysates that have incorporated puromycin were incubated on the ELISA plates for 2 hrs, washed with 0.05% Tween-20/PBS three times and incubated with monoclonal mouse-anti-Puromycin (12D10, Millipore) for 1 hr. Unbound antibody was washed exhaustively with 0.05% Tween-20/PBS and plates were incubated with Donkey anti-mouse-HRP secondary antibody. Unbound antibody was washed again with Tween-20/PBS four times and 100 µL of HRP substrate (TMB solution) was added to each well for 5–10 mins. The reaction was stopped by adding 100 µL of stop-solution (2N H2SO4), and absorbance was measured at 450 nm.
|
10.1371/journal.pntd.0001584 | Multiple Mitochondrial Introgression Events and Heteroplasmy in Trypanosoma cruzi Revealed by Maxicircle MLST and Next Generation Sequencing | Mitochondrial DNA is a valuable taxonomic marker due to its relatively fast rate of evolution. In Trypanosoma cruzi, the causative agent of Chagas disease, the mitochondrial genome has a unique structural organization consisting of 20–50 maxicircles (∼20 kb) and thousands of minicircles (0.5–10 kb). T. cruzi is an early diverging protist displaying remarkable genetic heterogeneity and is recognized as a complex of six discrete typing units (DTUs). The majority of infected humans are asymptomatic for life while 30–35% develop potentially fatal cardiac and/or digestive syndromes. However, the relationship between specific clinical outcomes and T. cruzi genotype remains elusive. The availability of whole genome sequences has driven advances in high resolution genotyping techniques and re-invigorated interest in exploring the diversity present within the various DTUs.
To describe intra-DTU diversity, we developed a highly resolutive maxicircle multilocus sequence typing (mtMLST) scheme based on ten gene fragments. A panel of 32 TcI isolates was genotyped using the mtMLST scheme, GPI, mini-exon and 25 microsatellite loci. Comparison of nuclear and mitochondrial data revealed clearly incongruent phylogenetic histories among different geographical populations as well as major DTUs. In parallel, we exploited read depth data, generated by Illumina sequencing of the maxicircle genome from the TcI reference strain Sylvio X10/1, to provide the first evidence of mitochondrial heteroplasmy (heterogeneous mitochondrial genomes in an individual cell) in T. cruzi.
mtMLST provides a powerful approach to genotyping at the sub-DTU level. This strategy will facilitate attempts to resolve phenotypic variation in T. cruzi and to address epidemiologically important hypotheses in conjunction with intensive spatio-temporal sampling. The observations of both general and specific incidences of nuclear-mitochondrial phylogenetic incongruence indicate that genetic recombination is geographically widespread and continues to influence the natural population structure of TcI, a conclusion which challenges the traditional paradigm of clonality in T. cruzi.
| Chagas disease, caused by the protozoan parasite Trypanosoma cruzi, is an important public health problem in Latin America. While molecular techniques can differentiate the major T. cruzi genetic lineages, few have sufficient resolution to describe diversity among closely related strains. The online availability of three mitochondrial genomes allowed us to design a multilocus sequence typing (mtMLST) scheme to exploit these rapidly evolving markers. We compared mtMLST with current nuclear typing tools using isolates belonging to the oldest and most widely occurring lineage TcI. T. cruzi is generally believed to reproduce clonally. However, in this study, distinct branching patterns between mitochondrial and nuclear phylogenetic trees revealed multiple incidences of genetic exchange within different geographical populations and major lineages. We also examined Illumina sequencing data from the TcI genome strain which revealed multiple different mitochondrial genomes within an individual parasite (heteroplasmy) that were, however, not sufficiently divergent to represent a major source of typing error. We strongly recommend this combined nuclear and mitochondrial genotyping methodology to reveal cryptic diversity and genetic exchange in T. cruzi. The level of resolution that this mtMLST provides should greatly assist attempts to elucidate the complex interactions between parasite genotype, clinical outcome and disease distribution.
| Mitochondrial genes are among the most popular markers for the reconstruction of evolutionary ancestries and resolution of phylogeographic relationships [1]. Their pervasive use in population genetics can be attributed to several intrinsic characteristics, notably, their high copy number, small size (∼15–20 kb) and faster mutation rate (compared with nuclear DNA). In addition, their widespread application is founded on the assumptions that mitochondrial genomes are homoplasmic, uniparentally inherited and lack homologous recombination [2]. However, with technological advances affording increased sensitivity and greater sample throughput, a growing number of reports of heteroplasmy (heterogeneous mitochondrial genomes in an individual cell), introgression and inter-molecular recombination are challenging what was previously regarded as a strict set of rules for eukaryotic mitochondrial inheritance.
Chagas disease remains the most important parasitic infection in Latin America, where an estimated 10–12 million individuals are infected, with a further 80 million at risk [3]. The aetiological agent, Trypanosoma cruzi, displays remarkable genetic diversity and is currently recognized as a complex of six lineages or discrete typing units (DTUs), each broadly associated with disparate ecologies and geographical distributions [4]. T. cruzi infection is life-long and can lead to debilitation and death by irreversible cardiac and/or gastrointestinal complications [5]. It has been suggested that the geographical heterogeneity in Chagas disease pathology is related to the genetic variation among T. cruzi DTUs [6], [7]. However, the relationship between parasite genotype and clinical outcome remains enigmatic. DTU nomenclature has recently been revised by international consensus to reflect the current understanding of T. cruzi genetic diversity [8]. Several evolutionary scenarios have been proposed to account for the emergence of two hybrid lineages (TcV and TcVI) and their parental progenitors (TcII and TcIII). However, the number of ancestral nuclear clades (two or three) remains controversial [9], [10].
TcI is the most abundant and widely dispersed of all T. cruzi lineages, with an ancient parental origin estimated at ∼0.5–0.9 MYA [11]. The distribution of domestic TcI, propagated by domiciliated triatomine vector species, principally extends from the Amazon Basin northwards, where it is implicated as the main cause of Chagas disease in endemic areas such as Venezuela and Colombia [12], [13]. TcI is also ubiquitous in sylvatic transmission cycles throughout South America and extends into North and Central America [14], [15]. Recent advances in new high resolution genotyping techniques have seen a resurgence of interest in unravelling TcI intra-lineage diversity. In Colombia, sequencing of the mini-exon spliced leader intergenic region (SL-IR) has subdivided TcI isolates from domestic and sylvatic transmission cycles, irrespective of geographical origin [16]–[18]. Other studies have demonstrated geographical clustering of TcI strains and an ecological association between specific genotypes and Didelphis hosts [19]. Higher resolution studies exploiting multiple microsatellite markers (MLMT) also report limited gene flow between sylvatic and domestic transmission cycles manifesting as genetic diversity between TcI isolates from sympatric sites [20], [21]. In addition, unexpectedly high levels of homozygosity in multiple clones from single hosts may be indicative of recombination between similar genotypes (inbreeding) or recurrent, genome wide, and dispersed gene conversion [20], [22]. The frequency and mechanism of natural intra-TcI genetic exchange are thus unknown, largely due to inappropriate or inadequate sampling. Evidence for such recombination is increasing and has already been documented among strains isolated from sylvatic Didelphis and Rhodnius in the Amazon Basin [23] and within a domestic/peridomestic TcI population in Ecuador [21]. Furthermore, the generation of intra-lineage TcI hybrids in vitro indicates that this ancestral lineage has an extant capacity for genetic exchange [24].
In kinetoplastids, the mitochondrial genome is represented by 20–50 maxicircles (20–40 kb) which, together with thousands of minicircles (0.5–10 kb), form a catenated network or kinetoplast (kDNA), comprising 20–25% of total cellular DNA [25]. Maxicircles are the functional equivalent of eukaryotic mitochondrial DNA, encoding genes for mitochondrial rRNAs and hydrophobic proteins involved in energy transduction by oxidative phosphorylation [26]. Previously, phylogenetic analyses of T. cruzi maxicircle fragments classified isolates into three mitochondrial clades A (TcI), B (TcIII, TcIV, TcV and TcVI) and C (TcII) [10], [27]. To date, maxicircle typing has been principally used to examine T. cruzi inter-lineage diversity, with sequencing efforts reliant on a limited number of genes [28] and often in the absence of any comparative nuclear targets [29], [30]. However, the inherent features of mitochondrial markers argue for their inclusion as principal but not solitary components of phylogenetic studies. Indeed, the caveats highlighted by other eukaryotes are especially pertinent with respect to T. cruzi. Mitochondrial introgression has been reported in North America where identical maxicircles circulate in sympatric TcI and TcIV from sylvatic reservoirs [27] and in South America where maxicircle haplotypes are shared between TcIII and TcIV strains with highly divergent nuclear genomes [11]. However, this phenomenon has not been described among South American TcI isolates. In addition, mitochondrial heteroplasmy, a possible confounder of phylogenetic studies, has not been examined in the coding region of the T. cruzi maxicircle but is not unexpected considering the presence of up to fifty maxicircle copies within an individual parasite.
The potential for mitochondrial DNA to reveal diversity hidden at the sub-DTU level in T. cruzi has been largely overlooked. To address this deficit, we first employed a whole genome approach to investigate the existence of maxicircle heteroplasmy and to resolve its role as a source of genotyping error. Secondly, we exploited the online availability of three complete T. cruzi maxicircle genomes [31], [32] to develop a high resolution mitochondrial multilocus typing scheme (mtMLST) in order to describe TcI intra-lineage diversity. Lastly, we investigated the extent of incongruence between mitochondrial and nuclear loci (SL-IR, GPI and 25 short tandem repeat (STR) loci) to detect incidences of genetic exchange.
The maxicircle genome from Sylvio X10/1 (TcI) was sequenced at 183X coverage using Illumina HiSeq 2000 technology as part of the Sylvio X10/1 Whole Genome Shotgun project [33]. A total of 66,882 reads were generated which covered the maxicircle coding region (15,185 bp). The consensus maxicircle genome sequence was derived from the predominant nucleotide present across multiple read alignments at each position. However, this criterion masks minor maxicircle haplotypes (evidence of heteroplasmy) by disregarding low abundance single nucleotide polymorphisms (SNPs). To assess the presence/absence of true minor SNPs, all 66,882 reads were re-aligned to the Sylvio X10/1 maxicircle genome using the alignment software SAMtools [34] and SNPs were called using the SAMtools mpileup commands. A SNP was defined as a nucleotide variant present in at least 5 independent reads (with parameters: 20X coverage; and mapping quality, 30). The final alignment was manually inspected using Tablet [35]. In parallel, ten maxicircle gene fragments, described below, were amplified by PCR and Sanger sequenced from Sylvio X10/1.
A panel of 32 TcI isolates was assembled for analysis (Table 1). Parasites (epimastigotes) were cultured at 28°C in RPMI-1640 liquid medium supplemented with 0.5% (w/v) tryptone, 20 mM HEPES buffer pH 7.2, 30 mM haemin, 10% (v/v) heat-inactivated fetal calf serum, 2 mM sodium glutamate, 2 mM sodium pyruvate and 25 µg/ml gentamycin (Sigma, UK) [23]. Genomic DNA was extracted using the Gentra PureGene Tissue Kit (Qiagen, UK), according to the manufacturer's protocol. Isolates were previously characterized to DTU level using a triple-marker assay [36] and classified into seven genetic populations by microsatellite profiling [20]: North and Central America (AMNorth/Cen), Venezuelan sylvatic (VENsilv), North-Eastern Brazil (BRAZNorth-East), Northern Bolivia (BOLNorth), Northern Argentina (ARGNorth), Bolivian and Chilean Andes (ANDESBol/Chile) and Venezuelan domestic (VENdom). Genotypes for additional TcI–TcVI strains were included for comparison in selected analyses as indicated (Tables S1 and S2).
Ten maxicircle gene fragments were amplified: ND4 (NADH dehydrogenase subunit 4), ND1 (NADH dehydrogenase subunit 1), COII (cytochrome c oxidase subunit II), MURF1 (Maxicircle unidentified reading frame 1, two fragments), CYT b (cytochrome b), 12S rRNA, 9S rRNA, and ND5 (NADH dehydrogenase subunit 5, two fragments) coding regions. Degenerate primers were designed in primaclade [37] using complete maxicircle reference sequences from CL Brener (TcVI), Sylvio X10/1 (TcI), and Esm cl3 (TcII) available online at www.tritrypdb.org [38]. Primer sequences and annealing temperatures for PCR amplifications are given in Table 2. Robust amplification was first confirmed across a reference panel of all six T. cruzi DTUs (see Table S1 and Figure 1).
Amplifications for all targets were achieved in a final volume of 20 µl containing: 1× NH4 reaction buffer, 1.5 mM MgCl2 (Bioline, UK), 0.2 mM dNTPs (New England Biolabs, UK), 10 pmol of each primer, 1 U Taq polymerase (Bioline, UK) and 10–100 ng of genomic DNA. PCR reactions were performed with an initial denaturation step of 3 minutes at 94°C, followed by 30 amplification cycles (94°C for 30 seconds, 50°C for 30 seconds, 72°C for 30 seconds) and a final elongation step at 72°C for ten minutes. PCR products were purified using QIAquick PCR extraction kits (Qiagen, UK) according to the manufacturer's protocol.
The mini-exon spliced leader intergenic region (SL-IR) and glucose-6-phosphate isomerase (GPI) were amplified as previously described by Souto et al. (1996) [39] and Lewis et al. (2009) [36], respectively. PCR products were visualized in 1.5% agarose gels and if necessary purified using QIAquick PCR and gel extraction kits (Qiagen, UK) to remove non-specific products. Bi-directional sequencing was performed for both nuclear and maxicircle targets using the BigDye® Terminator v3.1 Cycle Sequencing Kit (Applied Biosystems, UK) according to the manufacturer's protocol. Maxicircle PCR products were sequenced using the relevant PCR primers described in Table 2. Nuclear amplicons were sequenced using their respective PCR primers. When ambiguous sequences were obtained, PCR products were cloned into the pGEM® - T Easy Vector System I (Promega, UK), according to the manufacturer's instructions, and transformed into XL1-Blue E. coli (Agilent Technologies, UK), prior to colony PCR and re-sequencing. For strains that produced incongruent nuclear and maxicircle phylogenetic signals, PCR and sequencing reactions were replicated twice using DNA derived from two independent genomic DNA extractions.
Data from 25 previously described microsatellite loci [20], distributed among ten chromosomes [40], were included for analysis. Loci were selected from a wider panel of 48 microsatellite loci based on their level of TcI intra-lineage resolution. In addition, these 25 microsatellite loci were amplified across eight new unpublished biological clones (M16 cl4, SJM22 cl1, SJM39 cl3, USAARMA cl3, USAOPOSSUM cl2, 92090802P cl1, 93070103P cl1 and DAVIS 9.90 cl1). Primers and binding sites are listed in Table S3. The following reaction conditions were implemented across all loci: a denaturation step of 4 minutes at 95°C, then 30 amplification cycles (95°C for 20 seconds, 57°C for 20 seconds, 72°C for 20 seconds) and a final elongation step at 72°C for 20 minutes. Amplifications were achieved in a final volume of 10 µl containing: 1× ThermoPol Reaction Buffer (New England Biolabs, UK), 4 mM MgCl2, 34 µM dNTPs, 0.75 pmol of each primer, 1 U Taq polymerase (New England Biolabs, UK) and 1 ng of genomic DNA. Five fluorescent dyes were used to label the forward primers: 6-FAM and TET (Proligo, Germany) and NED, PET and VIC (Applied Biosystems, UK). Allele sizes were determined using an automated capillary sequencer (AB3730, Applied Biosystems, UK), in conjunction with a fluorescently tagged size standard, and were manually checked for errors. All isolates were typed “blind” to control for user bias.
Pair-wise distances (DAS) between microsatellite genotypes for individual samples were calculated in MICROSAT v1.5d [41] under the infinite-alleles model (IAM). To accommodate multi-allelic genotypes (≥3 alleles per locus), a script was written in Microsoft Visual Basic to generate random multiple diploid re-samplings of each multilocus profile (software available on request). A final pair-wise distance matrix was derived from the mean of each re-sampled dataset and used to construct a Neighbour-Joining phylogenetic tree in PHYLIP v3.67 [42]. Majority rule consensus analysis of 10,000 bootstrap trees was performed in PHYLIP v3.67 by combining 100 bootstraps created in MICROSAT v1.5d, each drawn from 100 respective randomly re-sampled datasets.
Nucleotide sequences were assembled manually in BioEdit v7.0.9.0 sequence alignment editor software (Ibis Biosciences, USA) [43] and unambiguous consensus sequences were produced for each isolate. Heterozygous SNPs were identified by the presence of two coincident peaks at the same locus (‘split peaks’), verified in forward and reverse sequences and scored according to the one-letter nomenclature for nucleotides from the International Union of Pure and Applied Chemistry (IUPAC). For both nuclear genes (SL-IR and GPI), edited sequences were used to generate Neighbour-Joining trees based on the Kimura-2 parameter model in MEGA v5 [44]. Bootstrap support for clade topologies was estimated following the generation of 1000 pseudo-replicate datasets. Once both trees were visualized independently to confirm congruent topologies, nuclear SNPs were re-coded numerically and concatenated with microsatellite data (see Dataset S1). DAS values were calculated for the concatenated dataset as described above and used to generate a single Neighbour-Joining phylogenetic tree encompassing all nuclear genetic diversity. Nucleotide sequences for GPI and the SL-IR are available from GenBank under the accession numbers JQ581371–JQ581402 and JQ581481–JQ581512, respectively.
Sequence data were assembled manually as described for nuclear loci. For each isolate, maxicircle sequences were concatenated according to their structural arrangement (12S rRNA, 9S rRNA, CYT b, MURF1, ND1, COII, ND4 and ND5) and in the correct coding direction (alignment available on request). Nucleotide sequences for all ten gene fragments are available from GenBank under the accession numbers listed in Table 2. Phylogenies were inferred using Maximum-Likelihood (ML) implemented in PhyML (4 substitution rate categories) [45]. The best-fit model of nucleotide substitution was selected from 88 models and its significance evaluated according to the Akaike Information Criterion (AIC) in jMODELTEST 1.0. [46]. The best model selected for this dataset was GTR+I+G. Bootstrap support for clade topologies was estimated following the generation of 1000 pseudo-replicate datasets. Bayesian phylogenetic analysis was performed using MrBAYES v3.1 [47] (settings according to jMODELTEST 1.0). Five independent analyses were run using a random starting tree with three heated chains and one cold chain over 10 million generations with sampling every 10 simulations (25% burn-in). Shimodaira-Hasegawa likelihood tests (SH tests) [48] were implemented in PAML v.4 [49] to statistically evaluate incongruencies between alternative tree topologies derived from the mitochondrial and nuclear data.
Across the 15,185 bp of the Sylvio X10/1 maxicircle coding region a total of 74 SNPs were identified among eight genes (12S rRNA, 9S rRNA, MURF5, CYT b, MURF1, MURF2, CR4 and ND4) and three intergenic regions (between 12S rRNA and 9S rRNA, between 9S rRNA and ND8 and between CR4 and ND4, respectively) (Figure 2 and Table S4). Average read depth for each SNP site was 163. At heterozygous sites, the minor nucleotide was present among an average of 12.2% (±9.1%) of sequence reads. In each gene, SNPs were clustered often <5 bp apart in pairs and triplets. The most common mutations were transversions from A→T (14/74), T→A (10/74), T→G (7/74) and G→T (6/74) and transitions from A→G (13/74). SNPs were bi-variable at all sites. The presence of different contiguous SNPs distributed across separate sequencing reads at overlapping positions suggests the occurrence of at least two minor maxicircle templates within the same sample. However, the short average length of Illumina reads (∼100 bp) prohibits the full reconstruction of minor maxicircle sequence types. No evidence of heterozygosity was observed in any of the ten maxicircle Sanger sequences (from the mtMLST scheme) that covered the corresponding areas of heteroplasmy identified in Sylvio X10/1, which is consistent with the low sensitivity of this method.
Degenerate primers were designed by reference to complete TcI, TcII and TcVI maxicircle genomes. Ten gene fragments from eight maxicircle coding regions were selected in order to sample genetic diversity present across the whole T. cruzi maxicircle. For two genes (MURF1 and ND5) two fragments were selected from each coding region to examine intra-gene variation. Reliable PCR amplification of all ten maxicircle fragments was first confirmed using a panel of T. cruzi reference strains from each DTU (see Figure 1).
The maxicircle gene targets were then sequenced across the TcI panel (Table 1) and seven additional TcIII/TcIV strains (Table S2). Relatively uniform substitution rates were observed among all genes (gamma shape parameter α = 0.8121, based on the GTR+I+G model). For each TcI isolate, gene fragments were concatenated according to their structural position and assembled into a 3686 bp alignment. Twenty-two unique haplotypes were identified from a total of 355 variable sites (∼9.6% sequence diversity). No evidence of heterozygosity (‘split peaks’) was observed.
Maximum-Likelihood (Figure 3, right) and Bayesian phylogenies were both constructed from the concatenated maxicircle data. No statistically-supported incongruence was observed between the two topologies (Bayesian tree L = −6770.21, ML tree L = −6768.85, P = 0.428). The presence of at least three incongruent haplotypes (see below) precludes the accurate clustering of their respective populations (AMNorth/Cen, VENdom and BRAZNorth-East). However, phylogenetic analysis does resolve two well-supported clades corresponding to VENsilv and ANDESBol/Chile (90.8%/1.0 and 100%/1.0, respectively). Once the two TcIV-type maxicircles were excluded from analysis, the mtMLST was re-evaluated with respect to intra-TcI discriminatory power. One hundred SNPs were identified among 3681 bp (∼2.7% sequence diversity), corresponding to twenty maxicircle haplotypes. Both Bayesian and Maximum-Likelihood topologies were congruent with those constructed previously for the entire TcI isolate panel.
The resolutive power of the mtMLST scheme was evaluated by comparison to current markers used to investigate TcI intra-DTU nuclear diversity, specifically, a housekeeping gene (GPI), a non-coding multi-copy intergenic region (SL-IR) and a MLMT panel of 25 loci. Sequences for GPI were obtained for 32 T. cruzi isolates (Table 1) and assembled into a gap-free alignment of 921 nucleotides. Of the 921 bp, a total of 911 invariable sites and 10 polymorphic sites were identified (∼1.1% sequence diversity). A 350 bp alignment corresponding to the SL-IR was generated for the same panel of samples. Strains from two populations (5/6 BOLNorth and 4/4 ANDESBol/Chile) presented sequences with multiple ambiguous base calls due to the presence of a GTn microsatellite at positions 14–24. For these nine isolates, haplotypes were determined by sequencing four cloned PCR products to derive a consensus sequence. In the 350 bp alignment, 323 conserved sites and 36 polymorphic sites were observed (∼10.3% sequence diversity). All samples were also typed at 25 polymorphic microsatellite loci yielding a total of 1612 alleles. The majority of strains presented one or two alleles at each locus. Multiple alleles (≥3) were observed at a small proportion of loci (1.5%).
Individual Neighbour-Joining trees were re-constructed for GPI, SL-IR and the MLMT data. No well-supported sub-DTU level clades were recovered using GPI sequences. The SL-IR phylogeny resolved two populations (VENsilv and ARGNorth) with strong statistical support (85% and 99%, respectively; data not shown). Three major clades were identified by MLMT (VENdom, ARGNorth and ANDESBol/Chile) with good bootstrap support (72.6%, 99.3% and 98.4%, respectively; data not shown). There was no bootstrap-supported incongruence between the three nuclear tree topologies. This justified their concatenation and these data were re-coded and analyzed in a single distance-based phylogeny (independent of mutation rate heterogeneity) (Figure 3, left and Dataset S1). The concatenated nuclear tree recovered three well supported clades corresponding to TcI populations (VENsilv, ARGNorth and ANDESBol/Chile) (96%, 100% and 77.9%, respectively, Figure 3). Isolates belonging to the VENdom population remained grouped together but with a minor reduction in bootstrap values (64.8%), compared to the MLMT tree. In addition, the concatenated tree also subdivided BOLNorth into two well defined sympatric clades each containing three isolates (99.8% and 82.2%). No nuclear targets (either individually or concatenated) were able to reliably identify AMNorth/Cen, or BRAZNorth-East as discrete clusters. However, AMNorth/Cen was more closely related to VENdom than any other population by MLMT (90.2%), the SL-IR (99%) and the concatenated nuclear tree (100%).
Comparison of the mitochondrial and nuclear phylogenies revealed clear incongruence at multiple scales. The nuclear topology was a significantly worse model to fit the maxicircle data (nuclear tree L = −7008.72, mtMLST ML tree L = −6554.50, P<0.001). Three individual isolates had unambiguously different phylogenetic positions between the nuclear and mitochondrial datasets: 9307, 9354 and IM48 (Figure 3). The maxicircle sequences from 9307, a sylvatic TcI AMNorth/Cen strain, and 9354, a human TcI strain from VENdom, were divergent from all other TcI strains. Comparison with sequences from other DTUs indicates that the maxicircle from 9307 was most closely related to those found in TcIV samples from North America (92122) (100%/1.0) while 9354 shared its mitochondrial haplotype with TcIV and TcIII strains from neighbouring areas of Venezuela, Bolivia and Colombia (ERA, 10R26, X106, Sairi3 and CM17) (97.8%/0.9). IM48 from BRAZNorth-East also had a distinct maxicircle haplotype that formed a long branch separated from the other members of this population whereas for nuclear data all BRAZNorth-East isolates, including IM48, clearly grouped together.
To test whether inclusion of these isolates could explain the overall incongruence, the SH analysis was repeated for alternative nuclear vs. mitochondrial topologies with each of these strains excluded individually and then collectively. In all cases, statistically significant incongruence persisted (no 9307 P = 0.004, no 9354 P = 0.002, no IM48 P<0.001 and without all three P = 0.008). This indicated that mitochondrial introgression was generally pervasive in the TcI panel beyond these three isolates. For example, ARGNorth samples, which formed a homogeneous monophyletic clade that was most closely related to ANDESBol/Chile by nuclear data, grouped paraphyletically amongst subsets of BOLNorth strains in the maxicircle tree. In addition, BRAZNorth-East is grouped with one of the BOLNorth clades in the nuclear tree, but receives a basally diverging position in the maxicircle phylogeny. In agreement with the nuclear data, AMNorth/Cen was most closely related to VENdom. However, two isolates from AMNorth/Cen (ARMA and OPOS) displayed an unexpected level of maxicircle diversity and are grouped separately with strong bootstrap support (96.6%/1.0).
Elucidating the complex epidemiology, phylogeography and taxonomy of T. cruzi requires a clear understanding of the parasite's genetic diversity [4]. One objective of this study was to develop the first mitochondrial (maxicircle) multilocus sequence typing scheme (mtMLST) to investigate T. cruzi intra-lineage diversity and to critically assess its resolutive power compared to the current repertoire of phylogenetic markers.
The presence of intra-strain maxicircle diversity within Sylvio X10/1 is the first demonstration of heteroplasmy in the coding region of a T. cruzi maxicircle genome. Seventy-four variable sites were identified by read depth analysis of Illumina sequence data but undetected by conventional Sanger sequencing. These SNPs indicate the occurrence of at least two additional maxicircle genomes, present at a ∼10-fold lower abundance compared to the consensus published Sylvio X10 maxicircle genome [32]. Most heteroplasmic SNPs were linked. This may indicate an older most recent common ancestor (MRCA) between the major and minor maxicircles than that expected to have emerged in culture post-cloning. Thus these minor maxicircle classes more likely represent heteroplasmy within a single parasite than within a subpopulation of cells. Furthermore, the presence of SNPs <3 bp apart on contiguous sequence reads may have non-synonymous coding implications, although their relative rarity, and a lack of indels suggest that minority and majority maxicircle variants would not differ phenotypically. Finally, the presence of heteroplasmy at less than 0.5% of sites indicates it is unlikely to represent a major source of typing error when using maxicircle Sanger sequencing to characterize isolates.
Several factors are likely to contribute to mitochondrial heteroplasmy. Mutation in length or nucleotide composition and/or bi-parental inheritance in genetic exchange events are both exacerbated by differential replication rates and inequitable cytoplasmic segregation of mitochondrial genomes during mitosis [50], [51]. In kinetoplastids, maxicircle intra-clone diversity in the non-coding region was previously reported in both T. cruzi [31] and Leishmania major [52], [53]. In addition, an earlier study attributed a change in T. cruzi maxicircle gene repertoire (elimination of one of two heteroplasmic ND7 amplicons) to sub-culture [54]. However, biologically cloned samples were not used and the possibility of a mixed infection was excluded on the basis of only four microsatellite loci. Sylvio X10/1 (a biological clone produced by micromanipulation) was first isolated from a Brazilian patient in 1979 [55] and has been in intermittent sub-culture ever since. The retention of minor maxicircle classes in Sylvio X10/1 for over thirty years suggests that a heteroplasmic state in T. cruzi is naturally sustained.
The observations that T. cruzi mitochondrial heteroplasmy is not present at sufficient levels to adversely disrupt phylogenetic reconstructions stimulated the development of the mtMLST scheme and its assessment against traditional nuclear targets. Initially, three types of nuclear marker were evaluated, each characterized by different rates of evolution. Unsurprisingly GPI was highly conserved across TcI and lacked sufficient resolution to discriminate between isolates. The slow accumulation of point mutations at housekeeping loci, which are generally under purifying selection, renders these targets more appropriate to describe inter-DTU variation. Thus they are valuable candidates for inclusion in traditional nuclear MLST schemes [56]. The mini-exon SL-IR is widely used as a TcI taxonomic marker in view of its heterogeneity and ease of amplification [57]. In this study, SL-IR variability manifested as a ten-fold increase in sequence diversity as compared to that of GPI, and supported the robust delineation of two nuclear populations (VENsilv and ARGNorth). However, there are several caveats associated with the SL-IR, notably the presence of multiple tandemly-repeated copies with undefined chromosomal orthology between strains [58]. Previous attempts to estimate the level of intra-isolate SL-IR diversity have reported >96% homology between copies [19]. However, only ten clones were sequenced from each sample, representing less than 10% of the ∼200 copies present per genome. Recent observations of substantial variation in gene copy number and chromosomal arrangement between T. cruzi strains further discourage the use of such targets for taxonomy [59]. In addition, numerous indels in the SL-IR prevent the sequencing of a suitable outgroup [39] and multiple ambiguous alignments, introduced by the microsatellite region, can disrupt phylogenetic signals [60]. Ultimately both GPI and the SL-IR suffer from the same fundamental criticism that single genes are inadequate to infer the overall phylogeny of an entire species [61]. Recombination, gene conversion and concerted evolution have all contributed to the genealogical history of T. cruzi [62] but remain undetectable using single loci.
The 25 microsatellite loci afforded the highest level of resolution from an individual set of markers, defining three statistically-supported groupings (VENdom, ARGNorth and ANDESBol/Chile). Their superior performance compared to GPI and the SL-IR is expected considering microsatellites are neutrally-evolving, co-dominant and hypervariable with mutation rates several orders of magnitude higher than protein-coding genes [63]. However, the use of these markers is not devoid of limitations. Most importantly, microsatellites are particularly sensitive to homoplasy, a situation where two alleles are identical in sequence but not descent, and thus fail to discriminate between closely related but evolutionarily distinct strains [64]. The three nuclear markers (GPI, SL-IR and microsatellites) were concatenated based on the assumption that no robust incongruence was observed between individual phylogenetic trees. However, concatenating these data did not have a significant additive effect on the level of resolution, with just three populations (VENsilv, ARGNorth and ANDESBol/Chile) emerging as well-supported groups. Importantly this dataset did reveal a subdivision in the BOLNorth group, which went undetected by all individual nuclear markers.
Gross incongruence between the mtMLST and nuclear phylogenies revealed two incidences of inter-DTU mitochondrial introgression, indicative of multiple genetic exchange events in T. cruzi. Introgression was detected in North America, where identical maxicircles were observed in sylvatic TcI and TcIV isolates. A 1.25 kb fragment (COII-ND1) of this TcIV maxicircle haplotype has been previously described in other TcI samples from the US states of Georgia and Florida [11], [27]. On the basis of the limited nuclear loci examined, and in line with previous work [27], only TcI derived nuclear genetic material appears to have been retained in these hybrids. The genetic disparity between North and South American TcIV isolates, coupled with their geographical and ecological isolation [65], implies that this event most likely occurred in North/Central America. A second, independent novel mitochondrial introgression event was identified in a Venezuelan clinical isolate. This TcI strain (9354) shares its maxicircle haplotype with a subset of human and sylvatic TcIV and TcIII isolates from Bolivia, Venezuela and Colombia, consistent with a local and possibly recent origin. Presumably TcIV, a known secondary agent of human Chagas disease in Venezuela, is a more likely donor candidate than TcIII, which is largely absent from domestic transmission cycles [4].
Nonetheless, evidence of homogeneous maxicircle sequences in multiple, geographically dispersed isolates from different transmission cycles implies the occurrence of several genetic exchange events. It is conceivable that the TcIV/TcIII-type maxicircle sampled in this study is a relic from a TcI antecedent, supporting a common ancestry between TcI, TcIII and TcIV [9]. Alternatively, this haplotype may have originated from a TcIV or TcIII strain and its distribution reflects a recent unidirectional backcrossing event into TcI. Introgression is a more parsimonious explanation than the retention of ancestral polymorphisms through incomplete lineage sorting, particularly in areas of sympatry or parapatry among DTUs [66]. However, the historical diversification of TcI [67] and TcIII [68]–[70], driven by disparate ecological niches [71], and the current separation between most arboreal and terrestrial transmission cycles of TcIV and TcIII, respectively, challenge the likelihood of secondary contact between these lineages, a prerequisite of introgressive hybridization. Resolving the donor DTU of this event is complicated by the presence of indistinguishable mitochondrial sequences and paradoxically divergent nuclear genes in TcIII and TcIV isolates. It is unclear whether this results from a mechanism acting to homogenize maxicircles while allowing nuclear genes to slowly deviate [11] (unlikely), repeated and recurrent backcrossing (more likely), or merely reflects the relative paucity of available TcIV and TcIII genotypes for comparison (a certainty).
Regardless of the underlying mechanisms, it is clear that genetic exchange continues to influence the natural population structure of T. cruzi TcI. In this study, the failure to detect reciprocal transfer of nuclear DNA using an array of loci readily demonstrates the importance of adopting an integrative approach, complementing traditional nuclear markers with multiple mitochondrial targets. In the absence of comparative genomics, it is impossible to establish whether mitochondrial introgression is entirely independent of nuclear recombination.
Another advantage of the mtMLST scheme is its ability to reveal cryptic sub-DTU diversity. The significantly different evolutionary histories of the nuclear and maxicircle genes from members of BOLNorth and ARGNorth are consistent with intra-lineage recombination. The low levels of diversity observed within this incongruent maxicircle clade are indicative of recent and possibly multiple exchange events. In addition, two divergent maxicircles from AMNorth/Cen have also exposed a level of diversity that conflicts with earlier reports of reduced genetic differentiation in this group resulting from their recent biogeographical expansion [18], [72]. Furthermore, the incongruent basal phylogenetic position of most of BRAZNorth-East in the maxicircle tree as well as the presence of another divergent maxicircle in one isolate (IM48) from this population highlights the extent to which intra-lineage diversity can be neglected by other genotyping methods. The phylogenetic placement of IM48 suggests it may be the product of an intra-TcI introgression event. However, IM48 is also a geographical outlier within the BRAZNorth-East population and it is difficult to determine the origin of this maxicircle haplotype in the absence of additional isolates from West-Central Amazonia.
The mechanisms governing maxicircle genetic exchange and the origins of heteroplasmy observed in Sylvio X10/1 are debatable. Currently, all reported maxicircle inheritance in natural [11] and experimental T. cruzi hybrids [24] is uniparental. However, the demonstration of heteroplasmy in this study suggests that, following genetic exchange, any minor maxicircle genotypes may be undetectable using conventional sequencing techniques. In addition, evidence of bi-parental transmission of both maxicircles [73], [74] and minicircles [75] in experimentally-derived T. brucei hybrids indicates that this phenomenon can occur in kinetoplastids as a result of recombination. The mechanism of genetic exchange in T. cruzi [24] differs from meiosis, which is observed in T. brucei [73], [76]. Current data suggest in vitro recombination in T. cruzi may be analogous to the parasexual cycle of Candida albicans where nuclear fusion creates a tetraploid intermediate, followed by genome erosion and reversion to aneuploidy [24], [77], [78]. It is not implausible to suggest that the process of cell fusion and nuclear re-assortment may be accompanied by asymmetrical kinetoplast distribution to progeny cells. Furthermore, the sequence redundancy observed among minicircle guide RNAs has been postulated to allow biparental inheritance to occur with no detrimental consequences to mitochondrial RNA editing and hybrid viability [79].
Most importantly, the phenotypic implications of mitochondrial heteroplasmy and introgression in T. cruzi are unknown. Maxicircles play a fundamental role in parasite metabolism and development in the triatomine bug vector. Therefore the relationship between genetic recombination and phenotypic heterogeneity may have important implications for disease epidemiology. mtMLST presents a valuable new strategy to detect directional gene flow and examine the dispersal history of T. cruzi at the transmission cycle level. Furthermore, mtMLST is an excellent tool to identify genetic exchange between closely related isolates in conjunction with nuclear MLMT data. By adopting a combined nuclear and mitochondrial approach, one can simultaneously address local, epidemiologically important hypotheses as well as robustly identify parasite mating systems. Thus in combination with adequate spatio-temporal sampling, we strongly recommend this methodology as an alternative to exclusively nuclear or mitochondrial population genetic studies in future work with medically important trypanosomes. Finally, the level of resolution that the mtMLST method provides should greatly facilitate attempts to elucidate the relationship between specific parasite genotypes and phenotypic traits relating to Chagas disease pathology.
|
10.1371/journal.ppat.1007842 | Separate domains of G3BP promote efficient clustering of alphavirus replication complexes and recruitment of the translation initiation machinery | G3BP-1 and -2 (hereafter referred to as G3BP) are multifunctional RNA-binding proteins involved in stress granule (SG) assembly. Viruses from diverse families target G3BP for recruitment to replication or transcription complexes in order to block SG assembly but also to acquire pro-viral effects via other unknown functions of G3BP. The Old World alphaviruses, including Semliki Forest virus (SFV) and chikungunya virus (CHIKV) recruit G3BP into viral replication complexes, via an interaction between FGDF motifs in the C-terminus of the viral non-structural protein 3 (nsP3) and the NTF2-like domain of G3BP. To study potential proviral roles of G3BP, we used human osteosarcoma (U2OS) cell lines lacking endogenous G3BP generated using CRISPR-Cas9 and reconstituted with a panel of G3BP1 mutants and truncation variants. While SFV replicated with varying efficiency in all cell lines, CHIKV could only replicate in cells expressing G3BP1 variants containing both the NTF2-like and the RGG domains. The ability of SFV to replicate in the absence of G3BP allowed us to study effects of different domains of the protein. We used immunoprecipitation to demonstrate that that both NTF2-like and RGG domains are necessary for the formation a complex between nsP3, G3BP1 and the 40S ribosomal subunit. Electron microscopy of SFV-infected cells revealed that formation of nsP3:G3BP1 complexes via the NTF2-like domain was necessary for clustering of cytopathic vacuoles (CPVs) and that the presence of the RGG domain was necessary for accumulation of electron dense material containing G3BP1 and nsP3 surrounding the CPV clusters. Clustered CPVs also exhibited localised high levels of translation of viral mRNAs as detected by ribopuromycylation staining. These data confirm that G3BP is a ribosomal binding protein and reveal that alphaviral nsP3 uses G3BP to concentrate viral replication complexes and to recruit the translation initiation machinery, promoting the efficient translation of viral mRNAs.
| In order to repel viral infections, cells activate stress responses. One such response involves inhibition of translation and restricted availability of the translation machinery via the formation of stress granules. However, the host translation machinery is absolutely essential for synthesis of viral proteins and consequently viruses have developed a broad spectrum of strategies to circumvent this restriction. Old World alphaviruses, such as Semliki Forest virus (SFV) and chikungunya virus (CHIKV), interfere with stress granule formation by sequestration of G3BP, a stress granule nucleating protein, mediated by the viral non-structural protein 3 (nsP3). Here we show that nsP3:G3BP complexes engage factors of the host translation machinery, which during the course of infection accumulate in the vicinity of viral replication complexes. Accordingly, we demonstrate that the nsP3:G3BP interaction is required for high localized translational activity around viral replication complexes. We find the RGG domain of G3BP to be essential for the recruitment of the host translation machinery. In cells expressing mutant G3BP lacking the RGG domain, SFV replication was attenuated, but detectable, while CHIKV was essentially non-viable. Our data demonstrate a novel mechanism by which viruses can recruit factors of the translation machinery in a G3BP-dependent manner.
| Viral infections are inevitably accompanied by a competitive crosstalk between the host and the pathogen, engaging a complex network of protein-protein interactions. Since exploitation of host resources is crucial for the viral replication cycle, host responses aim to interfere with such measures and to clear the threat. Consequently, viruses have evolved to target host proteins involved in cellular defence mechanisms. G3BP-1 and -2 (hereafter jointly referred to as G3BP) are homologous proteins with critical roles in the assembly of cellular stress granules (SGs), dynamic assemblies of stalled translation initiation complexes and RNAs [1, 2]. The proteins contain an N-terminal NTF2-like domain, which is involved in dimerization and interaction with other proteins, long stretches of intrinsically disordered protein sequence as well as RNA-recognition motifs (RRMs) and arginine-glycine rich RGG motifs at the C terminus [2]. SG formation requires the NTF2-like and RGG domains and likely involves G3BP-driven condensation of stalled mRNP complexes as well as numerous SG nucleator proteins including TIA-1. SG induction is triggered by inhibition of translation initiation caused by cellular stresses including virus infection. As viruses strictly depend on the host translation machinery it is not surprising that many viruses from diverse virus families have developed mechanisms to disrupt SG assembly [3]. For Old World alphaviruses, including Semliki Forest virus (SFV) and chikungunya virus (CHIKV), interaction between FGDF motifs in the C-terminal region of viral non-structural protein 3 (nsP3) and G3BP disrupts SGs and blocks their formation [4, 5].
Alphaviruses are a group of enveloped positive-sense single-stranded RNA viruses belonging to the family Togaviridae [6]. Two separate open reading frames within the viral genome encode the non-structural and structural proteins. The four non-structural proteins, nsP1-4, form the viral replicase. Each of the non-structural proteins fulfils distinct tasks during the synthesis of viral RNA [7], but some non-structural proteins are found in other subcellular locations [8–10]. The role of nsP3 has been less well defined, but appears to be linked to several important interactions with host proteins, recently reviewed in [11] and [12]. A well-established nsP3-host interaction of the Old World alphaviruses is its binding to G3BP [4, 13–16]. This interaction is mediated by two FGDF motifs in the C-terminus of nsP3, which target a hydrophobic groove within the N-terminal NTF2-like domain of G3BP [4, 17]. In the absence of nsP3/G3BP-interactions, SFV replication is compromised, but still detectable [4, 5], while CHIKV is profoundly attenuated [16, 17]. The functional relevance of G3BP-binding during alphavirus infection has been the subject of several studies. Early events in infection lead to protein kinase R (PKR)-mediated phosphorylation of eukaryotic initiation factor 2α (eIF2α) and subsequent SG assembly [18]. G3BP is essential for SG assembly in response to eIF2α phosphorylation [2]. During the course of infection, nsP3 levels rise and SGs are disassembled through nsP3-mediated sequestration of G3BP to viral replication complexes and other locations of viral protein aggregation [4, 5, 19]. Variants of SFV mutated at both FGDF motifs, which consequently do not bind G3BP, provoke a prolonged SG response and virus growth is attenuated in cell culture [5]. Other, pro-viral roles of G3BP have been suggested, with possible functions in the switch from viral genome translation to negative-strand RNA synthesis [20] or protection of viral genomic RNA and formation of viral replication complexes [16]. In a previous study, we solved the 3-dimensional structure of the NTF2-like domain of G3BP1 (residues 1–139) in complex with a 25-residue peptide from SFV nsP3, including both FGDF motifs [17]. The NTF2-like domain forms dimers and each of the FGDF motifs binds to G3BP monomers on separate dimers, thereby crosslinking G3BP-dimers and inducing the formation of an nsP3-G3BP poly-complex. We hypothesized that this poly-complex stabilizes viral replication complexes by binding them together, producing high local concentrations of viral RNA and subsequently promoting virus replication, but functions of downstream domains of G3BP in alphavirus replication remain unidentified. Apart from the alphaviruses, many other viruses from diverse families interact with G3BP and recruit it to sites of viral macromolecule synthesis or processing [4, 21–27] but mechanisms of G3BP function in viral replication have not been described.
In this study we use a panel of G3BP1 variants to show that the NTF2-like and RGG domains of G3BP1 confer separate pro-viral functions upon recruitment to the alphaviral RNA replication complexes. The former is as a structural component of the viral RNA replication complexes, while the latter directs the recruitment of 40S ribosomal subunits to the cytopathic vacuoles (CPVs), thereby promoting efficient translation of viral mRNAs. Our work also revealed that CHIKV replication requires both these domains, while SFV replication is fully efficient with only the structural contribution of the NTF2-like domain.
In order to investigate the functional relevance of G3BP-recruitment to the sites of viral replication we utilised a U2OS ΔΔG3BP1/G3BP2 double-knockout cell line (hereafter referred to as U2OS ΔΔ) that was stably transfected with a selected set of different GFP-fused G3BP1 mutants (Fig 1A), previously described in [2]. A U2OS ΔΔ cell line constitutively expressing GFP only (GFP) served as a control cell line. In addition to full-length wild type G3BP1 (GFP-G3BP WT), we used G3BP mutants or truncations that lacked specific domains or larger regions. The most extreme truncation results in expression of only the NTF2-like domain (GFP-G3BP 1–135), which preserves the ability to bind to viral nsP3 [17]. The previously described F33W mutation disrupts this interaction (GFP-G3BP F33W) [4], while maintaining SG-forming activity [2]. Additional deletions were made of distinct regions present in the C-terminal region of G3BP1, an RNA-recognition motif (GFP-G3BP ΔRRM) and an arginine/glycine-rich region (GFP-G3BP ΔRGG). G3BP mutants lacking the RGG domain fail to assemble SGs and to interact with 40S subunits [2]. The different G3BP1 variants were expressed to similar levels with exception of GFP-G3BP1 1–135, which showed variable and notably lower expression level (Fig 1B). Densitometric analysis revealed the level to be 27% +/- 8% relative to GFP-G3BP1 WT.
We used SFV and CHIKV as representative alphaviruses to investigate the effect of the different G3BP1 mutants on the viral life cycle in a multistep growth curve experiment. SFV replication was significantly impaired but still detectable in cells expressing GFP alone or GFP-G3BP F33W (Fig 1C), in agreement with previous observations [4, 5]. Remarkably, expression of only the NTF2-like domain of G3BP1 (GFP-G3BP1 1–135) rescued titres similar to those obtained using full length GFP-G3BP1, despite the much lower expression of the NTF2-like domain relative to the other mutants (Fig 1B). GFP-G3BP1 ΔRRM also promoted virus replication very similarly as did GFP-G3BP1 WT, while GFP-G3BP1 ΔRGG caused a slight delay in virus growth, but still enhanced virus replication as compared to cells expressing GFP alone or GFP-G3BP F33W. Taken together, these results indicate a pro-viral effect of G3BP1 on SFV titres that is independent of G3BP’s RNA-binding, 40S subunit interaction or ability to form SGs, but is mediated by the NTF2-like domain alone.
We compared SFV replication in the GFP-G3BP1 WT cell line with the parental U2OS and the U2OS ΔΔ cells. We found that 24 hours post infection, the reconstitution rescued the viral titres to a large extent, although not completely (S1A Fig). However, at early times post infection, when we analysed the intensity of dsRNA signal per cell, we found there to be very similar levels in the parental U2OS cells as in the GFP-G3BP1 WT cell line, but barely detectable dsRNA signals in the U2OS ΔΔ cells (S1B Fig). The lack of complete rescue of SFV replication in the GFP-G3BP WT cells could be due to specific functions of G3BP2 that are lacking in the GFP-G3BP WT cells, or other effects of overexpression of the tagged protein (S1C Fig). Nevertheless, this supported that the GFP-G3BP1 WT cells are an appropriate model for studies on G3BP1 function in infection.
We next tested the growth behaviour of CHIKV in the full panel of reconstituted cell lines (Fig 1D). In agreement with previous reports [16, 17], production of progeny virus was not detectable in absence of the nsP3-G3BP1 interaction. In contrast to SFV, the G3BP1 NTF2-like domain was unable to promote CHIKV replication. CHIKV replication was rescued by full length GFP-G3BP1 and less effectively by GFP-G3BP1 ΔRRM. Thus, CHIKV only detectably replicated in cells expressing G3BP1 variants containing both the NTF2-like and the RGG domains, suggesting that there is another G3BP-dependent proviral effect conferred on CHIKV replication by the RGG domain.
To further characterize the functional relevance of G3BP1 during the SFV and CHIKV life cycles, we performed immunofluorescence experiments to test if the different G3BP1 mutants affected the organization of viral replication complexes (RCs) within cells. For this, we first infected the cell lines presented in Fig 1A with SFV at MOI 10, fixed them 8hpi and stained for nsP3 (red) and double stranded RNA (dsRNA, blue) (Fig 2A). In cells expressing GFP-G3BP1 WT, the proteins colocalized with nsP3 and accumulated in close proximity to clusters of dsRNA-positive foci, indicative of sites of active viral RNA replication. In contrast, foci of nsP3 and dsRNA were detectable but weak in cells expressing GFP alone or GFP-G3BP1 F33W, and GFP signals were diffuse in the cytoplasm. In the GFP-G3BP1 F33W expressing cells, GFP signals were clearly present in large SGs in cells at early stages of infection, which were small in cells expressing high levels of nsP3. This was similar to observations made in previous studies in cells infected with SFV mutants, which were unable to bind G3BP [4, 28, 29]. Expression of the NTF2-like domain of G3BP1 (GFP-G3BP1 1–135) was sufficient to promote clustering of viral RCs in discrete foci in the cell. However, nsP3-G3BP1 1–135 aggregates appeared fewer in number and smaller in size compared to nsP3-G3BP1 WT, likely due to the lower expression level of the truncated protein. Cells expressing G3BP1 mutants lacking either the RRM (ΔRRM) or the RGG (ΔRGG) domain still produced strong dsRNA signals that colocalised with nsP3 and G3BP1, although they were less obviously clustered, and a fraction of GFP signal remained diffuse in the cytoplasm. To summarise, recruitment of G3BP1 via the NTF2-like domain is necessary for efficient formation of clusters of SFV replication complexes. This is also reflected by the quantitative analysis of dsRNA signals in images from 3 replicate experiments (Fig 2B). Production of dsRNA-positive replication complexes is strongly impaired in the absence of the nsP3:G3BP1 interaction.
When the same panel of cell lines was infected with CHIKV at MOI 1, only cells expressing GFP-G3BP1 WT or GFP-G3BP1 ΔRRM stained strongly for nsP3 and dsRNA (Fig 2C), as expected from the growth curve data (Fig 1D). In GFP-G3BP1 WT cells, nsP3 was found to colocalise strongly with the GFP signal, mostly in large foci, devoid of dsRNA signal, but noticeably also colocalised with dsRNA signals in patches at the plasma membrane (Fig 2C). This was largely also true for GFP-G3BP1 ΔRRM cells although in those cultures there were fewer infected cells, containing weaker dsRNA signals and smaller nsP3-G3BP colocalizing foci. Only in very few cases (1–5% of the cells) did we observe nsP3-positive cells in GFP-G3BP1 F33W, GFP-G3BP1 1–135 or GFP-G3BP1-ΔRGG cells (S2 Fig). Moreover, in those rare cases, the dsRNA signal was either absent or at the limit of detection. Despite these differences between SFV and CHIKV, CHIKV nsP3 remained largely diffuse in GFP-G3BP1 F33W, ΔRRM and ΔRGG expressing cells, but predominantly associated with GFP-G3BP1 WT and 1–135 (Fig 2C and S2 Fig), similar to the observations made with SFV (Fig 2A).
The extreme reliance of CHIKV on G3BP for viral RNA replication and gene expression makes it difficult to study the effects of mutation or truncation of the protein on viral replication. However, since SFV is less sensitive to the absence of the protein, we were able to use SFV to study the contributions of the individual domains of the protein mutated in our panel. The immunofluorescence data show that G3BP mutants influence the formation as well as organization of SFV replication complexes. To obtain a more detailed picture of this process in the different cell lines, we applied transmission electron microscopy to view infected cells. Alphavirus replication induces the formation of bulb-shaped membrane protrusions at the plasma membrane (PM), referred to as spherules, which for some viruses are later internalized to mature into cytopathic vacuoles (CPV) in the perinuclear area [30, 31]. This internalization requires another activity of nsP3, the activation of the PI3K-Akt-mTOR pathway [31, 32], found for SFV nsP3 but not CHIKV. We infected the G3BP1 mutant cell line panel with SFV and fixed at 3h and 8h post infection, to investigate the impact of GFP-G3BP1 variants on spherule formation at the PM and CPV internalization, respectively. At 3 hpi, we observed marked differences in the appearance of spherules at the PM between the different cell lines (Fig 3A). In the absence of the nsP3-G3BP1 interaction (GFP and GFP-G3BP1 F33W cells), spherules were observed to be very few in number and sparsely distributed along the PM. In contrast, cells expressing intact G3BP1 NTF2-like domain (GFP-G3BP1 WT, 1–135, ΔRRM and ΔRGG) contained larger quantities of spherules, readily detectable at the PM. At 8 hpi (Fig 3B), CPVs were apparent in SFV-infected GFP-G3BP1 WT, 1–135, ΔRRM and ΔRGG cells. In particular, the SFV CPVs in GFP-G3BP1 WT and 1–135 expressing cells frequently appeared in clusters of 3–5 CPVs, in some cases reaching numbers of 10 or more. As expected from the earlier time point analyses, CPVs were much less frequent in GFP and GFP-G3BP1 F33W cells, and mostly appeared isolated (Fig 3B). From these observations, we conclude that the interaction of SFV nsP3 with the NTF2-like domain of G3BP1 is required for efficient formation of viral replication complexes early in infection and is associated with the appearance of increased numbers of CPVs, often found in clusters.
Although domains of G3BP downstream of the NTF2-like domain were not necessary for clustering of SFV CPVs, we noticed a distinct feature of the CPV clusters restricted to cells expressing GFP-G3BP1 WT or GFP-G3BP1 ΔRRM. In these cells, the clusters of CPVs were mostly surrounded by large (1–2 μm) patches of electron dense material, indicative of the presence of high molecular weight molecules. The formation of these patches was likely not simply a direct result of efficient RNA replication, since they were absent in SFV-infected GFP-G3BP1 1–135 cells, in which SFV replication was efficient (Fig 1C) and induced efficient clustering of CPVs, replete with spherules (Fig 3). The patches were also observed in parental U2OS cells after 8h of SFV infection (S3A Fig), excluding artefacts of GFP fusion protein overexpression, and were occasionally also detected surrounding the cytoplasmic side of spherule-containing sections of plasma membrane (S3B Fig). The CPV clusters and associated electron dense patches were frequently observed in the vicinity of membranous structures reminiscent of endoplasmic reticulum (S3C Fig). Since the GFP-G3BP1 WT and GFP-G3BP1 ΔRRM cell lines in which these patches were observed were the only lines in our panel capable of supporting the replication of CHIKV (Fig 1D), we thought it likely that these patches represent some important pro-viral factor that is required for CHIKV replication and which warranted further investigation.
These electron-dense patches bear a striking similarity to sodium arsenite-induced SGs in HeLa and HEK293 cells [33], and also to SFV infection-induced SGs which we frequently observed in U2OS ΔΔG3BP cells expressing GFP-G3BP1 F33W (S3D Fig). Overexpressed G3BP1 forms spontaneous, stress-independent SGs that require both the NTF2-like domain and RGG regions [1, 2, 34]. It is thought that the high concentrations of this protein can act to recruit ribosomes, mRNAs and other proteins containing low complexity regions and prion-like domains, which can interact via multiple low affinity interactions and undergo liquid-liquid phase separations (LLPS) to a hydrogel state [2]. The presence of the electron-dense patches around CPVs depended on the recruitment of G3BP1 via the NTF2-like domain and on the presence of the RGG region, since they were not seen in cells expressing G3BP1 variants lacking these regions. We propose therefore that the patches are generated by LLPS transitions nucleated by high local concentrations of G3BP1 surrounding clustered CPVs.
To identify contents of these electron-dense patches, we performed immuno-electron microscopy, staining SFV-infected U2OS ΔΔ + GFP-G3BP1 WT cells for nsP3 and GFP. In agreement with our immunofluorescence experiments, GFP-G3BP1 and nsP3 were found enriched in proximity to CPV clusters (Fig 4A). Moreover, both proteins were not restricted to the boundaries of the CPVs, where the spherules are located, but rather stretched out over a larger area, enclosing nearby CPVs, thus likely corresponding to the patches seen in Fig 3B. This was particularly noticeable for GFP-G3BP1, but to a lesser extent also for nsP3.
SGs are condensates of translationally silent mRNP complexes, which include many proteins necessary for translation initiation [35, 36]. SG components are in dynamic equilibrium with polysomes, and cycle rapidly between translationally silent SGs and translationally active polysomes in cells recovering from stress [37]. We therefore considered the possibility that the recruitment of such SG components around the CPV clusters might benefit the virus, as molecules associated with translation initiation would be concentrated in the vicinity of viral mRNA production. We infected cells with SFV and stained for canonical translation initiation factors eIF3B (Fig 4B) and eIF4A (Fig 4D) and found that both were recruited to CPV clusters in SFV-infected GFP-G3BP1 WT- and GFP-G3BP1 ΔRRM expressing cells, but less so in those where the nsP3:G3BP1 interaction is absent or G3BP1 lacks the RGG domain. In order to quantify the degree of correlation between dsRNA and translation initiation factors, we calculated the Pearson correlation coefficient (Fig 4C and 4E). For this analysis only infected cells with detectable dsRNA signals were considered. The results support the observation that the correlation is stronger for GFP-G3BP1 WT & ΔRRM cells compared to all the other cell lines, although the quantification also indicates that deletion of the RRM domain of G3BP has a negative effect on the correlation.
To verify that the recruitment of translation initiation factors is no artefact due to abnormal overexpression of GFP-G3BP1 WT, we repeated the experiment and included parental U2OS WT and U2OS ΔΔ cells. As shown in S4A Fig, dsRNA-positive CPVs of SFV-infected U2OS WT cells strongly enrich eIF4A in close proximity, similar to U2OS ΔΔ + GFP-G3BP1 WT cells, but no such enrichment was observed in U2OS ΔΔ cells. Quantitative analysis of the degree of dsRNA-eIF4A correlation in infected dsRNA-positive cells complements this observation (S4B Fig).
It has previously been shown that initiation of translation on the subgenomic 26S mRNA of certain alphaviruses requires the initiation factor eIF2D (previously referred to as ligatin) [38, 39]. We found that this factor is also specifically recruited to the CPV clusters in SFV-infected GFP-G3BP1 WT- and GFP-G3BP1 ΔRRM expressing cells (Fig 4F and 4G).
The RGG domain of G3BP1 is important for association with 40S subunits during formation of SGs [2]. We hypothesised therefore, that in infected cells, nsP3:G3BP1 complexes associate with 40S subunits, recruiting them to the CPVs. To determine if nsP3:G3BP:40S subunit complexes exist in SFV infected cells, we infected our panel of cell lines with SFV, isolated GFP-bound complexes at 7 hpi, and immunoblotted for nsP3 and ribosomal proteins rpS6, rpS3 and rpL4 (Fig 5A). As expected, nsP3 strongly coprecipitated with GFP-G3BP1 WT, 1–135, ΔRRM and ΔRGG. 40S ribosomal subunit proteins rpS3 and rpS6 coprecipitated with GFP-G3BP1 WT and F33W, and even more robustly with ΔRRM, but did not detectably interact with 1–135 or ΔRGG, confirming the necessity of the RGG region for that interaction [2]. These results show that the G3BP1:40S subunit interaction remains intact in SFV-infected cells and can be found in complex with nsP3.
In order to verify that ternary complexes of nsP3, G3BP and 40S subunits assemble in alphavirus infected cells, we performed immunoprecipitation experiments using pull-down of either 40S subunits or of nsP3. For pull-down of 40S ribosomal subunits, we used a U2OS cell line stably expressing rpS6-GFP, which functionally integrates into ribosomes [2]. We infected these cells with either WT SFV or SFV-F3ANC, which contains an inactivating point mutation in each of its two FGDF motifs, thus disrupting the nsP3:G3BP interaction [4, 17]. Isolation of 40S ribosomal subunits via rpS6-GFP was confirmed by immunostaining for rpS3. Endogenous G3BP1 coprecipitated in all conditions, but was more readily detected in SFV WT-infected cells (Fig 5B). Importantly, nsP3 coprecipitated from SFV WT-infected cells but not from those infected with SFV-F3ANC-infection. This supports the observation that the nsP3:G3BP1 interaction does not prevent the association of G3BP1 with the 40S subunits, allowing the formation of nsP3:G3BP1:40S multi-protein complexes in SFV-infected cells.
The binding mechanism for CHIKV nsP3 to G3BP is very similar to that of SFV [17, 28], so we predicted that 40S subunits would also be present in CHIKV nsP3-bound complexes. We transfected a selected set of our reconstituted cell lines with biotin-acceptor peptide (BAP)-tagged CHIKV nsP3 [28, 40] and purified nsP3 complexes using streptavidin beads. Since each of the G3BP1 variants used contained an intact NTF2-like domain, all coprecipitated with nsP3, as expected (Fig 5C). The 40S subunit proteins rpS3 and rpS6 also co-precipitated with nsP3 from lysates of GFP-G3BP1 WT and GFP-G3BP1 ΔRRM expressing cells, but not from those expressing GFP alone or GFP-G3BP1 ΔRGG.
Taken together, the results in Fig 5 demonstrate that SFV and CHIKV nsP3 form complexes with G3BP1-bound 40S ribosomal subunits in infected cells and suggest that the electron dense regions surrounding the CPV clusters contain G3BP1:40S subunit complexes as well as translation initiation factors.
The G3BP-dependent enrichment of the translation initiation machinery in close proximity to viral replication complexes might represent a mechanism for alphaviruses to ensure efficient initiation of translation of newly produced viral RNAs. To determine whether translation of viral RNAs was more efficient when 40S subunits and initiation factors were recruited to CPVs, we used the ribopuromycylation method to visualize translational activity in our reconstituted cells after SFV infection. This method is based on the covalent attachment of puromycin (PMY) to nascent polypeptide chains produced by actively translating ribosomes and its subsequent detection using a puromycin-specific antibody [41, 42]. We infected each cell line with SFV for 8 hours, and then applied a 5-minute pulse with puromycin before fixation and staining with nsP3 and puromycin antibodies. For each cell line, representative cytoplasmic regions with characteristic nsP3-staining indicate the localization of viral CPVs (Fig 6A, boxes). In some images, non-infected cells can be observed, containing very strong and diffuse PMY signals indicative of uninhibited translation, but in all infected cells, PMY staining is much reduced due to the profound shut off of host protein synthesis. In GFP-G3BP1 WT expressing cells, strong PMY staining was evident in and around the CPV clusters and was also detected in more diffuse areas in the cytoplasm. In contrast, GFP-G3BP1 F33W expressing cells displayed fewer clustered nsP3 puncta with no detectable PMY staining, even upon longer exposures. SFV nsP3 was also clustered in GFP-G3BP1 1–135 cells and, although PMY staining was evident close to these clusters, it was noticeably weaker than in GFP-G3BP1 WT cells, for which we had verified the interaction with 40S subunits. In GFP-G3BP1 ΔRRM expressing cells, CPVs were less clustered, but usually displayed relatively strong PMY signals. However, In GFP-G3BP1 ΔRGG expressing cells in which the G3BP:40S interaction is absent, nsP3 signals were more diffuse than in GFP-G3BP1 ΔRRM expressing cells and even when detected in clusters in some cells, contained little or no PMY staining. To allow for statistical analysis and provide a representative measure of localised translational activity, we quantified the degree of nsP3-puromycin correlation in multiple cells using the Pearson correlation coefficient (Fig 6B). We found there to be a positive correlation in all cell lines, which was not surprising since all cells support replication and gene expression of SFV mRNA (Fig 1C). However, there was a significant reduction in correlation compared to GFP-G3BP1 WT cells in all cell lines except GFP-G3BP1 ΔRRM cells. The noticeably low levels of correlation in F33W cells may be due to the presence of SGs that are still frequently found in these cells, which would additionally restrict localized translation in these cells.
Taken together, these results support a model in which G3BP1 acts as a bridge between viral replication complexes and the 40S subunit to enhance translational activity in close proximity of viral RCs.
To further characterize the G3BP-dependent enhancement of viral mRNA translation on a cellular level, we monitored global translational efficiency using polysome preparations. This technique employs a linear sucrose gradient to separate ribosomal subunits (40S, 60S), 80S monosomes and ribosomes engaged in polysomes, and thereby allows for quantification of the relative amount of ribosomes engaged in efficient translation (i.e. associated with polysomes) [43]. We infected BHK-21 cells with either SFV WT or the G3BP-binding mutant SFV F3ANC and compared their polysome tracings to mock-infected cells. Both viruses altered the polysome tracing noticeably and we observed increased absorbance at 254nM in fractions corresponding to polysomes. This is consistent with a higher number of translating ribosomes in SFV WT infected cells as compared to mock-infected cells or cells exposed to SFV F3ANC (Fig 7A and 7B). Thus in absence of G3BP-binding the translational efficiency in SFV-infected cells is significantly reduced. This observation is in accordance with our PMY staining, showing that recruitment of G3BP to CPVs leads to enhanced translational activity.
The host translation machinery plays an indispensable role in viral life cycles. Despite their often very limited coding capacity, viruses have developed remarkable strategies to manipulate cellular translation to ensure synthesis of viral proteins [44, 45]. Here we describe a novel strategy by which Old World alphaviruses utilize the cellular ribosome-associated protein G3BP1 to enrich components of the translation machinery at the sites of viral RNA replication. We show that this enrichment depends on the NTF2-like and RGG regions of G3BP1. These are the same domains that are necessary for the formation of SGs under conditions of cellular stress, the NTF2-like domain is needed for homodimerisation as well as binding to caprin-1 and USP10, the positive and negative regulators of SGs, while the RGG region mediates binding to the 40S ribosomal subunits [2]. Under infection conditions, nsP3 binds to the N-terminal NTF2-like domain of G3BP1 and recruits it to viral replication complexes. We demonstrate that 40S ribosomal subunits remain associated with nsP3:G3BP1 complexes (Fig 5), promoting the condensation of electron-dense patches around viral spherules at the PM and around internal CPV clusters (Fig 3B and S3B Fig). These electron-dense patches show noticeable similarity to bona fide SGs under the electron microscope (S3D Fig and [33]) and are enriched for translation initiation factors that are also found in SGs [35]. In contrast to stress-induced SGs however, the nsP3:G3BP1:40S complex-dependent protein accumulations are sites of enhanced translation (Fig 6). We also show that SFV-infected cells, in which the SFV nsP3:G3BP1:40S complex can form (Fig 5), engage significantly more ribosomes in translating polysomes than in cells where this complex cannot form (Fig 7A). The sequestration of G3BP upon infection with Old World alphaviruses thus not only subverts the cellular stress response, but also more explicitly exploits a host mechanism to condense the translation machinery and target it for production of viral proteins. Our data suggest that recruitment of components of the host translation machinery to CPVs is important for the efficient synthesis of CHIKV proteins, particularly at early stages of replication and subsequently influencing all steps of the viral life cycle.
It is remarkable that the recruitment of G3BP1 results in recruitment of translational machinery to SFV CPVs, but that its absence doesn’t impair efficient production of progeny viruses, as long as the NTF2-like domain is present. There may be a number of potential explanations for this. Compared to CHIKV, SFV is highly adapted to replication in cell culture and generally reaches higher titres in our various human as well as rodent cell lines. More efficient replication and consequently higher amounts of viral RNA produced might bypass the need for active recruitment of the host translation machinery. Viral mRNAs present in excess, may not require translation initiation directly at the CPV since they can anyway be translated efficiently enough to reach the levels required to advance the replicative cycle. The subgenomic RNA of SFV and SINV contain a stem-loop structure, originally termed the translational enhancer, situated downstream of the initiator AUG codon, which facilitates translation initiation under conditions of low GTP–eIF2–tRNAiMet ternary complex availability [18, 46, 47]. A similar RNA structure has not been detected in CHIKV subgenomic RNA [48] which might explain the increased requirement for direct recruitment of translational machinery to sites of CHIKV mRNA production. Another difference between SFV and CHIKV that could explain why active recruitment of the translational machinery is not needed for SFV is the cellular localization of viral RCs. SFV RCs are internalized from the plasma membrane (PM) and accumulate in CPV in the perinuclear area, while CHIKV RCs remain largely at the PM [31]. Naturally, exposure to the cytoplasm is more restricted for CHIKV RCs and access to translation factors in close proximity might be more limited, especially as in the initial hours of infection translation factors first accumulate in SGs, which tend to move towards the perinuclear region, away from the PM [49], unless disrupted by nsP3:G3BP interaction. Another explanation for the remarkable rescue of SFV replication in the G3BP1 1–135 cells might be that the NTF2-like domain, when expressed alone, provides supplementary proviral advantages that full-length G3BP1 cannot provide, thereby compensating for the absence of the RGG domain. For example, the clustering of CPVs may be more efficient without steric interference of other G3BP domains in the full-length protein.
The recruitment of translation initiation complexes around the CPV clusters is consistent with earlier observations of large ribonucleoprotein networks extending outwards from SFV CPVs in BHK cells that stained positive for nsP3 in immuno-EM studies [50]. In that work, the authors proposed that the networks might represent sites of production of viral RNA and proteins and of encapsidation of viral genomic RNA. Our work is consistent with that model and shows that the molecular link between the viral CPVs and the cellular translation apparatus is G3BP1, recruited by nsP3. Other, more recent work has provided evidence that the newly produced alphaviral RNAs, upon exit from the spherule neck are protected from degradation by cellular RNases [51, 52]. Indeed, a role for G3BP in protection of new viral genomic RNAs has been proposed [16], mediated by either or both of the RRM and RGG domains. Our work demonstrates a vital role for the RGG domain, and suggests that the RNAs are protected by their immediate engagement with translation initiation complexes. RGG domains have been implicated in a range of nucleic acid and protein interactions, mediated by Arg residues in glycine-rich contexts and likely adopt a disordered and flexible structure [53]. The RGG domains of G3BP1 and 2 are classified in the Di-RG group, according to Thandapani et al [54]. Several of the Arg residues are methylated, a modification that is rapidly removed prior to SG formation [55]. It is not known whether the interaction between the G3BP1 RGG domain and the 40S ribosomal subunit is mediated by protein or RNA components of the 40S ribosome, but we previously reported that the interaction is partially resistant to RNase digestion [2]. We also note a distinct lack of aromatic amino acids in the G3BP RGG domains that would be characteristic of a nucleic acid binding domain [54]. Further work will be required to more exactly map the interaction regions of G3BP and the 40S subunit.
Our work also identified a minor role for the RRM domain, since CHIKV replication was delayed and final titres were reduced in its absence (Fig 1D). Interestingly, we found that association of translation initiation factors with SFV dsRNA-positive replication complexes was reduced in the absence of the RRM domain (Fig 4B–4G). However, deletion of the RRM domain results in stronger association of G3BP1 with 40S subunits in GFP-G3BP1 pull-down experiments (Fig 5A) and [2], although not in our CHIKV BAP-nsP3 immunoprecipitation (Fig 5C). The RRM domain of G3BP is a putative RNA-binding domain containing two conserved ribonucleoprotein motifs, RNP1 and RNP2 [56]. Several studies have reported RNA-binding properties of G3BP, including interactions with viral RNA [57–59]. It is therefore possible that the RRM domain engages alphavirus mRNA once released from viral replication complexes, and thereby readily provides transcripts for 40S subunits bound to the RGG domain. Alternatively, it may be involved in the sorting of genomic RNAs away from translation complexes, for example for encapsidation. Further work will be required to determine the role of the RRM domain.
The strict requirement for particular domains of G3BP1 for CHIKV strongly restricted our experimental abilities, as CHIKV is essentially non-viable in most of our mutant cell lines (Fig 1D). We therefore employed the closely related SFV, which replicates at a reduced level even in the absence of G3BP (Fig 1C), in order to investigate the influence of different G3BP1 domains during various steps of the alphavirus replication cycle. SFV is commonly used as a representative model virus in Old World alphavirus research. In fact, CHIKV and SFV nsP3 proteins both contain FGDF motifs [28], which bind to the NTF2-like domain of G3BP in an identical manner although surrounding sequences may impose a slightly different orientation of the G3BP monomers in the proposed polycomplex [17]. Although the extent to which the viruses depended on G3BP1 differed, both replicated best in the presence of full length G3BP1 and least well in the absence of the protein or in the presence of the F33W binding mutant, as expected. A striking difference between CHIKV and SFV is the ability of the NTF2-like domain alone (G3BP1 1–135) to rescue the replication of SFV, but not CHIKV. The rescue of SFV replication was especially remarkable considering that the expression level of the GFP-G3BP 1–135 construct was considerably lower than that of GFP-G3BP WT (Fig 1B). To our knowledge, no enzymatic function of the NTF2-like domain has been reported and we propose a solely structural pro-viral role. It is known that the NTF2-like domain of G3BP forms dimers [60] and in our previous work we have shown that the two FGDF-motifs of nsP3 link these dimers into a chain of nsP3-G3BP oligomers [17]. We hypothesized that these oligomeric structures could stabilize viral replication complexes by binding them together, thus ensuring that, upon internalization, each CPV contains a high concentration of spherules. The importance of this clustering is also supported by our previous observation that mutation of either one of the two FGDF motifs within SFV nsP3 results in the same level of attenuation, as if both motifs are mutated [17]. Even though single FGDF motifs can still recruit G3BP, both motifs are required to link G3BP dimers and consequently to facilitate clustering of RCs. In contrast to SFV, CHIKV replication was not supported by G3BP1 1–135 alone (Fig 1D), but in the few G3BP1 1–135 expressing cells in which CHIKV nsP3 was detected, it was also observed in foci together with GFP-G3BP1 1–135 (S2 Fig). However, dsRNA signals were very weak, and no detectable level of infectious virus was released from those cells indicating that this recruitment is not enough to promote replication of CHIKV.
In contrast to Old World alphaviruses, the New World alphavirus VEEV is not sensitive to G3BP ablation, but efficient replication depends on another set of SG-related proteins that bind to nsP3, the Fragile X syndrome (FXR) protein family, FXR1, FXR2 and FMR1 [16]. Just like G3BP proteins, FXR proteins also associate with ribosomes, though predominantly with 60S ribosomal subunits [61]. If this interaction is important during VEEV infection remains to be investigated, but it could suggest an analogous mechanism for recruitment of ribosomes and translation initiation factors to viral replication sites by New World alphaviruses.
Targeting SG formation is a common viral strategy to interfere with inhibition of translation, often involving manipulation of G3BP [62]. The condensation of cellular translation machinery around viral RCs via sequestration of G3BP is, however, a previously unknown mechanism to ensure efficient synthesis of viral proteins. Other viruses might have evolved similar strategies, either G3BP-mediated or through other RGG domain-containing proteins, or possibly SG-associated proteins. As demonstrated here for Old World alphaviruses, this could also be a crucial step for efficient replication of other viruses, and could represent a target for antiviral therapies.
Human osteosarcoma (U2OS) cells (ATCC HTB-96) were kept in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% foetal bovine serum (FBS) and 100 U/mL penicillin, and 100 μg/mL streptomycin. U2OS-derived double-null ΔΔG3BP1/2 KO cells constitutively expressing GFP-G3BP1 WT and mutants were obtained from Nancy Kedersha [2] and maintained in the same medium as wild-type U2OS supplemented with 0.5 mg/mL geneticin/G418 (Thermo Fisher Scientific). Baby hamster kidney (BHK-21) cells (ATCC CCL-10) were maintained in Glasgow’s modified Eagle’s medium (GMEM) supplemented with 10% FBS, 10% tryptose phosphate broth, 20 mM HEPES, 1mM L-glutamine and 100 U/mL penicillin, and 100 μg/mL streptomycin. All cells were cultured at 37°C in 5% CO2.
Construction of pEBB/PP-CHIKV nsP3, encoding CHIKV nsP3 N-terminally fused to the BAP tag, is described in [40]. All plasmids were verified by sequencing (Eurofins). Cells were transiently transfected using Lipofectamine 2000 (Thermo Fisher Scientific) according to the manufacturer’s instructions.
SFV was rescued by transfection of BHK-21 cells with the infectious plasmid pCMV-SFV4 [63]. Similarly, CHIKV was rescued from the pCMV-CHIKV-ICRES (CHIKV LR2006-OPY1) infectious clone. Virus titres for infection experiments on U2OS-based cell lines were determined on wild type U2OS cells by plaque assay. For infections, cell monolayers were washed with PBS and virus was added in infection medium (DMEM supplemented with 0.2% BSA, 2 mM L-glutamine, and 20 mM HEPES) with periodic shaking for 1 h at 37°C. Infectious media were then removed and cells washed with PBS before adding pre-warmed complete medium. For virus growth curve experiments, cells were grown in 6-well dishes to ~90% confluence, infected as described above and cells overlaid with 2mL complete medium and samples of supernatant taken at different time points post infection. Virus titres were then determined on BHK cells by plaque assay.
A 10-fold serial dilution of virus suspension was prepared and used to infect monolayers of BHK-21 cells for 1 h at 37°C. Cells were washed with PBS, kept in BHK-21 media supplemented with 0.8% Agarose (w/v) and incubated at 37°C. After 36 h, 10% formaldehyde (v/v) in PBS was added, incubated for at least 4 h at room temperature and plaques revealed by crystal violet staining. Statistical analysis was performed using an unpaired, two-tailed Student’s t-test with a 95% confidence interval.
Samples were resolved on NuPAGE 4–12% Bis-Tris polyacrylamide gels (Thermo Fisher Scientific) and transferred onto Amersham Hybond P 0.45 PVDF blotting membranes (GE Healthcare). Membranes were blocked with 5% skim milk powder in Tris-buffered saline with 0.05% (v/v) Tween 20 (TBST) and incubated with primary antibodies (16 h at 4°C) and horseradish peroxidase (HRP)-coupled secondary antibodies (1 h at room temperature) in 1% BSA/TBST as listed below. Chemiluminescence was detected using SuperSignal West Pico PLUS Chemiluminescent Substrate (Thermo Fisher Scientific) and a Curix 60 film developer (AGFA).
Primary antibodies: rabbit anti-GFP (290; Abcam; 1:10.000), mouse anti-rpS6 (74459; Santa Cruz; 1:2.000), mouse anti-rpS3 (66046-1-Ig; Proteintech; 1:2.000), rabbit antiserum against SFV-nsP3 (1:7.000; [64]), mouse anti-G3BP1 (365338; Santa Cruz; 1:1.000), goat anti-actin (1616; Santa Cruz; 1:500). Secondary antibodies: HRP-conjugated anti-mouse (Sigma A9044; 1:10.000), HRP-conjugated anti-rabbit (Cell Signalling; 1:5.000). Densitometry was performed using ImageJ.
Cells grown on cover glasses (VWR) were fixed with 3.7% formaldehyde (v/v) in PBS for 15 min at room temperature, immersed in methanol for 10 min at -20°C and blocked with 5% horse serum (Sigma) in PBS at 4°C overnight. Antibodies were diluted in blocking buffer as listed below and samples were incubated for 1 h with primary antibodies, followed by 30 min incubation with secondary antibodies at room temperature. Cover glasses were mounted on glass slides using vinol mounting media [65] and imaged by confocal laser scanning microscopy using a Supercontinuum Confocal Leica TCS SP5 X equipped with a pulsed white light laser and a Leica HCX PL Apo 63x/1.40 oil objective. Images were processed using Adobe Photoshop. Settings for image acquisition and adjustment were kept constant for all samples for dsRNA signals. Settings for nsP3 and GFP were varied slightly between samples to compensate for strong differences in localised signal intensities. Primary antibodies: mouse anti-dsRNA (English and Scientific Consulting; 1:200), rabbit antiserum against SFV-nsP3 (1:500; [64]), rabbit anti-eIF2D (Abcam ab108218; 1:200), goat anti-eIF3B (Santa Cruz Biotechnology N-20 sc-16377; 1:200), rabbit anti-eIF4A (Abcam ab31217; 1:200). Secondary antibodies: Alexa Fluor 488 (Molecular Probes; 1:200), Alexa Fluor 568 (Molecular Probes; 1:1.000), Alexa Fluor 647 (Molecular Probes; 1:500).
Immunofluorescence images were analysed in CellProfiler [66]. Settings for confocal image acquisition were kept constant among all cell lines to ensure comparability during the analysis procedure. In CellProfiler, cells were detected using the ‘IdentifyPrimaryObjects’, ‘IdentifySecondaryObjects’ and ‘IdentifyTertiaryObjects’ modules. For each identified cell object, dsRNA signals were quantified using the ‘MeasureObjectIntensity’ module and correlations between dsRNA and initiation factors (eIF2D, eIF3B, eIF4A) were calculated as Pearson’s correlation coefficient using the ‘MeasureColocalization’ module. Calculations of correlation were performed on infected cells only, based on the dsRNA signal. Correlations between nsP3 and puromycin were performed on selected nsP3-positive fields of fixed size. Statistical analysis was done using an unpaired, two-tailed Student’s t-test with a 95% confidence interval.
Near-confluent cells grown in a 100-mm dish were washed with cold phosphate-buffered saline without calcium and magnesium (PBS(-)) and scrape-harvested in 600μl cold EE-buffer (50 mM HEPES, pH 7.0, 150 mM NaCl, 0.1% NP-40, 10% glycerol, 4 mM EDTA, 2.5 mM EGTA, 0.1 mg/mL Heparin (H3149, Sigma), 1 mM DTT, and HALT protease inhibitors (Thermo Fisher Scientific)) [2]. Cells were sonicated twice for 2 min in an ice-water bath (BRANSON 1510), rotated for 15 min at 4°C and cleared by centrifugation (10.000 g, 10 min, 4°C). GFP-fusion protein-complexes were immunoprecipitated with 25 μl GFP-Trap (ChromoTek) for 60 min at 4°C under constant rotation. Beads were then washed three times with cold EE-buffer and eluted into 2x NuPAGE LDS sample buffer (Thermo Fisher Scientific) containing 50 mM DTT, heated at 80°C for 10 min and analysed by SDS-PAGE and western blotting.
Ribopuromycylation was modified from [42], as described in [41]. In brief, cells were left uninfected or infected with SFV WT at a MOI of 10 for 8h. 5 min before fixation, puromycin was diluted in 37°C prewarmed complete medium to a final concentration of 5 μg/ml, added to the cells and incubated for 5 min at 37°C, 5% CO2. Cells were then washed with PBS and fixed with 3.7% formaldehyde (Sigma) for 5 min and permeabilized for 5 min with ice-cold methanol (Sigma). Cells were then washed with PBS and blocked with 5% horse serum for 30 min at RT. After blocking cells were stained with an anti-puromycin antibody (MABE343; Millipore; 1:1,000) and anti-SFV nsP3 antibody (1:500; [64]) for 2h at RT. Cells without puromycin treatment were used as negative controls. Images were taken with a Supercontinuum Confocal Leica TCS SP5 X equipped with a pulsed white light laser and a Leica HCX PL Apo 63x/1.40 Oil objective and processed using Adobe Photoshop. Settings for image acquisition and adjustment were kept constant for all samples.
Cells were grown on no.1 cover glasses (VWR) and infected with SFV4 at a multiplicity of infection (MOI) of 100. At 3 h and 8 h post infection (pi), cells were fixed with 2% glutaraldehyde in 0.1M sodium cacodylate buffer for 30 min at room temperature in the dark. After fixation, cells were washed three times with 0.1M sodium cacodylate buffer, stained with reduced buffered osmium tetroxide and uranyl acetate and processed for flat embedding and ultrathin sectioning as described in [67]. Images were taken with a Jeol JEM-1400 microscope (80 kV) and a bottom-mounted camera, Gatan Orius SC 1000B.
Cells were fixed in 3% paraformaldehyde in 0.1 M phosphate buffer (PB). Samples were rinsed in 0.1 M PB and infiltrated in 10% gelatin. Specimens were then infiltrated into 2.3 M of sucrose and frozen in liquid nitrogen. Sectioning was performed at -95°C and placed on carbon-reinforced formvar-coated, 50 mesh Nickel grids. Immunolabelling procedure was performed as follows: grids were placed directly on drops of 2% bovine serum albumin (BSA) + 2% gelatin in 0.1 M PB for 1 hour to block non-specific binding. Sections were then incubated with the primary antibody diluted in 0.1 M of phosphate buffer containing 0.2% BSA and 0.2% gelatin overnight in a humidified chamber at room temperature. The sections were thoroughly washed in the same buffer and bound antibodies were detected with protein A coated with 10 nm gold (BBI solution, Analytic standard, Sweden) at a final dilution of 1:100. Sections were rinsed in buffer and fixed in 2% glutaraldehyde and contrasted with 0.05% uranyl acetate and embedded in 1% methylcellulose and examined in a examined in a Tecnai G2 Bio TWIN (FEI company, Eindhoven, The Netherlands) at 100 kV. Digital images were taken using a Veleta camera (Soft Imaging System GmbH, Műnster, Germany) [68].
BHK cells were seeded in 15cm plates and cultured for 48h until 60–70% confluence. Following an 8h infection, polysome fractionation was performed as previously described [69] and the absorbance at 254 nm was recorded along a 5–50% sucrose gradient. The polysome-tracings were normalized to the total absorbance signal (i.e. the area under the curve for the 60S ribosomal subunit, the 80S ribosome and polysomes). Global translational efficiency was quantified as the ratio of polysomes (area under the curve of polysomes) over 80S ribosomes (area under the curve of 80S ribosomes).
|
10.1371/journal.pbio.2005264 | Peripherally derived macrophages modulate microglial function to reduce inflammation after CNS injury | Infiltrating monocyte-derived macrophages (MDMs) and resident microglia dominate central nervous system (CNS) injury sites. Differential roles for these cell populations after injury are beginning to be uncovered. Here, we show evidence that MDMs and microglia directly communicate with one another and differentially modulate each other’s functions. Importantly, microglia-mediated phagocytosis and inflammation are suppressed by infiltrating macrophages. In the context of spinal cord injury (SCI), preventing such communication increases microglial activation and worsens functional recovery. We suggest that macrophages entering the CNS provide a regulatory mechanism that controls acute and long-term microglia-mediated inflammation, which may drive damage in a variety of CNS conditions.
| The immune and the central nervous systems are now thought to be inextricably linked. In response to injury, the immune system shapes CNS recovery through a complex of molecular and cellular mediators. However, it is unclear how the kinetics, magnitude, and components of this response can be harnessed to improve CNS restoration. The two immune cells that dominate CNS lesions are resident microglia—already present before the injury—and infiltrating macrophages, which enter from the blood after injury. Both cells are thought to be critical to the outcome, yet it is unknown if, or how, they interact. To investigate this, we used mouse and human cells in microglia–macrophage coculture systems and an in vivo model of traumatic spinal cord injury. We show that infiltrating macrophages suppress key functions of microglia, such as removal of tissue debris and propagation of inflammation. Preventing macrophage–microglia communication increases microglial activation and worsens recovery. We suggest that infiltrating macrophages from the blood provide a natural control mechanism against detrimental acute and long-term microglial-mediated inflammation. Manipulation of the peripheral macrophages may provide a therapeutic treatment option to target microglial-mediated mechanisms that cause or exacerbate CNS injury and disease.
| The immune system plays a pivotal role in development and homeostatic functions of the central nervous system (CNS) [1]. Immune system dysfunction can give rise to CNS disease [2] and its response to injury shapes recovery [3–5]. The cellular response to CNS injuries is stereotyped and involves the rapid reaction of tissue-resident microglia [6, 7] and the recruitment of myeloid cells such as neutrophils and monocyte-derived macrophages (MDMs) within days [5], and the activation of lymphocytes from the blood, meninges, and choroid plexus [8]. Two cell types that dominate CNS lesions are resident microglia and infiltrating MDMs. It is known that microglia and MDMs are ontogenetically distinct [9, 10], express cell type–specific transcripts and proteins [11, 12], and can thus potentially perform different functions at the site of injury [13–16]. However, their relative contribution to the injury response, and subsequent recovery, remains unclear. It is not known if these two cell types interact to specifically modulate each other’s function.
Two of the primary functions of microglia and MDMs during CNS injury are phagocytosis and propagation of inflammation [17]. We have shown previously that after traumatic spinal cord injury (SCI), cessation of microglia phagocytosis coincides with the infiltration of MDMs [14]. In animal models of stroke and CNS autoimmune disease, expression profiling of microglia after injury or at the onset of disease (at the point of MDM infiltration) shows that pathways involving core functions of microglia, such as inflammation, RNA transcription, and phagocytosis, are significantly down-regulated [16, 18, 19]. We therefore hypothesized that MDMs entering the CNS signal to resident microglia and modulate their function.
Three to five days after CNS injury, infiltrating macrophages are distributed across lesion sites and are therefore potentially able to interact with microglia [14, 19, 20]. After SCI, resident microglia and MDMs both increase in number around the lesion site [21]. We observed that MDMs and microglia are often in close proximity to one another (S1A Fig); however, whether they can regulate each other’s functions is not known. To assess whether the cessation of microglial phagocytosis seen after SCI correlates with the entry of MDMs into the CNS, as reported previously [14], we used antibody-conjugated magnetic-bead sorting to isolate CD11b+ cells from spinal cord lesions in LysM-eGFP knock-in mice. LysM-eGFP reporter mice strongly express eGFP in myelomonocytic cells [22] but in less than 3% of microglia after SCI and other CNS injuries [19, 20, 23]. Therefore, isolated CD11b+ positive cells could be defined as resident microglia (CD11b+/eGFP−ve) or infiltrating MDMs (CD11b+/eGFP+ve/Ly6G−ve) after extraction from the injured spinal cord. To assess the effect of infiltrating cells on resident microglia, cells were isolated from spinal cord lesions prior to significant MDM infiltration (one day) and after infiltration (three days). Immediately after isolation, cells were placed in vitro and incubated with pHrodo-labeled myelin for four hours. Flow cytometric analysis revealed that increasing numbers of MDMs at the lesion site correlates with a significant reduction in microglial phagocytosis (Fig 1A–1E). We therefore hypothesized that MDMs entering the CNS signal to resident microglia and modulate their function. As an additional note, we detected significantly more MDMs one day after injury, as compared to naïve spinal cord (S1B Fig). This is likely to represent macrophages in the process of entering the tissue, either through the vasculature or the meninges but before they have been reported to be seen in the parenchyma. It is possible that they could release signaling molecules from these locations that influence microglia, as seen in the reduction of phagocytosis at day one after SCI, compared to uninjured (naïve) mice (Fig 1E). There is a further significant increase in infiltration of MDMs by day three, compared to one day after SCI (S1B Fig).
To directly test the hypothesis that MDMs modulate microglial function, we created a bilaminar culture system by plating bone marrow–derived macrophages (BMDMs) on coverslips with small paraffin spacers, which were then placed into wells containing adult microglia (S2D Fig). Primary adult mouse microglia were cultured under conditions that retain a transcriptional profile more similar to their in vivo counterparts, as compared to other media conditions, primary microglial cultures from neonates or microglial cell lines [11] (S2A and S2B Fig). Other features, such as genes that reflect region-specific factors or function, may be altered when cells are placed in culture. The addition of Transforming growth factor (TGF)-β was not only necessary for a gene expression profile more similar to freshly isolated microglia but also showed greater ramified morphology in culture at seven days (S2C Fig). Microglia “signature” genes were down-regulated during lipopolysaccharide (LPS)-induced inflammation, which supports their description as homeostatic [11]. There is no further modulation of these microglial genes in the presence of macrophages (S2E Fig). As a control for cell numbers, we assessed modulation of these microglial genes by coculturing with microglia instead of BMDMs (S2F Fig). Suppression of inflammatory genes in adult microglia does not occur when cocultured with adult microglia. For these experiments, adult mouse microglia were cultured with or without adult microglia and stimulated with LPS (100 ng/mL) (S2F Fig).
Using this bilaminar in vitro system, we assessed if soluble factors released by these two cell types affect phagocytic function in one another. Microglia and BMDMs were cocultured in the bilaminar system for 24 hours, separated, and incubated with pHrodo-labeled myelin for 90 minutes. Phagocytic uptake was assessed with flow cytometry. Uptake of pHrodo-labeled myelin was significantly decreased in microglia cultured in the presence of BMDMs compared with microglia cultured alone (Fig 1F and 1G). Surprisingly, myelin phagocytosis by BMDMs was significantly increased after coculture with adult microglia (Fig 1H and 1I). These findings reveal direct communication between the two cell types divergently affecting phagocytic function.
We next investigated macrophage effects on inflammatory gene expression in microglia and macrophages from adult mice and humans. We recently described a mathematical model of cytokine signaling, which found four inflammatory cytokines, interleukin (IL)-1β, tumor necrosis factor (TNF), IL-6, and IL-10, to be key nodes in the inflammatory network [24]. After LPS stimulation, coculture with BMDMs significantly down-regulated these genes in mouse and human microglia (Fig 1J and 1L). In contrast, IL-1β expression was increased in LPS-stimulated macrophages cocultured with mouse microglia (Fig 1K). There were nonsignificant trends towards increases in inflammatory cytokines in human macrophages in the combined presence of LPS and microglial cells (Fig 1M), contrasting with the significant suppression of these genes in microglia in the same conditions (Fig 1L). These experiments show the direct suppressive effects of macrophages on microglia, and reciprocal but divergent effects of microglia on macrophages.
In the bilaminar culture system in mouse and human microglia, we found significant suppression of pro-inflammatory cytokines IL-1β, TNF, and IL-6 in the presence of macrophages (Fig 1J and 1L). We also found a significant reduction in human microglia of IL-10, a canonical brake on inflammation [25, 26]. Despite IL-10’s opposing function to the pro-inflammatory cytokines, all four cytokines are increased with LPS 24 hours after stimulation. We have recently described the complexities of cytokine networks over time and shown how the modulation of one cytokine in the system may result in varying temporal kinetics of the others [26]. Here, we examined a single time point, but as all four cytokines were suppressed by the presence of macrophages, we sought to examine whether microglial transcription, in general, became globally suppressed by macrophages. To understand the global effects of macrophage suppression on microglia, we transcriptionally profiled LPS-activated adult mouse microglia in the presence or absence of macrophages. A total of 1,076 genes were significantly differentially regulated in activated microglia in presence of macrophages, with approximately 50% up-regulated and 50% down-regulated. Ingenuity pathway analysis (IPA) revealed that the most dysregulated canonical signaling pathways were those related to nuclear factor (NF)-κB signaling, a master regulator of inflammation (Fig 2A) and apoptosis and cell death (Fig 2B). Network analysis revealed three major clusters of genes distributed across two distinct regions reflecting distinct gene co-expression patterns (Fig 2C). In the major gene cluster 1 (185 genes), which comprised genes down-regulated within microglia in the presence of macrophages, we found that the top upstream regulators, predicted to be inhibited with high confidence, included MyD88, IL-1β, and TNF (Fig 2D). These analyses support our findings that gene expression within the major inflammatory cascades in microglia is suppressed in the presence of macrophages.
As IL-1β was one of the key cytokines to be differentially regulated in microglia and macrophages in the coculture experiments (Fig 1J and 1K), we searched for factors known to regulate IL-1β in microglia. Prostaglandin E2 (PGE2) signaling via the EP2 receptor has been reported to reduce IL-1β expression in microglia [27]. In addition, EP2 receptor signaling has also been reported to reduce phagocytosis [28–30]. We therefore hypothesized that PGE2 signaling via microglial EP2 receptors could be responsible for the suppressive effects of macrophages. Inducible microsomal prostaglandin E synthase-1 (mPGES) and EP2 receptor were up-regulated during inflammation in mouse and human microglia and macrophages in vitro (Fig 3A–3D and S3A Fig) and in vivo in mice after SCI (Fig 3E). EP2 receptor expression was also up-regulated 22-fold when assessed by transcriptional array (S3A Fig). Transcript levels for EP1 and 4 were down-regulated in mouse microglia in vitro, suggesting they are not involved (S3A Fig). mPGES and hydroxyprostaglandin dehydrogenase (HPGD) work in concert to regulate PGE2 production and release as HPGD converts PGE2 to its biologically inactive metabolite [31]. In human macrophages, mPGES was significantly increased and HPGD was significantly down-regulated during inflammation (LPS) when cocultured with microglia, compared with macrophages cultured alone (without LPS) (Fig 3B), suggesting greater PGE2 production in stimulated macrophages in the presence of microglia. The expression of EP2 receptor in human microglia was not significantly up-regulated upon stimulation with LPS (Fig 3D). However, the data show that it is expressed, allowing the cells to detect PGE2.
Taken together, these experiments show that the components needed to allow PGE2 signaling at the EP2 receptor in microglia are up-regulated during inflammation and may be utilized for macrophage–microglia communication.
To functionally assess the role of the EP2 receptor, we treated adult mouse microglia with the EP2 specific agonist, Butaprost. Treatment of microglia significantly reduced TNF, IL-6, and IL-10 in the same manner as the macrophage-mediated suppression of these genes (Fig 3F). This contrasted with the effects of Butaprost on BMDMs, which only reduced TNF expression (S3E Fig). Butaprost also significantly reduced phagocytosis by microglia (Fig 3G). Macrophage suppression of microglial phagocytosis was rescued by the selective EP2 antagonist, PF-04418948 [32] (Fig 3H). In addition, unlike wild-type (WT) BMDMs, macrophages that lack mPGES (mpges −/−) do not suppress microglial phagocytosis (Fig 3I). Taken together, these results show that PGE2 produced by peripherally derived macrophages plays a major role in suppression of microglial phagocytic function via EP2 receptors.
To investigate this mechanism in vivo, we performed SCI in WT and mpges −/− mice and assessed the phagocytic microglial response. In addition, as our in vitro data show that PGE2 derived from macrophages suppresses microglial phagocytosis, we performed SCI in C–C chemokine receptor type 2 (CCR2) null mice to compare the microglial response in a lesion that contains very few MDMs (see Fig 5A and 5B) [33]. Three days after SCI in WT, mpges −/−, and CCR2 null mice, microglia were isolated from the lesion and ex vivo phagocytosis of pHrodo (Green)-myelin was quantified by flow cytometry (Fig 4A and 4B). Microglia from WT mice after SCI (i.e., with macrophages and PGE2 present in the lesion), showed low levels of phagocytosis (Fig 4A and 4B). In contrast, microglia from mpges −/− and CCR2 null mice showed significantly increased phagocytosis (Fig 4A and 4B), indicating that in the absence of PGE2, or the absence of macrophages in the lesion, microglial phagocytic function is increased.
To assess the effect of blocking the EP2 receptor, in vivo, on microglial phagocytosis, we injected pHrodo-myelin into the brain (corpus callosum) of WT mice, together with vehicle or the EP2 receptor antagonist (PF-04418948). We assessed brain tissue three days after injection with immunofluorescence and confocal microscopy. In the corpus callosum of mice injected with pHrodo-myelin and vehicle, the area containing pHrodo-myelin is mainly populated with CD11b+, Tmem119-negative cells (MDMs) (Fig 4C). In mice injected with pHrodo-myelin and EP2 antagonist, there is a significant increase in Tmem119+ microglial cells in the area containing pHrodo-myelin (Fig 4D and 4E). Many of these Tmem119+ microglia contain or are closely associated with the fluorescently tagged myelin (Fig 4D and 4F). Importantly, the percentage of total microglial cells in the corpus callosum that are in contact with or contain pHrodo-myelin is significantly increased in EP2 antagonist–injected mice compared with controls (Fig 4C, 4D and 4F). These data indicate that blocking the EP2 receptor pathway in vivo promotes microglial phagocytic activity and increases recruitment of microglia to the site of injury. In summary, in vivo and in vitro studies suggest macrophage production of PGE2 acts at the EP2 receptor to mediate suppression of microglial phagocytosis.
We also sought to investigate macrophage effects on microglial cell death and proliferation, as microglia proliferate at the sites of CNS injury [14, 34]. In vitro, the presence of macrophages did not significantly reduce microglial viability (S4A Fig). However, when combined with inflammation (LPS stimulation), macrophages significantly reduced viability, suggesting an increase in microglial cell death, compared with untreated microglia (S4A Fig). This fits with our transcriptional profiling data, which show that apoptotic and cell death pathways are significantly dysregulated under similar conditions (Fig 2B). These data suggest that macrophages can affect microglial apoptosis under inflammatory conditions in vitro and warrant further analysis.
To investigate macrophage effects on microglial proliferation, we used Click-iT EdU assay (Invitrogen), which incorporates 5-ethynyl-2′-deoxyuridine (EdU; a nucleoside analog of thymidine) to DNA during active DNA synthesis. We found no evidence that macrophages affect the proliferation of microglia in vitro with bilaminar cultures (S4B and S4C Fig). We also depleted circulating macrophages prior to their infiltration after SCI using clodronate liposomes in LysM-eGFP mice to assess microglial proliferation with Ki67. Clodronate significantly depleted eGFP+ infiltrating cells at the lesion site when assessed five days after injury (S4D and S4E Fig), but this had no effect on microglial proliferation (S4F and S4G Fig).
To investigate whether macrophages modulated another important microglial function, namely rapid process extension towards microlesions, we induced laser lesions in organotypic hippocampal slice cultures (OHSCs), as done previously [35], in the presence or absence of BMDMs. We also investigated the initial reaction of microglial morphologies in an in vivo model of traumatic brain injury (TBI) [36] with local administration of Butaprost versus vehicle control. Initial process extension and acute morphological changes that are dependent on purinergic receptor signaling [6, 36] are not affected in OHSCs by the presence of macrophages (S4H–S4M Fig) or altered by the EP2 agonist in TBI (S4N–S4O Fig).
In summary, coupled with our transcriptional profiling data, these results highlight that macrophages target specific pathways and functions in microglia, such as inflammation and apoptosis, but do not affect microglial proliferation or their rapid process extension response to injury. These findings may be of significance, as the rapid reaction of microglia in the early phases of injury are thought to be protective [35, 37, 38]. Our results suggest that peripheral macrophages do not interfere with this response.
Mice that lack CCR2 cannot successfully recruit MDMs to traumatic CNS lesions [33]. We showed that this leads to increased microglial phagocytic activity (Fig 4A and 4B). Therefore, we next assessed whether lack of MDM infiltration affects activation of microglia and functional recovery in vivo after SCI using CCR2 null mice [33]. Although CCR2 has been reported to be expressed in injured neurons, expression appears variable between species and between investigators [39–41]. We are not aware of other genetic evidence for CCR2 protein expressed in neurons [13, 42]. CCR2 is also reported to be expressed in a subset of T-regulatory cells (Tregs) [43]. Although it is unknown what role CCR2+ Tregs play after SCI, it has been shown that T cells may play a beneficial role in CNS repair [44]. Despite this, a major phenotype five days after SCI was that MDMs were almost absent in the lesioned spinal cord of CCR2rfp/rfp (CCR2 KO) mice, as compared with WT mice (Fig 5A and 5B). Neutrophil infiltration was not affected five days after injury (Fig 5C). Three days after injury, there was a trend to an increase, but this was not statistically significant (S4P Fig). To study the impact of the absence of MDMs on microglial-mediated inflammation, we isolated microglia from SCI lesions four and seven days after injury and assessed the expression profiles of 86 inflammatory genes using a PCR array. Four days after SCI in WT mice, approximately half of these genes were significantly down-regulated in microglia. Seven of the top 20 most down-regulated genes in WT mice were significantly less down-regulated in CCR2 KO mice, indicating that the absence of MDMs at the lesion resulted in less suppression of microglial inflammation (Fig 5D). Moreover, pathways identified as being suppressed by macrophages in our in vitro bilaminar system, such as inflammation driven by MyD88 and NF-κB and apoptotic pathways driven by Trp53 and Bcl2 were also significantly more suppressed when MDMs were present at the lesion (Fig 5D). Seven days after SCI, microglial inflammatory genes continued to be dysregulated (Fig 5E). Components of important inflammatory pathways continued to be significantly less suppressed in the absence of macrophages, such as MyD88, Il17a, and Cxcl2. However, dysregulation of microglial gene expression was less unidirectional than factors associated with pro-inflammatory response, such as Irf1, Tlr9, and Il12b, and growth factors such as Egf and Tgfb1 associated with recovery were significantly more down-regulated in microglia, in the absence of infiltrating macrophages (Fig 5E).
Our previous work shows that initial perturbation to inflammatory networks is likely to cause unpredictable patterns of expression at later time points [24, 26]. Therefore, to investigate the long-term consequence of the initial loss of microglial suppression and subsequent dysregulation, we assessed long-term activation of microglia and its impact on functional recovery after SCI in CCR2 KO versus WT mice (Fig 6A–6G).
Up-regulation of CD11b (αM integrin) is well established as a readout of microglial/macrophage activation [45–47], and CD86 is a costimulatory receptor up-regulated during inflammation in microglia in vivo [48, 49]. CD11b expression in microglial cells was already increased at a cellular level in mice lacking macrophage infiltration (CCR2 KO) seven days after SCI versus controls (Fig 6A). There was a trend to an increase in CD86+ microglia but it did not reach statistical significance (S4Q Fig). To assess microglial activation 28 days after injury, we quantified CD11b and CD86 expression by immunofluorescence of tissue sections caudal to the lesion epicenter (Fig 6C and 6D). Importantly, 28 days after SCI, CD11b and CD86 immunoreactivity was greater in area and intensity in CCR2 KO mice despite the lack of infiltrating MDMs, which also express CD11b and CD86 (Fig 6C and 6D). In other words, CD11b and CD86 expression is markedly increased in microglia in CCR2 KO mice 28 days after SCI. These results indicate that preventing the communication between MDMs and resident microglia contribute to long-term microglial activation after CNS injury.
We also investigated whether increased microglial inflammation in the absence of infiltrating macrophages in CCR2 KO mice influences functional recovery and histopathology. CCR2 KO mice showed greater myelin loss, an indicator of secondary tissue damage, caudal to the lesion 28 days after SCI, compared with controls (Fig 6E and 6F). The increased microglial activation associated with the absence of macrophage influx after SCI is associated with worse locomotor recovery in CCR2 KO mice compared with WT controls, as measured by the Basso Mouse Scale (BMS) (Fig 6g).
The role of microglia in CNS injury and disease is now considered critical to the pathological process [50]. Our work suggests a novel concept that macrophages from the peripheral circulation, which enter the CNS after injury, may act to modulate microglial activation, thus preventing microglial-mediated acute and chronic inflammation.
These findings support previous work that shows blocking CCR2-dependent macrophage infiltration with an anti-CCR2 antibody worsens locomotor recovery after CNS injury [51, 52]. However, these earlier papers [51] did not show how macrophages mediate these effects. Our work now shows that infiltrating macrophages suppress microglial activation by reducing their expression of inflammatory molecules and ability to phagocytose, thus preventing chronic microglia-mediated inflammation in the CNS. Other work has suggested that subsets of infiltrating macrophages are detrimental to SCI [53], and it has been reported that CCR2 antagonism, producing a 50% reduction in infiltrating macrophages, is beneficial after TBI [54]. Also, CCR2 KO mice showed acute and transient behavioral improvement after intracerebral hemorrhage (one and three days), but this was not sustained at seven days [55]. Here, our finding that inhibition of macrophage entry to the CNS results in a worse outcome after SCI is supported by work that defines specific beneficial macrophage populations in multiple CNS injury and disease contexts [56–59]. Our results now suggest a new mechanism by which infiltrating macrophages mediate their beneficial actions via the regulation of microglial activation. Such a mechanism will operate alongside macrophage-intrinsic mechanisms. We observed that macrophages regulate microglia in both mouse and human cells. This is important as it represents an independent replication of the concept in a different laboratory. It shows that macrophages derived from the blood (human) or bone marrow (mouse) appear to have similar effects on microglia and that the findings may be relevant to human disease.
It is still controversial as to whether the net effect of microglia is beneficial or detrimental to CNS injury [4, 5, 60]; however, the kinetics of the microglial response must be considered. There is evidence that the initial responses of microglia, which occur in the first few minutes to several hours after injury, are beneficial and limit the expansion of CNS lesions [35, 37, 38]. Conversely, prolonged microglial dysregulation and neuroinflammation are deleterious to the CNS [61]. Therefore, the initial microglial response to injury may be beneficial, but prolonged inflammation and activation are potentially detrimental to recovery. Our data suggest that macrophages play a role in mitigating this detrimental response by infiltrating the injury site and reducing microglial-mediated inflammation and chronic microglial activation. The absence of this protective mechanism may contribute to a worse outcome when infiltrating macrophages do not enter the CNS after SCI in CCR2 KO mice. To our knowledge, this is the first description of such a cellular mechanism to reduce deleterious consequences of CNS injury.
Microglial cells are now prime targets in drug discovery for CNS injuries and neurodegenerative diseases [62]. To properly assess the roles of microglia in CNS injury, our data suggest that the context, timing, and interaction with macrophages should also be considered. Attempts to target either of these two cell populations should be approached with caution and a better understanding is needed of their divergent and complex roles in injury and disease. The heterogeneity and region-specific differences in microglia [63] and macrophage populations [64] will also need to be considered.
Recent work has shown that peripherally derived macrophages can engraft the brain and maintain an identity distinct from microglia [65], thus opening the possibility for therapeutic engraftment of MDMs to the CNS and allowing macrophage–microglia cross talk in disease contexts. Cell-to-cell interactions between different brain resident cell types are now becoming evident [66–69]. During inflammation, microglia have also been shown to drive astrocyte-mediated toxicity [66], which, subject to the context, is dependent on microglial NF-κB signaling [67]. Our data show that peripheral macrophages regulate the NF-κB signaling pathway in microglia that, in turn, reduce inflammatory mediators, such as TNF, which can drive astrocyte-mediated toxicity [66, 67]. This raises the possibility that macrophage signaling to microglia may have subsequent effects in other CNS cells, such as astrocytes.
In summary, we suggest that infiltrating macrophages provide a natural control mechanism against detrimental acute and long-term microglial-mediated inflammation. Manipulation of peripherally derived infiltrating cells may provide a therapeutic treatment option to target microglial-mediated mechanisms that cause or exacerbate CNS injury and disease.
All animal procedures were approved by the Animal Care Committee of the Research Institute of the McGill University Health Centre and followed the guidelines of the Canadian Council on Animal Care and the ARRIVE guidelines for reporting animal research [70]. Before surgical interventions and cardiac perfusions, mice were deeply anesthetized by intraperitoneal injection of ketamine (50 mg/kg), xylazine (5 mg/kg), and acepromazine (1 mg/kg). Human brain tissue was collected during clinical practice, fully anonymized, and therefore available for use under the legislation of the Tri-Council Policy Statement two and Plan d'action ministériel en éthique de la reserche et en intégrité scientifique of Quebec and Canada. This study was carried out in accordance with the guidelines set by the Biomedical Ethics Unit of McGill University, approved under reference ANTJ2001/1, and conducted in accordance with the Helsinki Declaration.
C57BL/6 (Charles River, St-Constant, QC), heterozygote lysM+/EGFP mice (kindly provided by Dr. Thomas Graf and obtained from Dr. Steve Lacroix); homozygote CCR2RFP/RFP and their C57BL/6J controls (Jackson); heterozygote Cx3CR1+/gfp (Jackson) and Ptges−/− mice [71] (obtained from Dr. Maziar Divangahi, McGill University), aged 8–14 weeks, were kept under a 12-hour light/dark cycle with ad libitum access to food and water. The LysM-eGFP mouse was originally generated by Faust and colleagues, 2000. EGFP is expressed specifically in the myelomonocytic lineage by using homologous recombination. This was achieved by knocking the enhanced GFP (EGFP) gene into the murine lysozyme M (lys) locus and using a targeting vector, which contains a neomycin resistant (neo) gene flanked by LoxP sites and “splinked” ends, to increase the frequency of homologous recombination. Removal of the neo gene through breeding of the mice with the Cre-deleter strain led to an increased fluorescence intensity [22].
Female mice were anesthetized by intraperitoneal injection of ketamine (50 mg/kg), xylazine (5 mg/kg), and acepromazine (1 mg/kg) and a moderate contusion injury (50 kDa force; 500–600-μm tissue displacement) was made at the T11 thoracic vertebral level using the Infinite Horizon Impactor device (Precision Scientific Instrumentation, Lexington, KY), as previously described [72].
Male C57BL/6 mice (8–12 weeks) were, anesthetized, transcardially perfused and brains removed and kept in ice-cold Hanks Balanced Salt Solution (HBSS). Cerebellum and meninges were removed, and brain was cut into small pieces. Tissue was enzymatically dissociated using Neural Tissue Dissociation Kit (P) (Miltenyi cat # 130-092-628) according to the manufacturer’s instructions, with modifications. Following digestion, tissue was transferred to a 15-mL Dounce on ice and homogenized with 20× passes of a large clearance pestle. Tissue was resuspended in 35% isotonic percoll and overlaid with HBSS. Following centrifugation (400g; 45 minutes), myelin was removed, and pure populations of microglial cells were selected using CD11b microbeads (Miltenyi #130-093-634), as previously described [63, 73]. Pure (>95% CD11b-positive) adult microglia were resuspended at 8×105 cell−mL (approximately two brains per mL) in media (DMEM F12, 10% fetal bovine serum [FBS], 1% penicillin/streptomycin [P/S]), with 10% L-cell conditioned media, a source of macrophage colony-stimulating factor (M-CSF), or 10 ng/mL recombinant mouse M-CSF (R and D cat no 416-ML-010/CF), and 50 ng/mL recombinant human TGF-β1 (Miltenyi cat no: 130-095-067) to maintain their transcriptional profile, as previously described [11]. Cells were plated in pre-coated poly-L-lysine plates, media was changed at three days, and experiments were performed at seven days. At seven days, microglia were collected and microglial “signature” genes assessed by qPCR. Network analysis and Markov clustering (see below) were performed to assess conditions driving cells to a similar phenotype of their freshly isolated counterparts.
CD11b+ cells (myeloid cells) were collected by magnetic bead cell sorting, as above, from SCI lesions of lys-EGFP-ki mice (2.5 mm either side of the epicenter) at 1 or 3 days after injury, or from uninjured controls. The CD11b+ fraction was immediately plated into 96 well plates (pre-coated poly-L-lysine, one animal per well) in DMEM F12 media containing 10% FBS. Cells were incubated for four hours with pHrodo (Invitrogen)-labeled myelin and taken for FACS analysis.
BMDMs were generated as previously described [74] from adult male C57/BL6− or Ptges−/− mice. Mice were euthanized, and their femurs were removed. Bone marrow was flushed out and homogenized, and RBCs were hypotonically lysed. After washing, cells were cultured in RPMI media containing 10% FBS, 10% L-cell-conditioned media, and 1% P/S for seven days.
Adult microglia and BMDMs (from the same animal when possible) were cultured separately for seven days, as described above; BMDMs were then replated on poly-L-lysine–coated glass coverslips. Cells were plated at 2×105 per 25-mm coverslip (for insertion in 6 well plates) of 4×104 per 12-mm coverslip (for insertion in 24 well plates). These numbers were calculated to match the number of microglia in the same well as, three days after SCI in vivo, the proportion of microglia to macrophages around the lesion (54% ± 8% versus 46% ± 8%). Therefore, equal numbers of microglia and BMDMs were plated into wells in vitro. Coverslips were pre-mounted with 3 small paraffin droplets on the same surface as the cells were plated, as previously described for neuron-astrocyte cocultures [75]. BMDMs were allowed to adhere overnight. Bilaminar culture experiments began by inserting coverslips containing BMDM and wax paraffin nodule face down into culture plates containing adult microglia. All experiments were performed in DMEMF12 serum-free media.
Adult microglia and BMDMs were cocultured in the bilaminar system, as above, for 24 hours. At 24 hours, BMDM coverslips were removed from the microglial wells and placed in new cell culture wells. Following separation, both adult microglia and BMDMs were incubated with pHrodo Red (ThermoFisher, Mississauga, ON) labeled myelin for 90 minutes. pHrodo Red is weakly fluorescent at neutral pH but increasingly fluorescent as the pH drops, therefore an increasing signal can be detected because of pHrodo-myelin entering the lysosome. After treatment, cells were trypsinized, harvested, and labeled for FACS analysis.
Adult microglia and BMDMs were cocultured in the bilaminar system, as above. Wells containing microglia and/or BMDMs were treated with vehicle or LPS (100 ng/mL) for four hours. Cells were then washed and fresh, serum-free media was added to each well for a further 20 hours. At 24 hours, supernatants were harvested and cells lysed with 350 μL of RLT lysis buffer (Qaigen, Germantown, MD) and snap frozen until RNA extraction. For the in vitro proliferation assay, microglia rather than BMDMs were plated on coverslips. Cells were treated as above, with the addition of Click-iT EdU (20 μM) at the time of LPS administration. EdU is a nucleoside analog of thymidine that is incorporated into DNA during active DNA synthesis and was used according to the manufacturer’s instructions (Invitrogen, Carlsbad, CA).
To obtain MDMs, monocytes were isolated from healthy human venous blood. PBMCs were isolated from whole blood using Ficoll–Paque density gradient centrifugation (GE Healthcare, Piscataway, NJ). CD14+ monocyte isolation was done using immune-magnetic bead selection according to the manufacturer’s instructions to achieve 95%–99% purity (Miltenyi Biotec), as determined by flow cytometry. Cells were cultured in RPMI (Invitrogen, Carlsbad, CA) supplemented with 10% FBS, 0.1% P/S, and 0.1% L-glutamine. Cells were plated at a density of 5×105 cells/mL in 6-well tissue culture plates for four days and matured in vitro to become MDMs, using supplementation with recombinant M-CSF (25 ng/mL, PeproTech, Rocky Hill, NJ).
Human microglia were isolated from adult brain tissue using previously described protocols [76]. Adult microglia were derived from surgical resections of brain tissue from pharmacologically intractable nonmalignant cases of temporal lobe epilepsy. The tissue provided was outside of the suspected focal site of epilepsy-related pathology. Briefly, tissue was obtained in pieces <1 mm3 and treated with DNase (Roche, Nutley, NJ) and trypsin (Invitrogen, Carlsbad, CA) for 30 min at 37 °C. Following dissociation through a nylon mesh (37 μm), the cell suspension was separated on a 30% Percoll gradient (GE Healthcare, Piscataway, NJ) at 31,000g for 30 minutes. Glial cells (oligodendrocytes and microglia) were collected from underneath the myelin layer, washed, and plated at a density of 2×106 cells/mL in tissue culture flasks. After 24 hours in culture, microglia were separated by the differential adhesion properties of the cells. Microglia were grown for four days in flasks before gently collecting using 2 mM EDTA (Sigma-Aldrich, Oakville, ON); cells were then plated in minimum essential medium (MEM, Sigma-Aldrich, Oakville, ON) supplemented with 5% FBS, 0.1% P/S, and 0.1% L-glutamine at a density of 5×105 cells/mL on a glass 25-mm cell culture insert in a 6-well plate.
Following addition of microglia to the MDM-containing well, as described above in the bilaminar coculture system, cells were exposed to 100 ng/mL lipopolysaccharide (serotype 0127:B8, Sigma-Aldrich) for 24 hours before collection in Trizol Reagent (Invitrogen, Carlsbad, CA) for subsequent total RNA isolation using the Qiagen RNeasy mini kit following the manufacturer's instructions. RNA isolated was treated immediately with DNase (Qiagen, Germantown, MD). Reverse transcription and cDNA generation were performed using random hexaprimers (Roche) and the Moloney murine leukemia virus-RT enzyme (Invitrogen, Carlsbad, CA) at 42 °C. PCR reaction cycling was performed according to the ABI PRISM 7000 Sequence Detection System default temperature settings (two minutes at 50 °C, 10 minutes at 95 °C, followed by 40 cycles of 15 seconds at 95 °C, one minute at 60 °C).
TaqMan quantitative real-time PCR was used to measure mRNA expression levels for all mRNAs. Relative gene expression data were calculated according to the 2−ΔΔCt method.
For phagocytosis assay after treatment with EP2 receptor agonist, adult microglia were treated for one hour with Butaprost (1 μM) (Cayman Chemicals, Ann Arbor, MI) prior to incubation with pHrodo-myelin. For treatment with EP2 antagonist, bilaminar cocultures of adult microglia and BMDMs were treated with PF-04418948, (10 μM) (Sigma-Aldrich, Oakville, ON) prior to incubation with pHrodo-myelin. For treatment of adult microglia or BMDMs with Butaprost during inflammation, wells containing microglia or BMDMs were treated with vehicle or LPS (100 ng/mL) and Butaprost (1 μM) for four hours. Cells were then washed and fresh, serum-free media containing Butaprost (1 μM) was added to each well for a further 20 hours. At 24 hours, supernatants were harvested and cells lysed with 350 μL of RLT lysis buffer (Qaigen, Germantown, MD) and snap frozen until RNA extraction.
Injections were made into the right motor cortex of eight-week-old female C57/BL6J mice. A 2 × 2 mm opening was made in the skull just to the right side of the midline and just below bregma. A 26G needle attached to a 10-μL World Precision Instruments NanoFil syringe was used. The needle was inserted into the cortex just deep enough for the entire bevel of the needle to be inside the brain—a depth of approximately 0.46 mm. The EP2 receptor antagonist (PF-04418948) (Sigma-Aldrich, Oakville, ON) was initially dissolved in 100% DMSO to a 50-mM stock solution. This solution was diluted in PBS, pH 8.0, for a final injection concentration of 1 uM. The vehicle controls were a corresponding dilution of DMSO in PBS, pH 8.0. The total volume injected was 2 μL. Each injection contained 20 μg of pHrodo-labeled myelin. The pHrodo-myelin was prepared as follows: 18.5 μL myelin (15 mg/mL) + 25 μL of pHrodo was suspended in 206.5 μL of PBS, pH 8.0, and incubated 45 minutes at RT in the dark on the rocker. Samples were then spun down at 4000g for 10 minutes and resuspended in 25 μL of PBS plus either vehicle or antagonist. Three days after injection, mice were humanely killed and transcardially perfused, as described below.
Cultures were prepared as previously described [77]. Briefly, 300-μm hippocampal slices were prepared from P6-7 CX3CR1gfp/+ mice and cultured on semiporous nylon membranes (Millipore, Bedford, MA) for 9–11 days, with medium changes every two days. The culture medium contained 6.5 mg/mL glucose, 25% HBSS, 25% horse serum, 50% minimum essential medium with Glutamax, and 0.5% P/S (Gibco, Mississauga, ON). Five experimental conditions were assessed: (1) control slices in serum-free slice culture medium, (2) OHSCs incubated with a feeder layer of BMDMs for 24–36 hours, (3) OHSCs incubated with a feeder layer of BMDMs pre-stimulated with LPS (100 ng/mL, two hours) for 24–36 hours, (4) OHSCs incubated with BMDMs added directly onto the slice (5×105/slice), for 24–36 hours, and (5) OHSCs incubated with a LPS pre-stimulated (100 ng/mL, two hours) BMDMs added directly onto the slice (5×105/slice) for 24–36 hours prior to laser lesioning.
Laser lesions and imaging of microglial responses were performed with a customized Olympus FV1200MPE equipped with a MaiTai HP DeepSee-OL IR laser and a 25× XLPlan N objective. Slice cultures were perfused at 1–2 mL/minute with room temperature artificial cerebral spinal fluid containing (in mM) NaCl (126), NaHCO3 (26), KCl (2.5), NaH2PO4 (1.25), glucose (10), MgCl2 (2), and CaCl2 (2), corrected for osmolarity with sucrose. Laser lesions and imaging were performed sequentially at 820 nm, which provided excitation of both GFP and lipofuscin [35]. Lesions were performed in a 3-μm circular ROI at a depth of 40–70 μm in the central plane of the imaged volume, using 2–7 scans at 60% laser power, with a pixel dwell time of 10 μs. Post-lesion images were 30-μm z-stacks with a 1.5-μm step size, once per minute for 15 minutes. Analysis was performed with MTrackJ for ImageJ [78].
For mTBI experiments, CX3CR1gfp/+ mice were anesthetized with ketamine, xylazine, and acepromazine and maintained at a core temperature of 37 ºC. Hair was removed from the head using hair clippers and Nair. An incision was made in the scalp to expose the skull and a metal bracket was secured on the skull bone over the barrel cortex. The bone was quickly thinned to a thickness of about 20–30 μm. Once thinned, the blunt end of a microsurgical blade was used to compress the skull bone into a concavity without cracking the skull. Artificial spinal fluid (ACSF) or 1 μM Butaprost (Cayman Chemicals, Ann Arbor, MI) in ACSF was immediately applied to the skull after mTBI and kept on throughout the imaging experiment. Mice that had mTBI procedures were imaged using a Leica SP8 two-photon microscope equipped with a 12,000-Hz resonant scanner, a 25× color corrected water-dipping objective (1.0 NA), a quad HyD external detector array, and a Mai Tai HP DeepSee Laser (Spectra-Physics, Santa Clara, CA) tuned to 905 nm. Three-dimensional time-lapse movies were captured in z-stacks of 15–30 planes (3-μm step size) at 1–2-minute intervals. Signal contrast was enhanced by averaging 10–12 video frames per plane in resonance scanning mode. Three-dimensional time-lapse movies were imported for analysis into Imaris software (Bitplane). “Jellyfish” microglia were quantified by counting cells that had a process that was 20 μm or more in diameter at any time within about two hours of imaging immediately following mTBI. This number was then divided by the area of reactive microglia for each mouse.
Clodronate liposomes used to deplete the population of infiltrating macrophages were purchased through the website, www.clodronateliposomes.com, and were prepared by Nico Van Rooijen in the Netherlands. Clodronate liposomes were injected on the day of spinal cord contusion and at days 1, 2, 3, and 4 after the injury. For each treatment, 100 μL/10 g body weight was administered intravenously (IV) and 50 μL/10 g intraperitoneally (IP). Saline injections were administered to control animals at the equivalent time.
For analysis of cells from ex vivo experiments, cells were harvested with trypsin and stained with eflouro eFluor780 viability dye (1:1,000; eBioscience, Mississauga, ON), blocked with FC-receptor blocked (1:200; BD Bioscience), and stained with CD11b-V450, Ly6G-PerCP-Cy5.5 (all 1:200; BD Bioscience, Mississauga, ON). LysM-GFP and pHrodo Red labeled myelin were detected in the FITC and PE channels, respectively. CD11b+/Ly6G−/LysM-GFP− (microglia) CD11b+/Ly6G−/LysM-GFP+ (macrophages) were assessed for pHrodo-labeled myelin; the content was detected in the PE channel. For analysis of cells from in vitro experiments, adult microglia or BMDMs were harvested with trypsin and stained with eflouro eFluor780 viability dye (1:1,000; eBioscience, Mississauga, ON), blocked with FC-receptor blocked (1:200; BD Bioscience), and stained with CD11b-V450 (1:200; BD Bioscience, Mississauga, ON). pHrodo-labeled myelin was detected in the PE channel.
For analysis of cells from in vivo experiments, spinal cord tissue was harvested and transferred to a 15-mL Dounce homogenizer on ice and homogenized with 20× passes of a large clearance pestle. Tissue was resuspended in 70% and overlaid 30% isotonic Percoll. Following centrifugation (900g; 25 minutes), cells at the Percoll interface were collected, washed, and resuspended. Cells were blocked with FC-receptor blocked (1:200; BD Bioscience, Mississauga, ON) and stained with CD45-PE-Cy7, CD11b-V450, Ly6G-APC, or FITC, Ly6C-APC-Cy7 (all 1:200; BD Bioscience, Mississauga, ON). CCR2-RFP was detected in the PE channel. Cells were acquired using a BD FACS Canto II and analyzed using FlowJo software. For fluorescently activated cell sorting (FACS), a BD FACSAria Fusion (BD Bioscience, Mississauga, ON) was used to isolate CD11b+ve/CD45lo/Ly6G−ve/Ly6C−ve microglial cells, which were directly sorted in Trizol for subsequent RNA extraction. All gating strategies not contained within the figures are presented in the Supporting information (S5 Fig).
Total RNA from FACS-sorted microglia was extracted from WT and CCR2-KO spinal cord injured mice, four and seven days after injury, or genotype-matched uninjured controls, as described previously [79]. Briefly, CD45hi/CD11b+ve/Ly6G−ve/Ly6C−ve microglial cells were recovered in 700 μL of Trizol, vortexed, centrifuged, snap frozen in dry ice, and kept at −80 °C until further use. On the day of extraction, samples were thawed and 140 μL of chloroform (Fisher) was added, vortexed for 15 seconds, and centrifuged at 12,000g at 4 °C for 15 minutes. The clear aqueous layer was recovered and 1 μL of Glycoblue (20 mg mL–1, Ambion; AM9515) was added, followed by 1:1 volume of isopropanol (Thermofisher Mississauga, ON) (approximately 400 uL). Samples were kept overnight at −20 °C to precipitate the RNA. The next day, the samples were centrifuged at maximum speed for 20 minutes at 4 °C. At this point, a blue pellet of RNA is visible. The pellet was washed twice with 75% RNAse-free ethanol; it was let dry at room temperature and resuspended in 14 μL of RNAse-free H2O. To assess inflammatory and immune-related genes, an RT2 Profiler PCR array was used (Qiagen; PAMM-181Z), following instructions provided by the manufacturer. This array screened for 84 genes, and the data obtained were analyzed with the online Qiagen analysis software (RT2 profiler PCR array data analysis V3.5), http://www.qiagen.com/shop/genes-and-pathways/data-analysis-center-overview-page/.
We performed quantitative real-time polymerase chain reactions (RT-qPCRs) to assay for the expression levels of multiple transcripts. RNA was extracted from cultured adult microglial BMDMs using the RNeasy Mini Kit (Qiagen, Germantown, MD). For spinal cord tissue, samples were homogenized, and total RNA was extracted using the RNeasy Lipid Tissue Kit (Qiagen, Germantown, MD). Reverse transcription was performed with the Quantinova Reverse Transcription Kit (Qiagen, Germantown, MD), and qPCR was performed using 1 μL of cDNA with Fast SYBR Green Master Mix (Applied Biosystems, CA) on a Step-One Plus qPCR machine (Applied Biosystems). Peptidylprolyl isomerase A (PPIA) was used as an internal control gene. The 2−ΔΔCt method was used to calculate the cDNA expression fold change following standardization relative to PPIA [24, 80]. All primers had a Tm of 60°C. Primer sequences were as follows:
Mouse
p2ry12 forward: 5′ CTG GGA CAA ACA AGA AGA AAG G 3′
p2ry12 reverse: 5′ CCT TGG AGC AGT CTG GAT ATT 3′
mertk forward: 5′ CCT CCA CAC CTT CCT GTT ATA TT 3′
mertk reverse: 5′ TGT TGC TCA GAT ACT CCA TTC C 3′
itgb5 forward: 5′ GGA TCA GCC AGA AGA CCT TAA T 3′
itgb5 reverse: 5′ AAT CTT CAG ACC CTC ACA CTT C 3′
gas6 forward: 5′ AGG AGA CAG TCA AGG CAA AC 3′
gas6 reverse: 5′ TTG AGC CTG TAG GTA GCA AAT C 3′
fcrls forward: 5′ AAT CAC ATT CTC CTG GCA TAG G 3′
fcrls reverse: 5′ GCA TGG CTT TCC CTG ATA GT 3′
c1q forward: 5′ GAA AGG CAA TCC AGG CAA TAT C 3′
c1q reverse: 5′ GGT GAG GAC CTT GTC AAA GAT 3′
Tnf forward: 5′ TTG CTC TGT GAA GGG AAT GG 3′
Tnf reverse: 5′ GGC TCT GAG GAG TAG ACA ATA AAG 3′
Il6 forward: 5′ CTT CCA TCC AGT TGC CTT CT 3′
Il6 reverse: 5′ CTC CGA CTT GTG AAG TGG TAT AG 3′
Il1b forward: 5′ ATG GGC AAC CAC TTA CCT ATT T 3′
Il1b reverse: 5′ GTT CTA GAG AGT GCT GCC TAA TG 3′
Il10 forward: 5′ ACA GCC GGG AAG ACA ATA AC 3′
Il10 reverse: 5′ CAG CTG GTC CTT TGT TTG AAA G 3′
Ptges forward: 5′ CCA CAC TCC CTC TTA ACC ATA AA 3′
Ptges reverse: 5′ GCC AGA ATT GTA GGT AGG TCT G 3′
EP1 forward: 5′ CTC TCG ACG ATT CCG AAA GAC 3′
EP1 reverse: 5′ GTG GCT GAA GTG ATG GAT GA 3′
EP2 forward: 5′ GCC TTT CAC AAT CTT TGC CTA CAT-3′
EP2 reverse: 5′ GAC CGG TGG CCT AAG TAT GG-3′
EP4 forward: 5′ CAA GCA TGT CCT GTT GCT TAA C 3′
EP4 reverse: 5′ GTC GGT TCA GCT ACG CTT TA 3′
Total RNA was quantified using a NanoDrop Spectrophotometer ND-1000 (NanoDrop Technologies) and its integrity was assessed using a 2100 Bioanalyzer (Agilent Technologies). Sense-strand cDNA was synthesized from 100 ng of total RNA, and fragmentation and labeling were performed to produce ssDNA with the Affymetrix GeneChip WT Terminal Labeling Kit according to the manufacturer’s instructions (ThermoFisher-Affymetrix, Mississauga, ON). After fragmentation and labeling, 3.5 μg DNA target was hybridized on Mouse Clariom S Assay (ThermoFisher-Affymetrix) and incubated at 450 °C in the Genechip Hybridization oven 640 (ThermoFisher-Affymetrix) for 17 hours at 60 rpm. Clariom S were then washed in a GeneChips Fluidics Station 450 (ThermoFisher-Affymetrix) using Affymetrix Hybridization Wash and Stain kit according to the manufacturer’s instructions (ThermoFisher-Affymetrix). The microarrays were finally scanned on a GeneChip scanner 3000 (Affymetrix).
Microarray data sets were normalized by the robust multiarray averaging (RMA) method in Affymetrix Expression Console (Affymetrix). To assess whether there were transcripts differentially expressed between activated microglia and activated microglia in the presence of macrophages, normalized data sets were compared in Affymetrix Transcriptome Analysis Console (TAC) Software. To assess whether there were transcripts differentially expressed between LPS-stimulated microglia and LPS-stimulated microglia in the presence of macrophages, normalized data sets were compared by ANOVA with a p < 0.05 cutoff. Differentially regulated gene lists were then assessed with the IPA software tool (Qiagen). To investigate gene co-expression relationships between groups, a pairwise transcript-to-transcript matrix was calculated in Miru from the set of differentially expressed transcripts using a Pearson correlation threshold r = 0.86. A network graph was generated in which nodes represent individual probe sets (transcripts/genes), and edges between them the correlation of expression pattern, with Pearson correlation coefficients above the selected threshold. The graph was clustered into discrete groups of transcripts sharing similar expression profiles using the MCL algorithm (inflation, 2.2; minimum cluster size, 10 nodes).
After SPI in mice, locomotor recovery was evaluated in an open field test using the 9-point BMS [81]. The BMS analysis of hind limb movements and coordination was performed by two individuals who were trained in the Basso laboratory, and the consensus score was taken. The final score is presented as mean ± SEM. The 11-point BMS subscore was also assessed.
Animals were perfused with 4% paraformaldehyde in 0.1 M PBS, pH 7.4, at three, five and 28 days. Spinal cord segments containing the lesion site, or whole brain from pHrodo-myelin injected mice, were removed and processed for cryostat sectioning (20-μm-thick cross sections). Immunofluorescence was performed using rat anti-CD11b (1:250; Serotec, Raleigh, NC), rabbit anti Iba1 (1:500; Wako, Richmond, VA), anti-rat CD86 (1:200; BD Bioscience, Mississauga, ON), rabbit anti-P2ry12 (1:500; kindly provided by Dr. Oleg Butovsky), rabbit anti-Tmem119 (1:1 hybridoma supernatant; kindly provided by Dr. Mariko Bennett, Ben Barres Lab), rabbit anti-Ki67 (1:500; Abcam, Cambridge, MA), and chicken anti-GFP (1:500; Abcam, Cambridge, MA) and detected using the appropriate secondary antibodies at 1:500; anti-rabbit Alexa Fluor 568 or 647, anti-rat Alexa Fluor 568 or 647, and anti-chicken Alexa Fluor 488 (Invitrogen, Carlsbad, CA). All images were visualized using a confocal laser scanning microscope (FluoView FV1000; Olympus) using FV10-ASW 3.0 software (Olympus) and prepared with ImageJ. Histochemical staining with Luxol fast blue was used to assess myelin loss 28 days after injury. Myelin was quantified as a measure of Luxol fast blue (mean gray value, ImageJ) across the whole cross section, measured at 200-μm intervals over the 2-mm length of the cord.
In spinal cord sections 28 days after injury, single plane images at a depth of 8 μm were acquired at two sites either side of the midline and included white and gray matter, as depicted in Fig 4. Images were taken 600 μm caudal to the lesion epicenter to avoid overt tissue cavities but include putative areas of MDM and microglia cell interactions. Images were acquired and analyses performed whilst blinded to genotype. Area of fluorescence and fluorescence intensity (as measured by Integrated Density [IntDen], which is the product of area and mean gray value) were quantified with ImageJ. For quantification of pHrodo-myelin injected mice, images were acquired from three consecutive slices around the injection site and analyses performed whilst blinded to treatment. The corpus callosum provided a boundary at which microglia cells and their association with pHrodo were quantified.
All relevant data are within the paper and its Supporting information files (S1 Data). Microarray data are available at Gene Expression Omnibus (accession number GSE102482).
Data were analyzed using one-way and two-way ANOVAs with Bonferroni correction or Student t tests when appropriate and as indicated. Data were checked for compliance with statistical assumptions for each test, including normal distribution and equal variances across groups. Tests were two-tailed throughout. Statistical significance was considered at p < 0.05. Data show mean ± SEM.
|
10.1371/journal.pntd.0002061 | Iron Overload Favors the Elimination of Leishmania infantum from Mouse Tissues through Interaction with Reactive Oxygen and Nitrogen Species | Iron plays a central role in host-parasite interactions, since both intervenients need iron for survival and growth, but are sensitive to iron-mediated toxicity. The host's iron overload is often associated with susceptibility to infection. However, it has been previously reported that iron overload prevented the growth of Leishmania major, an agent of cutaneous leishmaniasis, in BALB/c mice. In order to further clarify the impact of iron modulation on the growth of Leishmania in vivo, we studied the effects of iron supplementation or deprivation on the growth of L. infantum, the causative agent of Mediterranean visceral leishmaniasis, in the mouse model. We found that dietary iron deficiency did not affect the protozoan growth, whereas iron overload decreased its replication in the liver and spleen of a susceptible mouse strain. The fact that the iron-induced inhibitory effect could not be seen in mice deficient in NADPH dependent oxidase or nitric oxide synthase 2 suggests that iron eliminates L. infantum in vivo through the interaction with reactive oxygen and nitrogen species. Iron overload did not significantly alter the mouse adaptive immune response against L. infantum. Furthermore, the inhibitory action of iron towards L. infantum was also observed, in a dose dependent manner, in axenic cultures of promastigotes and amastigotes. Importantly, high iron concentrations were needed to achieve such effects. In conclusion, externally added iron synergizes with the host's oxidative mechanisms of defense in eliminating L. infantum from mouse tissues. Additionally, the direct toxicity of iron against Leishmania suggests a potential use of this metal as a therapeutic tool or the further exploration of iron anti-parasitic mechanisms for the design of new drugs.
| Leishmania are important vector-borne protozoan pathogens that cause different forms of disease, ranging from cutaneous self-healing lesions to life-threatening visceral infection. L. infantum is the most common species causing visceral leishmaniasis in Europe and the Mediterranean basin. Iron plays a critical role in host-pathogen interactions. Both the microorganism and its host need iron for growth. However, iron may promote the formation of toxic reactive oxygen species, which contribute to pathogen elimination, but also to host tissue pathology. We investigated the effect of manipulating host iron status on the outcome of L. infantum infection, using the mouse as an experimental model. We found that dietary iron deprivation had no effect on L. infantum growth, and iron-dextran injection decreased the multiplication of L. infantum in mouse organs. The fact that this anti-parasitic effect of iron was not observed in mice genetically deficient in superoxide and nitric oxide synthesis pathways indicates that iron is likely to act in synergy with reactive oxygen and nitrogen species produced by the host's macrophages. This work clearly shows that iron supplementation improves the host's capacity to eliminate L. infantum parasites and suggests that iron may be further explored as a therapeutic tool to fight this type of infection.
| Leishmania are trypanosomatid protozoans that alternate between two forms: the extracellular motile promastigote in the gut of phlebotomine insects and the intracellular non-motile amastigote inside the macrophages of mammalian hosts. These parasites cause leishmaniasis, a spectrum of human diseases that range from self-healing cutaneous ulcers to fatal visceralizing infection. Every year, approximately 2.0 million people develop symptomatic disease (0.5 million of them the visceral form) [1]. In Europe, visceral leishmaniasis is caused almost exclusively by L. infantum, which is transmitted as a zoonosis. The domestic dog is one of the main reservoirs of this parasite and canine leishmaniasis is an important veterinary problem in European Mediterranean countries [2].
There are currently no effective vaccines to prevent human leishmaniasis [3]. Therefore, management of the disease relies on chemotherapy. However, available drugs are highly toxic and the frequency of resistant parasite strains is increasing worldwide [4], [5]. The improvement of our knowledge on the mechanisms of host resistance to Leishmania is important to contribute to the development of new therapeutic strategies.
The important role played by iron metabolism in the interaction between host and pathogens is being increasingly highlighted by recent research [6], [7]. Both the host and the pathogens absolutely need iron for survival and must have efficient mechanisms for its acquisition together with adequate mechanisms of cell defense to avoid iron toxicity. Data obtained in human patients, as well as in different animal infection models indicate that in many cases iron availability favors the multiplication of pathogens, whereas iron deprivation impairs their growth [8]. Interestingly, the vertebrates innate immune response to infection includes several mechanisms of iron with-holding such as lactoferrin, hepcidin, Nramp1 or lipocalin2 [6]. Still, adequate concentrations of iron are required to support macrophagic killing mechanisms during infection [9], [10]. Contrary to what happens with other pathogens, the growth of Brucella abortus inside macrophages [11] and that of L. major in the mouse [12] are decreased by host's iron overload. In both cases, killing was correlated to oxidative burst [11], [13]. These examples highlight that iron can be exploited, in some cases, by the host to strengthen its antimicrobial defense mechanisms.
Our group has previously shown that the infection by Mycobacterium avium, an intra-macrophagic pathogen, is clearly exacerbated by host's iron overload, either genetically determined [14] or caused by iron-dextran injection [15], [16]. Furthermore, we have also demonstrated that iron chelation can be used to inhibit the growth of M. avium [17]. In the present work, we evaluated the effect of iron on the growth of L. infantum, the agent of European visceral leishmaniasis. We found that iron consistently inhibited the replication of L. infantum both by a direct effect and through the activity of the host's macrophage.
BALB/c and C57BL/6 mice were purchased from Charles River (Madrid, Spain). Mice deficient on the p47 subunit of the NADPH oxidase complex, on a C57BL/6 background (p47phox−/−), were bred at IBMC from a breeding pair purchased from Taconic (Lille Skensved, Denmark). The p47phox−/− mice were administered trimethoprimsulfamethoxazole (Bactrim; 600 mgl−1) in the drinking water, as prophylactic treatment against bacterial infection. This treatment was ceased when infection experiments began. Mice deficient in the nitric oxide synthase 2, on a C57BL/6 background (NOS2−/−), were bred at IBMC from a breeding pair kindly provided by Drs. J. Mudgett, J. D. MacMicking and C. Nathan (Cornell University, New York, USA). Mice deficient in HFE, on a C57BL/6 background (Hfe−/−) were bred at IBMC from a breeding pair obtained from Centre Nationale de la Recherche Scientifique. All animals were housed at IBMC facilities under specific pathogen free conditions and fed ad libitum, except for the diet experiments indicated below. All animals were used at 8 to 16 weeks of age. Only female mice (average 20–25 g) were used for in vivo experiments. Mice were euthanized by isofluorane anesthesia followed by cervical dislocation and tissues were collected in aseptic conditions.
The experimental animal procedures were approved by the Local Animal Ethics Committee of IBMC and licensed by the Portuguese General Directory of Veterinary (DGV, Ministry of Agriculture, Rural Development and Fishing), in May 18, 2006 with reference 520/000/000/2006. All animals were handled in strict accordance with good animal practice as defined by national authorities (DGV, Law nu1005/92 from 23rd October) and European legislation EEC/86/609.
All experiments were performed with L. infantum strain MHOM/MA/67/ITMAP-263 (zymodeme MON-1). For each experiment, parasites were obtained from the spleens of infected mice. Promastigotes were differentiated from spleen amastigotes by culturing at 25°C in complete Schneider's medium (Sigma-Aldrich Co., St Louis, MO, USA) supplemented with 20% heat-inactivated fetal bovine serum (FBS), 100 Uml−1 penicillin, 100 µgml−1 streptomycin (all from Gibco, Life Technologies, Carlsbad, CA, USA), 2% human urine, 5 µgml−1 phenol red (Sigma) and 5 mM HEPES sodium salt (Sigma) pH 7.4. Promastigote cultures were expanded at 25°C, for a maximum of 5 passages, in RPMI 1640 GlutaMAX™-I medium (Gibco, Life Technologies), containing 20% FBS, 50 Uml−1 penicillin, 50 µgml−1 streptomycin and 25 mM HEPES sodium salt pH 7.4. Promastigote differentiation from the exponential to the stationary phase was promoted by culture at 25°C, without medium renovation, for 4 to 5 days.
Axenic amastigotes were derived from the above-mentioned L. infantum strain, by the culture at 37°C with 7% CO2 atmosphere, in a medium for axenic amastigotes supplemented with 2 mM L-glutamine (GlutaMAX™-I, Gibco, Life Technologies) and 20% FBS (MAA20 medium, adapted from [18]).
Promastigotes (2×106/well) and amastigotes (4×105/well) of L. infantum were cultured in complete RPMI (25°C) or MAA20 (37°C) medium, respectively, supplemented with iron-dextran (Fe3+ hydroxide-dextran complex, Sigma), iron citrate (Fe3+, Sigma) and iron sulphate (Fe2+, Merck) in the concentrations of 0.018, 0.035, 0.070, 0.14, 0.28, 0.56, 1.1, 2.2, 4.5, 9 and 18 mM (96 well plates). Equivalent concentrations of dextran (Sigma), tri-sodium citrate (Merck) and magnesium sulphate (Merck) were used as controls. After 24 h of culture, 20 µl of a 2.5 mM resazurin solution (freshly prepared and filtered in phosphate buffered saline, pH 7.4, Sigma) was added to each well. The fluorescence intensity (excitation at 560 nm and emission at 590 nm) was determined 24 h (for amastigotes) or 48 h (for promastigotes) after resazurin addition to allow conversion to fluorescent resorufin, with a fluorometer SpectraMAX GeminiXS (Molecular Devices LLC, Sunnyvale, CA). In order to exclude a possible interference of iron with resazurin conversion to resorufin, appropriate controls were included without cells. No resazurin conversion was detected in the absence of parasites. Complete RPMI and MAA20 medium contains approximately 6 and 7.7 µM of iron, respectively, according to supplier's information.
Mice were injected in the lateral vein of the tail with 2×107 L. infantum stationary promastigotes in 200 µl of phosphate buffered saline pH 7.4 (PBS). At defined time points, the animals were euthanized and total livers and spleens were removed and homogenized, respectively, in 3.5 ml and 3 ml of complete Schneider's medium. These suspensions were further diluted 1∶100 (liver) or 1∶10 (spleen). Four-fold serial dilutions of the homogenized tissue suspensions were performed in quadruplicate (96 well plates). After 7 to 14 days at 25°C, the wells were examined for viable promastigotes. The reciprocal of the highest dilution that was positive for parasites was considered to be the number of parasites per ml of suspension and was used to calculate the number of parasites per organ (parasite burden).
In the kinetics experiments, 1 mg of iron, as iron-dextran (Sigma), was administered every other day i.p. to each mouse, from day −20 to day −2 of infection (a total of 10 mg of iron per mouse). Control mice received equivalent amounts of dextran (Sigma) by the same route. Parallel studies allowed us to verify that the administration of 10 mg of iron in a single injection produced the same effect in terms of outcome of Leishmania infection and also that dextran alone had no effect on the course of infection. Consequently, in subsequent experiments, iron overload was achieved by one single injection with 10 mg of iron per mouse. Control animals were injected with saline solution pH 6.0. None of the iron-dextran doses tested were toxic to the mice, as treated but uninfected mice remained healthy.
Mice were fed iron-free chow (Mucedola, Milan, Italy), from weaning until the end of the experiment (24 weeks). Control mice were fed a chow that differed only at its iron content (180 mgKg−1, Mucedola, Milan, Italy). Chow was administered in plastic recipients to avoid metal contamination. Animals were allowed to drink deionized water.
Non-heme iron was measured in tissues by the bathophenanthroline method [19]. Briefly, tissue samples (30–100 mg) were weighted, placed in iron-free Teflon vessels (ACV-Advanced Composite Vessel, CEM Corporation, Matthews NC, USA) and dried in a microwave oven (MDS 2000, CEM Corporation). Subsequently, dry tissue weights were determined and samples digested in an acid mixture (30% hydrochloric acid and 10% trichloroacetic acid) for 20 h at 65°C. After digestion, a chromogen reagent (5 volumes of deionized water, 5 volumes of saturated sodium acetate and 1 volume of 0.1% bathophenanthroline sulfonate/1% thioglycollic acid) was added to the samples in order to react with iron and obtain a colored product that was measured spectrophotometrically at 535 nm. The extinction coefficient for bathophenanthroline is 22.14 mM−1cm−1. Iron content in tissues was expressed as µg non-heme iron/organ.
Liver (100 mg) and spleen (50 mg) samples were lysed in Bio-plex cell lysis buffer containing 2 mM phenylmethanesulfonyl fluoride (PMSF, Bio-Rad Laboratories Inc., CA, USA), sonicated for 5 min in a refrigerated bath and centrifuged at 4500 g for 4 min at 4°C to remove debris. Total protein was quantified in supernatants with the MicroBCA kit (Pierce, Thermo Fisher Scientific). Clear homogenates were diluted to 1.5 mgml−1 in PBS pH 7.4 containing 0.5% bovine serum albumin and again centrifuged at 16000 g for 10 min at 4°C. Cytokine quantification in liver and spleen homogenates was performed following Bio-plex assay (Bio-Rad) instructions.
Spleen cells were obtained by teasing these organs gently with forceps and incubating them in NH4Cl haemolytic buffer to lyse any remaining erythrocytes. Cell suspensions were then washed with HBSS and resuspended in DMEM/10% FBS.
For immunofluorescence staining, 106 splenic cells were incubated for 15 min at 4°C, in a 96 well plate, with fluorescein isothiocyanate (FITC)-conjugated anti-DX5 (1∶200), anti-Gr1 (1∶800) or anti-CD19 (1∶200) antibodies, phycoerythrin (PE)-conjugated anti-CD3 (1∶200), anti-CD11c (1∶200) or anti-CD11b (1∶400) antibodies and allophycocyanin-conjugated anti-mouse F4/80 (1∶200) antibodies (BD Pharmingen, San Diego, CA, USA), in PBS containing 1% FBS in order to analyze spleen cell populations during infection. The cells were washed twice with PBS/1% FBS. The analysis of the cell populations was based on the acquisition of 10 000 events in a Becton Dickinson (BD, Franklin Lakes, NJ, USA) FACSCalibur equipped with BD CELLQuest and FlowJo (Tree Star Inc., Ashland, OR, USA) software.
Liver samples (50 mg) were fixated in 4% buffered paraformaldehyde pH 7.4 and embedded in paraffin. Tissue sections (5 µm) were stained with Perls' blue stain for iron detection. Representative pictures were obtained with an Olympus CX31 light microscope equipped with a DP-25 camera (Imaging Software CellˆB, Olympus, Center Valley, PA, USA).
Statistical analysis was carried out using GraphPad Prism 5.0 software (GraphPad Software Inc., La Jolla, CA, USA). Student's t-test was used to estimate the statistical significance of the differences between groups. Multiple comparisons were performed with One-way ANOVA followed by Dunnett or Student Newman-Keuls post hoc test. Differences between groups were considered statistically significant when p value was less than 0.05 (*p<0.05; **p<0.01; ***p<0.001).
Iron withdrawal has been suggested as a means of controlling the growth of several unrelated pathogens [8], [20]. To evaluate the effect of iron deprivation on the growth of L. infantum, BALB/c mice were fed normal or iron-deficient chow for the first 120 days of life. They were subsequently infected and kept in the respective diets for the next 60 days, before being euthanized. Non-heme iron quantification in the liver and spleen confirmed that mice kept on an iron-deficient diet had less than half the amount of iron found in controls (Figure 1A). However, no differences in parasite load were observed between groups fed control or iron-deficient diets (Figure 1B), indicating that a mild iron deficiency has no impact on L. infantum replication in mouse tissues.
Iron overload correlates with increased susceptibility to a great variety of pathogens [8], [20]. In order to determine the effect of host's iron overload on the infection by L. infantum, BALB/c mice were injected with 10 mg of iron-dextran or an equivalent amount of dextran before infection. Parasite burdens were determined in the livers and spleens at 7, 15, 30 and 60 days after infection. The growth of L. infantum in the liver and spleen of non-treated mice followed kinetics similar to that previously reported [21], [22]. During the first 30 days of infection, the parasite numbers increased significantly in both organs, returning to baseline levels in the liver thereafter, while continuing to grow in the spleen (Figure 1C). Conversely, parasite load was significantly lower in both organs of iron-overloaded mice throughout the experiment period (Figure 1C). Non-heme iron quantification confirmed that iron-overloaded animals had 8 and 3 times more iron, respectively, in the liver and spleen than control mice at 60 days after infection (Figure 1D). Iron distribution was analyzed in the liver, by Perl's staining. In those animals that were not injected with iron-dextran, iron deposition was very rarely seen (Figure 1E, G), the exception being the faint staining of some cell infiltrates (Figure 1G). The administration of iron to non-infected mice led to its accumulation predominantly in Kupffer cells (Figure 1F, black arrows), as expected from previous reports [23], [24]. Infection with L. infantum, led to the appearance of heavily iron-loaded macrophages inside cell infiltrates, the areas presumed to correspond to parasite containment (Figure 1H) [25], [26].
In order to investigate whether the in vivo anti-leishmanial effect of iron could be attributed to a generalized host tissue oxidative damage, we measured DNA cleavage and lipid peroxidation in the hepatic parenchyma through the immunofluorescence staining for TUNEL and 4-hydroxy-nonenal (4-HNE), respectively. Moreover, we assessed the formation of protein carbonyl groups in liver protein lysates by western blotting. However, we could not find any differences between control and iron-treated mice (Figures S1–S3 in Text S1), indicating that iron supplementation in our model did not cause a generalized oxidative damage to the tissue.
These experiments revealed that the accumulation of iron inside macrophages at L. infantum infection foci correlates with reduced parasite's multiplication, but not to generalized tissue damage.
In order to clarify the mechanisms by which iron exerts its anti-leishmanial effect in our model, we first asked whether iron could be exerting a toxic effect directly on the parasites.
Axenic promastigotes (Figure 2A–C) or amastigotes (Figure 2D–F) of L. infantum were grown in the presence of increasing concentrations of iron (0.018–18 mM) in the form of either dextran (A,D) or citrate (B,E) complexes or sulphate (C,F) salt. L. infantum viability was measured, based on the parasite's capacity to metabolize the dye resazurin. Iron concentrations below 0.56 mM had no effect on the multiplication of the parasites (Figure 2). Amastigotes seemed to be more susceptible to iron toxicity, as iron decreased the viability of these parasites in a dose dependent manner from 0.56 to 18 mM irrespective of its molecular form (Figure 2D–F). Promastigotes were inhibited by iron-dextran (Fe3+) at 0.56 mM or above (Figure 2A), while iron citrate (Fe3+) and iron sulphate (Fe2+) were active against promastigotes only above 4.5 mM (Figure 2B, C). Promastigotes exposed to 9–18 mM of iron in any form displayed oval shape, atrophied cell body and reduced motility (not shown), changes which are characteristic of stress situations [27]. The results obtained by resazurin reduction were confirmed by the visual microscopic quantification of the parasites in a Neubauer chamber, on selected samples (not shown). Overall, these results indicate that iron can inhibit the growth or even kill L. infantum promastigotes and amastigotes, although the concentrations needed to achieve that effect are relatively high.
Since previous studies had suggested that host's iron-overload interfered with the development of a protective immune response [12], we evaluated the impact of iron-overload on the induction of protective cytokines and specific splenic cell populations in our model of visceral leishmaniasis.
BALB/c mice were treated with 10 mg of iron (given as iron-dextran) or saline solution 15 days prior to infection. They were infected with L. infantum and sacrificed 60 days later. Groups of non-infected mice were kept as controls. We performed a cytometric analysis of the number of splenic CD3+ (T cells), CD3−DX5+ (NK cells), CD19+ (B cells), CD11c+ (Dendritic cells), CD11b+F4/80+ (Macrophages) and CD11b+Gr1++ (Neutrophils) cells. The infection with L. infantum resulted in a significant increase in the numbers of CD19+ and CD11c+ cells, while other splenic cell sub-sets remained unaltered (Figure 3). More importantly, the number of cells belonging to each of the abovementioned populations was the same in control and iron overloaded groups (Figure 3). Additionally, the in situ production of a number of cytokines was measured, using a multiplex assay. The results of this screening revealed that infection with L. infantum did not have a dramatic impact on cytokine production. Only IL-1β, IL-6, TNF and IL-4 were significantly induced by infection in the spleen (the latter also in the liver, Figure 4), while the production of IL-12p70 and IL-13 decreased with infection, in the liver (Figure 4). No significant differences were found between iron-overloaded and control infected animals in any of the cytokines tested (Figure 4). The determination of cytokine mRNA expression in the tissues at earlier time-points did not reveal any differences between iron-overloaded and control infected mice (not shown).
Overall, these experiments indicate that the inhibitory effect of iron on the growth of L. infantum in the mouse does not result from an improvement of the activation of protective cells or increased production of protective cytokines.
In order to further understand the mechanisms of the iron inhibitory effect on L. infantum, we tested the importance of different experimental parameters.
First, we decided to assess if the same protective effect could be obtained with lower iron doses. BALB/c mice were injected with different amounts of iron (given as iron-dextran complex) prior to infection with L. infantum. Parasite loads were determined 30 days after infection. A significant inhibitory effect on the growth of L. infantum in the liver and spleen was seen only in mice that received 10 mg of iron (Figure 5A).
Next, we asked whether iron would still have an inhibitory effect on L. infantum growth when given after infection. We administered 10 mg of iron (-dextran) or saline solution to BALB/c mice, 1 or 15 days after infection. Parasite loads were determined 60 days post-infection. In the liver, similar levels of growth inhibition were seen when iron was given either before, 1 day after or 15 days after infection (Figure 5B). However, in the spleen and in contrast with iron pre-loading, no significant reduction on the growth of L. infantum was detected when the administration of iron was done after infection (Figure 5B).
These results suggested that high amounts of iron present in the host prior to infection favor the decrease of the multiplication of L. infantum. In the model used, iron accumulation is observed in macrophages. To assess the relevance of the cellular location of iron deposition at the time of infection, we used HFE-deficient mice which are used as a model of human hemochromatosis and accumulate iron in parenchymal cells rather than inside macrophages [28], [29]. We infected Hfe−/− mice with L. infantum and evaluated the parasite load and iron content in the livers and spleens at 60 days after infection. HFE-deficient mice had around 3 times more iron in the liver than controls and normal amounts of iron in the spleen (Figure 6A). As expected, in non-infected Hfe−/− mice, iron was found predominantly in the hepatic parenchyma (Figure 6D). However, in L. infantum-infected mice, strong iron staining was found inside cell infiltrates both in wild-type mice and (more intensely) in Hfe−/− mice (Figure 6E,F) showing that infection re-routed iron between the two cell types. Interestingly, the parasite loads in wild-type and Hfe−/− mice were the same (Figure 6B).
Overall, these results indicate that a decrease in the growth of L. infantum is observed when high amounts of iron are found inside the host's macrophages, with a stronger effect when this occurs before the infection.
Iron concentrations necessary to inhibit L. infantum growth in axenic conditions are relatively high. So, we hypothesised that iron synergizes with antimicrobial mechanisms of macrophages, such as the production of reactive oxygen species (ROS) by the NADPH oxidase (respiratory burst) and reactive nitrogen species (RNS) by the nitric oxide synthase 2 (NOS2) to decrease Leishmania viability.
Mice genetically deficient in the p47phox subunit of NADPH oxidase (p47phox−/−) and in the NOS2 enzyme (NOS2−/−) were used to test this hypothesis. Animals were treated with 10 mg of iron (-dextran) or saline solution and 15 days later were infected with L. infantum. Mice were sacrificed 15 days after infection and the parasite load was determined in the liver and spleen. Iron overload decreased L. infantum growth in wild-type but not in knock-out mice in the liver (Figure 7A, B) and spleen (Figure 7C, D), indicating that the mechanism through which iron exerts its inhibitory effect is dependent on the production of ROS and RNS by the host. A similar experiment in which animals were sacrificed 30 days after infection gave identical results (data not shown).
Iron is a central element in host-parasite interaction and several iron-depriving mechanisms are used by the host to inhibit pathogen proliferation [6], [7]. In contrast, previous work has shown that host's iron overload prevented the growth of L. major in BALB/c mice [12], [13]. In those studies, it is shown that iron overload correlates with increased production of ROS upon L. major infection [13], [30]. In the present work, we treated mice with iron-dextran and infected them with L. infantum, an agent of visceral leishmaniasis. We observed iron-loaded macrophages inside hepatic infiltrates, the areas of parasite containment, concomitantly with the decrease of tissue parasite loads. Such iron-associated decrease did not occur in p47phox- or NOS2-deficient mice, suggesting that iron exerts its effects through the combination with ROS and/or RNS produced by the macrophage. In fact, superoxide (O2•−) and nitric oxide (NO•), synthesized by the phagocytic NADPH oxidase and the nitric oxide synthase 2 (NOS2), respectively, have been implicated in the elimination of Leishmania by the host's macrophages [31], [32], [33], [34], [35], [36]. Moreover, both macrophagic NADPH oxidase and NOS2 require iron for proper function [37]. In a few models of macrophage infections with different bacteria, macrophages were shown to need iron to exert their antimicrobial activity. Iron increases the capacity of macrophages to eliminate or prevent the multiplication of B. abortus, by catalyzing the production of hydroxyl radical [11] and iron loading of Staphylococcus aureus prior to infection enhances bacterial killing by monocytes, most likely by promoting oxidative damage [38]. Also in the case of Salmonella, iron seems to be needed for intramacrophagic killing of these bacteria [10]. This suggests that iron, a well known pro-oxidant, promotes oxidative microbicidal mechanisms inside macrophages, to which Leishmania is sensitive [31], [33], [34], [35], [36], [39]. The fact that the anti-parasitic effect of iron is lost in mice deficient in only one of the two enzymes, either NOS2 or NADPH oxidase, suggests that both ROS and RNS are simultaneously required for iron to exert its anti-leishmanial effect. Iron can possibly favour the formation of peroxynitrite (ONOO−), a strong oxidizing species formed by the reaction of NO• with O2•− [40]. Since we could not find evidences of tissue oxidative damage (DNA damage, lipid peroxidation and protein oxidation) in iron-treated mice, we suggest that this formation of highly reactive ROS and RNS in combination with iron has a highly localized activity, inside the macrophage.
It was somewhat surprising that Hfe−/− mice, which have spontaneous iron overload, predominantly in the liver, had tissue parasite loads similar to those of wild-type mice, when infected with L. infantum. This could be justified by the fact that Hfe−/− mice develop spontaneous iron overload predominantly in hepatocytes, keeping macrophages relatively iron depleted [28], [29], [41]. When Hfe−/− mice were infected with L. infantum, we could see iron accumulation inside macrophages at the infection foci, similarly to what we had previously reported in M. avium infection [14]. However, as suggested by the experiments in which we treated mice with iron-dextran after infection, the inhibitory effect of iron is best accomplished when the macrophages are iron-loaded prior to infection. Another hypothesis to explain the lack of an impact of Hfe−/− iron overload on the growth of L. infantum is the level of iron overload in the tissues. Indeed, Hfe−/− mice had tissue iron levels that were significantly lower than those found in iron-dextran-injected mice. The fact that even with iron-dextran injection, we needed high iron doses to decrease the parasite burden in tissues, argues for this hypothesis.
In the mouse model of cutaneous leishmaniasis, iron-induced respiratory burst at the onset of L. major infection is coupled to later activation of the nuclear transcription factor NF- κB [30] and to the display of a protective immune response [12]. Iron and ROS can modulate the activation of NF-κB signaling pathways [42], known to regulate several genes involved in immune and inflammatory responses [43]. In the case of L. major infection, mouse resistance is clearly related to an IL-12-driven, IFN-γ-dominated Th1 immune response, whereas susceptibility correlates with an IL-4-driven Th2 response [44]. Experimentally iron overloaded BALB/c mice, infected with L. major, exhibited a Th1-type immune response, with increased levels of IFNγ and NOS2 and decreased levels of IL-4 and IL-10 transcripts compared to untreated mice [12]. In accordance, supplementation of rats with iron-dextran [45] or saccharated colloidal iron [46] potentiated the induction of hepatic NOS2 and the production of NO• by LPS. However, the decrease of NO• production has also been observed in mice [47] and macrophages [48] treated with different iron sources. Additionally, delayed Th1 immune responses and Th2 phenotypes have been observed in response to iron supplementation in mice infected with Cryptoccocus neoformans [49] and Candida albicans [50], indicating that each particular host-pathogen interaction responds differently to iron overload.
In the case of visceral leishmaniasis, an efficient control of infection is also dependent on Th1 responses, although a mixed Th1/Th2 cytokine profile is detected during the course of infection [51], [52]. So, iron supplementation in our model, besides exerting a direct toxic effect on parasites in conjunction with ROS and RNS, could be improving the host's capacity to control the infection, by modulating the adaptive immune response. When we evaluated the cytokine response to Leishmania infection, we saw a discrete induction of IL-4 both in the liver and the spleen of infected mice, together with increases in the expression of the pro-inflammatory cytokines, IL-1β, IL-6 and TNF. However, iron overload did not significantly alter the immune response profile induced by infection, leading us to conclude that the modulation of the adaptive immune response does not contribute significantly to the protective effect of iron.
Malnutrition is associated with susceptibility to visceral leishmaniasis in humans [53], [54] and mice [55]. Although iron deficiency is the most common micronutrient deficiency in the human population [56], the relationship between human iron deficiency alone and increased risk of acquiring visceral leishmaniasis has never been investigated. In our model, feeding mice with an iron deficient diet did not affect L. infantum growth. These nutritionally iron-deprived animals had half the normal iron stores in the liver and the spleen, but presented normal haematocrit and body weight (not shown). We hypothesize that the low levels of iron in tissue stores were probably sufficient to maintain the growth of Leishmania and not low enough to impact on the host's capacity to control the infection. Observations regarding the effects of iron chelators on Leishmania growth are conflicting. Treatment of mice with desferrioxamine (DFO) led to the decrease of L. infantum proliferation [57] but not that of L. major [12]. Also, in in vitro models of macrophage infection, DFO has shown either no effect [58] or an inhibitory effect [59], [60], [61] on Leishmania growth. Finally, when tested on Leishmania promastigotes growing in culture medium, hydroxypiridinone-derived chelators showed an inhibitory effect, which was higher than that of DFO [62]. Thus, the available data do not allow inferring that Leishmania infections are amenable to treatment by iron depletion. It may be valuable to further explore the effects of different iron chelating ligands or of more drastic iron depletion protocols.
In addition to the results obtained in vivo and discussed above, we found axenic cultures of L. infantum to be sensitive to the direct toxicity of iron. It is plausible that high concentrations of iron (>0.56 mM) may have promoted the endogenous generation or propagation of reactive species in both parasite stages. On the other hand, lower iron doses (<0.56 mM) may not have been sufficient to overcome the antioxidant capacity of axenic L. infantum. In this regard, L. infantum promastigotes have been shown to accumulate iron in catalytically active forms, which contribute to their sensitivity to killing by hydrogen peroxide [63], [64], possibly through the Fenton reaction. Besides, upon exposure to high doses of this metal, L. infantum promastigotes exhibited impaired motility and morphological changes identical to those reported to occur after exposure to antimony (III) and arsenic (III) [27] (not shown). These metalloids, used for a long time as first line treatments against trypanosomatid infections, were recently found to act through the induction of oxidative damage in L. donovani [27], [65]. Interestingly, increased intracellular iron levels directly correlate to the parasite sensitivity to these drugs [27]. The influence of iron on anti-leishmanial drug activity also includes non-metalloid drugs. Iron potentiates the leishmanicidal activity of artemisinin [66], by inducing oxidative injury that culminates in cell death of L. donovani promastigotes. Furthermore, iron treatment can induce accumulation of pentamidine in the mitochondria of L. enriettii promastigotes, hence increasing their sensitivity to the drug. This effect is probably due to the action of the multidrug resistance protein 1 (LeMDR1), a putative mitochondrial iron importer [67]. Hence, the increase of intracellular iron levels in Leishmania overall increases its vulnerability to chemotherapy.
The interaction between pathogens and their hosts are complex processes dependent not only on the genome of both, but also on nutritional factors. In most of the reported cases, iron excess increases and iron chelation decreases susceptibility to infection [8], [20]. However, this is not observed in murine models of infection by Leishmania. Moreover, despite several reports that iron supplementation (to correct nutritional iron deficiency) can significantly increase the risk of several infections [8], [56], no correlation between iron administration and susceptibility to human leishmaniasis has, to our knowledge, ever been described. Although iron chelation has been suggested as an effective therapeutic strategy against several infections, in the case of leishmaniasis and especially in areas where this disease and malnutrition coexist, iron chelation may be inappropriate. Iron overload decreases Leishmania proliferation and induces parasite death, probably by promoting oxidative reactions pernicious to the parasite. The further investigation of the molecular mechanisms of these effects will be fundamental to explore a potential utilization of iron itself as a therapeutic tool and also to understand and improve the mechanisms of action of other anti-leishmanial drugs.
|
10.1371/journal.pcbi.1004899 | HepatoDyn: A Dynamic Model of Hepatocyte Metabolism That Integrates 13C Isotopomer Data | The liver performs many essential metabolic functions, which can be studied using computational models of hepatocytes. Here we present HepatoDyn, a highly detailed dynamic model of hepatocyte metabolism. HepatoDyn includes a large metabolic network, highly detailed kinetic laws, and is capable of dynamically simulating the redox and energy metabolism of hepatocytes. Furthermore, the model was coupled to the module for isotopic label propagation of the software package IsoDyn, allowing HepatoDyn to integrate data derived from 13C based experiments. As an example of dynamical simulations applied to hepatocytes, we studied the effects of high fructose concentrations on hepatocyte metabolism by integrating data from experiments in which rat hepatocytes were incubated with 20 mM glucose supplemented with either 3 mM or 20 mM fructose. These experiments showed that glycogen accumulation was significantly lower in hepatocytes incubated with medium supplemented with 20 mM fructose than in hepatocytes incubated with medium supplemented with 3 mM fructose. Through the integration of extracellular fluxes and 13C enrichment measurements, HepatoDyn predicted that this phenomenon can be attributed to a depletion of cytosolic ATP and phosphate induced by high fructose concentrations in the medium.
| Despite the key role of hepatocytes in carbohydrate and lipid homeostasis, available dynamic models of hepatocyte metabolism tend to be limited to a single pathway and/or are based on assumptions of constant concentrations of key metabolites involved in redox and energy metabolism (ATP, NAD, NADPH etc.). Furthermore, most dynamic models are unable to integrate information from 13C based experiments. 13C based experiments allow us to infer the relative activity of alternative pathways and hence are highly useful for indicating flux distributions. To overcome these limitations, we developed HepatoDyn, a dynamic model of hepatic metabolism. HepatoDyn uses a large metabolic network including key pathways such as glycolysis, the Krebs cycle, the pentose phosphate pathway and fatty acid metabolism, and dynamically models the concentrations of metabolites involved in the redox and energy metabolism of hepatocytes. In addition, the model was coupled to the label propagation module of the package IsoDyn, allowing it to integrate data from 13C based experiments to assist in the parametrization process. These features make HepatoDyn a powerful tool for studying the dynamics of hepatocyte metabolism.
| No other organ performs as many physiological functions as the liver. The liver is responsible for detoxification, bile acid and blood proteins synthesis, plays a key role in the inflammatory response and, above all, it is a key regulator of glucose and lipid homeostasis in blood. Most of its functions and properties can be linked to hepatocytes, the most abundant cell type in liver, and therefore hepatocytes are often used as a model to study liver function and pathologies [1]. Accordingly, computational modelling of hepatocyte metabolism has received a great deal of interest.
Recently, genome scale metabolic reconstructions based on stoichiometric modelling techniques have been successfully used to model hepatocyte metabolism [2–4]. However, stoichiometric models provide a static picture of metabolism based on mass balance equations and the assumption that the system is under a strict steady state. In these models each reaction step is described by only one parameter, its steady state flux [5]. The alternative is to use dynamic metabolic models, usually referred to as kinetic models. They are based on building a system of ordinary differential equations (ODEs), with kinetic laws describing transport and chemical transformations for each reaction-step and parameters describing biochemical and biophysical constraints. Kinetic modelling has two main advantages over stoichiometric based modelling; firstly, it is capable of performing dynamic simulations, that is to say, it can predict the variation in metabolite concentrations and fluxes over time outside of the steady state. Secondly, it can follow the global effects of constraints emerging from the specific kinetic properties of enzymes, post-translational modifications and regulatory circuits, thus revealing the complex regulation of the system. Over the years, multiple kinetics models of hepatocyte metabolism have been developed [6–11]. The main limitation of kinetic models is that they are complex to build and parametrize. Due to this complexity, kinetic models of hepatocyte metabolism available in the literature contain only a small number of reactions and, with some exceptions [11], are often limited to a single pathway. Furthermore, with the exception of some models focused on mitochondria [8, 9], most of them assume a constant redox and energy state, which limits their application. In fact, despite the huge interest in hepatocyte metabolism, there are no models capable of adequately modelling the effects of the energy and redox dynamics on hepatocyte core metabolism. Additionally, while 13C experiments have proven their usefulness in studying the metabolism of hepatocyte under metabolic steady state [12–24], there was only one kinetic model of hepatocyte capable of integrating 13C data [10].
In this work, we present HepatoDyn (Hepatocyte Dynamics) a model of hepatocyte core metabolism capable of simulating the redox (NAD/NADH, NADP/NADPH, etc.) and energy (ATP/ADP/AMP, etc.) dynamics. The model includes glycolysis, gluconeogenesis, glycogen metabolism, the pentose phosphate pathway, the Krebs cycle and fatty acid metabolism as well as reactions associated with energy and redox metabolism (respiratory chain, malate/aspartate shuttle, glycerol phosphate shuttle, etc.). To our knowledge, no model of such size capable of dynamic redox and energy metabolism simulations exists in the literature. Furthermore, the model was coupled to the module for isotopic label propagation of the software package IsoDyn [25, 26]. This enables HepatoDyn to integrate data from 13C based experiments to assist in the parametrization process, regardless of whether experimental measurements correspond to an isotopic steady state. The latter is a key feature because the levels of isotopic label enrichment are often a non-steady phenomenon with long transition times [27]. Therefore, HepatoDyn is a very powerful tool capable of taking advantage of both the constraints derived from a detailed tissue-specific kinetic model and data derived from 13C based experiments to simulate hepatocytes.
In the last decades there has been a significant increase in fructose in our diets [28] and accordingly there is great interest in studying the potential effects of fructose in the metabolism [29–32]. To date, fructose-rich diets have been associated with many adverse metabolic conditions, such as nonalcoholic fatty liver disease, insulin resistance and obesity [28, 33, 34], most of which are directly or indirectly related to abnormal hepatocyte function. Therefore, we used HepatoDyn to study the short-term response of hepatocyte metabolism to different concentrations of fructose.
A metabolic network, including those pathways deemed necessary to accurately and dynamically simulate the core metabolism of rat hepatocytes in the study conditions, was constructed based on pathways that have been reported in the literature to be active in hepatocytes [42, 43].
Each reaction in the metabolic network was assigned a kinetic law. Kinetic laws describe the dependence of each reaction flux on metabolite concentrations. They take into account the affinity of substrates and products, the reaction mechanism and the effect of activators and inhibitors on reaction fluxes. The kinetic laws used were mostly derived from existing kinetic laws described in the literature [6, 11, 44]. The exceptions were the kinetic laws for aldolase activity, which catalyses eight related elementary reactions, which were built as described in the Supplementary Material (S1 Text).
Kinetic laws are integrated with the metabolic network topology, described by the stoichiometric matrix (N), to build a system of ordinary differential equations (ODEs) that predict the evolution of metabolite concentrations, and by extension the evolution of reaction fluxes, over time. Because fluxes are provided in units of mmol per cell per minute, but ODEs are solved in units of mmol per litre per minute, in order to build the ODEs, the cell number and the volume of the compartment at which each metabolite is located must also be taken into account. Therefore, the system of ODEs can be written as:
dc[t]dt=N·j(c[t],p)·ncellvol
(1)
Where j is a vector of reaction fluxes, which is a function of the vector of metabolite concentrations (c[t]) in mM, and a vector model parameter (p) as defined by the kinetic laws used in the model, ncell is the cell number and vol is a vector containing the volumes of the compartment at which each metabolite is localized in litres.
In reversible reactions, forward and reverse reaction rates are computed separately with different kinetic laws, albeit sharing most of the parameters. Additionally, the fluxes of invisible reactions, that is to say, reactions that can propagate labelled carbons even though they do not change the overall concentrations of metabolites, are also computed [10]. This is necessary in order to fully simulate the propagation of 13C.
To simulate the propagation of 13C through the metabolic network, fluxes are decomposed into isotopomer fluxes. Then, an ODE system is built using the algorithms from IsoDyn [25, 26]. The resulting ODE accounts for concentrations of all isotopomers, isomers with 13C substitution in specific carbon positions [24]. To avoid unnecessary complexity, isotopomers are not simulated for those metabolites where, according to the defined metabolic network, 13C from labelled substrates cannot be propagated. The process is briefly summarized in Fig 1.
The system of differential equations for metabolite and isotopomer concentrations is solved to predict metabolic fluxes, metabolic concentrations and isotopomer concentrations from the initial time to the defined end time.
Model predictions are for isotopomers but experimental measurements refer to isotopologues (or mass isotopomers), isomers with a specific number of 13C substitutions [24]. Thus, the resulting concentrations of isotopomers are converted into fractions of isotopologues, by adding up all isotopomers that correspond to each isotopologue and dividing by the total concentration of each metabolite (S1 Fig). The fractions of such isotopologues can then be compared with the experimental measurements obtained with GC coupled to MS.
Kinetic parameters representing enzyme activity (Vmax or equivalent) were fitted to the experimental data. For this process Vmax from the reverse reaction rate in reversible reactions are assumed to be a function of the Vmax of the forward reaction and of the equilibrium constant as described by the Haldane relationship [44]. To further reduce the number of parameters fitted, enzyme activities catalysing sequential reactions with no ramifications (the so called reactions chains) were fitted as a group. This is because in reactions chains the flux through the whole chain could be determined by any of the enzyme activities involved and consequently most of the activities of enzymes constituting the chain would be unidentifiable. Furthermore, other activities known to be unidentifiable are not fitted, such as the activities of reactions that are known to operate in rapid equilibrium in physiological conditions (glucose phosphate isomerase, triose phosphate isomerase, enolase, etc.). The remaining parameters of the kinetic model were assigned based on an extensive literature search, completed with data from Brenda [45] and UniProt [46] databases.
The fitting algorithm, a variant of the basic simulated annealing algorithm [47], seeks the set of m parameters (Ez) that minimizes the objective function. The objective function (X2) is the square deviation between the n experimentally measured values (Yi) and simulated values (Zi) for both isotopologue fractions and total metabolite concentrations, normalized by the experimental standard deviation (σi). To prevent a bias generated by very low standard deviations, a minimum threshold of 0.01 was used. Additionally, parameter sets where any metabolite reached concentrations greater than 50 mM were discarded.
Consequently, the fitting algorithm seeks the set of enzyme activities that minimize the difference between experimentally measured and simulated isotopologue fractions and metabolite concentrations in the experimental conditions considered.
The fitting procedure provides one set of fitted parameters, which minimizes the objective function, and is referred to as the best fit parameter set. However, other sets of parameter values might result in similar or equal objective function values and are therefore as valid as the best fit. The range of acceptable variation in parameters was evaluated through an identifiability analysis. Identifiability is a property that indicates whether unknown model parameters can be determined from the available experimental data. It depends both on the structure of the model and the quality and amount of experimental data. A parameter is defined as identifiable if the confidence interval for its estimated value at a given significance level is finite [48, 49].
If we define X2(θi) as the optimized square deviation if parameter i is fixed to a value of θi and the remaining parameters being fitted θj are readjusted to minimize the square deviation
X2(θi)=minθj≠i[x2(θj)]
(3)
then if experimental errors are assumed to follow a normal distribution, for a parameter i, the confidence interval can be defined as:
{θi|X2(θi)−Xbf2<Δα} with Δα=X2(α,1)
(4)
where X2bf is the best fit square deviation (optimized with no fixed parameters) and Δα is the significance threshold associated with a given significance level (α) with a Chi Square distribution with one degree of freedom. Accordingly, the upper and lower limit of the confidence intervals for a given parameter are estimated by respectively increasing and decreasing the value of the parameter until the square deviation difference obtained when optimizing the remaining parameters exceeds the threshold (Δα) [48].
Additionally, intervals for system dependent variables (fluxes, metabolite concentrations and isotopologue fractions at different time points) are estimated from the maximum and minimum parameter values of confidence intervals generated during the identifiability analysis.
We present HepatoDyn, the first detailed model of hepatocyte core metabolism capable of dynamically simulating energy and redox metabolism. It consists of 88 reactions and 81 metabolites distributed into three compartments (extracellular, cytosolic and mitochondrial). A schematic representation of the model can be found in Fig 2 and a complete list of metabolites, reactions and compartments can be found in S1, S2 and S3 Tables, respectively.
Each reaction has an associated kinetic law and the model has a total of 470 parameters associated to kinetic laws (S4 Table). 55 of these parameters correspond to enzyme activities that were fitted to experimental data, taking parameter groups (S5 Table) into account this results in 29 independent parameters that were fitted to experimental data. To the greatest extent possible, the kinetic laws and their parameters were specific to the enzyme isoforms active in the liver.
It is worth noting, that while most of the reactions included in HepatoDyn are also present in genome scale reconstructions of hepatocyte metabolism [2–4], HepatoDyn includes complete kinetic laws and regulatory loops, which allow for dynamic and regulatory studies. Nevertheless, HepatoDyn also has 2 reactions that are absent in genome scale reconstructions of hepatocyte. Specifically, the reactions aldolase 3 (Fru16bP + Gra ↔ Fru1P + GraP) and transketolase 3 (Fru6Pa + Rib5P ↔ E4P + Sed7P). Those reactions emerge because the enzymes aldolase and transketolase allow multiple combinations of substrates and products. Additionally, HepatoDyn also incorporates the channelling of hexose phosphates to glycogen in the form of two separate pools of hexose phosphates, a and b, as previously described in the literature [10].
The kinetic model, fully parametrized, can be found in SBML format in the Supplementary Material (S1 XML and S2 XML).
In addition, HepatoDyn is capable of simulating the propagation of 13C from isotopically labelled substrates to metabolic intermediaries and products. This allows HepatoDyn to integrate isotopologue enrichment measurements from 13C based experiments greatly enhancing the predictive capabilities of the model.
HepatoDyn is provided in the Supplementary Material as a C++ program (S1 Software).
The liver has a high capacity to metabolize fructose, it is estimated that up to 50% of fructose ingested is metabolized by hepatocytes [50]. Fructose metabolism in hepatocytes consists of phosphorylation of fructose to fructose 1-phosphate by fructokinase and the split of this metabolite by the liver aldolase isoform (aldolase B) into dihydroxyacetone-phosphate and glyceraldehyde, with the latter metabolite being phosphorylated by triokinase into glyceraldehyde 3-phosphate. Because fructose enters at the level of triose phosphate, bypassing the highly regulated glucokinase and phosphofructokinase steps of glycolysis, fructose uptake is largely unregulated. Consequently, the limiting step in fructose metabolism is assumed to be fructose uptake by hepatocytes, which is heavily dependent on the extracellular concentration of fructose due to the low affinity of the proteins mediating fructose transport into hepatocytes, GLUT2 and other carriers like GLUT8 [51–53].
As a proof of concept of the capabilities of HepatoDyn, we applied it to study the short term response of hepatocytes to incubation with 20mM glucose supplemented by either 3mM fructose or 20mM fructose. These concentrations were chosen because our experimental data showed that hepatocytes responded quite differently to them. While incubation with 20mM glucose supplemented with 3mM fructose resulted on a rapid glycogen accumulation, incubation with 20mM glucose supplemented with 20mM fructose resulted on almost no glycogen accumulation (Fig 3.A). While it has been reported that supplementation with low concentrations of fructose favours glycogen accumulation [19, 29, 54], the fact that supplementation with high fructose concentrations inhibits glycogen accumulation was not known. Furthermore, isotopologue analysis indicated that in the second condition, unlike the first condition, almost no 13C from labelled glucose was propagated to lactate (Fig 3.B). In both conditions lactate and glucose were produced from fructose at a similar rate. Hence it was an interesting case of study.
Specifically, HepatoDyn was used to integrate experimental measurements derived from rat hepatocytes incubated for 2 h with the following media: 20 mM glucose 50% enriched in [1,2-13C2]-glucose and 3 mM fructose (condition A1), 20 mM glucose and 3 mM fructose 50% enriched in [U-13C6]-fructose (condition A2) and 20 mM glucose 50% enriched in [1,2-13C2]-glucose and 20 mM fructose (condition B). The experimental data for condition A1 had been published previously [19]. This integration was achieved using the experimental measurements of extracellular concentrations and isotopologue fractions as input to fit the 29 independent parameters associated to enzyme activities in the model assuming that the enzyme activities, normalized by cell number (S2 Fig), were equivalent in the three conditions. Consequently, the fitting algorithm identifies a single set of parameters that allows reproduction of the three experimental conditions. It is worth noting that because conditions A1 and A2 only differ in the labelling pattern of substrates, the predicted fluxes and concentrations values will be the same in both conditions. The resulting values of the fitted parameters can be found in S6 Table. The resulting metabolites concentrations for condition A1/A2 and condition B can be found on S3 and S4 Figs respectively. The resulting fluxes for condition A1/A2 and condition B can be found on S5 and S6 Figs respectively. The resulting isotopologue fractions for key metabolites in condition A1, A2 and B can be found on S7, S8 and S9 Figs respectively. A comparison between the experimentally measured metabolite concentrations and isotopologue fractions and those simulated by the model with the best fit parameter set can be found in Fig 3.
High concentrations of fructose have been shown in vivo and in vitro to result in the depletion of ATP and phosphate in hepatocytes [52, 55]. This occurs due to an accumulation of fructose 1-phosphate caused by the elevated fructokinase activity [52, 55]. This phenomenon was predicted by HepatoDyn. The model predicted that a persistent cytosolic ATP and phosphate depletion would occur with an extracellular concentration of 20 mM fructose (Fig 4). This is mainly caused by an accumulation of fructose 1-phosphate, although the depletion can also be partially attributed to the accumulation of some other phosphorylated metabolites. In this context, the low glycogen synthesis observed at 20mM glucose supplemented with 20 mM fructose can be attributed to the depletion of cytosolic ATP and phosphate. Likewise, the almost non-existent propagation of 13C from glucose to lactate under this condition can mainly be attributed to the low glucokinase and phosphofructokinase activities caused by ATP depletion. Conversely, at 20mM glucose supplemented with 3 mM fructose, a persistent accumulation of fructose 1-phosphate does not occur. Accordingly, under this condition, ATP and phosphate are not persistently depleted (Fig 4).
Overall, 25 of the 29 independent parameters were identifiable with at least 95% confidence. This remarkable degree of identifiability can be attributed to the numerous feedback regulations through the redox and energy balances (ATP/ADP, NADH/NAD, etc.), the use 13C data and the integration of data from multiple metabolic conditions.
Concerning the non-identifiable parameters, the non-identifiability of the aldolase activity and the activities involved in the lactate production and malate aspartate shuttle reaction chains can be attributed to the fact that the reactions associated to those pathways are predicted to be close to the equilibrium in experimental conditions, hence the system is fairly insensitive to the value of the enzyme activities associated to them. On the other hand, the non-identifiability of the citrate synthase activity arises because in our model the flux through the citrate synthase reaction can depend solely on the two activities upstream, pyruvate dehydrogenase and β-oxidation, which catalyse the production of acetyl-CoA, the substrate of citrate synthase.
Compared to parameters, fluxes and to a lesser extent concentrations, show a much narrower range of variation (S3, S4, S5 and S6 Figs). This can serve as an indication of robustness, the capacity of the system to maintain its functional properties in the face of external and internal perturbations and uncertainty [56].
Interestingly, fluxes associated with the pentose phosphate pathway and fatty acid synthesis have fairly low upper bounds in both conditions (incubation with 3 mM fructose and 20 mM glucose and incubation with 20 mM fructose and 20 mM glucose). This is consistent with hepatocytes extracted from fasted rats, as they can be expected to have low activity in fatty acid synthesis, and thus only need to generate a small amount of reductive potential (NADPH) to maintain cell functions. However, with longer incubation times, an increase in the fatty acid synthesis and pentose phosphate pathway activities and fluxes should be observed as fructose is known to increase the expression of key lipogenic enzymes in hepatocytes[28, 57, 58].
It is also worth noting that the identifiability analysis further reinforces the notion that hexose phosphate metabolism in hepatocytes is compartmentalized into two different pools as previously reported [10]. This is because most of enzyme activities present in both hexose pools have a lower bound above 0 in the confidence interval, suggesting that the separation of hexose phosphates into two separate pools must be taken into account to adequately simulate the experimental conditions. If there was no compartmentalization, all activities present in both pools would have a lower bound of 0 because they would be made redundant by the activities in the other pool.
Metabolic modelling is based on applying constraints to limit the space of feasible solutions for system variables, such as reaction fluxes and metabolite concentrations. Constraints can arise from different components of the model including reaction stoichiometry and kinetic laws, and from the experimental measurements integrated by the model. Consequently, the use of a highly complete metabolic network, including the fundamental balances affecting redox and energy metabolism (ATP/ADP, NAD/NADH, etc.), serve as an important set of constraints. Furthermore, the inclusion of highly detailed kinetic laws and parameters derived from the literature further constrains the solution space. For instance, important constraints that emerge from kinetic laws are regulatory circuits, such as fructose 6-phosphate inhibiting glucokinase or fructose-1-phosphate disrupting such inhibition [59–61]. Other important constraints that emerge from the kinetic laws are thermodynamics constraints, which are in the form of equilibrium constants. Finally, integrating 13C based data provides additional constraints such as labelling enrichments which provide information on ratios among fluxes through alternative metabolic pathways. While numerous kinetic models of hepatocytes exist in the literature [6–11], HepatoDyn is the first that is capable of integrating all the aforementioned constraints in a single model.
As a proof of concept of the capabilities of the model, we applied HepatoDyn to study the metabolic effects of high fructose concentrations on rat hepatocytes. Experimental data showed that hepatocytes behaved quite differently depending on whether they were incubated with 20mM Glucose supplemented with either 3 mM fructose or 20 mM fructose. Using HepatoDyn, we managed to find a physiological explanation for this behaviour, which involved the rapid and persistent depletion of cytosolic ATP and phosphate at 20 mM fructose, which was in accordance with information reported in the literature [52, 55]. This phenomenon has a strong dynamic component, is dependent on the kinetic properties of enzymes and on the balances involved in energy metabolism. Additionally, it may be relevant for understanding the potential adverse effects of fructose-rich diets. This is because ATP depletion impairs protein synthesis and induces inflammatory and prooxidative changes and thus, in a fructose-rich diet, this depletion might result in increased susceptibility of hepatocytes to injury leading to adverse hepatic conditions such as nonalcoholic fatty liver disease [62].
Furthermore, HepatoDyn has countless applications that go beyond studying the effects of fructose. For instance, HepatoDyn can be used to study liver centric metabolic diseases such as diabetes. Given that HepatoDyn is capable of dynamically simulating the redox and energetic state of hepatocytes, it can be used to better understand the mechanism of action of anti-diabetic drugs like metformin which target the energetic and redox metabolism [63] as well as identifying new drug targets. HepatoDyn can also be used to study the relative contribution of different reactions to redox and energy balances in different conditions. Therefore, potential applications of HepatoDyn can be to analyse the ATP consumption or production associated to different pathways or the relative contribution of the glycerol phosphate shuttle and the malate aspartate shuttle to the transfer of reducing equivalents between the cytosol and the mitochondrial matrix. Last, but not least, new reactions can easily be added to HepatoDyn provided kinetic mechanisms and kinetic information such as affinity constants or inhibition constants are known for the enzymes catalysing those reactions. Likewise, through the modification of reactions and kinetic laws specific to hepatocytes, HepatoDyn can be adapted to other cell types.
|
10.1371/journal.pgen.1007202 | Ras/ERK-signalling promotes tRNA synthesis and growth via the RNA polymerase III repressor Maf1 in Drosophila | The small G-protein Ras is a conserved regulator of cell and tissue growth. These effects of Ras are mediated largely through activation of a canonical RAF-MEK-ERK kinase cascade. An important challenge is to identify how this Ras/ERK pathway alters cellular metabolism to drive growth. Here we report on stimulation of RNA polymerase III (Pol III)-mediated tRNA synthesis as a growth effector of Ras/ERK signalling in Drosophila. We find that activation of Ras/ERK signalling promotes tRNA synthesis both in vivo and in cultured Drosophila S2 cells. We also show that Pol III function is required for Ras/ERK signalling to drive proliferation in both epithelial and stem cells in Drosophila tissues. We find that the transcription factor Myc is required but not sufficient for Ras-mediated stimulation of tRNA synthesis. Instead we show that Ras signalling promotes Pol III function and tRNA synthesis by phosphorylating, and inhibiting the nuclear localization and function of the Pol III repressor Maf1. We propose that inhibition of Maf1 and stimulation of tRNA synthesis is one way by which Ras signalling enhances protein synthesis to promote cell and tissue growth.
| The Ras oncogene is one of the primary drivers of cell and tissue growth in both normal development and in diseases such as cancer. In this report, we identify the stimulation of tRNA synthesis as an important mechanism by which Ras functions. Using fruit fly genetics, we show that Ras promotes tRNA synthesis by inhibiting Maf1, a protein that normally inhibits RNA polymerase III, the enzyme complex that stimulates tRNA synthesis. We further show that stimulation of tRNA synthesis is required for Ras to promote growth in two important cell types—stem cells and epithelial cells. This work provides new insight into mechanisms that are important for growth and that may contribute to cancer.
| The Ras small G-protein is one of the key conserved regulators of cell growth and proliferation. Over three decades of research have defined the textbook model of how Ras is activated by growth factors to stimulate a core RAF kinase, MEK (Mitogen-activated protein kinase kinase) and ERK (Extracellular signal–regulated kinase) signalling cascade. Work in model organisms such as Drosophila, C elegans and mouse has shown how this Ras/ERK pathway coordinates tissue growth and patterning to control organ size during development and homeostatic growth in adults.
Given its central role in development it is not surprising that defects in Ras signalling contribute to disease. Most notably, activating mutations in Ras and RAF occur in a large percentage of cancers, and lead to hyper-activation of ERK, which drives tumour formation in both epithelial and stem cells [1]. Ras pathway mutations are also seen in several genetic developmental disorders–described collectively as RASopathies–often characterized by abnormal growth[2]. Understanding how Ras promotes cell proliferation and tissue growth is therefore an important concern in biology.
Drosophila has been a powerful model system to understand the biological roles of Ras signalling. In flies, Ras functions downstream of epidermal growth factor (EGF) and activation of its tyrosine kinase receptor (the EGFR). A series of genetic studies initiated over 25 years ago were pivotal in defining the canonical EGFR/Ras/ERK pathway in Drosophila (for reviews of this early work see:[3,4]). Extensive studies since then have established when, where and how the pathway is activated during the fly life cycle to control development. This work has emphasized the importance of Ras signalling in the control of cell growth and proliferation (e.g. [5–9]. Notably, during larval development Ras/ERK promotes EGFR-mediated cell proliferation and tissue growth in epithelial organs such as the imaginal discs, which eventually give rise to adult structures such as the legs, wings and eyes [10–14]. In addition, in the adult the EGFR/Ras/ERK signalling controls proliferation of stem cell populations to maintain homeostasis and promote regenerative growth [15–19].
How does Ras mediate these effects on cell and tissue growth? Most work on this area has focused on transcriptional effects of Ras signalling. Work in Drosophila has identified several transcription factors that are targeted by ERK such as fos, capicua, and pointed, and that regulate growth [19–22]. Ras signalling has also been shown to crosstalk with other transcriptional regulators of growth such as the hippo/yorkie pathway and dMyc [10,23–27]. These transcriptional effects control expression of metabolic and cell cycle genes important for growth[20,21]. Less is known, however, about how Ras/ERK may regulate mRNA translation to drive growth. The prevailing view, arising mostly from mammalian tissue culture experiments, is that ERK controls protein synthesis by stimulating the activity of translation initiation factors[28]. In particular, these effects are mediated via two ERK effector families—the MNK (MAP kinase-interacting serine/threonine-protein kinase) and RSK (ribosomal s6 kinase) kinases [28–30]. These kinases are important for cellular transformation and tumour growth in mammalian cells [31–34]. However, MNK and RSK mutants in mice and Drosophila have little growth or developmental phenotypes, and mouse MNK mutant cells show no alterations in protein synthesis [33–37]. These findings suggest Ras uses additional mechanisms to control translation and growth in vivo during animal development.
In this paper, we report that the Ras/ERK pathway can stimulate RNA polymerase III-dependent tRNA synthesis. We find that these effects are required for Ras to drive proliferation in both epithelial and stem cells. Finally, we show that ERK promotes tRNA synthesis by inhibiting the Pol III repressor Maf1. These findings suggest that stimulation of tRNA synthesis may be one way that Ras promotes mRNA translation to drive cell and tissue growth.
We first examined whether Ras signalling regulates protein synthesis in Drosophila S2 cells using a puromycin-labelling assay [38]. When a constitutively active Ras mutant (RasV12) was expressed in Drosophila S2 cells using an inducible expression vector, we found an increase in protein synthesis, which was blocked by treatment of cells with cycloheximide (CHX), an inhibitor of mRNA translation (Fig 1A). Also, using polysome profiling to measure mRNA translation, we saw an increase in polysome levels in RasV12 overexpressing cells when compared with control cells (Fig 1B). Conversely when we blocked Ras/ERK signalling by treating cells with the MEK inhibitor, U0126, protein synthesis was decreased (Fig 1C). Finally, we found that total protein content/cell increased after RasV12 was overexpressed in S2 cells (Fig 1D). Our findings suggest that one way that the Ras/ERK signalling pathway may drive growth in Drosophila is by promoting protein synthesis.
We previously identified the regulation of RNA Polymerase III and tRNA synthesis as a mechanism for controlling protein synthesis in Drosophila larvae [39,40]. We showed that these processes were regulated by TORC1 kinase signalling, and that they were important for driving tissue and body growth[39,40]. We were therefore interested in examining Ras signalling could also promote tRNA synthesis. We first used qRT-PCR to examine both pre-tRNA and total tRNA levels in S2 cells. We began by using the pharmacological MEK inhibitor U0126 to examine the effects of blocking Ras signalling. We found that treatment of S2 cells with UO126 lead to a decrease in levels of both pre-tRNAs and total tRNAs (Fig 2A and 2B). Also, using Northern blots, we saw that treatment with UO126 lead to reduced pre-tRNA and tRNA levels in S2 cells (S1A Fig). We also examined the effects of Ras pathway inhibition by using RNAi to knockdown Raf. We found that treatment of cells with dsRNA to Raf lead to reduced levels of both pre-tRNA and total tRNAs (Fig 2C and 2D). In contrast to Ras pathway inhibition, we found that RasV12 (constitutively active Ras) overexpression lead to an increase in both pre-tRNA and mature tRNA levels as measured by both qRT-PCR (Fig 2E and 2F) and Northern blot (S1B Fig), indicating enhanced tRNA synthesis. In contrast, to these effects on tRNA levels, we found no effect of inhibiting Ras signalling on expression levels of TBP (which was previously reported[41]) or on levels of Brf1 or Trf1 –both of which are components of TFIIIB complex, which is required for Pol III recruitment to tRNA genes (S1C Fig). We also found that altering Ras signalling (either by MEK inhibition or overexpression of RasV12) had no effect on levels of 5S or 7SL RNA in S2 cells (S1C and S1D Fig). Both of these genes transcribed by RNA polymerase III, but they are different Pol III gene types (type I and III respectively) that use a different set of core promoter factors compared to tRNA genes (which are type II Pol III genes). These finding suggest that the changes in tRNA synthesis we observed upon altering Ras signalling are not due to alterations in the levels of the basal transcriptional machinery required for tRNA transcription. These data also suggest Ras signalling may predominantly affect type II RNA pol III genes.
We also examined the effects of Ras signalling on tRNA levels in the developing wing imaginal discs. We used the temperature-sensitive escargot-Gal4 (esg-Gal4ts) system, which allows for inducible transgenes expression in all imaginal tissues. When we overexpressed UAS-Rafgof using this system, we found a marked increase in pre-tRNA levels in wing discs as measured by qRT-PCR on dissected wing discs (Fig 2G).
Brf1 is a conserved component of TFIIIB complex, which is required for Pol III recruitment to tRNA genes [42]. We previously showed that Brf1 is involved in controlling Pol III-dependent transcription, and tissue and body growth in Drosophila larvae [39]. Here we examined whether Brf1 is required for Ras-induced tRNA synthesis. We found that knocking down Brf1 using dsRNA in S2 cells (S2A and S2B Fig) suppressed the RasV12 induced increase in tRNA levels (Fig 3A) without altering the strong induction of ERK phosphorylation seen with RasV12 expression (S2B Fig). Hence, our data here suggest that the elevation of tRNA levels upon Ras activation is due to increased Pol III transcription. We then examined whether Brf1 is required for Ras-induced growth in Drosophila wing discs. We expressed UAS-driven transgenes in the dorsal compartment of the wing imaginal disc (using an apterous-Gal4 driver, ap-GAL4) and then, in each case, measured tissue size in wandering stage third instar larvae, Overexpression of UAS-EGFR (UAS-λtop) in the dorsal compartment of the wing imaginal disc stimulates Ras/ERK signalling and leads to tissue growth (Fig 3B and 3C). We found that RNAi-mediated knockdown of Brf1 by expression of a UAS-Brf1 inverted repeat line (UAS-Brf1 RNAi) in the dorsal compartment had little effect on tissue growth. However, expression of UAS-Brf1 RNAi blocked the overgrowth seen with UAS-EGFR expression. Expression of UAS-Brf1 RNAi with ap-GAL4 had little effect on tissue growth, suggesting we are not knocking down Brf1 to a level that cannot support any growth. We previously showed that Brf1 knockdown had no effect on ribosome synthesis, suggesting that its predominant effect was to block Pol III function [39]. Hence, these data indicate that Brf1 and Pol III transcription is required for EGFR/Ras/ERK-mediated increases in epithelial tissue growth in Drosophila.
A major role for the EGFR/Ras/ERK pathway is in the growth and maintenance of the Drosophila intestine. In larvae, activation of the pathway plays a central role in controlling the proliferation of adult midgut progenitor cells (AMPs), which eventually give rise to the adult intestine [9]. In the adult the EGFR/Ras/ERK pathway is required to promote stem cell proliferation and tissue regeneration [15–17,19,43]. We therefore examined whether Brf1-mediated Pol III transcription was required for these proliferative effects of Ras/ERK signalling. We first examined the larval intestine. During the larval period, AMPs proliferate and give rise to clusters of ~5–10 cells scattered throughout the larval intestine. These cell clusters eventually proliferate and fuse during metamorphosis to give rise to the adult intestinal epithelium. The EGFR/Ras/ERK pathway controls the proliferation of AMPS [9]. Overexpression of either UAS-EGFR or UAS-Rafgof in the AMPs using the temperature-sensitive escargot-Gal4 (esg-Gal4ts) system lead to a massive increase AMP proliferation and an increase in the numbers of AMP cells per cluster as previously reported. We found that expression of UAS-Brf1 RNAi (Fig 3D and 3E and S2C Fig) lead to a small reduction in the number of cells per cluster. However, we found that when co-expressed along with UAS-EGFR or UAS-Rafgof, UAS-Brf1 RNAi blocked the increase in AMP cell numbers. These data indicate Brf1 is required for EGFR/Ras/ERK mediated cell proliferation.
We next examined Brf1 function in homeostatic growth in the adult intestine. Damage to intestinal epithelial cells leads to an increase in expression and release of EGF ligands from both intestinal cells and underlying visceral muscle [15]. These EGF ligands then act on the intestinal stem cells (ISCs) to stimulate the Ras/ERK pathway, which triggers stem cell growth and division, and promotes regeneration of the intestinal epithelium. This damage-induced increase in ISC proliferation is dependent on EGFR/Ras/ERK signalling and can be mimicked by genetically activating the pathway specifically in the stem cells [15–17,19]. We tested a requirement for Brf1 in this Ras-mediated homeostatic growth response. We first examined the effects of intestinal damage. As previously reported [19], we found that feeding flies either DSS or bleomycin–two different gut stressors–leads to an increase in ISC proliferation. However, we found that this effect was inhibited when we knocked down Brf1 (using UAS-Brf1 RNAi expression) specifically in the ISCs and their transient daughter cells, the enteroblasts (EBs), using the inducible esg-Gal4ts system (Fig 4A and 4B). We next examined the effects of activation of the Ras/ERK pathway. We first overexpressed UAS-RasV12 in the adult intestine using the inducible esg-Gal4ts system, and observed an increase in pre-tRNA levels (Fig 4C). As previously reported, when we overactivated the pathway in stem cells by expressing UAS-Rafgof using esg-Gal4ts, we saw an increase cell proliferation as indicated by a marked increase in GFP labelled ISCs and EBs (Fig 4D). Expression of a UAS-Brf1 RNAi had little effect on GFP labelled cells, but when co-expressed with UAS-Rafgof it blocked the increase in cell proliferation. These results suggest that Brf1 and Pol III-dependent transcription is required for stem cell proliferation in the adult intestine.
We next wanted to examine how Ras signalling stimulates Pol III-dependent tRNA transcription. One candidate regulator we tested was dMyc. In both mammalian cells and Drosophila, Myc can interact with Brf1 and stimulate Pol III-dependent transcription [39,44,45]. Moreover, studies in both mammalian cells and Drosophila suggest Ras signalling can regulate dMyc levels and that Myc is required for Ras-induced growth[10,23,24,46,47]. Indeed, we found that the UAS-EGFR- and UAS-RasV12S35-induced proliferation of larval AMPs was blocked when we knocked down dMyc by expression of a UAS-dMyc RNAi construct (S3A–S3C Fig). We therefore examined whether dMyc functions downstream of Ras in the control of Pol III. Using S2 cells we found that the increase in tRNA levels seen following RasV12 expression was blocked when cells were treated with dsRNA to knockdown dMyc (Fig 5A). In contrast, we found that overexpression of dMyc in S2 cells was not able to induce tRNA synthesis when the Ras pathway was inhibited by treatment with the MEK inhibitor UO126 (Fig 5B). Under these conditions of Ras pathway inhibition, however, dMyc mRNA levels were not affected (Fig 5C) and overexpressed dMyc was still able to significantly stimulate expression of Nop60B, PPAN and NOP5—three dMyc Pol II target genes (Fig 5D)–although the effect on PPAN and NOP5 was somewhat reduced. Nevertheless, these data suggest that U0126 does not simply abrogate dMyc’s ability to stimulate transcription of its target genes, and that dMyc is required, but not sufficient, to mediate the effects of Ras signalling on tRNA synthesis. These data suggest that Ras/ERK signalling can use an additional mechanism to control Pol III transcription.
Another candidate that we considered as a mediator of Ras-induced tRNA synthesis was the conserved Pol III repressor, Maf1. Studies in yeast, Drosophila and mammalian cells have shown that inhibition of Maf1 is the main way that the nutrient-dependent TORC1 kinase pathway stimulates Pol III and tRNA synthesis [39,48–51]. Knockdown of Drosophila Maf1 (dMaf1) has been shown to promote tRNA synthesis, and to enhance tissue and body growth [40]. Here, we found that when we expressed UAS-dMaf1 RNAi in the Ras-responsive AMP cells during larval development using esg-GAL4ts, we observed a modest, but significant increase in the number of AMP cells per cluster (S4A Fig). Although considerably weaker than the effect of Ras pathway activation (e.g. see comparison with effect of UAS-EGFR, S4B Fig) this effect of dMaf1 knockdown was similar to the increase in AMP numbers seen with overexpression of dMyc, another stimulator of tRNA synthesis and mRNA translation (S4C Fig). We therefore next examined whether the Ras/ERK pathway functions to promote tRNA synthesis by inhibiting dMaf1. We examined pre-tRNA levels using qRT-PCR in S2 cells, and, as described above, we saw that treatment of cells with the MEK inhibitor UO126 led to reduced tRNA synthesis (Fig 6A and 6B). However, we found that this decrease in tRNA synthesis was reversed when cells were treated with dsRNA to knockdown dMaf1 levels (Fig 6A and 6B). We observed similar effects when we used Northern blotting to measure pre-tRNA and tRNA levels (S4D Fig). We also used treatment of cell with dsRNA to Ras to block Ras signalling, and saw a decrease in tRNA synthesis (S4E Fig). However, as with UO126 treatment, we found that this decrease in tRNA synthesis caused by dsRNA to Ras was reversed by co-treatment of cells with dsRNA to dMaf1. These data suggest that one main way that Ras/Erk signalling functions to promote tRNA synthesis is by inhibiting the Pol III repressor function of dMaf1.
Studies in both yeast and mammals indicate that Maf1can be regulated by controlling its nuclear localization (e.g [50,52]). We first tested this in S2 cells using an antibody to endogenous dMaf1. Under our normal media culture conditions, we observed that dMaf1 was localized throughout the cell (Fig 6C). When we carried out antibody staining in dMaf1 dsRNA-treated cells (which leads to a strong knockdown of both dMaf1 mRNA, S5A Fig, and dMaf1 protein, S5B Fig) we saw minimal background staining, suggesting that the antibody is specific for dMaf1 (S5C Fig). We found that treatment of cells with the MEK inhibitor U0126 lead to a significant increase in nuclear localization of dMaf1 (Fig 6C and 6D), without having any effect on overall dMaf1 protein levels (S5D Fig). We also found that genetic inhibition of Ras signalling in AMPs, by overexpression of dominant-negative Ras (UAS-RasN17), lead to an increase in nuclear localization of dMaf1 (Fig 6E and 6F). Similar results were seen when we used expression of either UAS-EGFR or UAS-Ras RNAi to block Ras signalling in AMPs (Supplemental Fig 6A). Thus, Ras/ERK signalling functions to prevent nuclear accumulation of dMaf1, hence blocking its Pol III repressor activity and promoting tRNA synthesis.
Previous studies showed that the TORC1 pathway can regulate Maf1 nuclear localization and repressor function via phosphorylation [48–51,53]. We therefore explored whether the Ras pathway could also control the phosphorylation status of dMaf1 in S2 cells. We used the phos-tag reagent, which slows the migration of phosphorylated proteins in SDS-PAGE gels, and hence helps resolve phosphorylated vs. non-phosphorylated versions of a protein on a western blot. For example, when we examined total ERK levels by western blotting following SDS-PAGE with Phos-tag, we observed two ERK bands. The relative levels of the upper band were reduced when we treated cells with the MEK inhibitor (Fig 7A), while levels of the upper band were increased in cells overexpressing RasV12 (Fig 7B), thus indicating this method can detect protein phosphorylation changes. We then examined dMaf1 protein levels in western blots following SDS-PAGE with Phos-tag. As with ERK, we observed two dMaf1 bands, and the relative levels of the upper band were reduced when we treated the sample with phosphatase prior to SDS-PAGE (S6B Fig), suggesting this upper band is a phosphorylated version of dMaf1. Also, like ERK, we found that relative levels of the upper band were reduced when we treated cells with the MEK inhibitor (Fig 7A), while levels of the upper band were increased in cells overexpressing RasV12 (Fig 7B). Together these data suggest that Ras signalling may regulate dMaf1 phosphorylation, and based on previous work with TORC1 signalling, this may be one way that Ras regulates dMaf1 nuclear vs. cytoplasmic localization (Fig 7C).
We propose that stimulation of RNA polymerase III and tRNA synthesis contributes to the ability of the conserved Ras/ERK pathway to promotes mRNA translation and growth. Our data indicate that Ras can control Pol III by inhibiting the Maf1 repressor, in part by preventing its nuclear accumulation. Maf1 is a phospho protein and studies in yeast and mammalian cells have described how phosphorylation can regulate Maf1 nuclear localization. For example, both TORC1 and PKA can phosphorylate Maf1 on several conserved residues[48–53]. This phosphorylation prevents Maf1 nuclear accumulation and allows both kinases to stimulate Pol III. In contrast, dephosphorylation of Maf1 by both PP2A and PP4 protein phosphatases leads to nuclear accumulation of Maf1 and Pol III repression [54–56]. Thus, it is possible that ERK may function by promoting Maf1 phosphorylation–either directly or indirectly–to prevent its function. Other mechanisms may also be important for Ras to simulate tRNA synthesis. For example, one study in mammalian cells showed that ERK could phosphorylate and regulate Brf1 function [57]. Also, Ras was shown to upregulate TBP, which can increase transcription by all three RNA polymerases [41], although we did not see a similar effect. Interestingly, we found that the decrease in tRNA synthesis caused by inhibiting Ras signalling could be completely reversed by dMaf1 knockdown. This result suggests that while Ras signalling may exert multiple effects to control Pol III transcription, inhibition of dMaf1 seems to be an important effector of Ras in the control of tRNA synthesis. Maf1 function is conserved suggesting that the Ras/ERK-dependent regulation of Maf1 and tRNA synthesis that we describe in Drosophila may operate in other organisms, particularly human cells.
Our data using the phos-tag reagent suggest that one way that Ras/ERK signalling may control dMaf1 is via phosphorylation. Previous studies in both yeast and mammalian cells have shown that the TORC1 pathway can control the nuclear localization and repressor function of dMaf1 via phosphorylation of several conserved residues [48–50,53]. One can therefore speculate that Ras signalling may work in a similar manner. Although further studies are required to identify if ERK directly phosphorylates dMaf1 and to identify the phosphorylated residues, it is interesting to note that two of the conserved TORC1 phosphorylation sites on dMaf1 are serine residues followed by proline, which are sites that are often phosphorylated by ERK, a proline-directed kinase.
We also show that the transcription factor dMyc is required for the effects of Ras on Pol III and tRNA synthesis. Previous work from both mammalian cells and Drosophila has shown that in some cells Ras can promote Myc levels and that Ras-mediated growth requires Myc function[10,23,24,46,47]. We previously showed that Drosophila Myc could stimulate expression of the Pol III transcription factor, Brf1, and also other Pol III subunits [39]. In addition, Myc can directly interact with Brf1 and localize at Pol III to directly stimulate tRNA transcription in Drosophila and mammalian cells [39,44,45]. We suggest that both these effects are under the upstream control of Ras/ERK signalling and may, in part, explain the requirements for Myc in Ras-induced growth in both animal development and cancer.
Given our findings with dMaf1 and dMyc, we attempted to address which of the two mechanisms—inhibition of dMaf1 or activation of Myc—might explain the main effects of Ras/ERK signalling on tRNA synthesis. To do this, we inhibited Ras/ERK signalling in S2 cells and then asked whether knockdown of dMaf1 or overexpression of dMyc could maintain tRNA synthesis. We found that, of these two manipulations, only dMaf1 inhibition could restore tRNA synthesis when ERK signalling was inhibited. We interpret these findings to suggest that, while dMyc is required for tRNA synthesis, it is the inhibition of dMaf1 that explains a substantial part of the mechanism of action of Ras/ERK signalling in the regulation of Pol III and tRNA synthesis. We previously showed that dMaf1 knockdown does not alter expression of dMyc target genes [39], suggesting that enhancement of dMyc function doesn’t explain why Maf1 knockdown can maintain tRNA synthesis in cells in which Ras/ERK signalling is inhibited.
Previous studies in mammalian cells have shown that Ras/ERK signalling can promote protein synthesis by stimulating translation initiation factor function. We suggest that inhibition of Maf1 represents another target of Ras/ERK signalling, and that the subsequent increase in tRNA levels may cooperate with enhanced translation initiation factor activity to promote maximal stimulation of mRNA translation. Most of the work on Ras-mediated gene expression has focused on the effect of several Pol II transcription factors identified downstream of Ras in Drosophila such as fos, pointed, and capicua [19–22]. Stimulation of Pol III transcription to enhance tRNA levels and mRNA translation may provide another layer of control on overall gene expression by Ras signalling. For example, translational control of cell cycle genes has been proposed as one way to couple growth signalling pathways to cellular proliferation [58,59]. Furthermore, selective translational regulation of certain mRNAs has been shown to regulate growth and metastatic behaviour of tumour cells [60–62]. It important to note though that we find that simply knocking down dMaf1 alone has only a modest effect on cell proliferation AMPs, compared to the strong hyperproliferative effect of overactivation of Ras signalling. This is likely because increasing Pol III is only one downstream effect of Ras signalling and that the full Ras effect on cell proliferation requires the coordinated increase in the expression of many genes. Indeed, it is likely that Ras stimulates the activity of all three RNA polymerases to drive cell growth and proliferation.
Ras is one of the most often overactivated or mutated pathways in cancer, hence our findings may also have implications for processes that contribute to tumour growth and metastasis. Indeed, there is increasing appreciation for potential roles for alterations in tRNA biology in cancer cells [63]. For example, tRNA expression profiling has revealed that levels of many tRNAs are elevated in different cancer types [64,65]. Interestingly, these changes in tRNA levels have been shown to correlate with codon usage in mRNAs whose expression also changes in cancer cells [66]. Several studies have reported that increasing the levels of specific tRNAs can promote tumour growth and metastatic behavior [67–70]. Previous work also showed that increasing tRNA levels alone is sufficient to drive growth in Drosophila [40,71]. Hence, an increase in tRNA levels caused by oncogenic Ras signalling may be a driver of tumour growth and progression, rather than simply a consequence of increased growth. Ras also controls other process such as cell fate specification, differentiation and cell survival. Many of these effects are mediated through translation and so may also rely on the effects of Ras on tRNA synthesis.
Flies were raised on standard medium (150 g agar, 1600 g cornmeal, 770 g Torula yeast, 675 g sucrose, 2340 g D-glucose, 240 ml acid mixture (propionic acid/phosphoric acid) per 34 L water) and maintained at 25°C, unless otherwise indicated. The following fly stocks were used:
For all GAL4/UAS experiments, GAL4 lines were crossed to the relevant UAS line(s) and the larval or adult progeny were analyzed. Control animals were obtained by crossing the relevant GAL4 line to either w1118 or yw depending on the genetic background of the particular experimental UAS transgene line. For the esg-gal4ts system, larvae and were flies were initially raised at 18°C and then for each experiment they were shifted to 29°C to inactivate the temperature sensitive GAL80 and to allow GAL4-mediated transgene expression.
Drosophila Schneider S2 cells were grown at 25°C in Schneider’s medium (Gibco; 11720–034) supplemented with 10% fetal bovine serum (Gibco; 10082–139), 100 U/ml penicillin and 100 U/ml streptomycin (Gibco; 15140). Stably transfected inducible RasV12 cells were a gift from the lab of Marc Therrien [73]. Stably transfected inducible dMyc cells were a gift from the lab of Paula Bellosta [74]. Both RasV12 and dMyc expression are under the control of a metallothionein promoter. For all experiments RasV12 or dMyc were induced by addition of copper sulphate to the culture media.
dsRNA Treatment of S2 cells: dsRNAs were synthesized with RiboMAX large-scale RNA production system (Promega) using PCR products from either cDNAs or genomic DNA (primer sequences in S2 Table). Cells were pretreated with 15 μg of dsRNAs in the absence of serum for 30 mins and then 2 mls of media plus serum was added, and cells were then incubated for 96 to 120 hrs. Control cells were treated with ds RNA to Green Fluorescent Protein (GFP). Cells were harvested by centrifugation at 4 °C and washed with cold PBS and frozen for RNA isolation or protein extraction.
MEK inhibitor (U0126) treatment of Drosophila S2 cells: S2 cells were cultured at 25°C in Schneider’s medium (Gibco; 11720–034) supplemented with 10% fetal bovine serum (Gibco; 10082–139), 100 U/ml penicillin and 100 U/ml streptomycin (Gibco; 15140). Cells were treated with either 10 μM U0126 (Promega Cat. No. V1121) or DMSO (Sigma; D2650) for 2 hours. Then cells were washed twice with ice-cold PBS. Cells were then used to isolate RNA or make protein extracts as described below.
Drosophila S2 cells were lysed with a buffer containing 20 mM Tris-HCl (pH 8.0), 137 mM NaCl, 1 mM EDTA, 25% glycerol, 1% NP-40 and with following inhibitors 50 mM NaF, 1 mM PMSF, 1 mM DTT, 5 mM sodium ortho vanadate (Na3VO4) and Protease Inhibitor cocktail (Roche Cat. No. 04693124001) and Phosphatase inhibitor (Roche Cat. No. 04906845001) according to the manufacturer’s instruction.
Drosophila S2 cells were lysed with a buffer containing 20 mM Tris-HCl (pH 8.0), 137 mM NaCl, 25% glycerol, 1% NP-40 and with following inhibitors 1 mM PMSF, 1 mM DTT and Protease Inhibitor cocktail (Roche Cat. No. 04693124001) and Phosphatase inhibitor without EDTA. Phos-tag SDS-PAGE was prepared according to the manufacturer’s instruction (Wako Chemicals USA, Inc). Cell lysates were separated on 12.5% SDS-polyacrylamide gel containing 20 uM Phos-tag acrylamide (AAL-107 Wako Chemicals USA, Inc), and transferred onto PVDF membranes (Bio Rad).
Protein concentrations were measured using the Bio-Rad Dc Protein Assay kit II (5000112). Protein lysates (15 μg to 30μg) were resolved by SDS–PAGE and electrotransferred to a nitrocellulose membrane, subjected to Western blot analysis with specific antibodies, and visualized by chemiluminescence (enhanced ECL solution (Perkin Elmer). Brf primary antibodies were against a C-terminal fragment of Drosophila Brf, alpha-tubulin (E7, Drosophila Studies Hybridoma Bank), dMyc [24], phospho-ERK (Cell Signalling Technology 4370) and ERK (Cell Signalling Technology 4695). Peptide antiserum against Drosophila Maf1 was raised by immunizing rabbits with synthetic peptide LADFSPNFRC corresponding to residues 65–74 (GL Biochem (Shanghai) Ltd).
10 μM puromycin was added to Drosophila S2 cell culture media and the cells were incubated with puromycin for 30 min at 25 °C. Cells were harvested by centrifugation at 4°C and washed with cold PBS. Cells were frozen on dry ice and then lysed according to the Western blot protocol described above and analyzed by SDS-PAGE and western blotting using an anti-puromycin antibody (3RH11) (Kerafast, Catalog No.EQ0001) at 1:2000 dilution.
Total RNA was extracted from Drosophila S2 cells using TRIzol. 5 μg total RNA was separated on a 5% denaturing polyacrylamide/urea gel and northern blotting was carried using alkaline transfer. Hybridization of tRNA probes were carried out as described in Roche DIG Easy Hyb (Cat. No.11603558001). Digoxigenin-labelled probes were made by in vitro transcription using either full-length cDNAs or PCR fragments as templates. Primers used for PCR are included in S1 Table.
Drosophila S2 cells were fixed in 4% paraformaldehyde at room temperature for 20 mins on cover slips. Cells were then washed with 1x PBS and permeabilized with 0.1% Triton X in PBS by washing 2x for 5 mins. Cells were blocked with 5% FBS, 0.1% Triton X in PBS for 2 hours. Primary dMaf1 antibody was diluted in 5% BSA in PBS at 1:500 dilution and incubated overnight at 4°C. Then washed 3x with 0.1% Triton X in PBS for 5 min each and Alexa 568 (Molecular probes) goat-anti rabbit secondary antibody was diluted at 1:400 in 5% BSA in PBS for 2 hours at room temperature. Then, cells were washed 3x with 0.1% Triton X in PBS for 5 min each and mounted using VectaShield mounting medium.
Drosophila larvae were inverted and fixed in 8% paraformaldehyde/PBS at room temperature for 45 mins. After blocking for 2hrs in 1%BSA in PBS/0.1% Triton-X 100, larval carcasses were incubated overnight in anti-dMaf1 antibody (1:1000). Primary antibody staining was detected using Alexa 488 (Molecular probes) goat-anti rabbit secondary antibodies.
For experiments looking at dMaf1 subcellular localization, we used Image J to measure dMaf1 staining intensity. Nuclear localization was measured was calculated by measuring the total intensity of signal in the nucleus and dividing this by the total intensity in the cytoplasm (calculated as total overall cellular signal intensity minus total nuclear signal intensity).
Total RNA was extracted using TRIzol according to manufacturer’s instructions (Invitrogen; 15596–018). RNA samples were DNase treated according to manufacturer’s instructions (Ambion; 2238G) and reverse transcribed using Superscript II (Invitrogen; 100004925). The generated cDNA was used as a template to perform qRT–PCRs (ABI 7500 real time PCR system using SyBr Green PCR mix) using specific primer pairs (sequences available upon request). PCR data were normalized to either actin or Glyceraldehyde-3-phosphate dehydrogenase (GAPDH) levels. Each experiment was independently repeated a minimum of three times. All primer sequences are in S3 Table.
Polysome gradient centrifugation was performed as described [40]. 100 million Drosophila S2 cells were lysed in 1 ml of lysis buffer (25 mM Tris pH 7.4, 10 mM MgCl2, 250 mM NaCl, 1% Triton X-100, 0.5% sodium deoxycholate, 0.5 mM DTT, 100 mg/ml cycloheximide, 1 mg/ml heparin, Complete mini Roche protease inhibitor (Roche), 2.5 mM PMSF, 5 mM sodium fluoride, 1 mM sodium orthovanadate and 200 U/ml ribolock RNAse inhibitor (Fermentas) using a Dounce homogenizer. The lysates were centrifuged at 15,000 rpm for 20 minutes and the supernatant was removed carefully. 150 to 250 g μg RNA was layered gently on top of a 15–45% w/w sucrose gradient (made using 25 mM Tris pH 7.4, 10 mM MgCl2, 250 mM NaCl, 1 mg/ml heparin, 100 mg/ml cycloheximide in 12 ml polyallomer tube) and centrifuged at 37,000 rpm for 150 minutes in a Beckmann Coulter Optima L-90K ultracentrifuge using a SW-41 rotor. Polysome profiles were obtained by pushing the gradient using 70% w/v Sucrose pumped at 1.5 ml/min into a continuous OD254 nm reader (ISCO UA6 UV detector) showing the OD corresponding to the RNA present from the top to the bottom of the gradient.
All qRT-PCR data and quantification of immunostaining data were analyzed by Students t-test, or two-way ANOVA followed by post-hoc students t-test where appropriate. All statistical analysis and data plots were performed using Prism software. In all figures, statistically significant differences are presented as: * p<0.05, ** p<0.005, *** p<0.0005, **** p<0.0001.
|
10.1371/journal.pcbi.1000351 | Towards Prediction of Metabolic Products of Polyketide Synthases: An In Silico Analysis | Sequence data arising from an increasing number of partial and complete genome projects is revealing the presence of the polyketide synthase (PKS) family of genes not only in microbes and fungi but also in plants and other eukaryotes. PKSs are huge multifunctional megasynthases that use a variety of biosynthetic paradigms to generate enormously diverse arrays of polyketide products that posses several pharmaceutically important properties. The remarkable conservation of these gene clusters across organisms offers abundant scope for obtaining novel insights into PKS biosynthetic code by computational analysis. We have carried out a comprehensive in silico analysis of modular and iterative gene clusters to test whether chemical structures of the secondary metabolites can be predicted from PKS protein sequences. Here, we report the success of our method and demonstrate the feasibility of deciphering the putative metabolic products of uncharacterized PKS clusters found in newly sequenced genomes. Profile Hidden Markov Model analysis has revealed distinct sequence features that can distinguish modular PKS proteins from their iterative counterparts. For iterative PKS proteins, structural models of iterative ketosynthase (KS) domains have revealed novel correlations between the size of the polyketide products and volume of the active site pocket. Furthermore, we have identified key residues in the substrate binding pocket that control the number of chain extensions in iterative PKSs. For modular PKS proteins, we describe for the first time an automated method based on crucial intermolecular contacts that can distinguish the correct biosynthetic order of substrate channeling from a large number of non-cognate combinatorial possibilities. Taken together, our in silico analysis provides valuable clues for formulating rules for predicting polyketide products of iterative as well as modular PKS clusters. These results have promising potential for discovery of novel natural products by genome mining and rational design of novel natural products.
| Polyketide synthases (PKSs) form a large family of multifunctional proteins involved in the biosynthesis of diverse classes of therapeutically important natural products. These enzymes biosynthesize natural products with enormous diversity in chemical structures by combinatorial use of a limited number of catalytic domains. Therefore, deciphering the rules for relating the amino acid sequence of these domains to the chemical structure of the polyketide product remains a major challenge. We have carried out bioinformatics analysis of a large number of PKS clusters with known metabolic products to correlate the chemical structures of these metabolites to the sequence and structural features of the PKS proteins. The remarkable conservation observed in the PKS sequences across organisms, combined with unique structural features in their active sites and contact surfaces, allowed us to formulate a comprehensive set of predictive rules for deciphering metabolic products of uncharacterized PKS clusters. Our work thus represents a major milestone in natural product research, demonstrating the feasibility of discovering novel metabolites by in silico genome mining. These results also have interesting implications for rational design of novel natural products using a biosynthetic engineering approach.
| It is well known that polyketide synthase (PKS) gene clusters can generate enormously diverse array of polyketide products by making use of various biosynthetic paradigms like, modular organization of sets of catalytic domains or iterative catalysis of condensation steps using single set of catalytic domains [1]. In view of the pharmaceutical importance of polyketides, there is tremendous interest in identifying PKS gene clusters capable of producing novel polyketides by genome mining. However, the relating the sequence of the various catalytic domains present in a PKS biosynthetic cluster to the chemical structure of the final metabolic product is a major challenge. The availability of the sequences of a large number of experimentally characterized PKS clusters and 3D structural information on homologous protein domains presents a unique opportunity to carry out in silico analysis for addressing structural and mechanistic issues concerning polyketide biosynthesis. A number of recent theoretical studies have demonstrated the utility of in silico analysis in providing novel insights into the mechanistic details of polyketide biosynthesis as well as in identifying novel natural products by genome mining. Computational analysis of polyketide synthase (PKS) and nonribosomal peptide synthetase (NRPS) proteins have provided valuable clues for development of knowledge-based methods for identification of catalytic domains in PKS [2],[3] and NRPS [4] proteins, prediction of the substrate specificity for AT domains [2],[3],[5] and adenylation domains [4],[6],[7]. Such predictions have also been experimentally validated by the recent successful reprogramming of the phthiocerol dimycocerosate (PDIM) biosynthetic pathway in Mycobacterium tuberculosis [8] and experimental characterization of a novel exogenous standalone enoyl reductase (ER) involved in PDIM biosynthesis [9]. Bioinformatics analysis of secondary metabolite biosynthetic pathways have also played a crucial role in discovery of novel natural products by genome mining [10]–[14]. Very recently it has also been demonstrated that, computational analysis of KS domains from trans-AT PKS clusters can give novel clues about the chemical structures of the final polyketide product [15]. Similarly, bioinformatics analysis of docking domain sequences (the original term applied to these regions was “interpolypeptide linker”, but the term docking domain is being increasingly used in recent literature) have given novel insight into the evolution of specificity in inter polypeptide interactions in modular PKSs [16]. Pioneering work at Ecopia BioScience using data mining approaches has also led to development of proprietary databases which can aid in genomics driven discovery of cryptic biosynthetic pathways [17] and utility of these databases have been demonstrated by identification of novel secondary metabolites [18].
Thus, these studies have established that knowledge based computational approaches can play a powerful role in elucidation of novel secondary metabolite biosynthetic pathways. However, for in silico identification of polyketide products of uncharacterized PKS clusters, the computational method should also take into consideration various different paradigms employed by PKS biosynthetic machinery [19]. Several excellent reviews [20],[21] describe the type I, type II and type III biosynthetic paradigms. Type I modular PKSs harbor distinct sets of catalytic domains, each set termed as a “module”. Each module is responsible for one condensation step and the number of modules in a modular PKS correlate directly with the number of ketide units in its biosynthetic product. In contrast, type I iterative PKSs are characterized by a single set of catalytic active sites which are used iteratively for several rounds of successive condensations till the final product is released. It was initially believed that bacterial PKSs are modular while fungal PKSs function in an iterative manner. However, discovery of mixed PKS clusters involving programmed iterative modules and several other deviations [22],[23] from conventional textbook PKS biosynthetic paradigms in various microbes indicate that PKS proteins are not amenable to simple classification based on species of their origin. Therefore, in silico methods should be capable of predicting from sequence information, whether a given PKS cluster is iterative, the number of iterative chain condensation steps catalyzed by it and crucial amino acids which control the number of iterations.
In contrast to type I iterative PKSs where a single multifunctional enzyme is involved in biosynthesis of the polyketide product, biosynthesis in type I modular PKS clusters often involve multiple ORFs, each containing several modules. Therefore, predicting the correct order of substrate channeling between various ORFs is crucial for deciphering the final metabolic product of a modular PKS cluster. Several lines of experimental evidence reveal that inter subunit interactions between C-terminal docking domain region of the upstream ORF and N-terminal docking domain region of the downstream ORF, play a crucial role in channeling of substrates from upstream domains to downstream domains [24]–[27]. Moreover, these interactions involving C-terminus and N-terminus amino acid stretches have been reported to increase the maximum velocity (kcat) of chain transfer of otherwise disfavored substrates by as much as 100-fold [28]. Structural studies using NMR suggest that, these terminal docking domain regions of PKS proteins adopt a specific 3-dimensional fold consisting of a four helix bundle structure [29]. In fact, after the elucidation of this NMR structure, the term ‘docking domain’ is being increasingly used in the recent literature to describe these terminal amino acid stretches, which were earlier called ‘inter polypeptide linkers’. Based on this structure, it has been proposed that recognition between upstream and downstream ORFs in a modular cluster is governed by formation of specific contacts in the docking domain. Several recent experimental studies [30],[31] have further validated the role of specific inter polypeptide contacts in controlling inter subunit communication in modular PKS clusters. Very recently NMR studies [32] have also elucidated the role of similar docking domains in governing protein-protein interactions in hybrid megasynthases. Even though these experimental studies have identified specific residue pairs involved in inter subunit recognition, no systematic analysis of experimentally characterized modular PKS clusters have been characterized to investigate whether correct order of substrate channeling in type I modular PKS clusters can be predicted based on these specific inter polypeptide contacts. It may be noted that, even though recent study by Thattai et al [16] has attempted to address this question, their algorithm for prediction of PKS multiprotein chain order has been tested on a hypothetical five ORF cluster with only six combinatorial possibilities.
In this work, we have carried out a detailed comparative analysis of the experimentally characterized modular and iterative PKS clusters with known polyketide products to address following major questions relating to in silico prediction of polyketide products. Is it possible to distinguish between modular and iterative PKS from their sequence alone? Can we predict the number of iterations a given iterative PKS protein would catalyze and identify crucial amino acid residues that control the number of iterations? Is it possible to predict the correct order of substrate channeling between various ORFs in a modular PKS cluster? We have carried out profile Hidden Markov Model (HMM) analysis of KS domains to identify signature profiles which can decipher whether a PKS protein is modular or iterative. Structural modeling of KS domains of iterative PKS proteins and analysis of their active site pockets have given novel insight into the structural features that dictate the number of iterations catalyzed by a PKS protein and crucial amino acids which control them. Similarly, comparative analysis of crucial inter polypeptide contacts between cognate and non-cognate pairs of ORFs based on the three dimensional structure of the docking domains have given novel clues for prediction of the correct order of substrate channeling.
KS domains are the most conserved among various catalytic PKS domains and are responsible of catalysis of the chain condensation step. We have analyzed them in detail to identify class specific conserved patterns which distinguish modular and iterative PKS systems. For KS domains, the total dataset comprised of 217 pure modular KS domains, 82 pure iterative domains, 19 enediyne, 43 trans-type and 34 KS domains from hybrid NRPS-PKS clusters. Apart from the sequences of 20 experimentally characterized bacterial type I modular clusters included in our earlier analysis [2], an additional set of 18 modular PKS clusters was used as described in Methods. Despite sharing a significant degree of homology ranging from 24% to 40% sequence identity, KS domain counterparts from modular and iterative PKSs and other PKS subfamilies, segregate into distinct clusters in a phylogenetic dendrogram (Figure S1). We have used profile Hidden Markov Models (HMMs) to quantify subtle position specific differences in the probability of occurrence of amino acids in various subfamilies of KS domains (See methods for description of various subfamilies). The available KS data set was divided into training and test set, and sequences belonging to the training set were used for building profile Hidden Markov Models by the HMMER package [33]. Benchmarking on the test set indicated that, these HMM profiles were highly sensitive, with a prediction accuracy of 100% for both enediyne and trans-AT sub families, 97% for pure iterative PKSs, 92% for modular KS domains and 88% for hybrid clusters. Therefore, using HMM profiles it is not only possible to distinguish between modular and iterative PKS with a very high accuracy, these profiles can also be used to classify an uncharacterized sequence of a KS domain into various subfamilies within modular and iterative systems. This result has interesting implications for genome sequencing efforts towards identification of novel PKS clusters, because from KS sequence alone, one can get clues about PKS family and decide whether to sequence the entire cluster or not.
The polyketide products of various iterative PKS proteins are biosynthesized by different number of iterative condensation steps and undergo varying degrees of reductions. Phylogenetic analyses of iterative KS domains revealed that the clustering of iterative PKS sequences is highly correlated with the number of iterations they perform and degree of reductions undergone by the metabolite during biosynthesis (Figure 1). The biosynthesis of polyketides, lovastatin and bikaverin involve eight condensation steps, but their final structures are different because of the different cyclization patterns. Our analysis suggests that, the sequence of KS domain encodes information about chemical structure of the polyketide product. Hence, KS sequences of lovastatin and bikaverin form two different clusters. Based on similar phylogenetic analysis, earlier reports have proposed that KS domains cluster into groups depending on whether the corresponding type I iterative PKS contains additional reductive domains [34]–[36]. We attribute this feature to a complex programming within the KS domains which enables specific molecular recognition of the products. The observed clustering in Figure 1 could thus be arising from sequence features, that control recognition of specific substrates which have undergone different degrees of chemical and structural modifications due to the presence of reductive domains. Therefore, we wanted to analyze the structural models of various iterative KS domains for identification of specific amino acids or sequence stretches that can potentially control substrate size and extent of unsaturation. The various iterative KS domains were modeled using comparative modeling approach (see Methods for details). The structural templates for various iterative KS domains were identified by BLAST search against PDB or by using threading approach. The E. coli KAS-II protein (pdbids 1KAS, 1B3N) were used as the templates for modeling these iterative KS domains. Since 1B3N was a ligand bound structure (Figure 2A), the putative active site pockets (Figure 2B) of various iterative KS structural models could be identified based on amino acids which were in contact with the bound ligand in 1B3N. The structural features of the active site pockets of different iterative KS domains were analyzed further to identify the cavity lining residues (CLRs) and cavity volumes following protocols described in the methods section. Active site residue patterns (Figure 2B) in these structural models allowed us to correlate the cavity volume and hydrophobicity of the active site pockets to the number of iterations and the degree of unsaturation of the polyketide products they synthesize.
The substrate binding cavity in the 1KAS is highly hydrophobic owing to its completely saturated substrate. Polyketides, on the other hand, may contain several hydroxyl groups and unsaturated double bonds. Accordingly, the catalytic pockets in the structural models of polyketide KS domains were found to be less hydrophobic compared to the FAS cavities. Table 1 compares PKS product characteristics with a variety of cavity features. We observed a distinct difference in pocket hydrophobicity within polyketides and it correlated negatively with the extent of unsaturation seen in the product (Figure 3A). For example, the T-toxin PKS model cavity is more hydrophobic than the methylsalicylic acid synthase (MSAS) model cavity and this correlates with the fact that T-toxin is a reducing PKS having a greater proportion of saturated carbons in its final product than the partially reducing MSAS polyketide. Interestingly, cavity volumes correlate positively with the number of iterations (or corresponding product size). We found that polyketide KS cavity volumes fall into three distinct groups; small, large and intermediate (Figure 3B and 3C). The smallest cavities (∼300Å3) belong to the MSAS type PKSs that perform three iterations. Intermediate sized cavities (∼800Å3) belong to the napthopyrone (NAP) like PKSs that iterate from five to eight times. The largest cavities, 1780Å3, were observed for the T-Toxin models that perform 20 iterations. Figure 2B depicts the residues that line the hydrophobic cavity of the template KAS-II protein (volume 934 Å3) and surround the ligand analogue cerulenin. A comparison of the modeled structures with the template FAS KS structure revealed that in case of MSAS and NAP, the backbones of the models had not altered significantly during modeling (Figure S2), and thus, their functional difference could be traced to specific cavity lining residues (CLRs) (Figure 4). Figure 5A and 5B show the surface topology of the small and intermediate sized cavities. Figure 5A depicts the modeled MSAS KS domain with two tyrosines protruding into the KS cavity from opposite walls and thus blocking the downward flow of the cavity along the dimer interface. These two cavity blocking residues correspond to positions 229 and 400 (1KAS numbering). Interestingly, the conservation profiles of the CLRs shown in Figure 4 revealed that these two Tyr residues are highly conserved in all PKSs which carry out three iterations. This further substantiates the important role attributed to these residues based on our structural modeling of the active site pocket. Remarkably, NAP type KS domains have an Ala at position 400, that allows the cavity to extend further down thus making their cavities similar to the FAS catalytic cavity, shown for reference in Figure 5C.
Structural analysis thus revealed how substrate binding sites of varying size and hydrophobicity can be generated in type I iterative KS domains by subtle variations of residues on similar backbone folds. The crystal structure of KS-CLF also highlights how specific residues can regulate chain length in type-II PKSs [37]. Our results on role of cavity volume in controlling number of iterative condensations or chain length of type I iterative PKS products is also supported by recent experimental studies involving swapping of KS domains in fungal iterative PKSs, where replacement of fumonisin KS domain by KS from lovastatin LDKS resulted in polyketides having short chain length [38]. Very recent experiments involving generation of altered fatty acid-polyketide hybrid products by rational manipulation of benastatin biosynthetic pathway [39] also suggest that number of chain elongations is dependent on the size of the PKS enzyme cavity. The in silico analysis of the sequence and structural features of iterative KS domains reported here provides a structural rationale for these experimentally observed variations in substrate specificities and further helps in identification of residues that can be specifically mutated to control the number of iterations in type-I PKSs. No experimental studies have as yet been reported on altering the number of iterations in type-I PKSs by site directed mutagenesis. The present in silico analysis gives crucial leads for such experiments.
In modular PKS clusters, the chemical structure of the product is governed by the order in which substrates are channeled between various ORFs. It has often been observed that the order of PKS ORFs during biosynthesis of a polyketide is not the same as the order of the corresponding ORFs in the genome. This complexity of module succession has been depicted in Figure S3 using schematic representation of a type I modular PKS cluster. This biosynthetic cluster has four polyketide synthase ORFs and their order in the genome is Orf1, Orf2, Orf3 and Orf4. But during the biosynthesis, Orf4 is the first to function and the product of Orf4 is transferred to Orf1. Orf2 functions at a later stage and its product is condensed with the rest of the polyketide. This inconsistency between ordering of ORFs in the genome and the order of substrate channeling is a commonly observed phenomenon, as is evident from the simocyclinone [40], nanchangmycin [41], microcystin [42], pimaricin, rapamycin and nystatin biosynthetic clusters. The prediction of the correct order of substrate channeling is essential for in silico identification of polyketide products of uncharacterized modular PKS clusters. Therefore, deciphering the cognate combination of ORFs in a modular PKS cluster from the large number of theoretically possible non-cognate combinations has been the major bottleneck in formulating predictive rules for in silico identification of polyketide products. Hence, we attempted to investigate whether predictive rules based on specificity of interaction between ORFs can be formulated for deciphering the correct order of substrate channeling in an uncharacterized PKS cluster.
Several experimental studies have suggested that inter protein interactions in modular PKSs are mediated by specific recognition between docking domains or the so called ‘interpolypeptide linker’ regions [24],[25],[29]. The amino acid stretches N-terminus to the first KS domain and C-terminus to the last ACP domain are referred as inter polypeptide linkers or docking domains. These have been extensively studied and it has been proposed that, the C-terminal (Cter) docking domains specifically pair with the N-terminal (Nter) docking domains of the succeeding ORF to facilitate cross-talk between the consecutive ORFs. Structural elucidation [29] of the cognate docking domains from erythromycin PKS (DEBS) has revealed that, unlike conventional linker sequences which join protein domains covalently within polypeptides, these docking domain regions are not non-structured, but adopt a relatively compact four helix bundle structure. It has been proposed that, this four helix bundle structure is the core fold of cross-talk [29] between ORFs of modular PKS clusters. These structures have been termed inter protein ‘docking domains’ to emphasize that they are responsible for the recognition and subsequent docking between successive protein modules. The C-terminal docking domain is reported to contain three helices (hereafter named helix 1, 2 and 3) whereas the N-terminal docking domain contains a single longer helix (hereafter named helix 4). This docking domain complex is a symmetrical dimer, consisting of two independent structural units called domain A and domain B. Domain A is an unusual intertwined α-helical bundle comprising helices 1 and 2. Domain B is also an α-helical bundle but with an entirely different topology and it comprises helix 3 (from Cter) and helix 4 (from Nter). Thus the actual docking interaction occurs in domain B, via several pairs of charged residues and a conserved set of hydrophobic residues. However, it has been proposed that, out of these various interacting residues, two pairs of appropriately placed charged residues at critical positions on the docking interface, form a kind of ‘docking code’ for DEBS [29] (Figure S4). When DEBS1 docks against DEBS2, the charges at these positions give rise to favorable interactions. However, in case of non-cognate combinations between DEBS1 and DEBS3, the resulting charge interactions are repulsive. The availability of DEBS docking domain structure provided us the opportunity to test, whether such a code exists in other PKS systems as well. We have carried out a structure based analysis of docking domain sequences to investigate if rules for identification of cognate ORF combination can be formulated based on key interactions found in DEBS docking domain structure.
It may be noted that, based on bioinformatics analysis of docking domains in type I modular PKS proteins, Broadhurst et al [29] had also proposed that DEBS-like docking domain structures would be present in other type I modular PKS clusters and they govern the cross-talk between ORFs. Since secondary structure analysis by Broadhurst et al [29] had clearly demonstrated propensity of docking domain sequences for four helix bundle structure similar to DEBS docking domain, inter polypeptide contacts were extracted for both cognate and non-cognate pairs of ORFs in various modular PKSs using the DEBS docking domain structure as a template. Since recent studies [16],[29],[43] suggest that PKS docking domains fall into at least three different phylogenetic classes, our assumption regarding docking domains from various phylogenetic groups adopting similar structural folds requires further justifications. It is well known that for a given protein family, structure is conserved to a much larger extent than sequence [44],[45]. There are many examples of proteins adopting similar three dimensional structural fold even in absence of detectable sequence similarity [44],[45]. Recently available structures [46] of mammalian type I FAS proteins also show remarkably high similarity to structures PKS protein domains even if they share only a limited sequence homology. Therefore, our assumption regarding myxobacterial PKS ‘docking domains’ adopting structural folds similar to docking domains from actinomycetes is not unreasonable. Hence, we extracted crucial interacting residues for various docking domain pairs based on alignment with DEBS docking domain structure. Figure 6 shows the alignment of cognate pairs of various PKS docking domain sequences with DEBS docking domain structure. The interacting residue pairs obtained from this alignment were ranked as favorable, unfavorable or neutral as per a simple scoring scheme (Table S1). The interactions between a pair of oppositely charged amino acids or between a pair of hydrophobic amino acids were ranked as favourable, while electrostatic repulsions between a pair of charged amino acids was called unfavourable. On the other hand, interactions between any other amino acid pairs, specifically the interactions between charged and hydrophobic amino acids was ranked as neutral. It may be noted that, this simplistic scoring scheme has been defined based on types of amino acid contacts found in interfaces of protein-protein complexes [47]. A total of 66 cognate pairs of docking domain sequences were checked for the two pairs of positions which give rise to favorable electrostatic interactions in the docking domain structure. Out of these, 54 pairs of ORFs were found to have at least one residue pair with favorable interaction. Moreover, there was no cognate pair where both of these interactions were unfavorable. Thus it can be concluded that cognate pairing of ORFs does generate energetically favorable contacts.
Since a good docking code interaction was observed in more than 80% cases, we investigated if these crucial inter polypeptide contact pairs could be used to predict the correct order of module succession in a given modular PKS. If all possible combinations of ORFs in a PKS cluster are considered together, there would be only one biosynthetically correct order of ORFs. This correct combination would in turn have a set of all cognate interfaces and therefore, the highest number of favorable interactions. The remaining combinations of ORFs would be incorrect and accordingly, they would have varying numbers of non-cognate interfaces, thus resulting in unfavorable interactions. It may be added here that, the identity of the first and last ORFs can usually be established by the presence of an initiating loading module and the terminal TE domain respectively. The presence of a very short C-terminal sequence beyond the conserved TE domain can also be used as a criterion for identification of the last module. Figure 7 shows the example of the Spinosad biosynthetic cluster which has ten modules arranged in five ORFs. These five ORFs can be combined in six different ways if the first and last ORFs are fixed. Each of the six combinations would have four interfaces. All the interfaces were scanned for favorable, unfavorable or neutral interactions at the positions corresponding to the DEBS docking code. As can be seen in Figure 7, the correct order of ORFs has the highest number of favorable interactions and no repulsive interaction at any of its interfaces. In contrast, each of the remaining five combinations has at least two repulsive interactions, and thus can be rejected in comparison with the correct combination.
A total of 39 characterized PKS clusters were analyzed in this manner to test the validity of this assumption. For a representative set of PKS clusters, Figure 8 shows in tabular format, the number of favorable, unfavorable and neutral contacts in the cognate combination and also the number of non-cognate combinations having a score better, equal or worse compared to the cognate combination. As can be seen from Figure 8, in several modular PKS clusters unfavourable interactions are present. However, the number of unfavourable interactions is much smaller than the favourable or neutral interactions present in the cognate interfaces. Thus analysis of cognate inter polypeptide contacts in 17 modular PKS clusters suggest that, both the interactions need not be favourable for effective docking domain interactions. However, non-cognate interfaces have more number of unfavourable interactions. Hence, there are relatively few non-cognate combinations having a score better than cognate combination. In ten out of 17 PKS clusters, no non-cognate combination has better score than the cognate combination. Even though there are non-cognate combinations having scores equal to cognate combination, the cognate combination can still be ranked among top few in these 10 cases. In case of four other PKS clusters, there are a significant number of non-cognate combinations having score higher then the cognate combination. However, the cognate combination can still be ranked within top 20% of all possible combinations. For example, in case of nanchangmycin 480 non-cognate possibilities have better score than cognate, 239 have scores equal to the cognate combination. Thus the cognate combination is ranked in top 720 combinations. However, the total number of combinatorial possibilities is 5040. Therefore, our computational method ranks the cognate combination in top 14% in case of nanchangmycin PKS cluster. It is important to note that, despite the large number of combinatorial possibilities, prediction based on docking domain sequences alone is able to reject a sufficiently high number of non-cognate combinations. Thus, our results on analysis of docking domain sequences indicate that, in more than 80% of the cases the cognate order of substrate channeling can be predicted correctly. However, we must clarify that, ‘correct prediction’ would mean eliminating significant number of non-cognate combinations and restricting the cognate combination to a relatively smaller number of possibilities. Such a relaxed definition of ‘correct prediction’ can be justified by the fact that, we are using a simple prediction method involving few crucial contacting residues rather than all the interactions present in the docking domain structure. Secondly, we are not taking into account role of other catalytic domains in preventing chain elongation in case of non-cognate associations.
Even though very recent theoretical studies [5],[16] have attempted to predict physical interaction between PKS proteins based on analysis of co-evolution of docking domain sequences, the prediction accuracy for order of substrate channeling has either not been studied in detail [16] or found to be low in cases involving clusters consisting of more than four ORFs [5]. However, in contrast to these purely sequence based methods, we have used a structure based approach. Using the conserved core structure of the docking domain as template, we have extracted crucial interacting residues which were suggested earlier by Broadhurst et al [29] to be determinants of specificity of inter subunit interactions. Exploitation of this crucial information in our study probably helps in improvement of prediction accuracy. Identification of specific interacting residue pairs also make the predictions easily amenable to experimental testing by site directed mutagenesis approach. Recent experimental studies [30],[31] have further established the feasibility of altering specificity of inter subunit interactions based on manipulation of putative interacting residues in the docking domain frame work. Apart from helping in deciphering the chemical structure of final polyketide product, our computational analysis of “docking code” in cognate and non-cognate interacting pairs in experimentally characterized modular PKS cluster can also provide knowledge base for fruitfully combining non-cognate ORF pairs for generation of novel aglycone structures. Our analysis of such interacting residues in docking domains of a mycobacterial PKS protein involved in biosynthesis of mycoketide has led to the discovery of a completely novel “Modularly iterative” mechanism of polyketide biosynthesis [48]. However, we must clarify that, apart from interactions between N-terminal and C-terminal docking domains of PKS proteins, the substrate specificity of various catalytic domains would also have a role in preventing chain elongation in case of non-cognate associations of PKS ORFs. Similarly, interactions between ACP and downstream KS will also discriminate non-cognate associations. In this work, we have only addressed the role of docking domains.
We have demonstrated that, the KS domains can be successfully classified into various functional subfamilies with high prediction accuracy using their HMM profiles. Structural modeling of the active site pockets of various iterative KS domains has revealed that certain key residues in the active site pocket can potentially control the size of final product by governing the total number of iterations. This result is in agreement with recent experiments [38],[39] which report cavity volume being a major determinant of substrate specificity of fungal PKSs. The major highlight of our work is that programmed iteration by fungal polyketide synthases may be rationally controlled by site directed mutagenesis of certain specific residues. These results also demonstrate that the number of chain extension reactions catalyzed by an iterative PKS protein can be predicted by computing the cavity volume of the active site pocket of its KS domain. This represents a major advance towards prediction of the polyketide products of iterative PKS proteins.
We have analyzed the docking domain sequences of various modular PKS clusters in detail to investigate if information contained in the docking domain sequences can be used to identify the correct order for channeling of substrates. Using the recently available NMR solution structure [29] of the docking domains from the erythromycin biosynthetic cluster as template, inter polypeptide contacts were analyzed for various types of cognate and non-cognate pairing of ORFs in various modular PKS clusters. Our investigation revealed that, cognate pairing of ORFs always generated energetically favorable inter polypeptide contacts, while in majority of cases non-cognate pairing resulted in energetically unfavorable contacts. The results of our benchmarking on known modular PKS clusters indicated that, using such inter polypeptide contact analysis, it is possible to narrow down the number of possible choices for the cognate order of substrate channeling. Thus our analysis of docking domain sequences would help in predicting the final polyketide products of modular PKS clusters.
In summary, the current work demonstrates that, in silico analysis of experimentally characterized PKS clusters can not only enhance our understanding of mechanistic polyketide biosynthesis, it helps in formulating rules for predicting, whether a given PKS protein is modular or iterative, the order of substrate channeling for modular PKSs, and the number of chain extension reactions catalyzed by iterative PKSs. Hence, our results can aid in identifying metabolic products of uncharacterized PKS clusters found in newly sequenced genomes.
In addition to the PKS gene clusters cataloged in the NRPS-PKS server, additional modular PKS clusters that were used for this analysis are ansamitocin [49], albicidin [50], Bacillus subtilis PKS, coronafacic acid, compactin CDKS [51], lovastatin LDKS [52], geldanamycin [53], leinamycin [54], lankacidin [55], microcytin (from two organisms) [56],[57], monensin [58], nanchangmycin [41], pederin [59], mupirocin [60], ta1 [61], bleomycin [62] and yersiniabactin [63]. The experimentally characterized fungal type I iterative PKS clusters used in this analysis are aflatoxin [64], avilamycin [65], bikaverin [35], C-1027 [66], calicheamicin (has two type I PKSs) [67], compactin [51], lovastatin [52], fumonisin [68], MSAS from four organisms [69]–[71], sterigmatocystin [72], THN from five organisms [73]–[76], T-toxin [77] and napthopyrone [78]. To this data, we added sequences analyzed in a previous phylogenetic analysis of fungal [79] type-I PKSs.
Profile HMM analysis [33] was carried out by HMMER package. The available KS dataset was divided into five different subfamilies. Apart from the major clusters of iterative and modular KS domains, the KS domain phylogenetic dendrogram showed further clustering into subfamilies like enediynes and non-enediynes within the iterative cluster. Similarly, modular KS domains have three clusters corresponding to pure modular PKSs, hybrid NRPS-PKSs and trans-AT systems. The enediyne family of antibiotics is structurally characterized by the enediyne core, a unit consisting of two acetylenic groups conjugated to a double bond or incipient double bond within the nine-membered or ten-membered ring. The enediyne cores bear no structural resemblance to any characterized polyketides, but precursor labeling experiments have unambiguously established that they are derived minimally from eight head-to-tail acetate units [80]. Natural products of hybrid peptide-polyketide origin have been known for a long time. These are metabolites that are assembled from amino acid and carboxylic acid precursors by hybrid NRPS-PKS gene clusters in which an NRPS-bound growing peptidyl intermediate is further elongated by a PKS module or vice versa [81]. Trans-AT clusters are also referred to as the AT-less clusters. These are complex PKSs where a single AT protein functions in trans- and charges the ACP domains of all the modules in the cluster [20]. Since the modular PKSs often have several KS domains on the same ORF, for building Hidden Markov Models of various subfamilies repartitioning of the various data sets into training and test set was done based on individual ORFs, rather than polyketide clusters or KS domains.
The various iterative KS domains were modeled using comparative modeling approach. The structural templates were identified by BLAST search against PDB or by using threading approach. Threading analyses were done using a local version of Threader package [82] (downloaded from the PSIPRED protein prediction server site) to identify the structural templates for modeling various KS domains. The various KS domains have been modeled using fatty acid KAS structure as template, which show only about 20% sequence identity with polyketide KS domains. However, availability of several structures of thiolase fold indicates that even at this low sequence identity, two KS proteins can adopt very similar structures. Since the overall active site architecture is conserved in this class of enzymes, our structural predictions are likely to be reliable even at low sequence identity between target and template. The crystal structure of the act KS-CLF protein and recently reported structure of DEBS KS have revealed that modular as well as iterative polyketide KS domains also adopt a thiolase fold, thus validating our assumptions.
Models of various polyketide KS domains were built using a local version of modeller V6.2 [83]. Structural mapping, ligand construction and pocket architecture visualization were done using different modules of InsightII package. The active site pockets of iterative KS domains were compared in terms of their hydrophobicity and cavity volumes to understand how binding pocket residues control chemical structure of the polyketide product. Cavity volumes were calculated using CASTp [84]. Only those cavities which contained the catalytic triad residues were chosen from the CASTp output for comparison across various models of a given KS domain. The cavity lining residues (CLRs) were identified from the selected CASTp pockets. The total number and total hydrophobicity of hydrophobic CLRs was tabulated for comparison with the FAS structural template. Hydrophobicity was calculated using Kyte and Doolittle's protein hydropathy scale [85]. Since cavity identification is often sensitive to small changes in orientation of residues, all the above mentioned parameters were calculated from at least five different homology models for the same sequence. Structural alignment of various KS structures was done using Combinatorial Extension (CE) server [86]. Visualization was also done using VMD [87].
Secondary structure propensities of various docking domain sequences were derived from the PredictProtein server [88]. ESPript service [89] from the predict protein server was used for structure based sequence alignment of docking domains. Interacting residues for each docking domain pair was identified by aligning their sequences with the docking domain structure. For each interface, the interacting residue pairs obtained from this alignment were ranked as favorable, unfavorable or neutral as per a simple scoring scheme (Table S1). A given combinatorial arrangement of a set of ORFs in a PKS cluster was assigned a score based on the favorable, unfavorable or neutral contacts present in all the interfaces. All the combinatorial possibilities were scored for each modular PKS cluster and score of the cognate combination was compared with scores of various non-cognate arrangements. The computational tool for carrying out inter subunit contact analysis involving docking domains and predicting the order of substrate channeling in modular PKS clusters is available as web server at http://www.nii.res.in/pred_pks_orf_order.html.
|
10.1371/journal.pgen.1005052 | A Genomic Portrait of Haplotype Diversity and Signatures of Selection in Indigenous Southern African Populations | We report a study of genome-wide, dense SNP (∼900K) and copy number polymorphism data of indigenous southern Africans. We demonstrate the genetic contribution to southern and eastern African populations, which involved admixture between indigenous San, Niger-Congo-speaking and populations of Eurasian ancestry. This finding illustrates the need to account for stratification in genome-wide association studies, and that admixture mapping would likely be a successful approach in these populations. We developed a strategy to detect the signature of selection prior to and following putative admixture events. Several genomic regions show an unusual excess of Niger-Kordofanian, and unusual deficiency of both San and Eurasian ancestry, which were considered the footprints of selection after population admixture. Several SNPs with strong allele frequency differences were observed predominantly between the admixed indigenous southern African populations, and their ancestral Eurasian populations. Interestingly, many candidate genes, which were identified within the genomic regions showing signals for selection, were associated with southern African-specific high-risk, mostly communicable diseases, such as malaria, influenza, tuberculosis, and human immunodeficiency virus/AIDs. This observation suggests a potentially important role that these genes might have played in adapting to the environment. Additionally, our analyses of haplotype structure, linkage disequilibrium, recombination, copy number variation and genome-wide admixture highlight, and support the unique position of San relative to both African and non-African populations. This study contributes to a better understanding of population ancestry and selection in south-eastern African populations; and the data and results obtained will support research into the genetic contributions to infectious as well as non-communicable diseases in the region.
| Genome-wide analysis of human populations is useful in shedding light on the evolutionary history of the human genome, with a wide range of applications from reconstructing past associations between different population histories to disease mapping. In this manuscript we report on the application of genome-wide data to southern African populations and the identification of genome-wide signatures of selection pre- and post-admixture. Several signals of selection, before and after admixture, were identified, some of which involved loci associated with human diseases, including malaria, influenza, tuberculosis and HIV/AIDS. These results may reflect adaptations of southern African populations to infectious diseases. Consistent with previous studies, this study highlights the significance of the San in the genetics of human populations, as they are distinct from the other populations in many respects i.e. haplotype structure, locations of recombination hotspots, copy number and population structure. Furthermore, our study demonstrates the admixture of the San, Bantu-speaking populations and populations of Eurasian ancestry in some of the southern and eastern African populations. It illustrates the value in correcting for this stratification in future genome-wide association studies, and suggests that a future admixture mapping in these populations would likely be warranted and successful.
| The analysis of high-throughput genotype data has revealed global patterns of human haplotype variation, casting light on the pre-history of human populations [1, 2, 3, 4, 5]. The International HapMap consortium [1,5]) and Human Genome Diversity Project (HGDP) [6], among others, have facilitated the analysis of human genome-wide variation, and linkage disequilibrium in disease association studies [1, 4, 5] and also helped refine estimates of recombination rates [7]. Comparative genome-wide genotype data among humans, Neanderthals and Chimpanzees have also shown that selection has played a significant role in human adaptation to the environment [8, 9, 10, 11]. These data have provided additional support for the African origin of modern humans [12,13] and highlight the effects of migration both within Africa and out of Africa. In general, African populations exhibit less linkage disequilibrium between adjacent markers than their non-African counterparts, consistent with a migratory bottleneck in the latter [1, 2, 5]. Such differences in the extent of linkage disequilibrium have a profound effect on the power of case-control association studies, since these studies depend largely on linkage disequilibrium between disease variants and genotyped single nucleotide polymorphisms (SNPs). Substantially more SNPs are required to capture genomic variation in African populations than populations of European ancestry [1, 5]. In addition, African populations are characterized by higher levels of genetic diversity [13, 14, 15, 16] and considerable population substructure [17, 18, 19], probably the combined result of several migration events, effective population size changes, population differentiation through genetic drift and local selective forces operating in ecologically diverse environments [18].
Hypotheses of migration within Africa based on mitochondrial DNA (mtDNA) suggest that at least three major migration events are plausible that could account for the patterns of mtDNA variation within Africa [17]; (1) the divergence of southern African San and east African populations who share the ancestral mtDNA haplogroup (L0d) and associated lineages in their maternal gene pool from an ancestral parental population circa 200 kya, (2) the establishment of west African maternal haplogroups (L1’5 & L0abf) from an east African source (circa 100 kya), and (3) the Bantu expansion from the Niger-Congo region into central, eastern and southern Africa (< 5 kya). Although a southern African versus east African origin of modern humans cannot be fully evaluated with current data, multiple lines of evidence from mtDNA [16], Y chromosomes [20], Alu insertions [21], and autosomal SNPs [3] place the divergence of the San at the root of modern humans with at least 100 ky of isolation from other non-San African populations [17, 22], and relatively recent (< 5 kya) admixture with Bantu-speaking populations [16, 23, 24, 25, 26, 27], followed by subsequent admixture (< 5 kya) in the region [16, 28, 29, 30]. Given this relative isolation of present-day San in southern Africa, it is expected that many SNPs ascertained in HapMap populations may not necessarily be polymorphic in San, unless the polymorphisms arose well before the divergence of these populations. Southern Africa was occupied exclusively by the San prior to the arrival of Bantu-speaking populations within the past 1,500 years, a consequence of the Bantu-expansion out of west Africa some 5000 years ago [16, 23, 24, 25, 26, 27, 31]. Migrations across equatorial central Africa to the region of the Great Lakes in east Africa, followed by southern African migrations [16, 25] established the eastern and southeastern Bantu-speaking groups, respectively. Migrations along the west coast of Africa contributed to western and southwestern Bantu-speaking groups, the latter, currently extending to Namibia [16, 25, 26, 27, 28, 29]. According to our findings, the label “Khoe-San” represent populations resulting from the mixture of predominately San, Eurasian and Bantu-speaking populations. Over hundreds of years, indigenous San and Khoe-San communities have undergone a sharp decline in population size, largely due to warfare and diseases such as smallpox which arrived with colonialists [29, 32]. It is estimated that the population decline (i.e. 90 percent) of both San and Khoe-San populations was due to smallpox [31, 32]. Recently, Lachance et al. [33] used the whole-genome sequences of five individuals in each of three different hunter-gatherer populations, including Pygmies from Cameroon, Khoe-San-speaking Hadza and Sandawe from Tanzania, and identified several genomic regions with evidence of archaic introgression in the hunter-gatherers. In addition, Lachance et al. [33] demonstrated that distribution of the time to the most recent common ancestors for these regions was similar to that observed for introgressed regions in Europeans [33]. Ancient and relatively recent contact between immigrants from Europe, Asia and Indonesia with sub-Saharan Africans [24, 26, 34] have resulted in varying degrees of admixture between these populations. Furthermore, a recent study by Gurdasani et al. [35] presented a broad survey of polymorphisms in a novel array genotyping data set of ∼1,481 individuals from 18 self-identified ethnic/linguistic and low coverage whole genome sequencing data set of 320 individuals from 7 self-identified ethnic/linguistic in Sub-Saharan Africa, and suggested that Eurasian back migrations to Africa and contributions to ancestry has a substantial impact on differentiation among some sub-Saharan African populations. These mixtures have also contributed to shaping the gene pool of the derived populations in south-eastern Africa [28, 35]. Other disciplines, such as archaeology, history and anthropology, have given us clues about the prehistory of African populations. The study by Pickrell et al. [16] convincingly demonstrated waves of two-way admixture between Niger-Congo-speaking African and west Eurasian (European or Middle Eastern) populations to form eastern and southern African (admixed) populations. However, the role of native indigenous San in the south-eastern African region and the genetic contribution of this population to the southern and eastern African admixed populations has not been elucidated. The present study makes use of genetic markers to investigate which factors, and to what extent, they have contributed in shaping the gene pools of extant southern and eastern African populations. More specifically, we used the Affymetrix Genome-Wide Human SNP Array 6.0, to examine ∼900K SNPs and copy number variants in five indigenous populations comprising 25 Ju\’hoansi San from Namibia (KHS), southeastern Bantu-speakers [25 Sotho-Tswana (STS), 36 Xhosa (XHS), 25 Zulu (ZUL)] as well as 25 Herero (HER), a southwestern Bantu-speaking group from Namibia. These data were used in conjunction with other published data to examine the genetic origins of southern African populations. Importantly, our study demonstrates the admixture of the indigenous San, Niger-Congo-speaking populations and populations of Eurasian ancestry in southern and eastern African populations. We have also developed two complementary approaches to identify signatures of selection prior to and following putative admixture events in the southern African populations.
The sample consisted of unrelated individuals belonging to the following five self-identified ethnic/linguistic populations of southern Africa: southeastern Bantu-speaking [25 Sotho-Tswana (STS), 25 Zulu (ZUL) and 36 Xhosa (XHS)], southwestern Bantu-speaking [25 Herero (HER)], and 25 Ju\’hoansi San (KHS). The Sotho-Tswana and Zulu samples were collected in Johannesburg, the Xhosa from Khayelitsha in Cape Town, the Herero from Windhoek, and the Ju\’hoansi from Tsumkwe [36]. The Blood samples were collected with the subject’s informed consent, and the use of DNA samples for population genetics research was approved by both the University of the Witwatersrand and University of Cape Town. DNA samples were shipped to Affymetrix (http://www.affymetrix.com) for genotyping using the Affymetrix Genome-Wide Human SNP Array 6.0, containing 906,600 SNPs and more than 946,000 probes for the detection of copy number variation. These data were used to examine patterns of migrations, genetic ancestry and effects of selection in this study. Other populations included in this study are listed in S1 Table.
The separation of Africans from non-Africans is clearly evident (Fig. 1 (A)); this has also been previously reported with both microsatellite data [37, 38] as well as with other SNP data [2, 3, 5]. From pairwise population genetic distance estimates, we find that there is little genetic difference among Bantu-speaking populations (S2 Table). In addition, Fig. 1 (A) shows a distinct separation of San populations (San (SAN) and Ju\’hoansi (KHS) and Khoe-San populations (Bushmen (BUS), ‡Khomani (KHO)), consistent with previous studies [16, 26, 33, 39, 40]. This result suggests Khoe-San, and both eastern and southern Bantu-speaking populations have undergone admixture. Furthermore, this result is consistent with the 3-population test [39, 40] result displayed in S3 Table, which shows clear evidence of admixture between Yoruba (YRI) and KHS in the southern Bantu (ZUL, STS, XHS). Furthermore, the ‡Khomani (KHO), and eastern Bantu-speaking populations also reflect a three-way admixture of Caucasian (CEU), Yoruba (YRI) and KHS. The results in Fig. 1 (A and D) suggest that the genetic make-up of the southeastern Bantu-speaking groups (ZUL, STS, XHS) includes ancestral contributions from Niger-Congo (26% ± 0.3%) and San populations (74% ± 0.4%). However, consistent with previous findings [40], the data in Fig. 1(B-C), suggests Niger-Congo ancestry (17% ± 1.2% and 57% ± 1.6%), San ancestry (70 ± 1.3% and 15% ± 0.4%), and notably Eurasian-related ancestry (13% ± 1% and 28% ± 2%) in the genetic make-up of ‡Khomani (KHO) and Sandawe (SAW), respectively. The admixture observed in the Khoe-San (KHO), and in the eastern African populations, (particularly) Sandawe (SAW) reflects the gene flow from Bantu-speaking agriculturalists and/or eastern African pastoralists within the past 1,200 years and sea-borne immigrants from Europe, Asia and Indonesia [33, 35, 39, 40, 41]. Our observation of Eurasian ancestry in both eastern (SAW) and southern (KHO) African populations is consistent with archaeological, genetic, climatological and linguistic data [24, 25, 26, 27, 28, 35]. Furthermore, Pickrell et al. [16] previously demonstrated multiple waves of population mixture in the history of many eastern and southern African populations, and that genetic material from Eurasians or related populations entered eastern Africa 2,700–3,300 years ago, and southern Africa 900–1,800 years ago [16, 41]. In addition, our study demonstrates the genetic contribution of the San population to the waves of admixture in the ancestry of the southern and eastern African populations.
Using the Mantel test with N = 10000 permutations (Materials and Methods), we found a significant positive correlation between genetic and geographic distance in the southern African populations (Pearson’s r = 0.64; p-value = 1.0 × 10−4; Fig. 2). To analyse more closely the outlier points in Fig. 2, we calculated the perpendicular distance between each point and the regression line. Analysing the concentration of points around the linear regression, we therefore defined outliers as points which are greater than 0.05 distance units from the regression line. When analysing the scatter plot (Fig. 2), there are 10 outlier points, which suggest possible obstacles to migration (S4 Table), assuming that populations have used the shortest path during their migrations. To assess patterns of migrations and to capture the genetic drift in southern African populations, we used a maximum likelihood tree and Gaussian approximation to the genetic drift model; implemented in Treemix [40]. We observed not only a major split between the African and European continent exhibited on this population tree, but also sub-lineages within African, and particularly within the southern African populations (S1 Fig.) which is consistent with previous results [16, 26, 34, 39, 40]. S1 Fig. (B) shows the inferred graph with three migration events, explaining the model for the relationship of southern and eastern Africans and non-Africans. This provides evidence for a shared origin for San-and Eurasian- and Bantu-related populations in Sandawe (SAW) and ‡Khomani (KHO). The latter possibility would be consistent with known south-east African admixture in the Sandawe (SAW) and ‡Khomani (KHO). We clearly see four population branches in southern Africa: (i) one formed from the southern Bantu-speaking populations, which are very distinct from the Niger-Congo and eastern Bantu-speaking populations, (ii) the second group formed with eastern Bantu-speaking populations, and (iii) the third, and (iv) the fourth group formed with San (KHS+SAN) and Khoe-San (BUS+KHO), both hunter-gatherers which are quite distinct, and are split into two distinct groups, including San populations (SAN and Ju\’hoansi (KHS)) and Khoe-San populations (BUS and KHO). This is also consistent with the admixture results shown in Fig. 1, reaffirming the concordance between genetic data with geographic origins of populations and their linguistic affinities.
Consistent with previous observations [13], the mean haplotype block lengths are substantially shorter in African populations than in non-Africans (Fig. 3 (A) and S5 Table). Mean block lengths are remarkably consistent across the southern African populations in this study and easily distinguishable from the non-African block lengths. Similarly, decay of linkage disequilibrium with physical distance along the genome is rapid in southern Africans when compared with non-Africans (Fig. 3 (B)). Ascertainment biases have been shown to result in faster decay of linkage disequilibrium compared to a sample of non-ascertained markers [42]. We performed coalescent simulations (S1 Text and S2 Text) in order to investigate the effects of ascertainment bias when markers are ascertained in a population divergent from that in which they are genotyped. Consistent with previous reports [42], we found the rate of decay of linkage disequilibrium to be greater with ascertained SNPs (S2 Fig. (A)). Similarly, haplotype block lengths are similar, irrespective of whether markers were ascertained in the genotyped population, or in a divergent population (S2 Fig. (A)). Frequency spectra, however, differ when SNPs are ascertained in a divergent population (S2 Fig. (A)). Indeed more monomorphic SNPs, and thus lower overall SNP diversity, are evident when markers are ascertained in a population divergent from that in which they are genotyped. This is further evident in distributions of minor allele frequencies from empirical data, in which the distribution of minor allele frequencies of San more closely resembles the theoretical expectation for a non-ascertained sample (S2 Fig. (B)), mostly due to the abundance of monomorphic SNPs. In addition to differences in demographic processes, such as bottlenecks, differences in the extent and pattern of linkage disequilibrium may be the result of differences in the patterns of fine-scale recombination rate. We assessed the impact of fine-scale recombination events to differences in linkage disequilibrium patterns using a coalescent-based method [7]. Interestingly, we found that the southern African Bantu-speaking populations share proportionally more recombination hotspots with both Yoruba (YRI) and Europeans (CEU) than with the Ju\’hoansi (KHS) (Fig. 4, S6 Table), where a shared hotspot is identified as a region with greater than five times the background recombination rate within a 10kb window. The proportion of hotspots shared between southern Africans and both European (CEU) and Yoruba (YRI) samples was generally low (Fig. 4). Our empirical analyses indicate that few recombination hotspots are shared between southern Africans and the HapMap populations, with San being the most extreme. More results on recombination hotspots and the test of whether increased frequency of low frequency and monomorphic SNPs improves the power to detect recombination hotspots are detailed in S4 Text and S7 Table.
To assess the accuracy with which missing SNPs in southern African populations can be imputed using Yoruba (YRI) or European (CEU) reference populations, we removed SNPs, imputed them and checked for correctness in imputation (detail in S1 Text and S3 Text). Our results show that YRI appears to be useful for imputation, at least for some of the southern Bantu-speaking groups included in the study, namely Sotho/Tswana (STS), Zulu (ZUL), Herero (HER) and Xhosa (XHS), but less so for the San, for whom imputation accuracy is significantly lower than for other African populations (S3 Fig.). Xhosa (XHS) also had lower imputation accuracy, compared with other Bantu-speaking groups.
We first developed an approach to select polymorphisms that exhibit large allele frequency differences between ancestral populations of Sandawe (SAW), Xhosa (XHS) and ‡Khomani (KHO) (see Materials and Methods). We constructed 3 different panels of AIMs [for Sandawe (SAW), Xhosa (XHS) and ‡Khomani (KHO)], where selected SNPs have a certain level of admixture LD with each other and with at least 1MB spacing between adjacent genetic markers on a chromosome (Materials and Methods). This was to avoid linkage disequilibrium (LD) in the ancestral population. Such background LD could contribute noise (or bias) to the estimation of ancestral allele frequencies and locus-specific ancestry [43]. Thinning down the SNPs to a 1Mb spacing may result in a reduction in power to detect cases of deviation in ancestry or allele frequency differences that result from selection. Consequently, our strategy to detect regions of unusual differentiation between the admixed southern African populations and their source populations, and unusual deviation in local ancestry, is conservative. We evaluated whether there is an excess of common SNPs with large allele frequency differences (expressed as a χ2 (1 d.o.f.) statistic under a model (see Materials and Methods) of neutral genetic drift) between putative ancestral populations of each admixed southern African population [‡Khomani (KHO), Sandawe (SAW) and Xhosa (XHS) (Table 1 and S5 Fig.)]. An unusual extent of population differentiation can suggest the action of population-specific natural selection. We observed several SNPs within chromosomal regions (Table 1) for which the evidence of unusual population differentiation was genome-wide significant between the Sandawe (SAW) and Caucasian (CEU) populations (S5 Fig.), and a small number of SNPs (on chromosome 17q25.1 and 12q24.21) showed unusual genome-wide significant differentiation between SAW and its two other putative ancestral populations, Yoruba (YRI) and Ju\’hoansi (KHS) (S5 Fig.). Chromosome region 3p11 yielded (to) a genome-wide significance of unusual differentiation between the Xhosa (XHS) and Ju\’hoansi (KHS) (p = 9.5e-10, lowest p-value), and between ‡Khomani (KHO) and Ju\’hoansi (KHS) (p = 7.6e-09, lowest p-value). Furthermore, unusual allele frequency differences between the Yoruba (YRI) and Xhosa (XHS) were identified on chromosome 1q41. No significant signal of unusual allele frequency differences between Yoruba (YRI) and ‡Khomani (KHO) were observed, which may be explained by the fact that the Niger-Congo contribution to admixture in the Khoe-San groups, in particular the ‡Khomani (KHO) (Khoe-San population) occurred too recently for it to have a significant impact on their allele frequencies. All these identified candidate SNPs of unusual allele frequency differences lie in or near known genes (Table 1). Their biological functions in the GeneCards database [44], are putatively linked with diseases of high prevalence in southern Africa; their detailed annotations are presented in Table 1.
We selected the best proxy parental populations of Xhosa (XHS) based on a pool of Click-speaking and Bantu-speaking populations using PROXYANC [45]. Yoruba (YRI) and Ju\’hoansi (KHS) were chosen as best proxy ancestral populations for Xhosa (XHS). Similarly, among the populations in the study, Yoruba (YRI), European (CEU) and Ju\’hoansi (KHS) were chosen as best non-San, European and San proxy ancestral populations for both ‡Khomani (KHO) and Sandawe (SAW) (Materials and Methods). Using AIMs panels, LAMP-LD [46] was employed to estimate the distribution of genetic contributions of ancestry across the genome (Materials and Methods) to provide additional reassurance from our data that we obtain unbiased results in the absence of possible background LD. The average locus-specific Ju\’hoansi (KHS) and Yoruba (YRI) ancestry proportions across the Xhosa (XHS) samples were estimated to be 27% ± 3.1% and 73% ± 3.1% (mean ± SD), respectively. We obtained 12% ± 0.8%, 77% ± 1.1% and 11% ± 0.9% (mean ± SD) locus-specific Yoruba (YRI), Ju\’hoansi (KHS) and Caucasian (CEU) average ancestry contributions, respectively along the genome of the ‡Khomani (KHO). For the Sandawe (SAW), the locus-specific ancestry proportions were 12% ± 0.9%, 70% ± 0.7% and 18% ± 1.0% for Yoruba (YRI), Ju\’hoansi (KHS) and Caucasian (CEU) average ancestry, respectively. The above estimates of average locus-specific ancestry are all consistent with the related genome-wide average proportion estimates in the admixture analysis section, indicating that there is no evidence of systematic distortion in our local ancestry estimates. The plots of these average locus-specific ancestries of these admixed southern African populations, namely Xhosa (XHS), ‡Khomani (KHO) and Sandawe (SAW) are in S6 Fig.. In the next two sections, we examined signals of selection, consisting of unusual deficiency or excess of ancestry in the admixed southern Xhosa (XHS), Sandawe (SAW) and ‡Khomani (KHO) populations. Such regions in admixed populations have served in previous studies as signatures of natural selection that occurred after admixture [43, 47, 48, 49, 50, 51]. Here, we considered not only the regions of strong deviation from ancestry, but we also implemented an approach that is now incorporated in PROXYANC [45] to test for unusual deficiency or excess ancestry using the inferred locus-specific ancestry across the genomes of admixed populations. The loci showing unusual ancestry patterns, i.e. four standard deviations above (excess ancestry) or below (reduced ancestry) the genome-wide average, were identified as candidates of post-admixture natural selection (Materials and Methods).
Examining the genome-wide distribution of ancestry in Xhosa (XHS), we detected the natural selection events post-admixture (Table 2). We identified a region on chromosome 3p11 (chr3: size: 17,184 (bp), p = 1.4e-10) with strongly reduced Ju\’hoansi (KHS) ancestry in Xhosa (XHS) (Table 2). This region yielded a genome-wide significance with an unusual difference of ancestry, suggesting a signal of selection after admixture. The SNP in the 3p11 region with the lowest p-value, rs4858960, is associated with POU1F1, which in turn interacts with five other genes [52], including ETS1, NR3C1, JUN, NR1I3 and MED1. These genes are known to play a role in a metabolic pathway that positively affects growth traits and hormone deficiency [53]. Furthermore, the 3p11 region showed strong differences in allele frequencies between Xhosa (XHS) and Ju\’hoansi (KHS) (p = 9.5e-10) (Table 1). Since San and Khoe-San communities have undergone a sharp population decline in their history, this differentiation suggests an environmental pressure that the San ancestors of the Xhosa (XHS) may have experienced before population admixture, and we speculate a possible adaptation of Xhosa (XHS) to the local environment. Mutations in the POU1F1/PIT1 gene, a pituitary-specific transcription factor, affect the development and function of the anterior pituitary and lead to combined pituitary hormone deficiency [53].
In spite of slight predominance of Ju\’hoansi (KHS), San ancestry in ‡Khomani (KHO) compared to Sandawe (SAW), and European (CEU) related ancestry in Sandawe (SAW) compared to ‡Khomani (KHO), consistent with previous findings [16, 26, 34, 40], our results from both admixture (Fig. 1) and locus-specific ancestry analyses (S6 Fig.) have shown a potential ancestral link between the admixed Sandawe (SAW) and ‡Khomani (KHO). Three chromosomal regions (12q24.1, 18p11.31 and 18p11.2), each within several SNPs with moderate and significant p-values, appear with excess of Yoruba (YRI) ancestry in both Sandawe (SAW) and ‡Khomani (KHO); an additional region (13q14.3) was also identified as an excess of Yoruba (YRI) ancestry in ‡Khomani (KHO), (Tables 2 and 3). These four candidate regions (Tables 2 and 3) showed strong unusual difference of ancestral contributions (p < 1.0 e-08, chi2 test), and have been associated with various important diseases, including malaria, T-cell leukemia, congenital muscular dystrophy, Noonan syndrome [53], and others listed in Tables 2 and 3. That some genes in these regions are associated with ‡Khomani (KHO)- and Sandawe (SAW)-specific high-risk diseases (such as malaria) [53], suggests a functional role these disease-related genes (or other genetic elements in these regions) might have played in their migration and particularly local adaptation due to such selective pressure resulting from shared gene-culture co-evolution and cultural practices in Bantu-speaking and Click-speaking populations. Overall, in the results of genome-wide allele frequency differences between Yoruba (YRI) and these two admixed populations (Tables 1, 2 and 3), only the 12q24.1 region was replicated significantly between Yoruba (YRI) and Sandawe (SAW). This may indicate different environmental pressures that the ‡Khomani (KHO) and Sandawe (SAW) experienced post-population-admixture.
We observed two other regions (12p13.31 and 14q13.2–14q13.3), with significant difference (Tables 2 and 3) of ancestry (p < 4.8e-08) showing a strong relative reduction of Caucasian (CEU) and Ju\’hoansi (KHS) ancestry in both ‡Khomani (KHO) and Sandawe (SAW). These regions were also identified as candidates of the natural selection after admixture (Tables 2 and 3). Importantly, these two regions (Tables 2 and 3) are also associated with some important diseases such as breast cancer, lung cancer, tumour inflammation, diabetes mellitus, Parkinson's and other diseases [44, 53], Although these regions have been associated with diseases, there is no indication of whether this points to any mechanistic association. However, it is tempting to speculate that factors such as food, pathogens, and life style, could also be responsible for such reduction in ancestry and may therefore play a role
Our approach to analyzing copy number variation in southern African populations involved the detection of known copy number polymorphisms (CNPs) using a Gaussian mixture model, and the identification of potential novel copy number variants (CNVs) using a Hidden Markov Model (HMM) (S5 Text). The number of CNPs (S5 Text) in Yoruba (YRI) is greater than that found in the European (CEU) and the southern African populations (Table 4). The former is probably the result of bottlenecks in non-Africans and subsequent loss of CNPs of low frequency [54, 55, 56], whereas the latter is likely the result of ascertainment bias. Given that CNP probes were ascertained in HapMap populations (including Yoruba (YRI)), lower levels of CNP diversity for populations that are divergent from ascertained populations is expected. However, southern African populations, which are approximately matched for sample size, show marked differences in the distribution of the number of CNPs, particularly in the San (Ju\’hoansi (KHS)) with fewer CNPs than other southern African populations (Table 4). Distributions of derived allele frequencies of CNPs suggest higher purifying selection on duplications (S7 Fig.). In contrast, however, there appears to be little difference in the degree of purifying selection on duplications and deletions in novel CNVs detected with the HMM (S7 Fig. (A)). We detected a total of 1873 CNVs (Table 5), of which 1231 were deletions. Only 137 of the CNVs were singletons, with 87 deletions and 50 duplications (Table 6). A total of 397 were novel with respect to the Database of Genomic Variants [55, 56, 57, 58]. At least 157 of these were unique CNVs, which occurred in only one population. The number of CNVs per individual is generally similar between populations (S7 Fig. (B)), except San which had significantly fewer deletions than other populations [e.g. Herero (HER) vs Ju\’hoansi (KHS)]: Student’s T-test, t20 = 22.4, P = 1.3e-15). Furthermore, distributions of derived allele frequencies of CNPs suggest purifying selection on duplications (S7 Fig. (A)). In contrast, however, there appears to be little difference in the degree of purifying selection on duplications and deletions in novel CNVs detected with the HMM (S7 Fig. (A)).
In this study, we have conducted a systematic population genomics survey and investigated demographic histories of indigenous southern African populations, making it possible to address questions about the signature of selection prior to and following purported ancient admixture events. Consistent with previous studies [16, 26, 33, 34, 35, 39, 40], we demonstrated stratification among indigenous southern African populations. Both the geographic distribution of genetic variations and the population structure, suggested a complex human population history generally within the African continent, and specifically in southern and eastern Africa. Incorporating the data from other Click-speaking populations from previous studies [16, 26, 33, 34, 39, 40] together with that from our 25 Ju\’hoansi (KHS) subjects, it was possible to investigate the relationship between Click-speaking and southern Bantu-speaking populations thought to represent an early diverging branch of modern humans.
The admixture analyses, particularly that of southern African populations, lends support of gene flow between San and Niger-Congo-speaking populations due to their contact following migrations of Bantu-speaking populations across the continent [17, 18, 26, 27, 33, 34, 35]. Consistent with previous studies [16, 26, 33, 34, 39, 40], our admixture (Fig. 1) and tree-mix analyses (S1 Fig.) suggested a division between south-west (San) and south-east (Khoe-San mostly admixed) populations. Our findings confirm an ancient link between San and some eastern African populations, including Sandawe, consistent with previous findings [16, 26, 35, 34, 39, 40]. The Eurasian ancestral components in south-east Khoe-San and some eastern Bantu speaking populations (such as Sandawe, Hadza) may be a consequence of an early Eurasian genetic contribution into Africa [16, 28, 35], Furthermore, the f-3 statistic test (S3 Table) confirms southern Bantu speaking populations, in particular Xhosa (XHS) to be two-way admixed, and both ‡Khomani (KHO) and Sandawe (SAW) are at least three-way admixed. The San (KHS) exhibit higher levels of homozygosity (S9 Table), increased relatedness (S9 Table) and higher proportions of monomorphic SNPs (S8 Table) than other African populations. However, we have shown that ascertainment of markers in a divergent population results in a reduction of diversity in the genotyped population, probably the result of polymorphisms arising after the divergence of the ascertained and genotyped populations, and the loss of polymorphisms in the genotyped population through fixation. Improved statistical models are therefore needed for the comparison of populations that have varying degrees of divergence from the population in which markers were ascertained.
Our copy number analysis included identification of both known CNPs, which are copy number loci previously identified in HapMap populations [55, 56, 58], and putatively novel CNVs. CNPs are highly ascertained, since they have been selected to be polymorphic and segregating at allele frequencies > 1% in HapMap populations [56]. CNVs, however, are less ascertained and should have more similar levels of polymorphisms in all of the studied populations [55]. In the case of CNVs, deletions are observed more frequently than duplications. This appears to be inconsistent with the proposal that deletions are under stronger purifying selection [58, 59, 60], which has also been inferred previously based on a lower degree of overlap between deletions and both genomic regions [59], and disease-related genes [59]. However, the disparity in the number of deletion and duplication CNVs probably reflects the relative difficulty of detecting the latter, due to a smaller relative change in copy number (3:2 versus 2:1) [59], rather than stronger purifying selection on duplications. In the southern African data, deletions and duplications have similar distributions to that of derived allele frequencies for CNVs, suggesting little difference in the relative degree of purifying selection. The number of deletion CNVs per individual differs markedly between the San (KHS) and other African populations. This may be an effect of sample size; however Herero (HER), with a similar sample size to San (KHS) for copy number calling, have no reduction in the number of deletions. In addition, copy number variants called for the Zulu (ZUL) panel with only 20 samples, were more than 99.9% concordant at normal, and 81.6% concordant at abnormal copy number regions, with those called in conjunction with other Bantu populations. Alternatively, some hybridization probes may have lower intensities in the San (KHS) due to probe-target mismatch mutations. However, such probe effects are likely to cause increased numbers of deletions in the San (KHS). Finally, population demographic and selective effects may cause differences in the number of deletion CNVs. In summary, copy number results suggest San (KHS) to be unique, although they should ideally be validated using trios, as shown previously [55, 56].
Haplotype blocks show very similar patterns of linkage disequilibrium between African populations, with this collective group having substantially shorter haplotype blocks, and less linkage disequilibrium, than Non-African populations. For instance, patterns of linkage disequilibrium surrounding the lactose tolerance (LCT) gene, known to have undergone a selective sweep in Europeans [7], have strong levels of linkage disequilibrium in Europeans, yet not in southern African populations (S2 Fig. and S4 Fig.). Khoe-San, however, appear to have increased levels of linkage disequilibrium associated with LCT than the other African populations [particularly the Sotho/Tswana (STS) and Zulu (ZUL); S2 Fig.]. This may be due to a weak selective sweep or the result of gene admixture with the San (KHS), a pastoral group from Namibia known to be lactose tolerant [29].
In addition, it was particularly interesting to examine the signature of selection in the indigenous and admixed southern African populations, including ‡Khomani (KHO), Xhosa (XHS) and Sandawe (SAW) due to the high mortality of the San population, historically. Following the recommendation of Bhatia et al. [61], we additionally implemented two strategies to detect possible evidence of population-specific natural selection in southern African populations. The first strategy, involved evaluating whether there is an excess of common SNPs with large allele frequency differences between admixed southern African populations, including ‡Khomani (KHO), Sandawe (SAW) and Xhosa (XHS) and their purported parental populations. The power of this analysis was based on an approach we developed to select three panels of 502 SNPs with at least 1MB spacing between adjacent genetic markers on each individual chromosome. Several SNPs on chromosomal regions for which there is evidence of unusual population differentiation between Sandawe (SAW) and Caucasians (CEU), are displayed in Table 1. Importantly, most of the signals of selection identified through this strategy are linked with specific high-risk diseases such as malaria, influenza, tuberculosis, and AIDs/HIV, which have a high prevalence in southern African populations (e.g. in the Sandawe, ‡Khomani and Xhosa populations) (Table 1). The allele frequency differences between southern African populations (including some putative parental populations) follow the null distribution predicted by neutral drift as a consequence of the recent origin of southern African population structure. This may yield a risk of false positive associations due to population stratification in disease association studies, despite the fact that there are differences between southern African populations [62].
The second strategy to detect possible evidence of population-specific post-admixture selection involved a signal of unusual excess or deficiency of ancestry in the admixed southern African populations [‡Khomani (KHO), Sandawe (SAW) and Xhosa (XHS)]. The recent studies by Bhatia et al. [61, 63] showed that loci with significant deviation in local ancestry (from the genome-wide average) may due to insufficient correction for multiple hypothesis testing and/or due to possible systematic errors in local ancestry inference. We have employed the minor allele frequencies from the correct proxy ancestral populations of the admixed population to correct for possible systematic errors on the inferred local ancestry that may lead to false positive deviations in local ancestry. Moreover our study did not only rely on the deviation (more than 4.0 standard deviations) in local ancestry from the genome-wide average; we additionally used the distribution of difference in locus-specific ancestry along the genome admixed population to evaluate the genomic regions showing unusual excessive or reduced ancestry which are likely to be signatures of natural selection after admixture [43, 48, 49, 50, 51].
Several recent studies have detected excessive or reduced ancestry contributions in admixed populations as signals of post-admixture selection, using reference ancestral parental populations [43, 48, 49, 50, 51]. Our study used selected best proxy ancestral populations and AIMs panels for our admixed southern African populations, and we extended previous approaches to test for unusually increased or decreased ancestry contribution along the genome. We identified three and four regions showing a significant excess of Yoruba (YRI) ancestry in Sandawe (SAW) and ‡Khomani (KHO), respectively (Tables 2 and 3). Three other regions showed unusually reduced Caucasian (CEU) and San (KHS) ancestry in both ‡Khomani (KHO) and Sandawe (SAW) (Tables 2 and 3). Since some of the genes in these regions are linked with specific high-risk diseases such as malaria in the ‡Khomani (KHO) and Sandawe (SAW), as has also been noted in the recent study by Gurdasani et al. [35], it is plausible that these disease-related genes might have played a role in population adaptation historically. Among the identified genomic regions, the 12q24.1 region was found in both strategies for detecting signals of natural selection, supporting evidence of environmental pressures that the ‡Khomani (KHO) and Sandawe (SAW) experienced. Furthermore, two other candidate regions pointing to natural selection were identified in both ‡Khomani (KHO) and Sandawe (SAW), showing strong deficiency of European and San ancestry components, and also an unusual population differentiation in these regions. These two regions are also linked with some important diseases such as breast cancer, lung cancer, inflammation, diabetes mellitus and Parkinson's disease [53], which are known to occur at a relatively higher prevalence in European populations, when compared to indigenous southern African populations [59].
African, and particularly southern and eastern African populations, face a heavy burden of diseases including HIV/AIDs, tuberculosis and malaria, and a growing burden of non-communicable diseases [17]. Of note, all the reported regions with signals of selection are in admixture LD and with significant deviation in average local ancestry (or unusual difference in allele frequency). In addition, our constructed AIMs panels for southern and eastern admixed populations may potentially be utilized for further admixture mapping studies in these populations. Nevertheless, further investigations are required to reveal the targets and agents of selection that have played important roles in shaping the admixed gene pool of these southern and eastern African admixed populations. With extensive admixture, both between none-San and San populations, and between African and non-African populations, southern and eastern African populations have a great potential for the identification of genes which determine susceptibility to both communicable and non-communicable diseases and to understand the African genetic variations with response to drugs/treatment variability.
The southern Bantu and Khoe-San populations are 'admixed' and future genome-wide studies will need to correct for this stratification or may need to use the locus-specific ancestry to increase power in association studies. Admixture mapping in the African-American and some other three-way admixed populations (such as Latinos, Puerto) has been successful for some disease traits [43, 51]. Since the admixed southern African populations have similar admixture proportions to admixed American populations, we hypothesize that admixture mapping would likely be a successful approach in many southern Bantu and Khoe-San cohorts, and particularly in the Xhosa, ‡Khomani and Sandawe.
A large proportion of the currently active genomic studies being conducted as part of the recently launched H3Africa programme (H3Africa, http://h3africa.org/) and the more recently described African Genome Variation Project [35], involve genome wide association studies [64]. A significant number of these studies involve large collections of sub-Saharan African subjects, and would benefit from this knowledge.
This study, investigating the genomic structure of indigenous southern African populations, was approved by the Research Ethics Committees of the University of Cape Town, and Witwatersrand University (REC Ref 305/2009 for the Project: Genome Wide Microarray Analysis of southern African Human Populations [65, 66].
Consider a pair of populations k and l from a pool of K ancestral populations of an admixed population and assume that the minor allele frequencies at SNPs i and j are greater than 0.005. Similar to Glaubitz et al. [67], we defined the admixture linkage disequilibrium as
Lij=mLijk+(1−m)Lijl+m(1−m)δikl×δjkl
(a)
Where m is the ancestral proportion, δi and δj are differences in allele frequency at SNPs i and j in population k and l, respectively. Assuming for each pair of SNPs i and j there is no linkage disequilibrium in ancestral populations, it thus follows,
Lij=m(1−m)δikl×δjkl
(b)
1=m(1−m)δikl×δjklLij
(c)
At a given pair of SNPs i and j in the admixed population, Equation (c) establishes a relationship between the observed linkage disequilibrium Lij in a recently admixed population and ancestral population differentiation. One can expected the ratio (part 2) in Equation c to be closer to 1 when the two reference ancestral populations contributed to the admixture of the related admixed population. Equation (c) is a total ancestry content (AC) at a pair of SNPs i and j. Let Iij denote the ration in Equation c, assuming a uniform ancestral proportion, and summing Equation (c) over all possible pairs of proxy ancestral populations, we can obtain the ancestry informativeness Iij of each pair of SNPs i and j as follows,
Iij=14K∑k≠lδikl×δjklLij
Let M be the total number of SNPs. For i ∊ {1,…, M}, let Ni be the total number of pair-wise LD j with i, where j ≠ i, ∀ j ∊ {1,…, M} within SNP i, we obtain the ancestry informativeness at SNP i as a weighted sum of Iij,
Ii=∑j=1NiIijM.
We applied this method to construct the AIMs panel for Xhosa, ‡Khomani and Sandawe. This approach of selecting ancestry informative markers (AIMs) is implemented in the PROXYANC program (http://web.cbio.uct.ac.za/proxyanc/).
We estimated the pair-wise genome-wide level of relatedness using a previously described relatedness statistic [67] applied to a random selection of 2500 putatively unlinked SNP markers with minor allele frequencies between 0.3 and 0.5. These SNPs were randomly selected across each chromosome, with a minimum spacing of 1 MB, to prevent inclusion of SNPs in strong linkage disequilibrium, which would violate the assumption of marker independence. Principal Component Analysis (PCA) was performed, using EIGENSOFT [68], on the combined HapMap3, HGDP, other African data from [26, 34, 39, 40] and southern African genotypes, which included a total of 50K SNPs shared between these different panels. In addition to the PCA analysis, an FST matrix using the smartpca program was generated. Admixture analysis [68, 69] was performed on combined panels based on 900K SNPs using the ADMIXTURE program [69]. To evaluate the genetic relationships among the above populations, we used the TreeMix software [40] to infer the structure of a graph from genome-wide allele frequency data and a Gaussian approximation to genetic drift. Furthermore, to identify some aspects of ancestry not captured by the tree, we also examined the residuals of the model’s fit and sequentially added the migration events to the tree. We also used copy number variants as a population marker in an additional population structure analysis, but only for HapMap3 and southern African samples for which the intensity data (CEL files) necessary for copy number calling were publicly available. Copy number variants, detected with a Hidden Markov model that identifies novel copy number variation [55], were preferred over previously described copy number polymorphisms, since these are affected to a lesser extent by ascertainment bias. We randomly selected a total of 2869 copy number variable positions, corresponding to 1 marker every 1Mb, across all chromosomes and specified copy number alleles as either a deletion, normal or duplicated state dependent on the copy number state called in the Birdseye algorithm [55]. We only selected simple copy number variants consisting of either a deletion or duplication, but not both.
Here, we used all available southern African population data, including HER, SAN, XHS, XHS, LWK, BUS, ZUL, SAW, a Niger-Congo-speaking population (YRI) and a non-African population, which included CEU. We made use of the Haversine formula to compute the geographic distance (in kilometre) between pairwise populations based on great circle distances using the way points between continents. The way-points used are Egypt (29.998392, 30.999751) and Turkey (41.015472, 27.986336). Thus, we computed the correlation between FST and Geographic distance using a linear regression equation as
FST= 1.298 × 10−5× Geographic distance + 1.709 × 10−2
We analysed the scatter plot of the relationship between FST and geographic distance. To address this, we computed the perpendicular distance between each point and the regression line. This enabled us to define outliers as points whose distance to the regression line is greater than or equal to 0.05 units.
To minimize deviation from the normality assumption, SNPs with minor allele frequencies < 0.05 are excluded. Thus, at a given locus i, the difference
(pik−pil)
between observed variant allele frequencies of two populations, k and l, can be approximated as a normal distribution under neutral drift with mean 0 and variance [60]
p(1−p)(2FST+1Nk+1Nl),
(d)
Where FST is the genetic distance between the population k and l. To avoid overestimating the degree of differentiation at single SNPs due to sample size difference, we used the estimator of FST in by Bhatia et al (63). Nk and Nl are total variant allele counts in each population, and p is the ancestral allele frequency that is commonly approximated as the average of the two observed variant allele frequencies. Similar to [60], we test unusual difference in allele frequency Ukl from population k and l as follows t
Ukl1=(pik−pil)2p(1−p)(2FST+1Nk+1Nl),
(e)
Ukl2=(pik−pil)2p(1−p).
(f)
Equations e and f are the χ2 distributed with 1 degree of freedom (d.o.f), and can be applied to unrelated (Equation b) and related samples (Equation c), respectively. An excess of large values of the χ2 statistic indicate deviations from the null model equation (Equation e and f), suggesting the action of natural selection [60]. We applied this method to the data from the Xhosa population using Ju\’hoansi and Yoruba as ancestral populations. We also applied this method to KHO and SAW using KHS, CEU and YRI populations. All gene annotations and associated diseases were obtained using both the GeneCards and MalaCards databases [44, 53].
We used LAMP-LD to infer locus-specific ancestry in admixed populations [46]. The model in LAMP-LD leverages the structure of linkage disequilibrium in the proxy ancestral populations. LAMP-LD achieved highest accuracy in both simulation and real data in the study of Puerto Rico and Mexico populations [43]. Here, we applied LAMP-LD to infer local ancestry in three potential southern African populations, including KHO, XHS and SAW. Following the population structure result and the proxy ancestry selection approach developed in PROXYANC [45], YRI, KHS and CEU was selected as reference ancestral populations from a pool of Bantu-speaking, Click-speaking and European populations, respectively. We obtained phased haplotype data by running Beagle software [70] on KHS, CEU and YRI data. To estimate the distribution of genetic contributions of ancestries to XHS across the genome, we used haplotypes of 80 YRI and 80 KHS. In addition, the haplotypes of 80 YRI, 80 CEU and 24 KHS were used to compute the locus-specific genetic contributions to KHO and SAW using the AIMs panel.
Admixed populations provide special opportunities for investigating recent selection. Prior to admixing, the ancestral populations have been isolated geographically, and their genomes may have evolved in distinct environments. Migration of previously isolated populations may have brought individuals of the ancestral populations into an unusual environment, and may consequently introduce life-style changes or changes in pathogens they are exposed to. This type of selection may differ from that faced by stationary populations, for which the local environmental changes may occur gradually, allowing for rare advantageous alleles to increase in frequency [43]. Here, we adopted an approach to detect ancestral signatures of selection by looking in an admixed population for genomic regions that exhibit unusually large deviations in ancestry proportions compared with what is typically observed elsewhere in the genome.
Given the genome-wide ancestral proportions, αk, from ancestral populations k ∊ {1, …, K} in N samples of an admixed population, let
φki,m
be the estimated locus-specific ancestry of individual i at genetic marker m ∊ {1, …, M}, from the kth ancestral population. We computed the deficiency or excess of ancestry, at each SNP using the estimated admixture proportion as a baseline. We thus define the deficiency/excess of ancestry from ancestral population k at marker m as,
δkm=(1N∑φki,m)−αk=φ¯km−αk
where
φ¯km
is the average locus-specific ancestry at SNP m.
δkm
can be approximated as a normal distribution under neutral drift with mean 0 and empirical variance, derived from the distribution of
φki,m
values among the N individuals [43, 51]. We can fit a chi-square on
φki,m
as follows,
Zkm=(δkm)2var(φki,m)
is a χ2 with 1 degree of freedom. A large value of the chi2 statistic indicates deviations from the null model and 4 standard deviations above (excess ancestry) or below (deficiency ancestry) the genome-wide average, suggests the action of natural selection post-admixture [51]. Summing-up the equation above over all SNPs assigned to a gene, we obtain the deficiency/excess of ancestry at the gene level. This allows us to assess the statistical significance of a deficiency/excess of ancestry at the SNP and gene level. To assess unusual difference in deficiency/excess of ancestry between a pair of ancestral populations given SNP m ∊ {1, …, M} within a gene, we compute
t˜kl=∑((δkm−δlm)2[var(φki,m)+var(φli,m)]/N)
Which is a two-sample t-statistic with M − 2 degrees of freedom, assuming equal sample size N. For a pair of populations, k ≠ l ∊ {1, …, K}, we compute the overall unusual difference in a deficiency/excess of ancestry,
t˜=∑∑((δkm−δlm)2[var(φki,m)+var(φli,m)]/N)
In order to summarize the types of loci and explore the potential adaptive genetic architecture implicated by our genome-wide selection scans, we identified all protein coding genes within 40 kb downstream or upstream of SNPs showing signatures of selection. To achieve this, we downloaded genomic coordinates for all genes from the NCBI ftp-server (ftp://ftp.ncbi.nih.gov/), retaining only entries for the human reference sequence and protein-coding genes. We updated genomic coordinates to the latest assembly using the Lift-Over tool on GALAXY (https://main.g2.bx.psu.edu/). We obtained the genomic predicted human genes from the GeneCard database [44]. We investigate the roles of genes and cells in disease processes using the MalaCard database [44; 53].
|
10.1371/journal.pgen.1007033 | The combinatorial control of alternative splicing in C. elegans | Normal development requires the right splice variants to be made in the right tissues at the right time. The core splicing machinery is engaged in all splicing events, but which precise splice variant is made requires the choice between alternative splice sites—for this to occur, a set of splicing factors (SFs) must recognize and bind to short RNA motifs in the pre-mRNA. In C. elegans, there is known to be extensive variation in splicing patterns across development, but little is known about the targets of each SF or how multiple SFs combine to regulate splicing. Here we combine RNA-seq with in vitro binding assays to study how 4 different C. elegans SFs, ASD-1, FOX-1, MEC-8, and EXC-7, regulate splicing. The 4 SFs chosen all have well-characterised biology and well-studied loss-of-function genetic alleles, and all contain RRM domains. Intriguingly, while the SFs we examined have varied roles in C. elegans development, they show an unexpectedly high overlap in their targets. We also find that binding sites for these SFs occur on the same pre-mRNAs more frequently than expected suggesting extensive combinatorial control of splicing. We confirm that regulation of splicing by multiple SFs is often combinatorial and show that this is functionally significant. We also find that SFs appear to combine to affect splicing in two modes—they either bind in close proximity within the same intron or they appear to bind to separate regions of the intron in a conserved order. Finally, we find that the genes whose splicing are regulated by multiple SFs are highly enriched for genes involved in the cytoskeleton and in ion channels that are key for neurotransmission. Together, this shows that specific classes of genes have complex combinatorial regulation of splicing and that this combinatorial regulation is critical for normal development to occur.
| Alternative splicing (AS) is a highly regulated process that is crucial for normal development. It requires the core splicing machinery, but the specific choice of splice site during AS is controlled by splicing factors (SFs) such as ELAV or RBFOX proteins that bind to specific sequences in pre-mRNAs to regulate usage of different splice sites. AS varies across the C. elegans life cycle and here we study how diverse SFs combine to regulate AS during C. elegans development. We selected 4 RRM-containing SFs that are all well studied and that have well-characterised loss-of-function genetic alleles. We find that these SFs regulate many of the same targets, and that combinatorial interactions between these SFs affect both individual splicing events and organism-level phenotypes including specific effects on the neuromuscular system. We further show that SFs combine to regulate splicing of an individual pre-mRNA in two distinct modes—either by binding in close proximity or by binding in a defined order on the pre-mRNA. Finally, we find that the genes whose splicing are most likely to be regulated by multiple SFs are genes that are required for the proper function of the neuromuscular system. These genes are also most likely to have changing AS patterns across development, suggesting that their splicing regulation is highly complex and developmentally regulated. Taken together, our data show that the precise splice variant expressed at any point in development is often the outcome of regulation by multiple SFs.
| Alternative splicing (AS) is highly regulated. Many genes have different splice patterns in different tissues and at different developmental stages, and splicing can also change in response to external cues (reviewed in [1]). AS plays a crucial role in the proper development of all animals [2–7] and AS is typically widely used to generate proteome diversity. For example, in humans ~95% of multi-exon genes are estimated to be alternatively spliced [8,9] and errors in regulation of AS can lead to a variety of human diseases, ranging from muscular dystrophy to cystic fibrosis to various neurological disorders [10].
Splicing requires the core splicing machinery but for the correct choice of splice site, specific regulatory splicing factors (we refer to these throughout as SFs) recognize short cis-regulatory elements in the pre-mRNA [11]—these SFs can either select for or repress the use of any specific exon-exon junction (reviewed in [12–14]). The precise combination of SFs that bind any particular pre-mRNA thus determine which exon-exon junctions are selected and hence which mature mRNA is made [15,16]. This is complex—each cell type expresses many different SFs and introns frequently contain binding sites for many SFs. To understand how AS is regulated thus requires us to know not only how individual SFs recognize and regulate any splice event but also how multiple SFs combine to affect splicing. Addressing this combinatorial regulation of splicing is central to the work we present here.
Several genomics technologies have transformed our ability to identify splice variants and this has given major recent insights into splicing regulation. In particular RNA-seq is an extremely powerful tool for identifying the splice variants that are present in any tissue or cell type and splicing profiles have been generated for many cell-types [17–20]. Other high-throughput technologies have also allowed the genome-scale characterization of the cis- and trans-acting factors involved in regulating AS events [13,21,22] and together these data have been used to define a regulatory ‘splicing code’ of RNA features that predict splicing patterns [16,23,24]. In addition, studies on individual SFs have resulted in RNA splicing maps [25] for several SFs, including Nova [26,27], RBFOX [28–30], PTB [31–33], hnRNP [34,35], TDP-43 [36], and TIA [37] proteins, which provide mechanistic insights into how binding positions correlate to regulatory effects on splicing [23,38,39].
Almost all these studies were carried out on individual cell-lines or isolated tissues and typically present a description of the splicing pattern in any specific cell-type—which isoforms are expressed and which SFs regulate these. During the development of an animal from one cell to the mature adult, splicing patterns change greatly however and this dynamic AS must be highly regulated and orchestrated. To begin to understand how AS regulation is coordinated across animal development, we use C. elegans as a simple animal model (its use as a model for studying AS regulation is reviewed in [3,40]). C. elegans development is well-characterized and its lineage is identical in every animal [41,42]. The overall splicing machinery in C. elegans is conserved with humans [43], and there is extensive AS: ~25% of genes are estimated to be alternatively spliced and many AS events show clear developmental regulation [44]. In C. elegans, key studies have demonstrated intricate co-regulation of splicing by multiple SFs [45–50], from examples of cooperative regulation (e.g. for ASD-1/FOX1 and SUP-12 [46,51]), to examples of tissue-specific splicing driven by differential tissue expression of SFs such as AS regulation by EXC-7 and UNC-75 [48]. In addition, high-throughput studies have shown that many AS events are co-regulated by multiple diverse SFs [52]. However, the cis- and trans- factors that regulate AS in the worm are underexplored—in particular only a handful of C. elegans SF targets are known. As a first step towards mapping the networks that regulate AS across C. elegans development, we thus aim to systematically identify targets that are regulated by each SF. In this initial study, we principally focus on 4 different SFs—ASD-1, FOX-1, EXC-7 and MEC-8.
Each of the SFs we chose to study is well-characterised—their involvement in C. elegans development and function is known, their expression patterns well described, and they all contain RRM domains that are one of the best understood RNA binding domains that appear frequently in SFs involved in AS regulation [53]. There are 183 RRM domains encoded in the C. elegans genome [54,55]–found within 105 genes–and they belong to distinct subfamilies. The SFs that we have selected here have RRMs that cover 5 of 6 distinct clades (Fig 1A). FOX-1 is a member of the RBFOX family of SFs [56]. It is involved in sexual differentiation by regulating splicing of xol-1 [57] and also functions to regulate splicing of a fibroblast growth factor receptor gene, egl-15 [45]. ASD-1 is a paralog of FOX-1, and was identified in a screen for other egl-15 splicing regulators [45], and redundantly regulates egl-15 splicing along with FOX-1. Both ASD-1 and FOX-1 are widely expressed in the neuromuscular system [45]. MEC-8 is a nuclear protein known to regulate alternative splicing of the unc-52, mec-2, and fbn-1 transcripts [58–61], and mutations in mec-8 leads to mechanosensory and chemosensory defects [62–64] as well as muscle defects. mec-8 is widely expressed in many tissues including both neurons and muscle cells in early development and becomes more restricted during development until it is confined to the six touch cells in later stages [59]. EXC-7 is a homolog of the Drosophila ELAV SF and is known to be involved in excretory cell formation [65], synaptic transmission [66] and neuron-specific AS events [48]. exc-7 is expressed in a number of tissues including the excretory cell and is widely expressed in neurons [65,67]. Our goal is both to identify the targets of each individual SF, but also to examine how multiple SFs might combine to affect splicing across development.
Using RNA-seq, we identified AS events that are affected by loss of each SF in vivo. Coupled with data on RNA-binding specificities of these SFs [68], we identify the AS events that are likely to be direct targets of each SF by querying for the presence of a conserved SF RNA-binding motif. Using these data, we find strong evidence for combinatorial regulation of AS by these SFs and identify different ways in which the SFs appear to combine to affect AS events. Finally, we find that multiple of these SFs appear to co-regulate the splicing of genes with similar functions, including genes with key neuronal functions such as ligand-gated ion channels and cytoskeleton-binding proteins.
In this study, we wanted to assess how multiple SFs regulate splicing in vivo in C. elegans. To do this, we combined two main approaches. First, we used RNA-seq to identify the AS events that are affected by loss of individual SFs. RNA-seq is a powerful method to survey splice variants [69,70]. Comparing the splicing pattern seen in wild-type animal with that in a strain containing a loss-of-function mutation in a specific SF identifies the splicing events that are affected by the activity of that SF. Some of these are direct targets of that SF; others will be indirect downstream consequences of loss of that SF. To distinguish between these, we used the experimentally-determined binding specificities of these SFs: direct targets of any SF contain binding sites for that SF in their pre-mRNAs whereas indirect targets do not. This approach has previously been shown to accurately identify direct targets of SFs in C. elegans [46,48,49] and in cell lines [27,30,71,72], and found to be predictive in identifying sequences that are bound by various RNA-binding proteins in vivo [68]. Thus for each studied SF, we combine the in vivo effects of loss of that SF on splicing patterns with the known in vitro binding specificity for that SF in order to identify likely direct targets of that SF. This is illustrated schematically in Fig 1B and provides a starting point to examine how multiple SFs combine to affect splicing in vivo.
We used RNA-seq to identify AS events that are affected by loss of each of 4 different SFs—ASD-1, FOX-1, MEC-8, and EXC-7. Each of these SFs is well-studied, there are well-characterized loss-of-function mutant alleles for each SF, and the in vivo functions of these SFs are known [63,65,73,74].
We sequenced polyA+ RNA isolated from wild-type and mutant worms harvested at the L4 developmental stage, obtaining approximately 200 million 100 bp, paired-end reads for each sample. Previous studies suggest that this read depth identifies the great majority of all splice variants present in the developing animal at this stage [44] and we observe similar results here (S1 Table). For each SF mutant, we identified differentially spliced exons (see Methods) that fall into several categories comprising major types of AS events (Table 1, S2 Table), such as cassette exons, alternative 5’ or 3’ splice sites, and mutually exclusive exons. We chose the L4 stage since both somatic tissues and germline are almost fully developed but, unlike adult samples, there is no risk of contamination with any fertilized embryos which can greatly affect transcriptomes. We also chose to examine the same developmental stage for all mutants so that we could easily compare any splice changes between mutants.
Overall, splicing is similar between wild-type animals and mutant animals (fox-1(e2643), asd-1(ok2299), mec-8(u218), mec-8(u303), and exc-7(rh252))—less than 10% of annotated cassette exons showed any significant differences in exon inclusion in any mutant (Table 1). For example, in the mec-8(u303) mutant we identified 73 single cassette exons that show differential inclusion levels compared with wild-type worms. This suggests that these mutations do not greatly affect global splicing patterns and this is similar to what has been observed by other groups using a similar approach [48]. We also note that a third of the altered splicing events that we see in the mutants are splicing events that were found to change significantly across development [44], suggesting that these transcripts are affected by AS during normal development.
We decided to focus on single and multiple cassette exon events, which comprise >70% of the AS changes identified in this study. To validate these AS changes that were identified in our RNA-seq data, we performed semi-quantitative RT-PCR on a subset of these events using independent RNA samples (Fig 2A). We were able to validate 80% (28/35) of the changes identified and found a strong correlation between percent-spliced-in (PSI) values obtained by the two methods (Fig 2B). Finally, to further validate our data, we also compared the PSI values of AS events between two different mec-8 mutants that exhibit varying mutation severities. mec-8(u218) is a relatively weak allele whereas mec-8(u303) is a more severe loss-of-function allele [63]. The PSI values between the two sequenced samples are strongly correlated (r = 0.975) (Fig 3A), suggesting that the variability between samples should not greatly affect our conclusions. Importantly, there is a considerable overlap between the splice sites affected in the two mec-8 mutants—as expected, the majority (30/44; 68%) of the changes found in the weaker mec-8(u218) strain reproduced in the more severe mec-8(u303) strain (Fig 3B) but more than half of the splice changes (51/81; 63%) were unique to the more severe strain. Moreover, of the AS changes detected in either strains, the changes in PSI values tended to be smaller in the mec-8(u218) mutant than in the mec-8(u303) mutant, consistent with differences in their mutation severities (Fig 3C).
So far, we have identified exons that are differentially spliced in the various SF mutants. However, these affected AS events could be due to either direct effects of the SF on these splicing targets or could instead be due to indirect, downstream effects, such as the aberrant splicing of other SFs or the mutant phenotype itself. We address this below.
SFs regulate splicing by binding at splicing regulatory elements in introns and exons to modulate selection of competing splice sites. We used this to distinguish between direct and indirect splicing regulation–we consider that any AS event that is affected by a SF but that has no consensus binding site is indirectly affected by that SF. We previously used the RNAcompete method [68,75] to define the in vitro binding specificities of our SFs of interest. We were able to obtain 7-mer binding motifs for each SF, reproducing both known [45,76,77] as well as novel consensus motifs (Fig 4A)—these data were previously published in a large survey of SF binding sites [68].We looked for the presence of these binding motifs at splice sites affected in the SF mutants, limiting our analysis to sequences 300nt proximal to the splice sites. Since these short sequence motifs are prevalent throughout the genome, we also used a phylogenetic approach to identify binding sites likely to be functionally relevant. We reasoned that just as functionally significant cis-acting transcriptional factor binding sites tend to be highly conserved across related Caenorhabditis species [78,79], functionally significant cis-acting SF binding sites would likewise be conserved. Similar approaches looking at sites conserved between C. elegans and C. briggsae have previously been used to identify motifs in C. elegans introns and exons that are important for splicing regulation [80,81]. Here, we used a multiple alignment of 5 Caenorhabditis species (C. elegans, C. briggsae, C. remanei, C. sp.11 and C. brenneri) [82], to identify conserved sequences.
We first examined whether the set of splice events that are affected by mutation of any individual SF are enriched for the presence of its consensus binding motifs. There is a weak, but not significant, trend in this direction (Table 2) if we simply consider all possible recognition sequences in the genome—if we refine this to only examine the conserved (and thus likely functionally significant) motifs, for exc-7 we find a significant enrichment of its binding motif at the affected splice sites relative to a control set of AS events (Table 2). For mec-8(u303), we only observe a significant enrichment at splice sites where we see increased exon inclusion in the mec-8 mutant (S3 Table). We note that while similar enrichments are seen for asd-1 and fox-1, they are not significant. Overall, we find that only a minority of the exons with conserved SF binding motifs were differentially spliced in the respective SF mutants—this ranges from ~5 (10/189 for asd-1) to ~20% (9/43 for mec-8). We conclude that, just as found in other studies [48], many of the splice changes seen in any mutant strain are indirect consequences of loss of that SF rather than direct effects of that SF on the splicing of those transcripts.
We next examined the locations of the binding motifs for each of the 4 SFs and find positional biases for all 4 SFs (Fig 4B). For example, ASD-1 and FOX-1 binding motifs are enriched in the introns flanking cassette exons that show altered splicing in asd-1 and fox-1 mutants. In the case of asd-1, we observed enrichment of binding motifs specifically in introns that flank exons that show increased inclusion in the asd-1 mutant (Fig 4C). We see a similar enrichment for EXC-7 motifs in introns that flank the cassette exons. MEC-8, however, shows a different pattern—MEC-8 sites appear to be enriched in the cassette exons themselves. In addition, for MEC-8, we observed more instances of increased exon inclusion in both mec-8 mutants compared with increased skipping (S1 Fig), and the enrichment for MEC-8 motifs is specific for cases of increased exon inclusion (S3 Table). This suggests that MEC-8 primarily functions to repress exon inclusion, at least at the L4 larval stage, which is consistent with previous characterizations of MEC-8 as a repressor of exon inclusion [58–60].
We thus find that the splice events affected by mutation of each SF studied are enriched for the presence of conserved (and thus likely functionally relevant) binding motifs for the SFs. In addition, the location of the conserved binding motifs can shed light on the likely way in which each SF regulates splicing and where they appear to act on the pre-mRNA transcript. In the rest of this paper, we now focus on combinatorial regulation of splicing by the 4 SFs studied here.
In the sections above, we describe how we combined RNA-seq analysis of splicing patterns in strains carrying mutations in specific SFs with in vitro binding site data to identify the likely direct targets of several different C. elegans SFs. We next used these data to explore whether different SFs might regulate splicing of similar targets.
We examined whether there were overlaps in the sets of splicing events affected in each of the mutant strains—if the same splice event is affected in more than one mutant, this might indicate that those SFs may combine to regulate that splice event. As expected, there was a significant overlap between splice sites affected in the two mec-8 mutant strains. We also found a significant overlap between the splice events affected in either asd-1 or fox-1 mutants—this is unsurprising since they are highly related paralogs. Unexpectedly, however, we also observed a significant overlap between the splicing events affected in other pairs of splicing factors (Fig 5, S2 Fig). This suggests that these four SFs may regulate many of the same splicing events.
To experimentally confirm this overlap in targets, we focused on MEC-8 and EXC-7. We identified individual splicing events that are affected by both the loss of EXC-7 and the loss of MEC-8 using our RNA-seq data and then used semi-quantitative RT-PCR to examine splicing at each individual splice site in strains lacking either EXC-7 or MEC-8 alone, or lacking both SFs at once. We identified several instances where loss of both splicing factors show an additive effect, suggesting distinct roles in regulating these AS targets (Fig 6). These include examples where both SFs either have the same effect on these targets, or differentially affect the target, with one resulting in increased exon inclusion and the other resulting in increased exon skipping. Our RNA-seq data show that SFs show much greater than expected overlaps in the splicing events that they affect and thus that multiple SFs often converge to regulate the same splicing events.
One caveat of these analyses is that both RNA-seq and RT-PCR approaches examine splicing at the level of the whole animal. For example, we find that loss of mec-8 results in increased inclusion of exons 17 and 18 of unc-52 (a well known result [58,59]), loss of exc-7 results in increased skipping of that exon (Fig 7A), and loss of both mec-8 and exc-7 together results in partial skipping and partial inclusion (Fig 7B), a so-called “antagonistic interaction” [48]. This partial inclusion in the double mutant could either be the result of some cells having complete skipping while others show complete inclusion or many cells could each be expressing a mixture of the isoforms. To examine this, we made a bichromatic splicing reporter construct that allows us to assess the level of unc-52 exon 17–18 inclusion or skipping in vivo in each cell. We find that in wild-type animals, hypodermal cells typically express either one or other isoform and that there is a mix of ‘inclusion’ cells and ‘skipped’ cells (Fig 7C). As expected [58,59], when mec-8 is mutated, there is a major shift towards exon inclusion—we cannot detect any exon-skipped isoforms but only see the exon-included form (seen as red signal in Fig 7C). When we target exc-7 by RNAi we see the reverse—cells mostly express the exon-skipped form (seen as green signal in Fig 7C). However, when we knock down exc-7 by RNAi in a mec-8 mutant, in addition to cells that appear to express either entirely one or other isoform, we also see cells that express a mixture of isoforms suggesting that MEC-8 and EXC-7 can combine to affect splice site choice in a single cell.
To explore the overlap in targets between the 4 diverse SFs in our study in greater detail, we examined whether the binding motifs for each of the 4 SFs co-occur more than randomly expected—we find this to be the case (Table 3). Since there are high quality binding motifs for more SFs in C. elegans than the 4 studied in depth here, we expanded our analysis to include all known or predicted C. elegans SFs that had high quality in vitro binding motifs that were defined using RNAcompete or inferred based on sequence similarity to RBPs with defined binding motifs [68]—these 16 SFs are listed in S4 Table and span a variety of biological functions. We find that many predicted SF binding sites co-occur more than randomly expected (see S5 Table) suggesting that there could be extensive co-regulation of splicing targets by SFs. To gain some insight into how such co-regulation might occur, we examined both the distance between co-occurring motifs and their relative orientation.
We first analyzed the average distance between co-occurring motifs—if motifs of two SFs tend to co-occur in close proximity, it suggests that they could cooperate to regulate splicing. Moreover, as MEC-8 and EXC-7, and ASD-1 and EXC-7, were found to physically interact [83], close proximities of binding sites for these SFs could indicate cooperative binding. We found the majority of EXC-7 and SUP-12, and FOX-1/ASD-1 and EXC-7 binding motifs in introns flanking cassette exons to co-occur within 60nt of each other (S3 Fig). More importantly, we also find that the closer their binding sites are together, the more likely the binding sites are to be conserved across multiple species at the sequence level (Fig 8A and 8B). This suggests that the proximity of the sites is likely to be key for their functions and for the way in which the SFs regulate these splicing events.
We next examined the relative positions of pairs of SF motifs—for example, if binding motifs co-occur, do they tend to show any bias in their relative order, and is this ordering conserved? We find that this is indeed the case for certain combinations of SF motifs (Table 4). For example, when an EXC-7 motif is upstream of a MEC-8 motif in the same intron, the EXC-7 motif also tends to be found upstream of the MEC-8 motif in other species even if the spacing between the motifs varies (Fig 8C). In some cases this ordering of SF sites appears to be unidirectional: for example the EXC-7 site always tends to be upstream of the MEC-8 site. In other cases, while the ordering is conserved for any individual pre-mRNA, the ordering varies between pre-mRNAs—for example, in cases where a GCAUG (FOX-1/ASD-1) motif is 5’ to a CUAAC (ASD-2) motif in C. elegans it will tend to be 5’ to a ASD-2 motif in other species, and where a FOX-1/ASD-1 motif is 3’ to an ASD-2 motif in C. elegans, it tends to be 3’ to a ASD-2 motif in other species. Our data thus suggest that SFs may combine to affect the same target splicing event in two distinct ways: either by binding closely located sites (such as for EXC7 and SUP-12, or FOX-1 and EXC-7) or by binding completely distinct regions of the pre-mRNA, often in some ordered manner (like EXC-7 and MEC-8).
To partly test the possible functional significance of any combinatorial interaction of SFs with their pre-mRNAs, we examined the effects of reducing the activity of either single SFs or combinations of SFs on the phenotype of C. elegans using a combination of genetic mutations and RNAi. We tested 48 pairs of SFs in this way and identified 6 clear functional interactions (Table 5, S6 Table). For example, EXC-7 and MEC-8 have overlaps in the splice events that they affect and in the transcripts that contain their binding sites. We find that reducing activity of either SF alone only has a weak effect on the fitness of the animals; however, reduction in activity of both together results in a much more severe fitness defect than expected with smaller brood sizes (Fig 9A) and growth defects (Fig 9B). This clear genetic interaction between these two SFs suggests that they functionally cooperate. We note that the molecular basis for this functional cooperation is not clear from these results—it might be that they co-regulate the same targets, and thus the double reduction has more severe consequences on a shared set of targets, or that the observed synergy is simply due to increased numbers of transcripts with altered splicing thus that the overlaps in their splice targets may indeed be functionally relevant. We find 5 other such genetic interactions (asd-1 and asd-2, asd-1 and etr-1, asd-1 and rsp-3, mec-8 and rsp-3; mec-8 and asd-2; Table 5). We note that the genetic interactions that we observe all occur between SFs that bind to distinct regions of the pre-mRNAs and that we observed no such interactions between SFs whose sites appear to be closely located on pre-mRNAs consistent with close cooperation.
In summary, we find multiple lines of evidence that indicate that many AS events are regulated by multiple SFs during C. elegans development. There is a significant overlap in the sets of splice events affected by mutation of each SF suggesting that they may regulate the same AS events. There is also clear enrichment for co-occurrence of the defined binding motifs of different SFs in the same transcripts. In addition (see analysis in next section below), the genes that contain many SF motifs in their introns are much more likely to show significant changes in their AS patterns across development [44]. Finally, we find that SFs may co-regulate AS events in two distinct ways. Some SFs have binding motifs that co-occur within close proximity in the same intron—this includes FOX-1/ASD-1 and EXC-7, EXC-7 and SUP-12 (S3 Fig) for example. The closer their sites, the more likely they are to be conserved and thus functionally relevant (Fig 8A and 8B). However, other SFs like EXC-7 and MEC-8 appear to functionally interact in a very different way—their binding sites often co-occur, but the distance between them is often quite long and the distance is not conserved. What is conserved is the orientation of their sites—EXC-7 sites tend to be upstream of MEC-8 sites when they occur in the same transcript (S4 Fig), and this orientation is maintained across long evolutionary distances. Taken together, these data all point to many transcripts having complex regulation of their splice patterns—we next examined whether this complex regulation is in any way restricted to particular classes of genes.
In the above analyses, we identified AS events with conserved consensus binding motifs for multiple SFs that suggest potential splicing co-regulation by these SFs. We noticed that many of these potentially co-regulated cassette exons are found in functionally coherent groups of genes that appear to be highly enriched in genes involved in proper neuromuscular or cytoskeletal functions (S7 Table). For instance, many of these genes are required for proper locomotion in the animal; mutations in many of these genes lead to an uncoordinated (unc) phenotype, such as unc-2 and unc-44, which have conserved binding motifs for FOX-1/ASD-1, MEC-8, EXC-7, SUP-12 and UNC-75 (Fig 10, S5 Fig).
To examine this further, we used Gene Ontology (GO) [84] to identify enriched functional categories among these potentially co-regulated targets. We found the set of genes with co-occurring motifs to be enriched for several classes of genes (Table 6, S8 Table). This includes genes involved in cytoskeletal structures that are key for neuron growth and function (e.g. unc-44, nab-1) [85–87] and genes that are key ion channels in the neuromuscular system (e.g. unc-2, unc-49, tmc-1, exp-2) [88–91], suggesting that neuromuscular genes tend to have more complex AS regulation. While some genes with key roles in the neuromuscular are known to have long and complex transcripts (e.g. ttn-1 has a 55kb coding sequence and 66 coding exons [55,92], unc-22 has a 20kb coding sequence and 34 exons [55]), this is not the reason for the observed enrichments: genes in these specific functional classes are more likely to contain motifs for multiple SFs than a random set of genes with similar total intron lengths (Table 7). For example, 36% of genes with the cytoskeletal protein binding GO term have both a conserved ASD-1/FOX-1 and EXC-7 motif, while only 8% of the intron length-matched set of genes without that GO term have the two motifs co-occurring, suggesting that the GO enrichment results were not merely due to those genes having larger motif search spaces and is instead likely biologically relevant.
These data suggest that the 4 SFs we studied tend to interact frequently to affect splicing of genes with roles in the neuromuscular system. To further explore this, we examined whether some of the splice factors that show genetic interactions have combinatorial effects on neuromuscular system function rather than simply on the broad phenotypes of growth rate or viability as we had done above. We find that when either asd-1 or asd-2 activity is lost alone (either through mutation or RNAi), the worms have near-normal movement. However reduction in activity of both asd-1 and asd-2 results in sterile adult worms (Fig 11A) that are almost completely paralysed (quantified in Fig 11B) indicating that these two factors do indeed show a strongly synergistic effect on movement. In the case of mec-8 and exc-7, we examined how these factors affect a specific aspect of neuromuscular system function, cholinergic transmission. Acetylcholine (ACh) is the major neurotransmitter at neuromuscular junctions (NMJs)—drugs such as aldicarb that alter ACh levels at NMJs result in paralysis [93,94]. Previous studies showed that exc-7 mutants show reduced sensitivity to aldicarb [48,66]–we decided to test whether mec-8 and exc-7, two factors that share many targets and whose binding sites co-occur frequently and that also show genetic interactions in our hands (see Fig 9) might have combinatorial effects on aldicarb sensitivity. We find this is indeed the case (Fig 11C)—reduction of activity of either mec-8 or exc-7 causes a small decrease in aldicarb sensitivity but reduction of activity of both results in a more severe reduction. We conclude that at least in these two cases, there is evidence that these factors combine to affect phenotypes that are relevant to neuromuscular system function.
Together, all these data combine to suggest that there is complex splicing regulation of genes involved in the neuromuscular system and in the cytoskeleton, whereas most other splice events in other sets of genes seem to be primarily regulated by just one of these SFs. This finding has a clear caveat however: the 4 SFs in this study all have characterized functions in neuromuscular development and this could explain the prevalence of potential shared targets with similar functions. For example, since MEC-8 and EXC-7 have key roles in proper neuronal function (in mechano- and chemo-sensation [62,64] and in synaptic transmission [66] respectively) it may not be particularly surprising that their splicing regulatory pathways may converge upon genes encoding ion channels, many of which are involved in neurotransmission. We therefore cast our net further and examined the set of all 16 SFs that had high quality in vitro defined or inferred binding motifs.
We first identified the sets of genes that had conserved intronic binding motifs for multiple of the 16 SFs. We find that these genes often show complex splicing regulation across development—for example ~40% of the genes with 4 or more distinct SFs motifs (43/108) showed significant changes to their alternative splicing patterns across development in a previous study [44]. This is a highly significant enrichment (p = 4.85×10−22, one-sided hypergeometric test) compared to 12.9% of background AS genes showing developmental regulation. This shows that the number of SFs that we predict to be able to bind and thus regulate any transcript correlates well with the complexity of the splicing changes for that transcript across development. We next identified GO terms that were enriched in these genes that have multiple SF motifs at introns flanking alternative exons (all results in S9 Table). Intriguingly, we found similar results to our initial analysis that only examined 4 SFs—many of the genes that are predicted to be targets of multiple SFs have GO terms consistent with a role in the neuromuscular system and regulation of the cytoskeleton. These include neurogenesis (GO:0022008), axon fasciculation (GO:0007413), regulation of locomotion (GO:0040012), cytoskeletal protein binding (GO:0008092) and transmembrane transporter activity (GO:0022857), and include many genes with well-characterized roles in neuronal functions such as unc-44, unc-2, and fust-1 (see Fig 10, S5 Fig). We thus suggest that, at least in C. elegans, genes with key roles in the neuromuscular system appear to have highly complex regulation of splicing. We note that the set of genes whose splicing was previously reported to change significantly across development [44] are enriched for similar neuronal and cytoskeletal GO terms (e.g. locomotion, microtubule cytoskeleton organization) (S10 Table) confirming the highly dynamic and complex splicing regulation of these functional classes of gene.
In summary, we find that genes with key roles in the neuromuscular system and in the cytoskeleton appear to be have more complex splicing regulation: they tend to have more dynamic changes in splicing across development and they are enriched for the presence of multiple binding sites for SFs of widely differing biological functions. We suggest that this may allow the subtle fine-tuning of neuronal functionality across development.
During development, there are extensive and highly regulated splicing changes. How are these AS events regulated? Is any AS event regulated by a single SF or do multiple SFs combine to regulate individual AS events? In this study, we studied four splicing factors (SFs) in C. elegans to try to determine their individual targets and in particular to examine whether there was any evidence that they cooperate to regulate splicing.
We used three pieces of data to examine how any individual SF affects splicing. We first used RNA-seq to examine how a loss-of-function mutation in each of the four SFs affects splicing patterns. We could confirm ~80% of the identified splice changes using RT-PCR and showed a strong overlap in the AS events affected in two different strains that each has a loss-of-function mutation in the SF MEC-8, suggesting that our data are of high quality. Overall, we found that <10% of cassette exon AS events were affected in any individual mutant showing that the effect of losing any single SF has specific and limited effects on splice patterns. The AS events that are affected by the loss of any individual SF are likely to be a mix of direct targets of the SF and indirect downstream effects. To distinguish between these, we used a second dataset–the in vitro binding specificities of each SF which we had measured previously [68]. We reasoned that AS events that change in any SF mutant that have binding motifs for the SF are likely to be direct targets whereas AS events that are affected by the activity of any SF but that have no discernable binding motif for that SF are likely to be indirect. This approach has been used by other groups and successfully identifies direct SF targets [27,30,71,72,95,96]. Finally, we used evolutionary signals to identify the SF binding motifs that are most likely to be functional—binding motifs for any SF that are present in 5 diverse Caenorhabditis species have been conserved across long evolutionary time periods and are thus likely to be important whereas those that turn over rapidly are less likely to be functionally significant.
We thus combined direct measurements of how loss of each SF affects splicing patterns, in vitro measurements of each SF’s binding specificity, and evolutionary signatures, to identify the direct targets of each of the four SFs. Surprisingly, we find that there are significant overlaps in the targets of the four SFs. We use RT-PCR to confirm that multiple transcripts are indeed affected by multiple SFs and that loss of function of multiple SFs has additive effects on individual AS events. Finally, at least in the case of EXC-7 and MEC-8, we show that these combinatorial effects on splicing are likely to be functionally significant—a loss of function mutation of either mec-8 or exc-7 alone has only weak phenotypic effects whereas mutation of both genes together results in more severe fitness defects, as well as reduced sensitivity to aldicarb that suggests more severe defects in synaptic transmission. We note that MEC-8 and EXC-7 have also been shown to physically interact [83], underlining the likely relevance of the combinatorial effects we observe on splicing. We also show, using a bichromatic splicing reporter, that these SFs can regulate splicing of exons (e.g. exons 17–18 in unc-52) in the same cells.
We also find that the positions and orientation of the SF motifs in any transcript indicate that SFs may combine to regulate AS events in two distinct ways. For example, EXC-7 and FOX-1/ASD-1 binding motifs in co-regulated targets tend to occur close together—indeed the closer the motifs sit, the more likely the binding motifs are to be conserved, suggesting that the proximity is functionally relevant. For other pairs of SF, the basis for the combinatorial interactions appears to be different, relying not on proximity of the binding sites but rather on their relative orientation. For example, EXC-7 and MEC-8 sites tend to have a conserved ordering around any target AS events: when EXC-7 sites are upstream of MEC-8 sites in C. elegans, this same ordering is frequently seen in all 4 other species although the precise spacing between the sites may vary considerably. We thus suggest that some SFs (such as ASD-1/FOX-1 and EXC-7) may combine to regulate individual AS events through binding to closely located binding motifs in the same transcript whereas others (e.g. EXC-7 and MEC-8) may interact in more complex spatial ways that are dependent on ordering rather than proximity. While we do not know the specific mechanisms by which these SFs combinatorially regulate AS events, we find that these SFs may act together through longer distances to regulate the same splice site, distinct from examples of cooperative binding that were found for other pairs of SFs [46,49,51].
Our data thus suggest that the splicing of many transcripts are regulated by multiple SFs. Intriguingly, this combinatorial regulation by multiple SFs particularly affects transcripts from genes that play key roles in the neuromuscular system. In particular genes that encode ligand-gated ion channels and the many key cytoskeletal components that are required for neuronal and muscle cell function are potentially regulated by multiple of the SFs we studied. We do not believe that this is due to the initial selection of the 4 SFs that we studied since we find a similar trend in a wider set of 16 known or predicted SFs. For genes with binding sites for 4 or more SFs, we also find an enrichment of GO terms associated with functions in the neuromuscular system (e.g. cytoskeletal protein binding, transporter activity). We thus suggest that genes with key roles in the neuromuscular system appear to have complex splicing regulation, and their splicing is regulated by many individual SFs. We also note that the genes with binding sites for multiple SFs are enriched for having dynamic regulation of their splicing patterns across development [44] and we suggest that this high complexity in the regulation of neuronal splicing genes may allow fine-tuning of neuronal functions across development.
In summary, we used a combination of RNA-seq and in vitro RNA-binding specificities to systematically identify likely direct AS targets for 4 C. elegans SFs. These data provide a resource that contributes to our understanding of how these SFs regulate AS of specific C. elegans genes. The identification of candidate AS events that may be co-regulated by several SFs also provides a starting point for future studies to look at the variety and intricacies of combinatorial splicing regulation in C. elegans.
All C. elegans strains were provided by the CGC, which is funded by NIH Office of Research Infrastructure Programs (P40 OD010440), and maintained at 20°C on NGM plates with OP50 bacteria. PolyA+ RNA was isolated from synchronized L4 worms, synthesized into cDNA and prepped for RNA sequencing as previously described [44].
All RNA-seq files in this study are available from NCBI/SRA under accession #PRJNA412927.
Each sequenced sample yielded approximately 200 million, 100bp paired-end reads, which were mapped to the WS220 version of the C. elegans genome using the RNA-Seq unified mapper (RUM) program [97] at default parameters. Uniquely mapped reads that spanned exon-exon junctions, and were annotated by RUM as “high-quality junctions”, were used to calculate splice junction usage.
The database of AS events that we used to compare splice site usage was generated using a combination of two methods. First, we generated an exon isoform database for C. elegans using SpliceTrap v.0.90.1 [98] and transcript annotations from Ensembl. SpliceTrap generates this database by subdividing each transcript isoform into exon trios to query for alternative splicing of the middle exon. To supplement the SpliceTrap-generated database, we also created our own custom database of known and possible C. elegans AS events from WS220 exon-level annotations downloaded from BioMart, by using a similar method that had been used to create an AS database for Drosophila [99]. The non-redundant set of possible AS events from these two databases were used to query for AS differences between samples.
For comparisons of splice site usage, we tested all events that have at least a combined 10 reads from junctions corresponding to the two different splice forms. For cassette exon and mutually exclusive events, we excluded instances where there is a greater than 30-fold difference between the numbers of reads mapping to the two adjacent junctions.
For cassette exons, we calculated percent-spliced-in (PSI) values as
PSI=100×12(C1A+AC2)12(C1A+AC2)+C1C2),
with C1 and C2 representing constitutive exons, A representing the alternative exon, C1A and AC2 corresponding to the number of reads mapping across the adjacent junctions, and C1C2 to the number of reads mapping across the alternative junction.
For instances with alternative donor or acceptor splice sites at the flanking exons, we calculated PSI values by summing all the combinations of adjacent junctions, similar to previously described methods [100]:
PSI=100×12(∑CiA+∑ACj)12(∑CiA+∑ACj)+∑CiCj,
where Ci and Cj are detected alternative donor and acceptor splice sites for the C1A and AC2 junctions respectively.
We then used Fisher’s exact test (2-sided) to test for significant differences in junction ratios between wild-type N2 and mutant samples. To identify splice changes, we require these events to have an accompanying p-value < 0.05, and a change in percent inclusion of at least 15%.
The same method was used to identify splice changes in other types of AS events.
For preparation of L4-staged samples, worms were synchronized using the bleaching method and an additional sorting step was done on a Union Biometrica COPAS worm sorter using an empirically determined window of TOF and EXT values that captures L4 worms. RNA was isolated using TRIzol reagent (Invitrogen) using standard RNA extraction protocols. After digestion with DNase I (Sigma), random nonamers (Sigma) were used to synthesize total RNA into cDNA using SuperScript III reverse transcriptase (Invitrogen). PCR amplification cycles used ranged from 27–35 cycles, and PSI values were estimated by densitometric analysis using ImageJ [101].
We used a quantitative method to assay for overall fitness of worm populations as previously described [102,103]. L1 worms were collected after putting worms through an 11μm filter. Bacteria expressing dsRNA were grown in LB media (supplemented with 1mM Carbenicillin) overnight at 37°C. Cultures were then induced with 4mM IPTG at 37°C for 2h, then spun down and resuspended in the same volume of NGM media (supplemented with 1mM Carbenicillin and 4mM IPTG). Approximately 10 worms were dispensed into each well in a 96-well plate along with 65μl of resuspended bacteria. Worms were then grown in a 20°C shaking incubator at 200rpm. Worms were harvested after 4.5 days and the population of worms in each well was assayed using a COPAS worm sorter, recording the time-of-flight (TOF) and extinct (EXT) values of each worm. In each independent experiment (4 total), 2–3 technical replicates were plated, with at least one blank well separating each different condition.
GFP RNAi clones were used as a non-targeting control, with 16–24 technical replicates plated for each experiment. Population averages were taken across the technical replicates and used as a proxy value for fitness. For both the wild-type and mutant exc-7 strain, the relative fitness of worms subjected to mec-8 RNAi in each strain was calculated by normalizing the mec-8 RNAi population size to the control GFP RNAi population size. The expectation value for fitness of the mutant strain subjected to mec-8 RNAi was then calculated using a multiplicative model (relative fitness in the wild-type strain targeted with mec-8 dsRNA x relative fitness in the exc-7 mutant strain with non-targeting dsRNA). The differences in actual fitness of the mutant strain subjected to mec-8 RNAi relative to expected was then calculated. As control, the differences in actual fitness of the mutant strain subjected to control GFP RNAi relative to expected was also calculated. The significance of difference in log ratio of actual and expected values between the mec-8 RNAi and control RNAi conditions was then calculated using a student’s t-test.
The same method was used screen for RBP interactors in other mutant SF backgrounds—asd-1(ok2299), fox-1(e2643) and mec-8(u218)/mec-8(303). S6 Table lists the primer pairs used to target various genes encoding known or putative SFs in each of the 4 SF mutant strains. Interaction scores were calculated as the actual fitness score relative to expected. Similarly, an interaction score of 0 implies a maximal negative interaction and a score of 1 implies no genetic interaction.
Gene pairs with an average interaction score < 0.7 across independent experiments are listed in Table 5. Scores involving mec-8 mutants were calculated as the average interaction score between the two mec-8 mutants tested.
For each RNAi knockdown, approximately 50 L4 worms were grown on NGM plates (supplemented with 1mM IPTG and 1mM Carbenicillin) seeded with dsRNA-expressing bacteria. GFP-targeting dsRNA-expressing bacteria was used as a negative control. L4-staged progeny were then harvested using a COPAS worm sorter and RNA samples were obtained as described above.
To search for binding motifs for each splicing factor, we looked for the occurrence of GCAUG, GCACA and U(A|G|U)(A|G)GUU for FOX-1/ASD-1, MEC-8 and EXC-7 respectively. For each AS exon, these motifs were searched in the alternative exon, flanking introns and flanking constitutive exons at regions up to 300nt proximal to each splice site. To filter for conserved motifs, we downloaded the 7-way multiple alignment of Caenorhabditis genomes from the UCSC Genome Browser database (http://genome.ucsc.edu/) [82] and used alignment data from the 5 Elegans group species—C. elegans, C. briggsae, C. remanei, C. sp.11 and C. brenneri. Motifs were first searched in the C. elegans sequence and then, using the UCSC multiple alignment, we consider a motif as conserved if it is present in 2 or more other Elegans group species within 25nt of the motif position in C. elegans.
The same method was used to identify conserved motifs corresponding to other known or putative SFs. All binding motifs (directly determined as well as inferred) were taken from the CISBP-RNA database [68]. The binding motifs for each SF that were used to find potential binding sites in this paper are listed in S4 Table.
To test for enrichment of binding motifs, a set of exons that are alternatively spliced in our data at the L4 stage (5 to 95 PSI) was used as a background control to the set of differentially spliced exons. From this background, we then picked a random set of non-SF-regulated cassette exons with a matched N2/wild-type PSI distribution to the set of differentially spliced exons. Enrichment of each motif was tested using Fisher’s Exact Test. The final background/control value presented was the average value taken from 100 randomized samples. For differentially spliced exons in fox-1, asd-1 and exc-7 mutants, as well as the matched control sets, we considered each event to have a conserved motif if at least one is present within flanking introns. For differentially spliced exons in mec-8 mutants and their matched control sets, we also considered conserved motifs in the alternative exons in addition to the flanking introns.
To calculate the number of exons with potential binding motifs for each SF, we looked for conserved motifs in the same manner among the set of exons that are alternatively spliced in our data at the L4 stage (5 to 95 PSI).
For testing co-occurrence of motifs, we considered different motifs to co-occur if one or more of each motif is present within either flanking introns in the same AS event. We then used a cumulative hypergeometric distribution function to test for significance of co-occurrence [104,105]. We used a multiplicative model to determine expected co-occurrence values, where the expected proportion of co-occurring motifs are calculated based on the observed proportions of AS events with one or more of either motif. We used the same method for finding conserved motifs for each SF as described above. While dinucleotide content could over-estimate rates of motif co-occurrences—e.g. occurrences of two motif may not be independent due to biases in GC content [106]—we find that, while there are differences in dinucleotide composition in regions where we find co-occurring SF motifs (S6 Fig), these differences are relatively small (<1.2 fold), and should not greatly affect our conclusions.
The R GOstats package (v.2.32.0) [107] was used to look for GO term enrichment among the set of genes with motifs co-occurring at sites surrounding one or more of its cassette exons. A background set of AS events with PSI values between 1% and 99% at the L4 stage was used for each comparison. For analyses involving genes that contain temporally regulated AS events, the sets of genes that were previously identified to be differentially spliced across development [44] were used. This includes genes that were identified using both RNA-seq and microarray methods (listed in S2 and S3 Tables in Ramani et al.) [44]. Significance of overlap between the set of genes with multiple SF motifs and the aforementioned set of genes that are differentially spliced across development was calculated using a one-sided hypergeometric test.
To normalize for differing total intron lengths between genes with specific GO terms and the set of background genes, we picked a random set of genes from the background set with a similar intron length distribution to each set of genes with a specific enriched GO term. Each random sample was taken from a background set of genes that have been annotated with a GO term (MF or BP depending on the GO term tested) but not annotated with the specific GO term or any of its child terms. The proportion of genes with co-occurring motifs at one or more AS event was then calculated for both the random sample and the gene set with the specific GO term or its child terms. The final values presented for the control set was taken from the total of 1000 randomized samples. All GO annotations were taken from the GO.db (v.3.0.0) and org.Ce.eg.db (v.3.0.0) packages in R.
For analyzing distances between pairs of SF binding motifs, we first looked for motifs at intronic regions proximal to splice sites. We split up each intron to denote two distinct intronic regions as either proximal to the 5’ or 3’ splice site. As previously, we restricted our analysis to regions up to 300nt proximal to each splice site.
We then carried out the distance analysis on each intronic region (5’ and 3’ ends of upstream introns and the 5’ and 3’ ends of downstream introns), where the distance between motifs was represented as the spacing between the end of the upstream motif and the start of the downstream motif. For the conservation analysis, we queried for motifs in other species at sequences that aligned with each of the respective intronic regions. For cases in which multiple instances of the same motif are present in an intronic region, we used the instance of that motif that is closest in distance to the second distinct motif to calculate the spacing between motifs in C. elegans. For spacing comparisons in other species, if multiple possible motif combinations exist, we used the comparisons of pairs of distinct motifs that result in a spacing distance that is most similar to the spacing distance in C. elegans. For plotting distribution of conserved spacing between motifs, we define a spacing between motifs as conserved if the motifs are present in 2 or more other Caenorhabditis species besides C. elegans, and if the length of the spacing in those other species are within 20% of the spacing observed in C. elegans. For plotting positions of co-occurring EXC-7 and MEC-8 motifs, we considered all instances in which both motifs are present (anywhere) in either introns, and in which these motifs are also present in 1 or more other species within a 25nt sliding window.
The aldicarb acute assays were all performed in liquid M9 in 96-well plates. Wild-type and mutant L4 worms were first grown on plates seeded with bacteria expressing dsRNA targeting either exc-7 or GFP as a negative control. After 2 days, L1 progeny from these plates were isolated by filtration and used for the drug assay. Approximately 100 L1 worms were added to each well containing 2mM aldicarb or DMSO in M9 (total volume of 200μl). After 3h, the movement of worms after aldicarb treatment was compared to the movement of worms in the no drug control. Worms were scored as ‘moving’ if they exhibit wild-type-like movement similar to those in the control. 4–6 wells of worms were scored for their responses to aldicarb across 4 biological replicates.
|
10.1371/journal.pntd.0003885 | Involvement of Tetraspanin C189 in Cell-to-Cell Spreading of the Dengue Virus in C6/36 Cells | Dengue virus (DENV) is naturally transmitted by mosquitoes to humans, infecting cells of both hosts. Unlike in mammalian cells, DENV usually does not cause extremely deleterious effects on cells of mosquitoes. Despite this, clustered progeny virions were found to form infection foci in a high density cell culture. It is thus interesting to know how the virus spreads among cells in tissues such as the midgut within live mosquitoes. This report demonstrates that cell-to-cell spread is one way for DENV to infect neighboring cells without depending on the “release and entry” mode. In the meantime, a membrane-bound vacuole incorporating tetraspanin C189 was formed in response to DENV infection in the C6/36 cell and was subsequently transported along with the contained virus from one cell to another. Knockdown of C189 in DENV-infected C6/36 cells is shown herein to reduce cell-to-cell transmission of the virus, which may be recovered by co-transfection with a C189-expressing vector in DENV-infected C6/36 cells. Moreover, cell-to-cell transmission usually occurred at the site where the donor cell directly contacts the recipient cell. It suggested that C189 is crucially involved in the intercellular spread of progeny viral particles between mosquito cells. This novel finding presumably accounts for the rapid and efficient infection of DENV after its initial replication within tissues of the mosquito.
| Dengue fever is one of the most important mosquito-borne viral infectious diseases in the world. Its etiological agent is naturally transmitted via blood feeding by Aedes mosquitoes. An ingested virus can replicate and be disseminated within and between tissues in mosquitoes. In this study, we found that infection of DENV in C6/36 mosquito cells can stimulate the up-regulation of tetraspanin C189, which usually co-localizes but does not directly interact to form C189-containing membrane-bound vacuoles (C189-VCs). Our results also showed that the virus can be delivered to a neighboring cell along with C189-VCs, frequently through cell contact with filopodia extended by the donor cell that touch the recipient cell. Knockdown of C189 can reduce the efficiency of virus delivery, indicating its crucial role in cell-to-cell transmission of DENV in C6/36 cells. Cell-to-cell transmission may thus be an alternative route for the efficient intercellular spread of progeny viruses within tissues of the mosquito.
| Dengue virus (DENV) consists of four serotypes that manifest similar symptoms, ranging from a mild febrile illness to a life-threatening dengue hemorrhagic fever [1]. Taxonomically, DENV is one of some 70 members of the family Flaviviridae and is transmitted between humans by Aedes mosquitoes [2], particularly Aedes aegypti [3]. Dengue fever (DF) and dengue hemorrhagic fever (DHF)/dengue shock syndrome (DSS) have become increasingly important public health problems in over 100 countries in tropical and subtropical regions [4]. It is estimated that 2.5–3 billion people are risk of dengue infection in the world [5]. As DENV is naturally transmitted to humans by mosquitoes, indicating the virus can also infect and replicate in the mosquito cell during its journey from the midgut to salivary glands [6]. In humans bitten by an infected mosquito, DENV inoculated with mosquito saliva initially infects Langerhan cells and keratocytes residing in the epidermis where it begins to replicate [7]. Subsequently, the virus can infect other organs including circulatory macrophages/monocytes, lymphoid tissues, liver, spleen, kidneys, and lungs [8], as well as the brain in a few cases [9]. DENV has also been detected in megakaryocyte progenitors and circulating platelets [10], suggesting that thrombocytopenia in dengue patients is closely associated with DENV infection [11, 12]. Such host cells are usually infected by DENV through receptor(s)-mediated endocytosis [13] and mostly end up undergoing apoptosis in response to dengue virus infection [14]. A huge number of viral particles from infected cells burst into the blood stream or a culture to become the source of infection for other cells.
Since mosquito cells can be protected from dengue virus infection by way of an induced antioxidant defense as well as anti-apoptotic effects [15, 16], infected cells usually remain intact even when abundant progeny viral particles have been produced within the cell [17]. In mosquito cell cultures, progeny viral particles are also released from infected cells into the medium as in mammalian cells [17]. Like other insects, the mosquito possesses an intestine composed of a monolayer of epithelial cells resting on an extracellular basal lamina that is morphologically divided into three parts; i.e., the foregut, midgut, and hindgut [18]. Normally, DENV must infect epithelial cells in the midgut, which is the first site of stay for engorged blood meals [19]. DENV contained within the blood meal initially infects a few epithelial cells, followed by the formation of infection foci involving multiple cells. Infection foci gradually expand and the entire organ may become infected within several days post-infection (pi) [20]. This suggests that the virus spreads laterally from the initial site of infection to neighboring cells rather than through a process of “release-and-entry,” leading to an increased number of infected cells that successfully establish the infection in the mosquito [21]. In turn, there might be a unique mechanism other than the process of virus release and entry such as occurs in the mammalian circulatory system or cell cultures for the intercellular spread of the virus in the mosquito vector [22].
Cell-to-cell spread has been reported as a route of rapid dissemination implemented by a variety of animal viruses [23]. Herpesviruses (HSV) are the typical example of spread by direct cell-to-cell transmission [24], which can promote immunity evasion deriving from the effect of a neutralizing antibody that blocks reinfection of other cells by an extracellular virus [23]. Hepatitis C virus (HCV) establishes infection via cell-to-cell transmission in the liver [25], particularly in patients with chronic infection [26]. In the case of human immunodeficiency virus (HIV), infection of CD4+ memory T cells are initially driven by cell-free virions. However, direct transfer of infection in the cell-to-cell mode is believed to occur more efficiently and rapidly [27]. Moreover, tetraspanin CD81 is involved in cell-to-cell transmission by such viruses as HIV [26, 28], raising the question of whether DENV can efficiently spread or disseminate via an intercellular mode in tissues, such as the midgut, of a mosquito host.
As mentioned above, most mosquito cells infected by DENV do not exhibit obvious damage but can cause persistent infection [17]. Thus, release-and-entry of the virus in mosquitoes may not be the primary way between cells within a tissue. In contrast, spread by way of cell-to-cell transmission seems to be more logical for extensive and efficient viral infection. In this study, we designed different cell culture systems to reveal how DENV can spread between C6/36 cells in a cell-to-cell transmission mode. As a novel tetraspanin C189 was previously identified from DENV-infected C6/36 cells and was shown a close association with a vacuole containing viral proteins [29], this study also aimed to demonstrate the possibility of its involvement in virus spread between mosquito cells.
The protocol for cell culture used in this study mostly followed the description reported previously. In brief, DENV type 2 (New Guinea C) used in this study was propagated in Ae. albopictus C6/36 cells that were grown in minimal essential medium (MEM) (Invitrogen, Carlsbad, CA) with non-essential amino acids and 10% fetal bovine serum (FBS) at 28°C in a closed incubator. Titration of the virus was carried out by plaque assay on BHK-21 cells, which were cultured at 37°C in an incubator with a 5% CO2 atmosphere [15].
C6/36 cells were dispended to culture in 48-well plates overnight; ~70 μl virus suspension (MOI = 1 or less) was then added to each well. After adsorption for 1 h at 28°C, the virus suspension was removed from each well and replaced with fresh medium containing 5% FBS or with 1.1% methylcellulose medium (a mixture of 2.2% methylcellulose and fresh medium). Subsequently, 1 ml of 4% paraformaldehyde was added to each well after incubation for 24 or 48 h and then the medium or methylcellulose was removed and fixed for 30 min. The plate was washed with PBS twice and then treated with 0.1% Triton X-100 at 4°Cfor 2 min to increase cell membrane permeability. The plate was washed again twice with PBS and then 0.2 ml of 1% BSA was added to each well to block at room temperature (RT) for 1 h or at 4°C overnight. An immunofluorescence assay (IFA) was subsequently implemented as described below.
Total RNA was isolated from the pellet of 4 x 106 ~ 1 x 107 C6/36 cells using Trizol reagent (Invitrogen). First-strand cDNA was synthesized using total RNA as the template, and a SuperScript First-Strand Synthesis kit (Invitrogen) was used according to the manufacturer instructions.
DENV-infected donor cells (MOI = 1, 24 hpi) and eGFP-expressing recipient cells were co-cultured together or separated by transwell either with or without treatment with antiserum for 6 h. Recipient cells were subsequently sorted by flow cytometry and viral RNA was analyzed by RT-PCR. The viral RNA expression level was detected with the primer pair located at the 5’-UTR, including DV2-5UTR-F (TGGACCGACAAAGACAGATTCTT) and DV2-5UTR-R (CGYCCYTGCAGCATTCCAA). The internal control gene 18S was detected with primers 18SF (AGGTCCGTGATGCCCTTAGA) and 18SR (TACAATGTGCGCAGCAACG). The C189 expression level was normalized to the 18S expression level.
Total RNA extraction followed the procedure in the previous report using Trizol reagent (Invitrogen, Carlsbad, USA) [29]. The gene expression level was quantified by real-time RT-PCR using SYBR Green dye (Applied Biosystem, Carlsbad, USA). The C189 expression level was detected with primers C189F (CTGCATGACCACGACCTATGG) and C189R (AGAGCGGCAACGACGATTT). 18S was used for the internal control as above.
Infected and uninfected C6/36 cells were smeared on a cover glass and washed with phosphate buffered saline (PBS; pH 7.4) three times. Cells were subsequently fixed in 4% paraformaldehyde for 10 min, washed again with PBS, and then blocked with 1% BSA in PBS for 1 h at 37°C. Primary polyclonal antibodies (1:100 in dilution) against C189 LEL (S1 Text) were added onto the cover glass and incubated at 37°C for 1 h. The cover glass was subsequently incubated with secondary antibodies (1:100) conjugated with FITC at 37°C for 1 h after being washed with PBS. The cover glass was finally mounted with a mixture of glycerol and PBS (3:7) and observed under a laser scanning confocal microscope (Zeiss LSM 510, Carl-Zeiss, Jena, Germany). Negative controls were incubated with diluents without primary antibodies; otherwise, samples were subjected to all the procedures described above.
The expression vector used in this study was based on insect-cell-expression vector pAC5.1-V5-His A (Invitrogen), from which pAC5.1-eGFP expression vector was previously constructed (Lin et al., 2007). For the expression of C189 tagged with eGFP fusion proteins, the open reading frame of C189 was amplified by the indicated PCR primers and cDNA derived from C6/36 cells, which were subsequently inserted into the pAC5.1-eGFP expression vector in the N-terminal domain of eGFP [29]. To express red Rhodamine fluorescent protein (RFP), the RFP gene was amplified from pTagRFP-C (Evrogen, Moscow, Russia) and inserted into pAC5.1-V5-His A. To express HAeGFP and HAC189, primers HA-F and HA-R were hybridized and ligated with pAC5.1-V5-His A to generate pAC5.1-HA. eGFP and C189 genes were amplified using the indicated primers and C6/36 cDNA, then inserted into pAC5.1-HA to form HAeGFP and HAC189 expression vectors. The primers to amplify the corresponding genes are listed in S1 Table.
C6/36 cells were seeded into 6-well plates and grown to 70–80% confluence for transfection. X-tremeGene HP DNA transfection reagent (Roche, Indianapolis, IN, USA) was mixed with vectors (ratio = 3:1 μl/μg, 1 μg plasmid DNA per well in most experiments) in basal medium (MEM, 2% non-essential amino acid, 0.0375% sodium bicarbonate, 0.2% hepes) at RT for 15 min. Cells were incubated with the transfection mixture for 5 h before replacing with complete medium.
A stable knockdown system was based on the miRNA system [30]. The target sequence (83–103 bp) was selected by BLOCK-iT RNAi Designer (Invitrogen). The sequence design and predicted pre-miRNA sequence structure of miC189 are shown in Supplementary S1 Table. After hybridization, design DNA was inserted into miRNA expression vector, pcDNATM6.2-GW/EmGFP-miR (Invitrogen). The pAC5.1-V5-His A was combined with pIE1-neo (Novagen, Cambridge MA, USA) to generate insect-cell-stable-expression vector, pAC5.1-neo. Sequentially, the trans-element of miRNA expression vector was amplified and inserted into pAC5.1-neo. A schematic process for construction of stable knockdown systems was shown in S2 Text. To generate those stable knockdown cell clones, miC189 and miN (negative control provided by Invitrogen) expression vectors were transfected into C6/36 cells and selected with G418 (Invitrogen). After sorting by flow cytometry and serial dilution, 5–10 stable knockdown clones were acquired. The different stable clones had been tested and shown similar knockdown effects.
The miC189 stable clone of C6/36 cells were transfected with a mixture (0.5 μg pAC5.1-C189eGFP and 4 μl FuGENE HD), forming miC189/C189 cells. Subsequently, 1.5 × 105 C6/36 cells, as well as stable clones of miN, miC189, and miC189/C189, were dispensed to each well of the 48-well plates and incubated for 24 h. DENV-2 (1.5 × 101 PFUs) was then inoculated with cells in each well at MOI = 0.0001. After incubation at 28°Cfor 1 h, the virus suspension was removed from each well and 1.1% methylcellulose containing 5% FBS medium was added to the wells. Plates were washed with PBS at 72 hpi and then 1 ml 4% paraformaldehyde was added to each well to fix cells for 20 min. Cells were washed again with PBS and then treated with 0.1% Triton X-100 at RT for 2 min in order to increase cell membrane permeability. Cells in plates were blocked with 0.2 ml 3% BSA at RT for 1 h or at 4°C overnight after another wash with PBS before being subjected to an immunofluorescence antibody test and observed under a fluorescent microscope.
C6/36 cells (2 × 105 cells/well) transfected with pAC5.1-RFP (or-eGFP), used as recipient cells, were seeded onto the 24-well plat. Another batch of C6/36 cells were infected by DENV-2 (MOI = 1) for 24 h serving as donor cells, those cells were then scratched out and transferred to upper layer of the transwell system (pore size is 4 μm) after being washed five times with PBS. In this system, recipient and donor cells were separated but not limited for virus diffusion movement between layers. To scavenge released virus particles, complement-inactivated serum containing neutralizing bodies diluted to 1:200 from human dengue patients was added, incubated at 4°C for 1 h and then 28°C for 18 h (or 6 h for that referred to RT-PCR). Those not treated with neutralizing bodies served as controls. Recipient cells were subsequently washed with PBS three times, fixed with 4% paraformaldehyde, and then reacted with monoclonal anti-DENV2 NS3 antibodies (1:100). This was followed by a reaction with anti-mouse Alexa Fluor 633 nm (1:100) and observed under a laser confocal microscope (Zeiss LSM510). For RNA identification, recipient cells harvested from the plate were subjected to RNA extraction using Trizol reagents under the protocol mentioned above. RT-PCR was thus carried out using primers (D1 and D2) [31] to detect viral RNA while the detection of 18S rRNA was utilized as the internal control as mentioned above.
To establish DENV2-infected donor cells, C6/36 cells were infected by the virus at an MOI of 1 for 24 h after transfection with eGFP expression vectors, C189eGFP expression vectors, miRNA-based knockdown vector miN and miRNA-based knockdown vector miC189, respectively. Another batch of C6/36 cells transfected with RFP expression vector was used as recipient cells. Donor cells washed by PBS to remove cell-free virus were co-cultured with recipient cells in the antiserum-containing medium (the ratio of donor cells to recipient cells was 1 to 3). At 18 h after co-culture, cells were conducted with immunofluorescence assay by fixing and staining with anti-NS3 antibody. As miRNA-based knockdown vectors contain the reporter gene EmGFP which is able to be the marker of donor cells. For viral RNA detection, infected cells scratched out from co-culture containing recipient cells soon after treatment with trypsin. Those cells that positively expressed eGFP were sorted out via flow cytometry and then subjected to RNA extraction, and viral RNA detection was carried out by RT-PCR as described above.
An anti-HA immunoprecipitation kit (Sigma, MO, USA) was utilized for this experiment. In brief, C6/36 cells were cultured in a 6-well plate until about 80% of a monolayer was formed. After transfection with pAC5.1-HAC189 (0.5 μg/well), cells were washed with culture medium once and then infected with a virus suspension at MOI = 1. Plates were incubated at 28°C for 1 h with periodical, gentle shaking for adsorption. Subsequently, 3 ml culture medium was added into each well of the plate, which was incubated again at 28°C for 48 h. After the medium was removed from wells and washed with PBS twice, 100 μl CelLytic-M Cell Lysis Reagent (Sigma) was added into each well and incubated at 4°C for 20 min. Cell lysate was transferred to a microtube and kept at -80°C for 1 h and then centrifuged at 12000 g for 10 min at 4°C after re-melting. The supernatant was transferred to a spin column in which 20 μl anti-HA-agarose was added and gently shaken at 4°C for 24 h. The spin column was then put in the collection tube. Liquid in the collection tube was discarded after centrifugation at 4°C and 13000 g for 30 sec. Then, 700 μl 1X IP buffer was added into the spin column, centrifuged at 4°C and 13000 g for 30 sec, and the liquid in the collection tube was discarded. After repeating this step for five times, 700 μl PBS was added and centrifuged again at 4°C and 13000 g for 30 sec. The spin column was then transferred to a new microtube. After discarding the liquid in the collection tube, an equal volume of sample buffer was added into the tube, heated at 95°C for 5 min in a dry bath, and then centrifuged at RT and 13000 g for 2 min. The collected product was subjected to analysis by Western blot as described below.
Harvested proteins were electrophoretically separated by 12% (w/v) SDS-PAGE in non-reducing conditions and transferred to Immobilon-P Transfer Membrane (Millipore, Darmstadt, Germany). After blocking with 5% milk-TBS-0.1% Tween 20 buffer at RT for 1 h, the membrane was stained with the indicated primary and secondary antibodies at RT for 1 h, as done previously in this lab. After final washing, the membrane was treated with Western Lightning Chemiluminescence Plus Reagent (PerkinElmer, Waltham, MA, USA) and signals were detected by FUJI X-ray film.
This experiment followed a method previously reported [32]. Briefly, C6/36 cells were cultured in a 6-well plate until an ~80% monolayer was formed. After transfection with pAC5.1-HAC189 or pAC5.1-eGFP (1 μg/well), cells in each well were infected with a virus suspension for 48 h at MOI = 1. Cells were subsequently scratched from each well and then added into the break buffer (10% w/v sucrose, 1 mM EDTA, 10 mM HEPES, 5 mM MgCl2, pH 7.4), and then homogenized on ice for at least 100 strokes in a homogenizer. The homogenized cell lysate was centrifuged at 4°C and 800 g for 10 min to remove nuclei and cell debris. The supernatant was topped up with a 10~60% sucrose gradient (10~60% w/v sucrose, 1 mM EDTA, 10 mM HEPES, 5 mM MgCl2, pH 7.4) in a tube and ultra-centrifuged at 4°C at 14440 g for 18 h. After centrifugation, 17 fractions (0.5 ml each) were collected from the top to the bottom. Of which, the first 7 fractions were pooled to one fraction as their similarity in density. As a result, a total of 11 samples were subject for further analysis by Western blot as described above.
C6/36 cells expressing C189eGFP with dengue 2 virus infection (MOI = 1) for 24 h were used as donor cells. These were washed with PBS and then scratched out to co-culture with recipient cells prepared as above with medium contain neutralizing antibodies derived from human serum. Co-cultured cells were adsorbed on a coverslip after incubating at 4°C for 1 h and then transferred to an 28°C incubator for another 6 h. Ultimately, antiDENV-2 E antibody (1:100) and anti-mouse Alexa Fluor 633 nm (1:100) were applied to IFA and observed under a laser scanning confocal microscope (Zeiss LSM510).
Uninfected C189eGFP-expressing C6/36 cells (donor cells) and RFP-expressing cells (recipient cells) were co-cultured for 30 h and then subjected to real-time observations on C189 transmission between live cells under a multi-photon confocal microscope (Zeiss LSM510 meta) at an interval of 15 sec. Donor and recipient cells were prepared as above. To observe the transmission of DENV/C189 complexes, infected C189eGFP-expressing (donor) and RFP-expressing (recipient) cells were co-cultured at 4°C for 1 h and then transferred to a 28°C incubator for another 6 h. Co-cultured cells were then observed for C189-mediated cell-to-cell transmission under a confocal microscope.
For electron microscopy, C6/36 cells (harvested 6, 12, 18, or 24 h post-infection with DENV-2) seeded on a dish or scraped from a culture dish (centrifugation at 4°C and 3000 rpm for 10 min) were immediately fixed with a mixture of 2% (v/v) glutaraldehyde and 2% paraformaldehyde in 0.1 M cacodylate buffer overnight at 4°C. After post-fixing in 1% (w/v) osmium tetroxide in 0.1 M cacodylate buffer for 2 h at room temperature, cells were washed with 0.2 M cacodylate buffer three times. Cells were again washed with 0.2 M cacodylate buffer three times and then dehydrated through an ascending series of ethanol grades. Cells were finally embedded in Spurr's resin (Electron Microscopy Science, Hatfield, PA, USA) and polymerized at 70°C for 72 h. Trimmed blocks were sectioned with an ultramicrotome (Reichert Ultracut R, Leica, Vienna, Austria). Immunocytochemistry embedding used LR White resin (London Resin Co. Ltd., Basingstoke, Hampshire, England) followed by treatment with anti-C189 antibodies and protein A tagged with 10 nm colloidal gold particles in sequence. All ultrathin sections were sequentially stained with saturated uranyl acetate in 50% ethanol and 0.08% lead citrate. Selected images were observed and photographed under an electron microscope (JEOL JEM-1230, Tokyo, Japan) at 100 kV.
Observation of the intercellular spread of DENV in C6/36 cells was first carried out by using the monolayer overlaid with 1% methylcellulose-containing semi-solid medium as the conventional plaque assay for virus titration. Infected cells sporadically appeared widely on the monolayer at 24 hpi either with or without a methylcellulose overlay (Fig 1a and 1d). Aggregates of infected cells started to appear on the overlaid monolayer at 48 hpi (Fig 1b and 1e), which became more evident at 72 hpi (Fig 1c and 1f). This implied the occurrence of DENV cell-to-cell transmission under conditions of high cell density even though the possibility of a limited rate of diffusion by released virions could not be completely excluded. Therefore, this phenomenon was further investigated by utilizing transwell or co-culture system methods to confirm the intercellular spread between virus-infected donor and RFP-transfected recipient cells (Fig 2a). A few recipient cells were infected by DENV at 18 hpi in the transwell-containing medium only while no infected recipient cells were observed in the same system in the presence of NeuAb (Fig 2b). On the other hand, most recipient cells became infected at 18 hpi in the co-culture system even when NeuAb was added to the culture medium (Fig 2b). Additionally, RT-PCR revealed that a large copy number of viral RNA was detected only from recipient cells in the co-culture system treated with NeuAb (Fig 2c). Though a low amount of viral RNA was detected in cells from the transwell system without NeuAb treatment, no sign of viral infection appeared in the presence of NeuAb (Fig 2c), further implying that the efficiency of diffuse movement by the virus released from donor cells is relatively low. All results suggested that DENV spreads from one cell to another by direct contact or cell-to-cell transmission.
Co-culture and transwell systems were further used to quantify the cell-to-cell spread of dengue virions in C6/36 cells by flow cytometry. By this quantitative assessment, eGFP-transfected recipient cells from both systems were gated by flow cytometry to detect the existence of viral NS3 antigen (Fig 3a). It was found that less than 0.01% of recipient cells in the transwell system at 2 dpi were infected while infection was 0.35% in the co-culture system. Meanwhile, 0.10% of recipient cells were NS3-positive at 3 dpi, which was significantly lower (13.15%) than for the co-culture system (Student’s t-test, p<0.01). This provided further evidence that cell contact significantly increased the efficiency of DENV intercellular spread in C6/36 cells. Further investigation was subsequently carried out in the above systems with and without specific neutralizing Abs (NeuAb), revealing that NS3 can be mostly detected in recipient cells from co-culturing (Fig 3b). When NeuAb was applied to prevent diffusion to recipient cells by released virions, no NS3 was detected in cells from the mock-infected co-culture or the lower well in the transwell system (Fig 3c). NS3-positive recipient cells from the transwell system without NeuAb (41.40%) were significantly lower than that in the co-culture system under the same conditions (63.02%) (Student’s t-test, p<0.05) (Fig 3d). This revealed that cell-to-cell transmission of DENV was more efficient than diffuse movement for cell-free virions released by donor cells.
The RNA level of C189 increases in response to DENV-2 infection in C6/36 cells [29]. This was confirmed herein at the protein level, which increased at 18 hpi and persisting until 48 hpi in infected C6/36 cells (Fig 4a). Imaging using double staining IFA showed that C189-containing spots that were formed in response to DENV-2 infection appeared to co-localize with the viral E protein in the cytoplasm of infected cells at 24 hpi (Fig 4b). This suggested that there is an intimate association between the two proteins during DENV-2 infection in C6/36 cells. To further confirm the role of C189 involved in the cell-to-cell spread of the virus, a microRNA-based system (miC189) to knockdown C189 was constructed and transfected into C6/36 cells with DENV-2 infection for 24 h. This revealed that C189 increased at the RNA level as shown before, while significantly declining to 27.34% and 15.23% in cells of miC189 with mock-infection and—transfection, respectively (Student’s t-test, p<0.01). It was even reduced to 15.23% when compared with the cell control (Student’s t-test, p<0.01) (Fig 5a). Our results also showed that the expression of C189 at the RNA level were not significantly different in infected cells with mock-transfection or those transfected with a scramble sequence (miN) (Fig 5a), indicating that no deleterious effect was derived from transfection in this experiment. The effect of C189 in DENV cell-to-cell spread was further evaluated by the overlay of 1% methylcellulose on C6/36 cells. Virus spread was evidently restricted in the miC189 group compared to the two control groups as above while a reverse increase was observed when a recovery system using a C189 overexpressing plasmid was applied to miC189 cells (Fig 5b), suggesting that virus spread was possibly associated with C189. In addition, we have obtained results from the co-culture assay that revealed enhanced efficiency of viral proteins transfer from donor cells to recipient cells but reduced when C189 was knocked down in C6/36 cells (S1 Fig). The diameter of the aggregated cell area in miC189 (170 μm on average) was significantly smaller than in the two control groups (290 μm for the untransfected group and 280 μm for the miN group), even in the group receiving C189 recovery (270 μm) (Student’s t-test, p<0.05) (Fig 5c). This indicated that C189 might be essential in cell-to-cell DENV spreading in C6/36 cells.
The spatial distribution of C189 and viral E protein in DENV2-infected C6/36 cells overexpressing eGFP-tagged C189 was investigated using confocal microscopy with pinhole narrow-down, acquiring more z-axle optical sections thinner than 0.5 μm for further observation on higher resolution images. Interestingly, viral E protein was found to be surrounded within C189-bound vacuoles (C189-VCs) that were formed in the cytoplasm in response to DENV infection in C6/36 cells (Fig 6a). An alternative approach using HAC189 fusion protein (C189 fused with HA-tag at the N-terminal domain) separately applied to DENV-infected cells further supported the viral E protein being localized in C189-VCs (S2 Fig). An ultrastructural study revealed that vacuoles containing numerous virions were usually formed in DENV-infected C6/36 cells, usually at 24 hpi (Fig 6b), which were peripherally positive to staining with anti-C189 antibodies according to the immunocytochemistry study (Fig 6c). This indicated that C189-VCs might be a reservoir not only for viral protein but also for virus particles. In order to clarify the relationship between the viral protein and C189-VCs formed in response to infection, sucrose gradient density ultracentrifugation was used to separate intracellular components from cell lysates of C6/36 cells infected by the virus and subsequently transfected with HA-tagged plasmid containing the C189 insert (HA-tagged C189). No viral protein appeared in the lysate of uninfected cells co-transfected with eGFP- and HAC189-expressing vectors according to Western blot results for proteins from each collected fraction. However, envelope (E) and capsid (C) proteins of the virus and C189 (positive to HA) were simultaneously detected in fractions with higher concentrations of sucrose (Fig 6d). In further investigation through immunoprecipitation (IP) performed with HA-tagged C189, none of the three viral proteins (E, NS3, and C) directly interacted with C189, although a weak band occurred close to the position of the E protein (Fig 6e). One possibility of this weak band reflects the heavy chain of mouse anti-HA monoclonal antibody used in this study as mentioned elsewhere [33]. These observations revealed that C189-VCs were usually formed in C6/36 cells in response to DENV infection, during which C189 spatially coexists with viral proteins or virions. However, there might be no direct binding occurring between them.
In culture for C189eGFP-expressing cells infected by DENV2, viral E protein was usually seen to co-localize with C189 at 24 hpi, which were frequently seen filopodia extended from donor cells (Fig 7a). The filopodia usually touched the recipient cell; through the contact site virus-containing C189 CVs were thus delivered to neighboring recipient cells (S1 Movie). For clarification, C189eGFP-expressing donor cells infected by DENV-2 were further co-cultured with RFP-expressing recipient cells for no longer than 6 h in the presence of NeuAb. Observations on the distribution of viral proteins and C189 by confocal microscopy showed that both the viral E protein and C189 were simultaneously detected in filopodia and were delivered to recipient cells even when cell-free virions have been blocked to release into the medium by treatment with antiserum (Fig 7b).
Arboviruses naturally infect vertebrate hosts through biting by hematophagous arthropod vectors, requiring the ability to replicate in the cells of both hosts. Like most arboviruses in humans, free DENVs are usually released into the circulatory system and subsequently infect susceptible cells via endocytosis, which are usually receptor(s)-mediated and clathrin-dependent [34, 35]. DENVs can also infect mosquito cells via a mode similar to that which occurs in vertebrate cells [36], although their specific receptors have not yet been clearly identified [13]. DENV dissemination within the mosquito has recently attracted increased attention. It seems that DENV-2 can infect various organs (e.g., neural tissue and salivary glands [20]) disseminating from the midgut, which is the original lodging and replication site of the virus when ingested along with blood meals [37]. The virus travels from the midgut to salivary glands before being transmitted to another host.
The virus probably does not depend highly on the release-and-entry mode due to the close proximity of epithelial cells within a tissue such as the midgut in the vector host [18]. Since mosquito cells do not die from DENV infection [15, 16], the release-and-infect mode of extending infection into mosquito tissues is less likely. In the present study, intercellular DENV spread in C6/36 cells was first detected by limiting the diffusion of released viruses within methylcellulose-containing medium. As the aggregation of infected cells became more evident, it brought up the possibility of DENVs being transmitted to neighboring cells by a cell-to-cell mode. A similar result also occurred in assays with transwell and co-culture systems, even when NeuAb was applied to neutralize released virions to eliminate the effect of diffusion. Recipient cells were more easily infected by donor cells infected with DENVs, especially when the possibility of contact between the two types of cells was high.
Increasing evidence reveals that specific host genes may be up-regulated, playing important roles in regulating viral infection and corresponding responses by infected cells [38]. The gene effects may cover most steps of the infection, from the viral entry, RNA replication, as well as assembly and release of progeny virions [39]. For instance, a total of 305 host proteins have been identified from human cells infected by West Nile virus, suggesting a close association between the virus and host cells [40]. Some of the identified host factors are important in cell-to-cell transmission of the virus [28]. In fact, the novel tetraspanin C189 from C6/36 cells is usually up-regulated in response to DENV-2 infection and is suspected to associate with the intercellular spread of the virus among mosquito cells [29]. Tetraspanins such as CD9 and CD81 play a role in modulating cell-to-cell transmission during HIV infection [41], consistent with our theory [29]. In this study, we have further observed co-localized endogenous C189 and viral E protein in C6/36 cells. When C6/36 cells were transfected with a C189-expressing construct, highly expressed C189 was incorporated into the membrane of virus-responsive vacuoles, called C189-containing membrane-bound vacuoles (C189-VCs), within which virions and/or viral proteins were confined. This further implicates the importance of tetraspanin involvement in the intercellular spread of arboviruses such as DENVs in mosquito cells. It is consistent with some extracellular vesicles that contain clusters of tetraspanins in their membranes to form microdomains [42].
Tetraspanins naturally existing in a broad spectrum of organisms possesses various biological functions including the regulation of cell proliferation, cell-cell or cell-extracellular matrix interactions, and cell motility [43]. They may also be involved in promoting metastasis in carcinomas [44] and act as transmembrane linkers [45, 46]. Further, tetraspanins were found to be involved in different kinds of microbial infections [47] such as T lymphocyte infection by HIV-1 and HTLV [48]. More specifically, in hepatitis C virus (HCV) infection of hepatocytes, CD81may serve as a receptor or co-receptor presumably via its association with the viral E2 protein [49, 50]. Nevertheless, according to the present study, C189 may not serve as the receptor of C6/36 cells during infection by DENVs as virus production evidently did not change in C6/36 cells with knockdown of C189 and/or further recovery by overexpression of the same protein. It seems that this tetraspanin more likely participates in intercellular transmission of C189-VCs formed in cell cytoplasm, which is supposed to be beneficial for virus delivery between C6/36 cells having resistance to host defense as in hepatocytes infected by HCV [50, 51]. To confirm the specific role of C189 involvement in the intercellular transmission of associated vacuoles, C6/36 cells were over-expressed with a panel of different proteins, including tetraspanins C189 and ER-related proteins including GRP94/endoplasmin and GRP78/BiP, as well as controls containing only eGFP (S3 Text and S3 Fig). Among these, only C189 was efficiently transferred from donor cells to recipient cells, revealing the essentialness of C189 on this process. Since C189 overexpression did not cause significant apoptosis in transfected cells, there was no possibility that artifacts originating from apoptotic bodies were engulfed by recipient cells [52]. C189 overexpressed in uninfected C6/36 cells can also translocate to neighboring cells, indicating that transmission is relatively dependent on cell-cell contact. In a culture containing a lower density of C6/36 cells, C189-VCs with or without virions were also delivered to neighboring cells along filopodia extended by the cell (S1 Movie). As cell death usually does not occur in mosquito cells infected by DENVs, at least for certain strains [17], release-and-entry may not be a normal way for viruses to spread in infected tissues of those cases. As a result, cell-to-cell transmission can be a direct mode for infecting neighboring cells [22]. Generally, tetraspanin proteins are able to mediate cellular penetration, invasion, and fusion events and define a membrane microdomain [53]. They may also modulate virus-induced membrane fusion [41]. In fact, HIV and perhaps also HCV launch cell-to-cell transmission by way of cell contact, in which the tetraspanin CD81 is believed to be involved [26, 28].
In addition to endogenous C189, we have also extensively used fluorescence-tagged C189 as a proxy in this study. Although it is not a completely natural condition, this approach provided convenience and advantages in observing how C189 is associated with DENV, particularly the intercellular spread of the virus. Taken together, DENV generally establishes infections through a release-and-entry mode in cultures, while they are more likely to employ cell-to-cell transmission as an alternative route for spreading progeny virus particles [54]. Tetraspanin C189 was up-regulated and incorporated into the membrane of virus-induced vacuoles, which are presumed to become rapid and efficient vehicles in the delivery of the virus from one cell to another. Our findings are interesting and may be responsible for lateral spread of DENV within the same tissue in a mosquito, and further understanding of its mechanism would increase our ability to intervene in the spread of DENV infections.
|
10.1371/journal.pgen.0030075 | Being Pathogenic, Plastic, and Sexual while Living with a Nearly Minimal Bacterial Genome | Mycoplasmas are commonly described as the simplest self-replicating organisms, whose evolution was mainly characterized by genome downsizing with a proposed evolutionary scenario similar to that of obligate intracellular bacteria such as insect endosymbionts. Thus far, analysis of mycoplasma genomes indicates a low level of horizontal gene transfer (HGT) implying that DNA acquisition is strongly limited in these minimal bacteria. In this study, the genome of the ruminant pathogen Mycoplasma agalactiae was sequenced. Comparative genomic data and phylogenetic tree reconstruction revealed that ∼18% of its small genome (877,438 bp) has undergone HGT with the phylogenetically distinct mycoides cluster, which is composed of significant ruminant pathogens. HGT involves genes often found as clusters, several of which encode lipoproteins that usually play an important role in mycoplasma–host interaction. A decayed form of a conjugative element also described in a member of the mycoides cluster was found in the M. agalactiae genome, suggesting that HGT may have occurred by mobilizing a related genetic element. The possibility of HGT events among other mycoplasmas was evaluated with the available sequenced genomes. Our data indicate marginal levels of HGT among Mycoplasma species except for those described above and, to a lesser extent, for those observed in between the two bird pathogens, M. gallisepticum and M. synoviae. This first description of large-scale HGT among mycoplasmas sharing the same ecological niche challenges the generally accepted evolutionary scenario in which gene loss is the main driving force of mycoplasma evolution. The latter clearly differs from that of other bacteria with small genomes, particularly obligate intracellular bacteria that are isolated within host cells. Consequently, mycoplasmas are not only able to subvert complex hosts but presumably have retained sexual competence, a trait that may prevent them from genome stasis and contribute to adaptation to new hosts.
| Mycoplasmas are cell wall–lacking prokaryotes that evolved from ancestors common to Gram-positive bacteria by way of massive losses of genetic material. With their minimal genome, mycoplasmas are considered to be the simplest free-living organisms, yet several species are successful pathogens of man and animal. In this study, we challenged the commonly accepted view in which mycoplasma evolution is driven only by genome down-sizing. Indeed, we showed that a significant amount of genes underwent horizontal transfer among different mycoplasma species that share the same ruminant hosts. In these species, the occurrence of a genetic element that can promote DNA transfer via cell-to-cell contact suggests that some mycoplasmas may have retained or acquired sexual competence. Transferred genes were found to encode proteins that are likely to be associated with mycoplasma–host interactions. Sharing genetic resources via horizontal gene transfer may provide mycoplasmas with a means for adapting to new niches or to new hosts and for avoiding irreversible genome erosion.
| Organisms belonging to the Mycoplasma genus (class Mollicutes) are commonly described as the simplest and smallest self-replicating bacteria because of their total lack of cell wall, the paucity of their metabolic pathways, and the small size of their genome [1,2]. In the 1980s, they were shown to have evolved from more classical bacteria of the firmicutes taxon by a so-called regressive evolution that resulted in massive genome reduction [3,4].
One of the models attempting to improve understanding of the evolution of bacteria with small genomes proposes that erosion of bacterial genomes is more prone to occur in bacterial populations that are spatially isolated and sexually deficient [5]. In restricted habitats, the environment is rather steady and natural selection tends to be reduced, resulting in the inactivation of many genes by genetic drift [5,6]. In this scenario, DNA acquisition would be strongly limited, resulting, after losses of large genomic regions and accumulation of mutations, in genome stasis [7]. This evolution scheme is relevant for a number of obligate intracellular bacteria, including insect endosymbionts (e.g., Buchnera and Wigglesworthia spp.), and arguably for Chlamydia, and Rickettsia spp. The recent findings of a putative conjugative plasmid in Rickettsia felis [8] and of a substantial number of prophage, transposase and mobile-DNA genes in the insect endosymbiont Wolbachia pipientis challenged this model and it was proposed that gene inflow by horizontal gene transfer (HGT) may occur in some obligate intracellular species depending on their lifestyles [9].
Mycoplasmas share with obligate intracellular bacteria a small genome size with marked AT nucleotide bias and a low number of genes involved in recombination and repair, but forces driving their evolution may not be quite the same, as they do have a very different lifestyle. Indeed, mycoplasmas mainly occur as extracellular parasites [10] and are often restricted to a living host, with some species having the ability to invade host cells [11]. They have a predilection for the mucosal surfaces of the respiratory and urogenital tracts, where they successfully compete for nutrients with many other organisms, establishing chronic infections (Table S1). Therefore, mycoplasma populations are far from being isolated and inhabit niches where exchange of genetic material may take place. The none-to-rare occurrence of HGT reported so far for mycoplasmas [12] is therefore surprising and seems to conflict with their lifestyle. On the other hand, HGT may depend on several other factors [9] that were described as limited or lacking in most mycoplasma species and that include an efficient machinery for recombination, genetic mobile elements such as prophages or conjugative plasmids, and a means for DNA uptake. However, this view of mycoplasma biology is changing, since homologous recombination has been demonstrated in these bacteria [13,14] and some new means of exchanging DNA are being discovered [15,16]. Indeed, several pathogenic mycoplasma species relevant to the veterinary field and the murine pathogen M. pulmonis were recently shown to form biofilms [17,18], structures that have been proposed to promote DNA exchange among bacteria. This finding, together with previous evidence for DNA transfer under laboratory conditions in M. pulmonis via conjugation [19], raises the exciting question of whether some mycoplasmas species are sexually competent. Subsequently, this would suggest that mycoplasma species which co-infect the same host niches might exchange genetic material. Remarkably, biofilm formation and the occurrence of an integrative conjugative element (ICE) have both been newly described in the M. agalactiae species [16,18]. This pathogen is responsible for contagious agalactia in small ruminants [20], a syndrome that includes mastitis, pneumonia, and arthritis and that is also caused by some members of the so-called mycoides cluster, such as M. capricolum subsp. capricolum and M. mycoides subsp. mycoides Large Colony. Although producing similar symptoms in the same host, these species belong to two distinct and distant branches of the mollicute phylogenetic tree (Figure 1). Their relative phylogenetic positions are irrespective of whether the tree is constructed from aligned 16S rDNA (Figure 1A) or from 30 aligned proteins shared by all living organisms [21] (Figure 1B). M. agalactiae belongs to the hominis phylogenetic branch, together with a closely related ruminant pathogen, M. bovis, while the six members that comprise the “mycoides cluster” belong to the spiroplasma phylogenetic branch [22]. Whole-genome sequences are available for two members of the mycoides cluster; M. mycoides subsp. mycoides SC [23], which is responsible for contagious bovine pleuropneumonia [24], and M. capricolum subsp. capricolum [25]. In contrast, there is a limited amount of sequence data available for M. agalactiae and M. bovis. Mycoplasmas that have been fully sequenced in the hominis phylogenetic group are a murine pathogen M. pulmonis [26], a swine pathogen M. hyopneumoniae (strain 232 [27]; strains 7748 and J [12]), an avian pathogen M. synoviae [12], and a mycoplasma isolated from fish, M. mobile [28] (Figure 1B).
Mechanisms underlying ruminant mycoplasma diseases have yet to be elucidated and very little is known regarding the mycoplasma factors that are involved in virulence and host interaction. Genes thus far identified in M. agalactiae and for which a function in relation to virulence has been predicted are (i) a family of phase-variable related surface proteins, designated as Vpma, which are encoded by a locus subjected to high-frequency DNA rearrangements and could be involved in adhesion [29,30], (ii) the P40 protein, which is involved in host–cell adhesion in vitro but is not expressed in all field isolates [31], and (iii) the P48 protein, which has homology to an M. fermentans product with a macrophage-stimulatory activity [32]. Several of these gene products have homologs in M. bovis but not in mycoplasmas of the mycoides cluster.
Whole-genome comparison between phylogenetically distant mycoplasmas that colonize the same host could provide a basis from which to comprehend the factors involved in mycoplasma host adaptation. With this initial goal, we sequenced the M. agalactiae genome of the pathogenic type strain PG2. Results revealed a classical mollicute genome with a coding capacity of 751 CDSs, half of which are annotated as encoding hypothetical products.
Unexpectedly, comparative analysis of the M. agalactiae genome with that of other mollicutes and bacteria suggests that a significant amount of genes (∼18 %) has been horizontally transferred to or acquired from mycoplasmas of the mycoides cluster that are phylogenetically distant while sharing common ruminant hosts. In light of these data, we re-examined mollicute genomes for HGT events with a particular focus on those that occurred after mycoplasmas branched into three phylogenetic groups (see Figure 1 for the hominis, pneumoniae, and spiroplasma phylogenetic groups). Our analyses confirm data so far reported regarding the low incidence of HGT between Mycoplasma species with the exception of that described in this study, between M. agalactiae and members of the mycoides cluster and, to a lesser extent, between M. gallisepticum and M. synoviae. To our knowledge, this is the first description of large-scale horizontal gene transfer between mycoplasmas.
The genome of the M. agalactiae type strain PG2 consists of a single, circular chromosome; general features are summarized in Table 1. The genome sequence was numbered clockwise starting from the first nucleotide of the dnaA gene, which was designated as the first CDS (MAG0010). This gene is involved in the early steps of the replication initiation process [33] and is typically located near mycoplasma origins of replication. Indeed, dnaA boxes flanking the dnaA gene were shown in M. agalactiae to promote free replication of the ColE1-based E. coli vectors in which they were cloned [34]. Although these experiments clearly localized the M. agalactiae oriC in the vicinity of the dnaA gene, whole-genome analysis did not indicate a significant GC-skew inversion [35] in this region (unpublished data). In contrast to other mycoplasma genomes [36], a high level of gene-strand bias was not observed, even when restricting the analysis to the dnaA vicinity.
Overall, M. agalactiae strain PG2 possesses a typical mollicute genome, with a small size (877,438 bp), a low GC content (29.7 moles %), a high gene compaction (88% of coding sequence), and UGA preferentially used as a tryptophan codon over UGG (Table 1). Its GC% value is slightly higher than that observed for some other mycoplasma species but is close to the average GC content (28%) calculated from the 16 available mollicute genomes. Using the CAAT-box software package, 751 CDSs were identified, 404 (53.8%) of which had a predicted function. The genome also contains 34 tRNA genes and two nearly identical sets of rRNA genes with two 16S–23S rRNA operons (MAG16S1-MAG23S1 and MAG16S2-MAG;23S2) and the two 5S rRNA genes (MAG5S1 and MAG5S2) clustered in two loci separated from each other by ∼400 kb (Figure 2).
Prediction of M. agalactiae CDS function was based on BLAST searches against SwissProt, trembl, and MolliGen databases. For CDSs showing significant similarities with database entries, most best BLAST hits (BBH) were found with M. synoviae and M. pulmonis, which belong, together with M. agalactiae, to the hominis phylogenetic group (Figure 1). Unexpectedly, a large number of BBH were also obtained with M. mycoides subsp. mycoides SC or M. capricolum subsp. capricolum, which both belong to the mycoides cluster (Figure S1). Since this cluster is exclusively composed of ruminant pathogens and is relatively distant from M. agalactiae in the mollicute phylogenetic tree (Figure 1), this prompted us to closely examine the corresponding CDSs. A total of 136 M. agalactiae CDSs were then identified as having their BBH with organisms from the mycoides cluster, with 50 having no significant similarity outside of this cluster (Table S2). Of the remaining 86, 73 also had a homolog in at least one in the four available genomes of the hominis group (M. pulmonis, M. mobile, M. synoviae, and M. hyopneumoniae) (Table S2) and 13 in other mollicutes or bacteria (Tables S2 and S3). Further phylogenetic tree reconstruction showed that 75 out of 86 CDSs display highly significant bootstrap values (≥ 90%) supporting HGT with homologs of the mycoides cluster. Among the 11 CDSs with low bootstrap values, six belong to gene clusters in which synteny is conserved in the mycoides cluster, three belong to an ICE element (see below) found in M. agalactiae and M. capricolum subsp. capricolum and two others were not further considered, suggesting that ∼134 CDS have undergone horizontal gene transfer in between mycoplasma(s) of the mycoides cluster and M. agalactiae or its ancestor.
Of the predicted transferred CDSs, nine and 22 have a homolog either in M. mycoides subsp. mycoides SC or in M. capricolum subsp. capricolum, respectively, while 102 have homologs in both species. Phylogenetic analysis and similarity comparisons of the 102 CDSs did not allow us to conclude whether they were more similar to M. mycoides subsp. mycoides SC or to M. capricolum subsp. capricolum (Figure S2). Additionally, one CDS (MAG4270) had a homolog only in M. mycoides subsp. capri, for which only a limited number of sequences are available. The occurrence of HGT was further supported by the genomic organization in M. agalactiae of 115 of the predicted transferred genes that occur as clusters containing two to 12 elements with approximately half of them displaying the same organization as in M. mycoides subsp. mycoides SC and M. capricolum subsp. capricolum genomes. Eleven of these clusters, which are distributed all over the M. agalactiae genome, are shown in Figure 2.
As previously mentioned, 73 of the predicted transferred CDSs have an ortholog in genomes of the hominis group. In a hypothesis regarding transfer from the mycoides cluster to M. agalactiae, one might expect to detect pseudo-paralogs [37] in the M. agalactiae genome, with one inherited from an ancestor of the hominis group, while the other was acquired by HGT. Indeed, in 17 unambiguous cases, vertically and horizontally inherited pseudo-paralogs were found. As an example, the gene encoding the glucose-inhibited division protein is present as a single copy in the genomes of M. pulmonis, M. synoviae, M. mobile, and M. hyopneumoniae. In M. agalactiae, two copies of this gene were found; one, MAG2970, has a BBH in M. pulmonis, while the other, MAG1470, has a BBH in M. mycoides subsp. mycoides SC. The oligopeptide ABC transporter locus (opp genes) is another interesting example, since opp genes occur twice in M. agalactiae, at two distinct loci. As shown in Figure 3, one opp locus (designated as the type 1) is composed of four opp genes (B–D and F), the sequences of which are highly similar to those of one of the two M. pulmonis opp loci. The other opp locus of M. agalactiae (type 3, Figure 3) is composed of five opp genes (A–D and F), the sequences and organization of which are closer to one of the two opp gene loci of M. capricolum subsp. capricolum and M. mycoides subsp. mycoides SC. Phylogenetic analyses of the oppB genes of types 1 and 3 with homologous sequences of other mycoplasma species suggest different origins for the two M. agalactiae opp loci. While the type 1 was inherited from a common ancestor of the hominis branch, the type 3 was laterally acquired from the mycoides cluster. A third, isolated, copy of the oppB gene (MAG4700) was predicted in the M. agalactiae genome, and might represent a relic of a displaced opp operon, as its best orthologs were found in mycoplasmas of the hominis group.
For CDSs found only once in the genome of M. agalactiae, the situation might be more complex, as illustrated by the glycerol kinase/glycerol uptake facilitator operon, glpK–glpF (MAG4470–MAG4480), which was unambiguously found to originate from a mycoides ancestor (Figure S3). This operon occurs as a single copy in all mycoplasma genomes of the hominis group but is absent from M. synoviae. Because of the relative phylogenetic closeness of M. agalactiae and M. synoviae (Figure 1B), the question arises as to whether glpK–glpF was lost in their common ancestor and acquired later on by M. agalactiae from the mycoides cluster.
While examining M. agalactiae candidates for HGT, sequence alignments showed that 38 are truncated versions of their homologs in M. capricolum subsp. capricolum and M. mycoides subsp. mycoides SC, or were annotated as pseudogenes (Table S2 and Figure S2).
Additionally, only 14 CDSs were suspected to have undergone HGT between M. agalactiae and species of the pneumoniae phylogenetic group or non-mollicute bacteria (Table S3).
Since restriction–modification (RM) systems serve in bacteria as a tool against invading DNA [38], it was of interest to specifically search for these systems in light of the high level of HGT in M. agalactiae. One locus encoding a putative RM system is composed of six genes with homology to type I RM systems (Figure S4) and was designated hsd. It contains (i) two hsdM genes (MAG5650 and MAG5730), coding for two almost identical modification (methylase) proteins (94% identity), which would methylate specific adenine residues; (ii) three hsdS genes (MAG5640, MAG5680, and MAG5720), each coding for a distinct RM specificity subunit (HsdS) that shares homology with the others (between 50% to 97% similarities); and (iii) one hsdR pseudo-gene (MAG5700/MAG5710), which is interrupted in the middle by a stop codon and would otherwise encode a site-specific endonuclease (HsdR). Finally, the hsd locus contains two hypothetical CDSs (MAG5660 and MAG5670) and one gene (MAG5690), whose product displays 76.9 % similarity to a phage family integrase of Bifidobacterium longum [39] and motifs found in molecules involved in DNA recombination and integration. In M. pulmonis, the hsd locus has been shown to undergo frequent DNA rearrangements but the gene encoding the putatively involved recombinase is located elsewhere on the genome [26,40].
Apart from this locus, only three other unrelated M. agalactiae CDSs display similarities with the restriction–modification system, one of which was annotated as a pseudogene.
Mycoplasma lipoproteins are of particular interest because they have been proposed to play a role in the colonization of specific niches and in interaction with the host [11,41]. In order to identify the putative lipoproteins encoded by the M. agalactiae genome, we combined results obtained by PS-SCAN analysis with the detection of a signature that was defined by using MEME/MAST software and a set of previously identified mycoplasma lipoproteins (see Material and Methods). This strategy resulted in the prediction of 66 lipoproteins, 85% of which were annotated as hypothetical proteins. The remaining 15% correspond to the previously characterized Vpmas, P40, P30, and P48; and to two CDSs homologous to the substrate-binding protein of an oligopeptide (OppA, MAG0380) and to an Alkylphosphonate ABC (MAG5030) transporter, respectively.
Among the genes encoding the 66 predicted lipoproteins, our analyses indicated that the corresponding genes of 19 have undergone HGT with the mycoides cluster (see Tables S2, S5, and S6). These 19 CDSs were annotated as hypothetical proteins, however, four (MAG2430, MAG3260, MAG6480, and MAG7270) share a high level of similarity, and constitute, with nine other polypeptides (MAG0210, MAG0230, MAG1330, MAG1340, MAG3270, MAG4220, MAG4310, MAG6460, and MAG6490), a protein family. A MEME/MAST analysis indicated that the 13 proteins of this family shared one to ten repeats of a 25 amino-acid motif A ([KN]W[DN][TV]SNVT[ND]MSSMFxGAK[KS]FNQ[DN][IL]S) (Figure S5). This motif is highly similar to the DUF285 domain of unknown function predicted in a large number of mycoplasma lipoproteins and found only in the mycoides cluster and in some non-mollicute bacteria (i.e., Listeria monocytogenes, Enterococcus faecalis, Lactobacillus plantarum, and Helicobacter hepaticus). A second motif B ([FM]PKN[VT][KV]KVPKELP[EL][EK][IV]TSLEKAFK[GN]) was also found in most of the family proteins. Of the 13 members of the family, whose corresponding genes are distributed all over the chromosome, five were predicted to be lipoproteins; the others may constitute a reservoir of sequence to generate surface variability. Altogether, these data suggest that M. agalactiae has inherited a family of genes encoding potentially variable lipoproteins that are otherwise specific to the mycoides cluster.
Another remarkable lipoprotein family is found in the portion of the genome (MAG7050–MAG7100; Figure S4) that encodes the phase-variable, related Vpma products. The Vpma family has been extensively described [29,30] and was previously shown to present typical elements of mobile pathogenicity islands [29]. However, comparison of the Vpmas coding sequences with other mycoplasma genomes indicate that they are specific of the M. agalactiae species, although their variation in expression and genetic organization closely resembles the Vsp system found in the close relative M. bovis [42–44]. No similar system or coding sequences was found in the mycoides cluster.
To our knowledge, attempts to naturally transform M. agalactiae or other mycoplasma species have failed, suggesting that HGT, if it occurs, is mediated via another mechanism. Only a limited number of viruses or natural plasmids have been described so far in mycoplasmas that could account as vehicles for HGT, apart from a new ICE that has been described in a few Mycoplasma spp. [12,15]. In a recent study, we documented the occurrence of such an element in M. agalactiae strain 5632 (ICEA5632) as chromosomal multiple copies and as a free circular form [16].
One copy, ICEA5632-I, was fully sequenced and Southern blot analyses suggested that it occurs in a minority of strains that did not include the PG2 type strain [16,45]. However, detailed sequence analyses performed in this study revealed that 17 CDSs of the M. agalactiae PG2 genome display different levels of similarities to CDSs present in ICEA5632-I and in other ICEs (Table S4) found in M. capricolum subsp. capricolum (ICEC), M. fermentans (ICEF-I and –II) [15], and M. hyopneumoniae strain 7448 (ICEH) [12]. These seventeen CDSs are clustered in the PG2 genome within a unique 20-kb locus, ICEAPG2 (Figure 4), and those with an ortholog in M. fermentans ICEF and/or M. agalactiae ICEA5632-I were designated as in previous reports [15,16]. Surprisingly, best alignments for ICEA products of the PG2 strain were consistently obtained with M. capricolum subsp. capricolum ICEC counterparts, with an average of 40% identity and 75% similarity, whereas alignments with ICEA5632-I or ICEF gave lower values. This close relationship between ICEAPG2 and ICEC was confirmed by bootstrap values of the phylogenetic trees inferred from the amino acid sequence of TraG, TraE, ORF19, and ORF22 (Figure S6). Moreover, ICEAPG2 and ICEC share three homologous CDSs (noted as x, y, and z in Figure 4) lacking in ICEA5632-I and other ICEs. All these results indicate a close relationship between ICEAPG2 and ICEC, and suggest that the ICEs found in strains PG2 and 5632 have a different history.
In strain PG2, the gene encoding TraE (MAG3910/MAG3920), a major actor in DNA transport across the conjugative pore, was found to be disrupted. In addition, a total of 11 out of the 20 ICEAPG2-CDSs might represent pseudogenes (hatched arrows in Figure 4), due to the presence of stop codons and/or frameshifts. Finally, regions directly flanking ICEAPG2 do not display the typical motifs found on each side of integrated ICEF and ICEA5632. These data strongly suggest that ICEAPG2 is unlikely to be functional.
In M. agalactiae strain 5632, ICEA5632-I excision leads to a chromosomal site that is reorganized into an “empty” locus carrying remnant motifs that cover a 476-bp sequence [16]. Interestingly, in the PG2 chromosome, a 476-bp sequence located ∼ 270 kb upstream from ICEAPG2 was found that is 94% identical to the sequenced “empty” ICEA5632-I locus, and includes the putative remnant motifs in the same order and spacing (Figure S7). Unfinished sequence data from the strain 5632 reveals that this 476-bp sequence is actually part of a larger (∼40 kb) synthenic region between PG2 and 5632.
The high number of CDSs predicted to have undergone HGT between M. agalactiae and organisms of the mycoides cluster prompted us to examine possible HGT events among other mycoplasma species whose genomes have been sequenced. For each mycoplasma genome, the CDSs with a BBH in a phylogenetic group different from that of the query were then identified (see Materials and Methods). Phylogenetic analyses, when possible, were applied to detect which, among the identified CDSs, were candidates for HGT (Table 2). Overall, this analysis clearly pointed out two cases of significant HGT levels, between the mycoides cluster and M. agalactiae and between M. gallisepticum and M. synoviae. Detailed examination of the data revealed a clear picture for M. synoviae, in which all identified CDSs but one designate M. gallisepticum as the HGT partner (Tables 2, S8, and S9). This is confirmed by the reciprocal data in M. gallisepticum, although in several cases the phylogeny was not strong enough to support with certainty a direct association with M. synoviae. These data are consistent with a previous study in which HGT between those two species was suspected [12]. No significant HGT was detected among other mycoplasma species across phylogenetic groups apart from that described above between M. agalactiae and mycoplasmas of the mycoides cluster (see also Tables S5 and S6).
For the human mycoplasma M. penetrans, which has the largest genome of the dataset, a fairly large number of CDSs had BBH in a phylogenetic group other than the pneumoniae group. However, none of these candidates for HGT were confirmed by further phylogenetic analysis.
Sixteen genome sequences from different mycoplasma species are now available in public databases and provide comprehensive data for comparative genomic studies that will, for instance, contribute to the understanding of their intriguing regressive evolution (by loss of genetic material) from Gram-positive bacteria with low GC content. Indeed, mycoplasmas are thought to be fast-evolving bacteria, as supported by their positioning on some of the longest branches of the bacterial phylogenetic tree [21]. This observation is in agreement with their small genome size, and hence with their limited DNA-repair capabilities [46]. Consequently, mycoplasma genomes would be prone to accumulate mutations that would contribute to further downsizing. In this scenario, acquisition of new genes by HGT was not considered to play a major role in mycoplasma evolution. Indeed, statistical analyses predicted that the smallest proportion of HGT occurred among bacteria in symbiotic or in parasitic species, including mycoplasmas [47]. Nonetheless, a few remarkable cases of HGT involving mycoplasmas have been described that include the independent displacements of the rpsR and ruvB genes with orthologs from ɛ–Proteobacteria [48,49] and the horizontal transfer of the surface-protein VlhA encoding gene among three phylogenetically distant mycoplasmas (M. gallisepticum, M. imitans, and M. synoviae), which are respiratory pathogens of gallinaceous birds [50,51]. More recently, sequencing of the M. synoviae genome suggested that ∼3% of the total genome length has undergone HGT in between M. gallisepticum and M. synoviae [12]. Analyses performed in this study confirmed this trend using a different approach, which estimated that ∼3%–8 % of their CDS have been involved in HGT in between the two avian species. However, these values are much lower than the ones found for M. agalactiae, in which 10%–18% of its coding genome was predicted to have undergone HGT with mycoplasmas belonging to the mycoides cluster. This proportion represents, to our knowledge, the highest extent of HGT for a bacterium with a small genome size (<1 Mb). The scattering of the HGT loci all over the M. agalactiae genome suggests the occurrence of multiple HGT events and/or the shuffling via intrachromosomal recombination events of alien genes after integration. Although HGT events could be confirmed by phylogenetic analyses, it was not possible to identify significant biases in the GC composition of the transferred genes that would distinguish them from ancestral genes. It is likely that the HGT events in M. agalactiae did not take place recently and/or that the acquired sequences quickly adjusted to their new genome pattern. In fact, it has been shown that the bias in GC content is not a reliable indicator for detecting HGT events [52,53].
Demonstrating the acquisition of genes by HGT is not trivial, especially among mycoplasma species that share a number of genetic features and are phylogenetically clustered. Analyses of the M. agalactiae genome with respect to HGT with mycoplasmas of the mycoides cluster revealed roughly two categories of CDSs: one composed of CDSs with several homologs and their BBH within the mycoides cluster, and one composed of CDSs that have few or no homologs but are highly similar to CDSs of the mycoides cluster. While for the first category, phylogenetic tree reconstruction can demonstrate or refute HGT, the issue is more delicate for the second. For instance, 50 CDSs of M. agalactiae have no homolog other than in the phylogenetically distinct mycoides cluster, raising the question of whether these genes were laterally acquired from these mycoplasmas or from a third common partner that has yet to be identified. In addition, sharing the same host might have resulted in M. agalactiae and mycoplasmas of the mycoides cluster retaining a common ancestral set of genes that were lost in all other species that do not colonize ruminants. Although these alternative hypotheses cannot be formally ruled out, they all imply a series of parallel, independent events. Taking into account that M. agalactiae, when compared to other sequenced mycoplasmas species of the same phylogenetic group (Figure 1B), is located on one of the most ramified branches of the phylogenetic tree, this scenario seems unlikely.
The more global analyses performed on the available genomes from mollicutes (with the exception of phytoplasmas) and on M. agalactiae identified four species in which HGT has taken place. Detailed results clearly identified only two pairs of partners, each from a different phylogenetic group: (i) M. agalactiae and the mycoplasmas of the mycoides cluster, and (ii) all mycoplasma pathogens of ruminants and M. gallisepticum and M. synoviae, two pathogens of poultry. This striking observation tends to indicate that mycoplasmas sharing a common host have the capacity to exchange genetic material. These mycoplasma species are the only ones sequenced thus far that are located in different phylogenetic groups but share the same lifestyle in terms of ecological niches (Table S1).
Indeed, other sequenced species that share the same host all clustered into the same phylogenetic group (human mycoplasmas of the pneumoniae group) and therefore our approach will not detect HGT among these mycoplasmas. For one human mycoplasma, M. penetrans, a number of putative HGTs were found (see Tables 2 and S7) but none were supported by phylogenetic analyses. The occurrence of HGT among human mycoplasma species cannot be dismissed by this study and remains to be investigated.
A striking feature of the HGT in this bacterium is that nearly all the events were predicted to have occurred with species of the mycoides cluster, which are, with M. agalactiae, pathogens for ruminants. Sharing this common environment would have favored the transfer of genetic material between these mycoplasmas and the fixation of genes leading to an increased fitness as parasites of ruminants. Interestingly, ∼30 % of the genes that have undergone HGT with a mycoides ancestor correspond to membrane-associated proteins, including several transporters or lipoproteins (Table 3). As surface proteins such as lipoproteins are supposed to play a major role in the mycoplasma–host interaction, this finding supports the proposal that genes acquired by HGT may have significantly favored the colonization of ruminants by the mycoplasma. Noticeably, a family of 13 CDSs of M. agalactiae has undergone HGT with the mycoides cluster. The predicted proteins contain repeats of a domain of unknown function (DUF285). The distribution of this domain in mycoplasmal proteins is strictly restricted to species belonging to the mycoides cluster. Whether this family, which includes several lipoproteins, participates in the interaction between the mycoplasma and its ruminant host remains to be elucidated.
At present, it remains rather difficult to evaluate the selective advantage that could have provided the acquired genes, especially because half of them encode proteins with unknown functions. A possible exception could be the oligopeptide transport system Opp, of unknown specificity, for which two loci have been found in the M. agalactiae genome. In bacteria, Opp transport systems participate in a wide range of biological events including biofilm formation [54], antimicrobial-compound production [55], and adaptation to specific environments [56–58], including milk [59,60]. The substrate specificity of Opp systems is determined by the OppA subunit and it is apparent that the predicted OppA proteins from the two M. agalactiae systems do not share any sequence similarity, in contrast to the other Opp subunits. Interestingly, one of them (MAG1000) shows 44% similarity with an M. hominis ortholog that is a lipoprotein involved in adherence to host cells and proposed to be a major ATPase [61,62]. The other OppA subunit (MAG0380) shows 83% and 82% similarity with M. capricolum subsp. capricolum and M. mycoides subsp. mycoides SC OppA, respectively. Although further studies are required to determine the role of the two Opp systems in M. agalactiae, it is reasonable to propose that the Opp system inherited from the mycoides ancestor could be directly involved in the adaptation to their ruminant hosts.
In virulent M. mycoides subsp. mycoides SC strains, cytotoxic effects towards host cells have been correlated with the ability of the bacteria to produce high amounts of hydrogen peroxide during the catabolism of glycerol [63,64]. Glycerol is imported and phosphorylated via two alternative systems, the GlpK/GlpF system and the GtsABC transporter. The glycerol-3-phosphate enters glycolysis after oxidation by the l-alpha-glycerophosphate oxidase (GlpO); this step results in production of H2O2 as a toxic by-product.
In M. agalactiae, it is noteworthy that gene clusters encoding the GlpK/GlpF and GtsABC systems have probably been inherited from a mycoides ancestor, suggesting an ability to import glycerol for energetic metabolism. However, in M. agalactiae, there is no glpO gene upstream of glpK/glpF, as found in M. mycoides subsp. mycoides SC and M. capricolum subsp. capricolum genomes, or elsewhere in the genome. As in several other mollicutes, a gene encoding a glycerol-3-phosphate dehydrogenase (gpsA, MAG0500) is present in M. agalactiae. This suggests that M. agalactiae is able to efficiently import glycerol and to use it as a carbon and energy source but that glycerol catabolism is not coupled with H2O2 production.
The mechanism of gene transfer among ruminant mycoplasmas remains to be elucidated but some recently published data raise interesting possibilities. Indeed, an ICE has been described in M. agalactiae strain 5632 [16] and a decayed ICE is also predicted in strain PG2 (ICEAPG2). ICEs, also designated CONSTINS or conjugative transposons, are widespread amongst prokaryotes, and are viewed as modular scaffolds with diverse genetic organizations and encoded functions that are able to confer various metabolic or resistance traits to their host and to disseminate in bacterial populations [65].
A striking feature is that ICEAPG2 CDS products displayed a closer degree of relatedness with the M. capricolum subsp. capricolum ICEC than with the M. agalactiae ICEA5632-I, which appeared to be more related to M. fermentans ICEF. The finding of higher sequence similarity between ICEC and ICEAPG2 suggested that these elements are or have been functional for lateral gene transfer between the mycoides cluster and M. agalactiae ancestors. The finding in PG2 of a sequence that is known to be generated by excision of the ICEA5632-I in strain 5632 can be viewed as a trace of a past excision or as a mere potential integration site for an ICE. Although mycoplasma ICEs display a modular structure and some species-to-species variations, they constitute a very homogenous set of genetic elements that appear to be specific to this class of bacteria. In particular, certain conserved CDSs (CDS19 or CDS22) that are present in all sequenced mycoplasma ICEs (Figure 4) do not have any homologs outside of the mollicutes. Only CDS5 and CDS17 have homologs (TraG and TraE, respectively) in certain other ICEs or conjugative plasmids.
Insertion sequences (ISs) are another type of mobile element that may be involved in genome plasticity. The IS ISMag1 was identified in several strains of M. agalactiae [66] and has most probably been exchanged between strains of M. agalactiae and M. bovis [67]. Although a complete ISMag1 could not be found in the sequenced genome from strain PG2, sequence analysis revealed that a fragment of this IS is located between positions 391476 and 391626. Moreover, other ISs are shared by M. bovis and M. mycoides subsp. mycoides SC, suggesting again that HGT events occur between these ruminant mycoplasmas [67]. Finally, our analysis of the M. agalactiae genome also revealed two genes encoding a putative prophage protein (MAG6440) and a phage family integrase (MAG5690). All of these genetic elements can be regarded as vestiges of potential shuttles that may have been involved in the transfer of genome fragments between ancestors of M. agalactiae and of mycoplasmas of the mycoides cluster.
As mentioned earlier, it is not known whether mycoplasmas are naturally competent and whether they can uptake naked DNA in their host environment. Viruses or natural plasmids that could serve as vehicles for HGT have thus far been described only in a limited number of mycoplasma species that do not include M. agalactiae. The presence of conjugative elements in M. agalactiae and in a phylogenetically distant member of the mycoides cluster, together with evidence of large-scale gene transfer in between those species strongly suggests that these simple organisms are being “sexually competent,” most likely via a conjugation-like mechanism.
Overall, data obtained in this study shed new light on the phenomenon that may underline the plasticity and evolution of mycoplasma genomes. While members of the genus Mycoplasma infect a wide range of hosts, individual species are thought to have strict host specificity. However, examples that have recently emerged from the literature may begin to challenge this idea [68]. Two examples are the isolation of the human-infecting mycoplasma M. fermentans from small ruminants [69] and the discovery in birds of M. capricolum-like strains closely related to the ruminant pathogen M. capricolum subsp. capricolum [70]. Whether HGT plays a role in adaptation to new hosts or in virulence has yet to be discovered, but understanding the mechanisms underlying HGT in mycoplasmas and their role in reshaping their reduced genomes is the next, exciting challenge.
The M. agalactiae PG2 type strain was originally isolated from a goat in Spain (1952). In previous studies, the PG2 strain was shown to contain a locus designated as vpma that encodes a family of abundant related lipoproteins [30] and that undergoes frequent DNA rearrangements [29]. The entire vpma locus was previously sequenced for the 55–5 clonal variant from the PG2 strain [29]; this variant was selected in this study for genome sequencing. The 55–5 clone was propagated in SP4 medium [71] at 37 °C. Genomic DNA was isolated as previously described [72].
Three genomic libraries were constructed for sequencing purposes. Two were obtained by mechanical shearing of M. agalactiae total DNA and subsequent cloning of the resulting 3–4–kb and 8–10–kb inserts into plasmids pcDNA2.1 (Invitrogen, http://www.invitrogen.com) and pCNS (pSU18 derived), respectively. From these libraries, DNA inserts from approximately 5,700 and 1,500 clones, respectively, were sequenced from both ends. For the third library, DNA fragments of ∼20 kb were generated by partial Sau3A digestion and introduced into the miniBAC plasmid pBBc (pBeloBac11 derived). From this library, DNA inserts of approximately 1,100 clones were sequenced from both ends. Plasmid DNAs were purified and end-sequenced using dye-terminator chemistries on ABI3700 sequencers (Applied Biosystems, https://www2.appliedbiosystems.com). About 16,800 reads led to an average 12-fold coverage. The Phred/Phrap/Consed software package was used for sequence assembly and quality assessment [73–75]. About 380 additional reactions were necessary to complete the genomic sequence. The integrity of the assembly was confirmed by comparing the in silico restriction map with restricted DNA fragments (SmaI, XhoI, and EclXI) previously analysed by PFGE [76].
The genome annotation was performed using the CAAT-Box platform [77], which was customized to facilitate the annotation process. CDSs were first detected using the Genemark software [78], implemented in the CAAT-Box environment. Putative CDSs of more than 300 amino acids were used to train the Markov model (order 5). The three codons AUG, UUG, and GUG were used as potential start codons, whereas UAG and UAA were defined as stop codons. Once trained, the Markov model was applied to the complete genome using 80 bp as a cut-off value for the smallest CDSs. Prediction of CDSs with CAAT-Box also integrates results of BLAST searches [79] in order to discriminate highly probable CDSs from false ORFs. Databases used for this purpose were SwissProt (http://www.ebi.ac.uk/swissprot/index.html), trembl (http://www.ebi.ac.uk/embl/index.html), and MolliGen (http://cbi.labri.fr/outils/molligen), a database dedicated to the comparative genomics of mollicutes. In order to determine the extent of sequence similarity, alignments between predicted proteins and best BLAST-hit sequences were performed using the NEEDLE software [80] implementing the Needleman-Wunsch global alignment algorithm and using the BLOSUM62 matrix. During the annotation process, proteins were considered to be homologs when the similarity in these alignments exceeded 40%. Predicted proteins with lower or only local similarities with previously characterized proteins were annotated as hypothetical proteins. Start codons were most often chosen according to CAAT-Box recommendations that resulted from both Genemark coding state prediction and BLAST results analysis. For CDSs showing neither obvious homology relationships nor clear coding curves, the most upstream start was chosen, with a preference for the most frequently used AUG codon.
Other tools incorporated into CAAT-Box were also used to improve annotation and function predictions: among them, InterProScan [81] and PrositeScan [82] for domains detection and TMHMM for trans-membrane segments prediction [83]. In order to recover small CDS or gene fragments that could have been discarded during the CDS prediction process, intergenic sequences of more than 80 bp were systematically compared to reference databases using BLASTX. The annotation of each CDS was manually verified by at least two annotators.
The tRNAs were located on the chromosome using the tRNAscan software [84] and the rRNA genes were searched using BLASTN by homology with the rRNA genes from M. pulmonis [26]. Precise boundaries were established after comparisons with the sequences stored in the European Ribosomal RNA Database (http://www.psb.ugent.be/rRNA) [85] and the 5S Ribosomal RNA Database (http://www.man.poznan.pl/5SData) [86].
For phylogenetic analyses based on 16S rDNA sequences, aligned 16S rDNA sequences were recovered from the RDPII database (release 9.46; [87]). From this alignment, 649 sites were informative. Phylogenetic analyses were performed using MEGA3 [88]. The three methods implemented in the version 3.1 of this integrated software were used: Neighbor-joining, Minimal Evolution, and Maximum Parsimony. The reliability of the tree nodes was tested by performing 500 bootstrap replicates.
For phylogenetic analyses based on selected shared proteins for species with sequenced genomes, supertree constructions were obtained using 30 shared proteins (COGs; [89]). These were selected because they have been shown not to be horizontally transferred in a large dataset [21]. Protein sequences corresponding to the selected COG (Table S10) were retrieved from the 17 mollicute genomes available in the MolliGen database. These are M. agalactiae; Mycoplasma capricolum subsp. capricolum; Mycoplasma mycoides subsp. mycoides SC; Mesoplasma florum; Ureaplasma urealyticum/parvum; Mycoplasma penetrans; Mycoplasma gallisepticum; Mycoplasma pneumoniae; Mycoplasma genitalium; Mycoplasma mobile; Mycoplasma hyopneumoniae strains 232, 7448, and J; Mycoplasma pulmonis; Mycoplasma synoviae; Onion yellows phytoplasma; and Aster yellows witches-broom phytoplasma. Separate multiple sequence alignments of each COG were built for all 17 mollicute genomes using ClustalW. These individual alignments were manually concatenated, which resulted in a super matrix with 5,257 informative sites. Phylogenetic analyses were performed using MEGA3 [88]. The three methods implemented in the version 3.1 of this integrated software were used: Neighbor-joining, Minimal Evolution, and Maximum Parsimony. The substitution matrix used was JTT, the sites with gaps were ignored, and the reliability of the tree nodes was tested by performing 500 bootstraps replicates. We also derived Maximum Likelihood phylogenetic inferences using PhyML [90], applying the JTT matrix and other options as reported by others [21]. The sites with gaps were ignored and support for the hypothesis of relationships was assessed using 100 bootstrap replicates.
To identify M. agalactiae putative lipoproteins, two methods were used. In the first one, M. agalactiae CDSs were scanned for the presence of the PROSITE Prokaryotic membrane lipoprotein lipid attachment site motif (PROKAR_LIPOPROTEIN), the sequence of which is [DERK](6)-[LIVMFWSTAG](2)-[LIVMFYSTAGCQ]-[AGS]-C. Twenty-five CDSs were determined to encode lipoproteins using this method. Because some already characterized lipoproteins of M. agalactiae did not match this motif [30], a second approach was devised using MEME/MAST [91,92]. Specific motifs within the first 35 amino acids of a set of 14 characterized lipoproteins from M. agalactiae (six sequences) and other mycoplasmas (eight sequences) were analyzed using MEME. This first step resulted in the identification of two motifs, corresponding to the charged N terminus and to the lipobox. These motifs were then searched for in all of the M. agalactiae CDSs using MAST, resulting in a set of 42 proteins displaying one or both of the two motifs. Six proteins were excluded from this set because the motifs were located too far from the N terminus of the polypeptides. A second round of MEME/MAST motif search was performed using the N-terminal sequence of the 36 remaining proteins as a seed. A total of 81 CDSs were recovered. Of the recovered CDSs, those without (i) motifs located in the N-terminal region, (ii) cysteine within the lipobox motif, or (iii) charged amino acids in the N-terminus were eliminated. After a manual check, a total of 55 proteins were predicted to be lipoproteins using this method. These proteins display a region that is composed of an N-terminal sequence of 3–10 positively charged amino acids followed by a hydrophobic segment of 10–17 aa, from which K, D, R, E, and H are excluded. At the very end there is a lipobox of 4 aa, the consensus sequence of which is either (V/I)AAKC (Type KC) or (I/L)(A/S)ASC (Type SC). Finally, combining PS-SCAN and MEME/MAST methods, a total of 66 lipoproteins were predicted, among which 14 were detected by both methods. Interestingly, none of the CDS exhibiting a “KC” lipobox was detected by PS-SCAN.
Best Blast Hits (BBH) were identified for every predicted protein using a BLASTP threshold E-value of 10−8. Five databases were searched: UniProt [93], and four other databases consisting of proteomes predicted from sequenced genomes of mollicute species belonging to distinct phylogenetic groups; spiroplasma (M. capricolum subsp. capricolum, M. mycoides subsp. mycoides SC, Me. florum and S. citri), Pneumoniae (U. urealyticum/parvum, M. penetrans, M. gallisepticum, M. pneumoniae and M. genitalium), hominis (M. mobile, M. hyopneumoniae strain 232, M. pulmonis and M. synoviae), and phytoplasmas (Onion yellows phytoplasma and Aster yellows witches-broom phytoplasma).
E-values were automatically compared and data were filtered to identify putative horizontal transfers. CDSs displaying a BBH with a mollicute sequenced genome belonging to a phylogenetic group other than that of the query were further analysed as follows. Pairwise alignments [94] between each query protein and the best hits from UniProt and each of the four mollicute databases were calculated. From these alignments, the percentage of similarity of the query with its BBH obtained with mollicutes belonging to the same phylogenetic group was compared to that obtained with mollicutes belonging to a different phylogenetic group. A 5% difference in favour of an ortholog not belonging to the query phylogenetic group was considered as the minimal threshold for further investigations.
Protein phylogeny tree reconstructions were performed using the MEGA3 software [88]. Trees were obtained using the distance/neighbor-joining method and the gaps complete deletion option; bootstrap statistical analyses were performed with 500 replicates. Bootstrap values lower than 90% were not considered to be significant. When supported by significant bootstrap values, incongruence between protein and species phylogenies was understood as a potential HGT.
When very few homologs were identified or when branches were only supported by low bootstrap values, the possibility of an HGT was not recorded except when other independent results support it. These were a particularly high similarity value (>80%) and conservation of gene synteny.
The genome sequence from M. agalactiae PG2 strain, as well as related features, were submitted to the EMBL (http://www.ebi.ac.uk/embl), GenBank (http://www.ncbi.nih.gov/Genbank/index.html), and DDBJ databases (http://www.ddbj.nig.ac.jp) under accession number CU179680. All data are also available from the MolliGen database (http://cbi.labri.fr/outils/molligen).
The National Center for Biotechnology Information (NCBI) (http://www.ncbi.nlm.nih.gov) accession numbers of other genomes mentioned in this manuscript are: M. capricolum subsp. capricolum, NC_007633; M. gallisepticum, NC_004829; M. genitalium, NC_000908; M. hyopneumoniae 232, NC_006360; M. hyopneumoniae 7448, NC_007332; M. hyopneumoniae J, NC_007295; M. mobile, NC_006908; M. mycoides subsp. mycoides SC, NC_005364; M. penetrans, NC_004432; M. pneumoniae, NC _000912; M. pulmonis, NC_002771; M. synoviae, NC_007294; Me. florum, NC_006055; and U. urealyticum/parvum, NC_002162.
The NCBI locus tags of the genes and gene products mentioned in this manuscript are M. capricolum subsp. capricolum OppA, MCAP_0116; M. hyopneumoniae gidA, mhp003; M. mobile gidA, MMOB1540; M. mycoides subsp. mycoides SC OppA, MSC_0964; M. mycoides subsp. mycoides SC GlpF, MSC_0257; M. mycoides subsp. mycoides SC GlpK, MSC_0258; M. mycoides subsp. mycoides SC GlpO, MSC_0259; M. mycoides subsp. mycoides SC GtsABC transporter components, MSC_0516/MSC_0517/ MSC_0518; M. pulmonis gidA, MYPU_2530; and M. synoviae gidA, MS53_0515.
The Pfam database (http://www.sanger.ac.uk/Software/Pfam) accession numbers for the protein motifs/domains mentioned in this paper are phage integrase motif, PF00589; and DUF285 domain of unknown function, PF03382.
The PROSITE database (http://www.expasy.ch/prosite) accession number for the prokaryotic membrane lipoprotein lipid attachment site motif is PS51257. |
10.1371/journal.pntd.0004979 | Urinary Biomarkers KIM-1 and NGAL for Detection of Chronic Kidney Disease of Uncertain Etiology (CKDu) among Agricultural Communities in Sri Lanka | Chronic Kidney Disease of uncertain etiology (CKDu) is an emerging epidemic among farming communities in rural Sri Lanka. Victims do not exhibit common causative factors, however, histopathological studies revealed that CKDu is a tubulointerstitial disease. Urine albumin or albumin-creatinine ratio is still being used as a traditional diagnostic tool to identify CKDu, but accuracy and prevalence data generated are questionable. Urinary biomarkers have been used in similar nephropathy and are widely recognised for their sensitivity, specificity and accuracy in determining CKDu and early renal injury. However, these biomarkers have never been used in diagnosing CKDu in Sri Lanka. Male farmers (n = 1734) were recruited from 4 regions in Sri Lanka i.e. Matara and Nuwara Eliya (farming locations with no CKDu prevalence) and two CKDu emerging locations from Hambantota District in Southern Sri Lanka; Angunakolapelessa (EL1) and Bandagiriya (EL2). Albuminuria (ACR ≥ 30mg/g); serum creatinine based estimation of glomerular filtration rate (eGFR); creatinine normalized urinary kidney injury molecule (KIM-1) and neutrophil gelatinase-associated lipocalin (NGAL) were measured. Fourteen new CKDu cases (18%) from EL1 and nine CKDu cases (9%) from EL2 were recognized for the first time from EL1, EL2 locations, which were previously considered as non-endemic of the disease and associated with persistent albuminuria (ACR ≥ 30mg/g Cr). No CKDu cases were identified in non-endemic study locations in Matara (CM) and Nuwara Eliya (CN). Analysis of urinary biomarkers showed urinary KIM-1 and NGAL were significantly higher in new CKDu cases in EL1 and EL2. However, we also reported significantly higher KIM-1 and NGAL in apparently healthy farmers in EL 1 and EL 2 with comparison to both control groups. These observations may indicate possible early renal damage in absence of persistent albuminuria and potential capabilities of urinary KIM-1 and NGAL in early detection of renal injury among farming communities in Southern Sri Lanka.
| Chronic Kidney Disease (CKD) is a challenging global health issue around the world. Impairment of kidney function with time is eminent, but indications of CKD may not be seen until considerable damage to kidney functions. Two main causes of CKD are diabetes and high blood pressure. However recently new form of CKD has been reported among agricultural works in the tropics other than known factors. The causes for this mysterious CKD are unknown and termed as Chronic Kidney Disease of uncertain etiology (CKDu). Clinical diagnosis depends on urinary markers and conventional creatinine based markers may underestimate the prevalence of the disease. Therefore, development of new sensitive markers for the early detection may certainly improve the treatment and patient management around the world.
| Chronic Kidney Disease of unknown etiology (CKDu) is an endemic disease among dry zone farming communities in Sri Lanka. First cases were reported in early 1990s in North Central Province (NCP) predominantly among male farmers [1]. It has reached epidemic proportions with ever increasing numbers of patients and deaths, thus becoming a new and emerging health issue that would eventually inflict adverse consequences on food security, merely for the fact that affected populations constitute the major rice farming communities in Sri Lanka. It is a global epidemic as similar types of kidney diseases are reported from Andhra Pradesh in India [2] and in Central America including Nicaragua [3], El Salvador [4] and Costa Rica [5].
Hypertension, diabetes, glomerulonephritis and other traditional causes are not associated with CKDu. However, multiple causes have been suggested such as chronic low dose exposure to multiple heavy metals and agrochemicals [6, 7], heat stress and recurrent dehydration [8–11], heat driven pathophysiologic mechanisms [12], nephrotoxic drugs [13], hyperuricemia and hyperuricosuria [14–17], leptospirosis [18, 19] and genetic susceptibility [20, 21]. Based on clinical and pathological studies, CKDu cases in Sri Lanka show glomerular and tubulointerstitial injury in kidneys [1, 22, 23] and similar glomerular and tubulointerstitial injury have also been reported in Mesoamerican nephropathy [24, 25].
A recent study by the World Health Organization (WHO) has used albumin creatinine ratio (ACR) as the diagnostic criteria for CKDu in Sri Lanka [26]. However, ACR ≥ 30 mg/g Cr may not detect early renal injury [27–29] hence may underestimate disease prevalence. Several novel biomarkers such as human neutrophil gelatinase-associated lipocalin (NGAL), kidney injury molecule 1 (KIM-1), N-acetyl beta glucosaminidase (NAG), interleukin 18 (IL-18), insulin like growth factor-binding protein 7 (IGFBP7) and tissue inhibitor of metalloproteinases-2 (TIMP-2) are being used to diagnose acute kidney injury [29] and their use in chronic kidney disease was evident in Mesoamerican nephropathy [30, 31]. However, tubular markers such as KIM -1 and NGAL have not been used for CKDu diagnosis in Sri Lanka.
KIM-1 also known as T-cell immunoglobulin and mucin-containing molecule is a type 1 trans-membrane protein with molecular weight approximately 100 kDa. Proximal tubular cells in kidneys are the main source of KIM-1 in the urine and it is up-regulated during acute kidney injury [29]. NGAL, also known as siderocalin, lipocalin-2 (lnc2) or lipocalin 24p3 is a 22–25 kDa glycoprotein that belongs to superfamily “lipocalin” [32]. Leucocytes, loop of Henle and collecting ducts are some of the major sources of NGAL in the body [29]. NGAL is released from lysosomes, brush-border and cytoplasm of proximal tubular epithelial cells during chronic kidney disease [29]. Urinary KIM-1 and NGAL are good predictors of renal injury prior to detectable changes in eGFR [33–35]. Moreover, urinary KIM-1 and NGAL are potential biomarkers in predicting chronic kidney disease due to tubulointerstitial damage [36].
The first objective of this study was to determine the prevalence of CKDu using case definition by Jayathilaka et al, [26] mainly focussing on albuminuria (ACR ≥ 30mg/g) and eGFR in disease emerging locations in Hambantota district (EL1 & EL2) and non-endemic areas Matara and Nuwara Eliya (CM & CN) in Sri Lanka. The second objective was to determine the levels of tubular markers KIM-1 and NGAL in the same study populations to assess potential early renal injury among CKDu subjects and healthy farmers from the selected locations.
A cross-sectional study was conducted at three locations namely Angunakolapelessa (EL1), Bandagiriya (EL 2) and Matara (CM) in the Southern Province and one location, Nuwara Eliya (CN) in the Central Province, Sri Lanka (Fig 1). EL1 and EL2 in Hambantota district situated in the dry zone of Sri Lanka (annual mean temperature more than 27.2°C) with similar farming practices to CKDu endemic areas in NCP. New patients with CKDu are recently reported from EL1 and EL2 farming locations. We also recruited male farmers from non-endemic locations in the wet zone, Matara (CM, annual mean temperature more than 27°C) and Nuwara Eliya (CN, annual mean temperature less than 15°C).
Male farmers over 20 years (n = 1734) were recruited from all four selected farming locations. Participants were screened to exclude farmers with less than 10 years of farming and lower working hours (less than 600 hours per year). Therefore, based on above exclusion criteria 1295 participants were excluded and remaining 439 (EL1 = 106, EL2 = 127, CN = 104 and CM = 102) were screened for co-morbid diseases (Fig 2). An interviewer administered, pre-tested survey questionnaire was used to collect data from the farmers. During the interview, information about co-morbid diseases (i.e. diabetes, hypertension, arthritis, gastritis, renal calculi etc.) was obtained. 76 farmers had comorbidities and 363 farmers were selected for the study. However, 140 participants were absent during the urine collection and 223 urine samples EL1 (n = 56), EL 2 (n = 61), CN (n = 52) and CM (n = 54) were used for the biomedical analysis. Details of the study population were illustrated in (Fig 2).
Ethical approval was obtained from the ethics review committee of the Faculty of Medicine, Rajarata University, Sri Lanka. Farmers from all four locations were made aware about the study verbally and with information leaflets. A written consent (n = 1701) was obtained from each farmer. There were 33 farmers that were unable to write, in such cases thumb print consent was obtained following ethical guidelines. The study was conducted in accordance with Helsinki declaration.
Urine and blood samples were collected from individuals to measure creatinine, eGFR, urine ACR, KIM-1, NGAL and HbA1c. Fresh morning first void urine samples were collected in sterile containers from each farmer and stored temporarily at 4°C. In the lab, urine samples were centrifuged for 15 minutes at 4000 rpm, 4°C and samples were stored in aliquots at -80°C until analysis. Blood samples were collected in labelled serum separator tubes and allowed to rest and clot at room temperature for 30 minutes. Blood sample tubes were then centrifuged at 2500 rpm for 10 minutes at 4°C. Later, the supernatant (serum) was labelled and transferred into cryogenic vials and stored at -80°C until analysis.
Creatinine was measured in urine and serum by modified kinetic Jaffe reaction to minimize interference of non-creatinine and jaffe-positive compounds [37, 38] in Dimension clinical chemistry system (Siemens, New York, U.S.A.). Picrate reacts with creatinine to produce a red chromophore in the presence of a strong base (NaOH). Absorbance was measured at 510 nm (assay range: 0 mg/dl– 20 mg/dl). Urinary or serum creatinine levels were expressed in mg/dl.
eGFR was calculated by using both CKD-EPI (CKD Epidemiology Collaboration) and modified MDRD equation [39]. GFR was expressed in mL/min/1.73 m2 of body surface area. A medical doctor in the study group measured the blood pressure after fifteen minutes’ rest, using a mercury sphygmo-manometer. The average of two readings taken five minutes apart was used.
Microalbumin was measured in urine by particle-enhanced turbidimetric inhibition immunoassay (PETINIA) and Dimension clinical chemistry system (Siemens, Newark, U.S.A.). In the presence of human albumin bound particle reagent (PR), albumin present in the sample competes for monoclonal antibody (mAb) and reduces the rate of PR–mAb aggregation. Therefore, rate of aggregation was inversely proportional to albumin concentration in urine samples. Rate of aggregation was measured using bichromatic turbidimetric reading at 340 nm (assay range: 1.3 mg/L– 100 mg/L).
Ion-exchange high-performance liquid chromatography (HPLC) principle based separation of HbA1c on a cation exchange cartridge method was used for the measurement of HbA1c in human anti-coagulated whole blood samples. The absorbance of separated HbA1c was then measured at 415 nm using Bio-Rad D-10 as per manufacturer’s instructions.
Human KIM-1 was measured in early morning urine samples using ELISA (CUSABIO, P.R. China; Cat#: CSB-E08807h) according to the manufacturer’s instructions. KIM-1 ELISA kit employs quantitative sandwich enzyme immunoassay technique for high sensitivity and specificity for human KIM-1 detection. Minimum detectable dose of human KIM-1 was typically less than 0.043 ng/ml. Intra-assay precision was (CV%: <8%) while inter-assay precision was (CV%: <10%). Detection range of the kit was (0.312 ng/ml—20 ng/ml). Absorbance was measured at 450 nm using micro plate reader (Utrao microplate reader–SM600, Shanghai Yong Chuang, P.R. China).
Human NGAL/Lipocalin-2 was measured in early morning urine samples using ELISA (Ray Biotech, Inc. Norcross, GA; Cat: ELH-Lipocalin2-001) according to the manufacturer’s instructions. NGAL ELISA kit employs quantitative sandwich enzyme immunoassay technique for specific detection of human Lipocalin-2 or NGAL. Minimum detectable dose of human Lipocalin-2 was 4 pg/ml. NGAL ELISA kit intra-assay precision was (CV%: <10%) and inter-assay precision was (CV%: <12%). Detection range of the kit was (4 pg/ml—1000 pg/ml). Absorbance was measured at 450 nm using microplate reader (Utrao microplate reader—SM600, Shanghai Yong Chuang, P.R. China).
Data were analysed using IBM statistics (version 22.0). Continuous variables were reported as means (SEM) whereas categorical variables were reported as proportions. Renal biomarkers (KIM-1 & NGAL) were adjusted for urine creatinine concentrations prior to analysis. All comparisons between groups were performed by one-way ANOVA test with normally distributed parameters or transformed to natural log parameters. Kruskall–Wallis test and the Mann–Whitney U-test were performed to compare the significance between groups when deviated from the normality. Association between renal biomarkers with Albumin- creatinine ratio (ACR) and eGFR were performed using linear regression models. In all analysis, P < 0.05 was considered as significant.
Baseline characteristics and information of the study sample were given in Table 1. Overall, 439 male farmers (age ≥ 20 years) participated in the cross-sectional study representing four different farming locations of Sri Lanka. Age of individuals ranged between 26–49 years in CM, 20–83 years in CN, 27–70 years in EL1 and 36–79 years in EL2. Most participants were paddy Farmers. Farmers from CN and CM were involved in vegetable farming in addition to paddy. Lower education level hence low socio—economic status has been reported in emerging locations EL1 and EL2 than CM and CN. The most significant factor was the previous source of drinking water where EL1 and EL2 farmers mainly consumed well water that has been categorized as very hard (≥ 181 ppm). However, most of the farming locations now have access to tap water. Co-morbid diseases (i.e. diabetes, hypertension, arthritis, gastritis, renal calculi etc.) were reported in certain farmers from all four farming locations with CM (4%), CN (11%), EL1 (27%) and EL2 (25%). Diabetes and hypertension were not reported within the study sample from CM. However, cases of diabetes and hypertension were reported within certain farmers from CN (1% and 6%), EL1 (6% and 12%) and EL2 (10% and 8%) respectively. Other co-morbid diseases such as arthritis, gastritis, renal calculi etc., were also reported in CM (4%), CN (4%), EL1 (9%) and EL2 (8%) respectively. Healthy subjects (i.e. no diabetes, hypertension, other kidney disease etc.) from CM, CN, EL1 and EL2 were 96% (n = 102), 89% (n = 104), 81% (n = 63) and 90% (n = 86) respectively.
Albuminuria (ACR ≥ 30 mg/g Cr) in repeated occasions was not reported in both non-endemic locations (CM and CN) and therefore no CKDu cases were identified under WHO case definition. However, fourteen (18%) and nine (9%) individuals from emerging locations EL1 (Angunakolapelessa) and EL2 (Bandagiriya) were reported with repeated ACR ≥ 30 mg/g Cr. These cases were defined as CKDu patients and grouped into EL1-CKDu and EL2-CKDu in further biomarker analysis. The range reported was 31–124.4 mg/g Cr in EL1-CKDu and 35.2–82.2 mg/g Cr in EL2-CKDu. Healthy subjects (ACR ≤ 30 mg/g Cr) in EL 1 and EL 2 were considered as controls from the emerging locations and grouped into C-EL1 and C-EL2. ACR of farmers from CM, CN, C-EL1 and C-EL2 was not significant with each other (p>0.05) except isolated two CKDu groups. Mean eGFR in both CM and CN was 108 ml/mi/1.73m2 and slightly lower eGFR was recorded at emerging locations EL1 and EL2. The lowest eGFR was recorded in EL2-CKDu. No cases of eGFR ≤ 60 ml/min/1.73m2, were reported in non-endemic locations (CM & CN) but three individuals (4%) from EL1 and two individuals (2%) from EL2 had eGFR ≤ 60 ml/min/1.73m2 Table 2.
Summarised statistics of KIM-1 normalized to urinary creatinine along with respective ACR, eGFR and HbA1C are presented in Table 3. The lowest mean values of urinary KIM-1 were reported in CN (0.17 μg/g Cr) followed by non-endemic controls in CM (1.26 μg/g Cr). Higher urinary KIM-1 was reported in CKDu emerging locations EL-1 and EL–2 with compared to both CM and CN. The highest KIM-1 levels (64.27 μg/g Cr and 48.53 μg/g Cr) were reported by CKDu subjects in both EL-1 and EL-2. As seen in Fig 3, urine-KIM-1 concomitantly increased in CKDu cases identified in farming locations (EL1- CKDu and EL2-CKDu). KIM-1 values reported at EL1-CKDu and EL2-CKDu were 50 fold and 38 fold higher than the control farming location (CM). They also reported significantly higher concentrations in comparison with CN (P < 0.001). Higher KIM-1 was also reported in healthy farmers (ACR ≤ 30 mg/g Cr) at EL1 and EL2. The highest urinary KIM-1 (16.07 μg/g Cr) recorded in C-EL1 was twelve times higher than the control farming location (CM). However, a lower KIM-1 concentration (10.52 μg/g Cr) was recorded at farming location C-EL2. KIM-1 levels in C-EL1 and C-EL2 were also significantly higher than the KIM-1 levels in non-endemic farming controls in CN (P < 0.001).
Creatinine adjusted NGAL levels in non-endemic farming locations (CM, CN) and emerging farming locations (EL1 and EL2) were given in Table 3. The lowest mean NGAL (0.30 μg/g Cr) was reported in non- endemic farmers in Matara (CM) followed by non-endemic farmers in Nuwara Eliya (CN; 0.47 μg/g Cr). Both CKDu groups in emerging locations, EL1-CKDu (2.46 μg/g Cr) and EL2-CKDu (6.95 μg/g Cr) reported the highest NGAL. Pairwise comparisons revealed that NGAL levels at both CKDu groups were significantly higher than the control farming locations (CM & CN, P < 0.001; Fig 4). NGAL levels of C -EL1 and C- EL2 were approximately 2.5 fold and 4 fold higher than the control group CM and 1.6 fold and 2.4 fold higher than the CN. However, NGAL levels in C -EL1 and C- EL2 were not significant (P > 0.05) with comparison to both control groups CM and CN. CKDu cases in both emerging locations (EL1-CKDu and EL2-CKDu) were also reported significantly higher NGAL when compare to healthy subjects (C -EL1 & C- EL2) in the same location (P < 0.001).
Analysis of urinary biomarkers (KIM-1 and NGAL) following the isolation of CKDu cases (based on urinary albumin-to-creatinine ratio (ACR) ≥ 30 mg/g Cr) revealed a significant correlation between increased urinary renal biomarker levels and albuminuria. Elevated levels of urinary KIM-1 were positively correlated with increased ACR in EL1 and EL2 (rs = 0.57, P < 0.001; Fig 5A) however, no significant correlation was found between Urinary KIM-1 and eGFR (rs = -0.12, P = 0.30; Fig 5B). Similarly, elevated levels of urinary NGAL was positively correlated with increased ACR in CKDu emerging locations (EL1 and EL2) (rs = 0.49, P < 0.001; Fig 5C) and significant negative correlation was observed between urinary NGAL and eGFR in EL1 and EL2 (rs = -0.37, P < 0.001; Fig 5D).
The current study contributes to reveal the role of novel urinary biomarkers (KIM-1 and NGAL) in CKDu detection for the first time in Sri Lanka. This is also the first cross-sectional study exploring chronic kidney disease of uncertain etiology (CKDu) adopting WHO guidelines based CKDu case definition in Hambantota district, Southern Province, Sri Lanka. No studies have been previously reported using combination of urinary ACR, KIM-1 and NGAL together in determining early CKDu cases among Sri Lankan farming community. New CKDu cases (6%, 23/363) were identified during the study based on WHO study group definition. Albuminuric groups (EL 1-CKDu & EL 2 –CKDu) reporting the highest, non-endemic control subjects (CM & CN) showing the lowest and healthy subjects in emerging locations (C-EL1 &C-EL 2) showing intermediate values of KIM -1 and NGAL indicate occurrence of sub-clinical renal injury. A gradient with clear separation of KIM -1 and NGAL values in albuminuric, non-endemic controls and endemic controls was evident. Overall, higher levels of urinary KIM-1 may indicate the proximal tubular damage whereas higher levels of urinary NGAL might be likely due to detectable damage occurring in loop of Henle and distal convoluted tubule.
In Sri Lanka, men are more vulnerable to CKDu than women. A study conducted in NCP showed CKDu prevalence is higher in males (6%) than in females (2.9%) [1]. Similarly, in El Salvador, the prevalence of CKDu was reported higher in men (25.7%) than in women (11.8%) [4]. A meta-analysis based investigation using 68 studies concluded that males with non-diabetic renal disease showed significantly rapid kidney function deterioration over time than females [40]. Rapid progression of males from early stages of renal damage to chronic stages of kidney injury was most probably due constant exposure towards occupational or environmental factors [3–5, 26]. As a consequence, the current study was precisely focused on male farmers excluding females and children. In 2011, a single suspected case of CKDu (0.025%) was reported in Hambantota district despite it was previously considered as a non-endemic region [27]. In the same year, six (0.43%) CKDu cases were identified in Hambantota district using non-specific qualitative dipstick test followed by sulfosalicylic acid test [1]. Hambantota, located in the dry zone, shares similar socio-economic background and identical farming practices to the CKDu endemic NCP in Sri Lanka. Therefore, there is a looming possibility of emergence of CKDu in Hambantota district, Southern Province, Sri Lanka.
Here, we report distinct 23 CKDu cases from EL1 and EL2 in Hambantota district, Southern Province, Sri Lanka, based on both CKDu definition by WHO [26] and increased levels of KIM-1and NGAL. Co-morbid diseases (i.e. diabetes, hypertension, pyelonephritis, renal calculi etc.) may influence levels of urinary ACR, KIM-1 and NGAL [41–45]. Consequently, we utilized questionnaire and medical history of individuals based assessment to identify and eliminate cases with co-morbid diseases. HbA1c was also measured in individuals with ACR ≥ 30 mg/g in EL1 and EL2 to exclude diabetes cases. Therefore, reported 23 new cases can be confirmed as CKDu. All CKDu cases (EL1-CKDu and EL2-CKDu) had ACR ≥ 30 mg/g on repeated testing. Measurement of albumin levels is a well-known early non-invasive biomarker to detect CKD [46, 47]. An epidemiological study also supports the notion of testing albuminuria as an indicator of renal disease in general population [48]. Urinary albumin levels define the glomerular integrity and proximal tubule function in kidneys [49].
CKDu cases were also confirmed by significantly higher levels of urinary KIM-1 and NGAL. Both markers could easily isolate the suspected cases. Apparently healthy farmers at emerging locations (C -EL1 & C- EL2) who had ACR < 30 mg/g, healthy eGFR and normal serum creatinine also showed elevated levels of KIM-1 and NGAL compared to both non-endemic control groups (CM & CN) indicating possible early renal damage. It may suggest that tubular damage expressed by KIM-1 is present before albuminuria appeared in farmers from C -EL1 & C- EL2 Similar cases have been reported among sugarcane cutters in Nicaragua with elevated urinary NGAL, IL-18 and NAG [30]. Urinary KIM-1 detection using micro-urine nanoparticle detection technique has been recently reported in Sri Lanka [50] but comparison is not possible due to smaller sample size of the study. KIM-1 is markedly up regulated in kidneys due to ischemic insult [35, 51]. Up regulation of KIM-1 is a well-known consequence of proximal tubular damage in the nephron. Until more recently, detecting glomerular KIM-1 expression could also be a useful tool in identifying glomerular injury [52]. Increased levels of KIM-1 may also represent its involvement in phagocytosis of damaged proximal tubule epithelial cells by converting epithelial cells into semi-professional phagocytes [53, 54]. KIM-1 up regulation may also be responsible in restoring functional and morphological integrity of kidneys following ischemic insult [35]. Our study shows that KIM-1 may be used to detect early CKDu cases in susceptible farming communities in Sri Lanka other than the conventional markers.
However, NGAL elevation was only notable in EL1-CKDu and EL2-CKDu with 8 fold and 23-fold increase with compared to CM and 5 fold and 14-fold increase with CN. Similar results have been reported in El-Salvador where 26% higher NGAL was reported in CKD cases [55]. Laws et al., reported 1.49 times higher NGAL among sugarcane farmers in Nicaragua [30]. However, no studies have been reported using NGAL in Sri Lankan population. NGAL elevation suggest, re-epithelialisation of damaged tubules and reabsorption of iron that was leaked due to damage of proximal epithelial tubule cells and also to induce iron-dependent nephrogenesis [51, 56]. NGAL was not significantly increased in C—EL1 & C—EL2 when compared to CM and CN. This suggests elevation of KIM-1 may be more sensitive in detecting early tubular damage when compared with NGAL. Therefore, this study does not support the use of NGAL to detect early cases of CKDu in susceptible populations however, further studies are necessary.
There are some limitations in our study. We initially recruited 1734 farmers. We used precise inclusion criteria to limit the study population i.e., continuous farming (> 10 years) with long working hours (> 600 hours per year). These inclusion criteria at the beginning of the study lead to a smaller sample size (n = 439). Some farmers (n = 140, 38.6%) were not present at the time of urine collection therefore the study left with modest sample size for the biomarker analysis (n = 223). Other main limitation of the study was lack of established urinary biomarker levels that reflect sub clinical damage in Sri Lankan nephropathy. No previous studies have been done under local conditions and comparable occupational cohorts are even difficult to find in Mesoamerican nephropathy except a few recent studies [30, 31]. Short term individual variation within the subjects was also unknown and a follow up study is required. The current study was conducted only on native male farmers in selected farming locations in Sri Lanka ignoring children and females therefore might hinder generalized applicability of the findings in other geographical locations and general population.
In conclusion, this study reports 23 new CKDu cases for the first time in Hambantota district, Sri Lanka in spite of previously being considered as a non-endemic location. This is the first study to identify CKDu suspected cases and detection of early kidney damage in Sri Lankan farming communities using urinary biomarkers KIM-1 and NGAL. New cases were defined by WHO study group criteria for CKDu diagnosis in Sri Lanka along with urinary markers KIM-1 and NGAL. The results of our cross-sectional study shows that the tubular damage predicted by urinary KIM-1 and NGAL were significantly correlated with high urinary ACR levels. Strikingly, early tubular damage as seen by higher urinary KIM-1 and NGAL was also observed in healthy farmers despite normal ACR levels (< 30 mg/g). Urinary tubular markers reconfirm tubulointerstitial disease with repeated tubular injury in CKDu among farming communities in Sri Lanka. However, longitudinal cohort studies are needed to predict use of tubular markers for precise prognosis, optimized treatment and patient management.
|
10.1371/journal.pntd.0003595 | Modeling the Dynamics of Plasmodium vivax Infection and Hypnozoite Reactivation In Vivo | The dynamics of Plasmodium vivax infection is characterized by reactivation of hypnozoites at varying time intervals. The relative contribution of new P. vivax infection and reactivation of dormant liver stage hypnozoites to initiation of blood stage infection is unclear. In this study, we investigate the contribution of new inoculations of P. vivax sporozoites to primary infection versus reactivation of hypnozoites by modeling the dynamics of P. vivax infection in Thailand in patients receiving treatment for either blood stage infection alone (chloroquine), or the blood and liver stages of infection (chloroquine + primaquine). In addition, we also analysed rates of infection in a study in Papua New Guinea (PNG) where patients were treated with either artesunate, or artesunate + primaquine. Our results show that up to 96% of the P. vivax infection is due to hypnozoite reactivation in individuals living in endemic areas in Thailand. Similar analysis revealed the around 70% of infections in the PNG cohort were due to hypnozoite reactivation. We show how the age of the cohort, primaquine drug failure, and seasonality may affect estimates of the ratio of primary P. vivax infection to hypnozoite reactivation. Modeling of P. vivax primary infection and hypnozoite reactivation provides important insights into infection dynamics, and suggests that 90–96% of blood stage infections arise from hypnozoite reactivation. Major differences in infection kinetics between Thailand and PNG suggest the likelihood of drug failure in PNG.
| Plasmodium vivax is one of two major parasite species causing human disease. This parasite can lie dormant in the liver as a hypnozoite, before later reactivating to cause blood-stage infection. Treatment to eliminate the dormant hypnozoite stage relies mostly on a single drug—primaquine. Understanding the rate of primary infection versus hypnozoite reactivation is important to understanding primaquine efficacy and drug resistance, as well as the development of new drugs targeting hypnozoites. Here we use mathematical modeling to analyse data from two clinical cohorts and show that up to 96% of infections may be caused by hypnozoite reactivation. We also use modeling to understand the impact of drug resistance, seasonal infection and subject age.
| Plasmodium vivax is one of the major agents of malaria infection, with around 2.5 billion people living in areas at risk of infection, and more than 70 million estimated annual infections [1–3]. P. vivax is generally less pathogenic than Plasmodium falciparum infection due to the absence of sequestration and cytoadherence, and all blood-stage forms can be detected in peripheral circulation[4]. P. vivax also differs due to the development of dormant hypnozoite forms in the liver, which serve as reservoir of infection after the clearance or treatment of the acute blood stage of infection. P. vivax shows a preference for reticulocytes as its host cell in the blood-stage of its life-cycle [5].
The potential for reactivation of dormant hypnozoites creates a number of difficulties in understanding the transmission dynamics of P. vivax infection[6], however the majority of diagnosed infections are thought to be due to hypnozoite reactivation rather than new primary infection [7–9]. After treatment to eliminate blood stage infection, new infections may occur as a result of recrudescence (failure of treatment of the blood stage), hypnozoite reactivation, or new primary infection. Although comparison of P. vivax genotypes may be useful in distinguishing recrudescence of the blood stage parasites after treatment, it is not always useful in differentiating reactivation of a dormant hypnozoite from new primary infection [10–12]. This is because the parasites causing relapses are often genetically different from those observed in the most recent blood-stage infection, so it is not possible to differentiate reactivation from new primary infection using genotyping[13, 14]. Therefore, it is difficult to know the proportion of blood stage infections due to hypnozoite reactivation versus new primary infection by P. vivax.
In this study, we aim to apply mathematical modeling to quantify the relative contribution of new primary infection, and infection initiated due to the reactivation of dormant liver stage hypnozoites. We analyse data from two published prospective studies where individuals were treated to eliminate blood-stage infection, and a subset were treated with primaquine, a licensed radical treatment for hypnozoites, to eliminate preexisting hypnozoites [15, 16]. By comparing the rate of observed blood stage infection in the two groups, we can estimate the contribution of primary infection versus hypnozoite reactivation in P. vivax infection.
The field study data were from a published prospective study with recruitment from July 1995 to July 1996, on 342 individuals of different ages (68% (258) <15 years of age) living on the western border of Thailand where P. vivax is endemic[9]. Individuals with asexual forms of P. vivax on a blood smear were enrolled, treated with chloroquine (25mg base/kg over 3 days) and followed up until presentation of pure or mixed P. vivax blood stage infection. Individuals with reappearance of pure P. vivax infection were retreated with either chloroquine only (70 individuals with mean age (range) 12 (1–50)) or chloroquine and primaquine (0.25 mg/kg daily for 14 days, 43 individuals with mean age (range) 13 (5–43)), and were followed up by microscopy until detection of P. vivax blood stage parasites. Each dose of chloroquine was supervised and the patient observed for 1 hour after dosing. The criteria for enrollment and method of detection and quantification of P. vivax parasites in the blood smears are detailed elsewhere[9]. From the fact that primaquine can kill liver stage hypnozoites[15, 16] we classified the data in two groups: individuals retreated with chloroquine+primaquine (CQ+PQ) group and individuals retreated with chloroquine only (CQ only) group.
We also analysed published data on a treatment-time-to-infection study in PNG [8]. In this study, the contribution of relapse to the risk P. vivax infection and disease was studied in 433 PNG children 1–5 years of age. Children were randomized into one of three groups: (1) artesunate (4mg/kg/d for 7 days) plus primaquine (0.5 mg/kg/d for 14 days), 149 individuals, (2) artesunate only (4 mg/kg/d for 7 days), 150 individuals or (3) no treatment (control), 150 individuals, and were followed up for infection and the presence of febrile illness. Every dose of treatment was administered as direct observed therapy. The criteria for enrollment, method of randomization, treatment procedures and method for the detection and quantification of P. vivax parasites in the blood smears were detailed in the original publication [8]. For our analysis we extracted data on infection rates from Fig. 2 of Betuela et al [8], using Grafula 3.0 (Knowledge Probe Inc, Aurora)to extract data on cumulative proportion of infections detected by light microscopy. The data was classified into two groups; those receiving artesunate+primaquine (AS+PQ group), and artesunate only (AS only) group).
The dynamics of P. vivax infection are characterized by both primary infection and reactivation. For individuals living in P. vivax endemic regions both primary infection and activation of hypnozoites occur throughout the year. For modelling purposes we will label the two groups of subjects as i) the B+H group, comprising individuals who received drugs against both blood stage parasites and hypnozoites and ii) the B group, comprising individuals who received drugs against blood stage parasites only. We initially assume 100% drug efficacy against both the blood and liver stage parasites. Thus for individuals receiving in the B+H group, all observed infection is due to new primary infection. For individuals in the B group, infections can arise either from new primary infection, or from reactivation of hypnozoites. Thus, we can model the time to first P. vivax infection after treatment, and fit this to the ‘survival curve’ of time to detection of P. vivax infection after treatment. The proportion of treated individuals in the B+H group remaining uninfected at a given time t is indicated as (SB+H(t)) and follows the equation:
SB+H(t)={1,t≤d1e−k(t−d1),t>d1,
(1)
and the proportion of treated individuals in the B group remaining uninfected at time t is indicated as (SB(t)) and follows the equation
SB(t)={1,t≤d2e−k(1+c)(t−d2),t>d2
(2)
Here k is the rate of initiation of new primary infections, c is the relapse to reinfection ratio and d1 and d2 are delays to detection of blood stage P. vivax parasites in the groups respectively. Fig. 1A and 1B show the schematic representation of Equation (1) and (2) where individual received treatments for either both blood stage and liver stage parasites or blood stage parasites only.
The modeling above assumes that all primaquine is completely effective in all individuals. However, if primaquine fails to kill hypnozoites from a proportion of strains, or in a proportion of individuals, then we expect altered dynamics. If primaquine kills only a fraction of hypnozoites, then the equation for SB+H(t) becomes:
sB+H(t)={1,t≤d1e−k(1+Rc)(t−d2),t>d1
(3)
Where R is the fraction of hypnozoites that are resistant to primaquine. We note that in our study we cannot differentiate between this and Equation (2) with a different c.
If primaquine only kills hypnozoites in a subset of individuals, then primaquine resistant individuals will behave as if they have not received primaquine. Thus, if primaquine is only effective in a proportion of individuals, the rate of infection of individuals in the B+H group will be;
sB+H(t)={1,t≤d1Pe−k(t−d1)+(1−P)e−k(1+c)(t−d1),t>d1,
(4)
where P is the proportion of individuals in whom primaquine is effective. The dynamics is illustrated in Fig. 1C. Equations (1), (2) and (4) were fit to the data using the lsqnonlin function in MATLAB R2012 (Release M (2012) The MathWorks Inc, Natick, MA, USA), which uses a nonlinear least-squares method, to understand the contribution of hypnozoites reactivation to the dynamics of P. vivax infection.
To understand the different contributions of primary infection and reactivation, we need to understand how the force of infection from primary infection and hypnozoite reactivation evolve over time / with exposure starting with a naïve host. If we consider a fixed rate of infectious mosquito inoculation (M), and that each inoculation infects a number of liver cells (S) that will eventually go on to produce a blood stage infection, then the total rate of successful infection of liver cells is simply MS. Assuming that the time spent in the liver stage is negligible for primary infection, the rate of new primary infections (Ip) is simply the overall rate of infection of liver cells MS multiplied by the fraction of liver cell infections that result in primary infection (f):
IP=fMS
(5)
We note that fMS is equivalent to k in Equations 1–4. Importantly, the rate of successful infection of liver cells may not be the same as the rate of successful mosquito inoculation (as one successful inoculation may infect one or more liver cells, see Fig. 1). Moreover, the definition of successful infection of a liver cell is not simply that a cell becomes infected, but that the infected cell gives rise to a blood stage infection at some stage (either immediately or with some delay).
If we consider an individual that starts with no hypnozoites (for example a neonate, or someone successfully treated with primaquine), the number of hypnozoites and rate of infection from hypnozoite reactivation (IH) evolves over time as hypnozoites accumulate upon repeated infection. Hypnozoites accumulate dependent on the rate of infection of liver cells (MS), and the fraction of liver cells that become hypnozoites (1-f). The dynamical equation for the number of hypnozoites (H) over time is
dHdt=(1−f)MS−aH
(6)
where a is the reactivation rate of hypnozoites. We note that this assumes that hypnozoites never infect more than a tiny proportion of total liver cells. The rate of infection due to hypnozoite reactivation is then simply:
IH=aH
(7)
The solution of Equations (6) and (7) is
IH=(1−f)MS(1−e−at)+aH0e−at
(8)
where H0 is the initial number of hypnozoites. At steady state, the rate of infection from hypnozoites (IH) in Equations 7 and 8 is equivalent to kc in Equations 2 and 4. Similarly, the total rate of new blood stage infections at steady state is simply MS (the rate of successful infection of liver cells). Fig. 2 illustrates the mechanisms of infection and the relationship between parameters in Equations 1–4, and 5–7.
Plasmodium vivax infection often follows a seasonal pattern, with higher infection in the wet season[17, 18]. To understand the role of fluctuations in the force of infection and the effect of seasonality, we modified Equation (6) by allowing the rate of infectious mosquito inoculation to vary seasonally;
M=Ms(1+Mfcos(2πt/365))
(9)
where Ms is the mean rate of infectious mosquito inoculation and Mf is a parameter for the degree of seasonal fluctuation. The periodicities of these functions are such they divide the season into dry and wet seasons. We examine the effect of seasonality and EIR on the contribution of primary infection to reactivation by numerical simulations of Equations (5) and (6) with seasonality terms included.
We first estimated using Equation (1) the rate of infection in individuals receiving chloroquine plus primaquine treatment in Thailand, in whom all infections are assumed to be due to new primary infection. Assuming a constant force of infection, we found that a rate of infection of approximately 0.0017 per day (equating to an average time to primary infection of 588 days) provided the best fit to the data (Fig. 3A). We next estimated the rate of infection and reactivation occurring in the individuals that received chloroquine alone using Equation (2), where infection can arise due to both new primary infection as well as hypnozoite reactivation. In this case, we observed the rate of 0.043 infections per day (equivalent to an average of 23 days to infection). Since the individuals receiving chloroquine alone experience both primary infection and reactivation from hypnozoites, we can subtract the rate of primary infection (estimated in the CQ + PQ group) to estimate the rate of hypnozoite reactivation. Thus, we estimate a rate of hypnozoite reactivation of 0.0413 per day (equivalent to a hypnozoite reactivation every 24 days). There was no evidence for a difference in the delays from treatment to first detection in the CQ+PQ group and the CQ group (p = 0.8563, F-test).
The relative rates of new primary infection and hypnozoite reactivation estimated in this data suggest that approximately 4% of infection events in individuals receiving chloroquine occurred due to primary infection, and approximately 96% of infection events occurred due to hypnozoite reactivation. This implies a ratio of primary infection to hypnozoite reactivation of approximately 1 to 24.
Once hypnozoites are laid down, they will reactivate at some later time either spontaneously, or following some form of stimulation (such as fever or concurrent infection). So, for a single bite we might imagine there is some average time to reactivation, and a distribution in the probability of reactivation with time. The delay between initial inoculation and subsequent hypnozoite reactivation is highly variable[19, 20]. However, even in the absence of knowing the precise schedule of reactivation of individual hypnozoites, we can still understand the dynamics of reactivation in an endemic setting. That is, in an endemic setting we do not observe reactivation from a single inoculation, but in fact from a long series of past inoculations. We have a probability of reactivation from an inoculation 6 months ago, which is dependent on the rate of inoculation six months ago, the proportion of hypnozoites surviving 6 months, and the rate of reactivation of 6 month old hypnozoites. The same is true for hypnozoites inoculated a month ago, or a year ago. In this circumstance if we have a constant rate of inoculation for individuals of a particular age group who have been exposed to Pv infections for a sufficient time to reach ‘steady state’ of infection, where each the average rate of reactivation of hypnozoites reflects the rate at which they were laid down (from Equation (6), when the number of hypnozoites is constant
(dHdt=0)
, then the rate of hypnozoite reactivation (aH) equals the rate of new hypnozoite infection ((1-f)MS). Since treatment for blood-stage infection is relatively short-lived compared to the history of infection (and has no impact on accumulated liver stages), this should not significantly affect this steady state.
Although we can estimate the ratio between primary infections and reactivations, we cannot directly estimate the ‘rate of infection’ (rate of new infectious inoculation from mosquito bites) from this data. That is, for an inoculation to be infectious, it must eventually produce either primary infection, or a hypnozoite that later reactivates. It is not clear that all inoculations must produce a primary infection (some may produce only hypnozoites). Thus, the rate of new primary infection in CQ+PQ individuals may or may not reflect infectious inoculation rate, since some fraction of inoculations may produce only hypnozoites. However, the minimal rate of infectious inoculation would occur when every new inoculation produced a primary infection. If we assume that every new inoculation must result in an early, acute blood stage infection, then the minimal inoculation rate is simply the rate of new blood stage infection in the CQ + PQ group (and both inoculation and new primary infection would be experienced approximately every 588 days). If this were the case, it would also require that each new inoculation must lay down approximately 24 hypnozoites (in order to account for the observed high rate of hypnozoite reactivation in the CQ group)(Fig. 3B).
The alternative scenario occurs if some infectious inoculations do not produce a primary infection, and instead only lay down hypnozoites. Since to be an ‘infectious inoculation’ the inoculation must produce either one primary infection or one hypnozoite, the highest rate of infectious inoculation would be when each infection only produced exactly one infected liver cell, which could either produce one primary infection or one hypnozoite. In this case, the rate of infection observed in the CQ group is exactly the rate of infectious inoculation. If this were the case, then it would require that only 4% of infectious inocula presented as a primary infection, and the rest became dormant (laid down their one hypnozoite) without being observed as a primary infection (Fig. 3C). This seems unlikely, as there is experimental evidence that more than one reactivation event (and thus more than one hypnozoite) can arise from a single inoculation[21].
The analysis above describes the maximal and minimal rates of infectious inoculation, which imply very different numbers of liver cells infected per inoculation. The maximal rate suggests that 24 liver cells are infected from each infectious bite, but that this only occurs every 588 days. The minimal rate suggests that each infectious bite infects at most one liver cell, but this happens as frequently as every 23 days. The rate of infection could also be considered in terms of the rate of infection of liver cells (either as primary infection or hynozoites), even without knowing how many liver cells are infected per mosquito inoculation. The infection rate in the CQ treated group is driven by both the rate of primary infection plus the rate of hypnozoite reactivation. As discussed above, the rate of reactivation reflects the sum of all previous inoculations (at all times) and their probability of reactivating, and is thus reflective of the rate of ‘laying down’ hypnozoites. The infection rate of the CQ group is thus the total rate of infection of liver cells (either destined for primary infection or hypnozoites), assuming that only one cell initiates each primary infection or reactivation (assuming chloroquine is effective). Thus, the minimum rate of liver cell infection can be derived from the CQ group, and is 0.043 infected cells per day (a new infected liver cell every 23 days). However, unless we know the number of liver cells infected on each inoculation, or the proportion of inoculations that cause primary infection, we cannot directly estimate the inoculation rate.
From the analysis above we can estimate the minimal and maximal inoculation rates, and the minimum rate of production of new infected liver cells (which is the same as the maximal inoculation rate). Fig. 3D illustrates one of many possible scenarios in between these extremes. For example, if we had an infection rate of 0.0086 per day (equating to a new infection being established every 120 days), this would require that approximately 20% of infection were observed as a primary infection, and each infection produced on average five hypnozoites (Fig. 3D). However, it is clear that although we can estimate primary infection and reactivation rates, we cannot directly estimate the rate of infectious inoculation from the data unless we assume that each infection event always produced an observed primary infection. In addition, we cannot estimate the rate of infection of liver cells unless we assume that each infection arises from a single infected liver cell.
In order to compare these dynamics in another population, we investigated the infection rates of patients treated with either artesunate alone, or artesunate + primaquine by analyzing a published data set from Papua New Guinea. In order to make our analysis directly comparable with the Thai study, we first restricted our analysis to infections detected by microscopy in the first 60 days since treatment (Fig. 4A). In individuals receiving artesunate plus primaquine, we observed using Equation (1) a rate of blood stage infection of 0.0102 / day (equivalent to 98 days between new blood stage infections). In individuals receiving artesunate alone, we observed using Equation (2) a rate of blood stage infection of 0.0344 / day (equivalent to a new infection every 29 days). Using the same approach as discussed above, we would estimate that in individuals treated with artesunate alone, 30% of infections occur due to primary infection and 70% due to hypnozoites reactivation (a ratio of 2.37 to 1). This ratio of reactivation to primary infection is 10 fold lower than the ratio observed in the Thai study. There was no evidence for a significant difference in the estimated delays to first detection of infection in AS+PQ group and AS only groups (p = 0.0807, F-test). The large difference in the ratio of primary infection to hypnozoite reactivation between the Thai and PNG studies raises a number of questions, which we explore below.
One reason for the difference between the Thai and PNG studies could arise from primaquine resistance. That is, either particular strains of parasites or particular individuals may be resistant to primaquine. If primaquine were only effective in a subset of parasite strains, then only a fraction of hypnozoites (from sensitive strains) would be killed by treatment. Hypnozoite reactivation would then be reduced by this fraction, and we would see simply an increase in the (exponential) rate of infection in the primaquine treated group. Alternatively, if primaquine was only effective in a proportion of individuals, then we might expect to see a rapid rate of infection in primaquine-resistant individuals (at the same rate as those receiving artesunate alone, due to both primary infection and hypnozoite reactivation), and a slow rate of primary infection in those in whom primaquine was effective. This would produce a different (non-exponential) infection curve, characterized by two populations and two infection rates.
In order to assess whether primaquine resistance might either affect a proportion of strains, or a proportion of individuals, we modeled the full time course of infection in the PNG data (comparing Equations (1), (2) and (4)). We found that a model in which a proportion of individuals are primaquine-resistant (Equation 4) provides a significantly better fit to the data (p<0.0001, F-test). In the case of primaquine resistant individuals, the infection rate in those in whom primaquine was ineffective should be the same as the infection rate in those receiving artesunate alone, and we can estimate the rate for those with successful primaquine therapy independently. When we fit the data to this function (Equation 4), we can estimate that the best fit to the data occurs if primaquine is effective in ≈60% percent of individuals, and the rate of primary infection (in the group in which primaquine was effective) was 0.0032 / day (equivalent to a new primary infection every 313 days). Interestingly, this gives a ratio of primary infection to reactivation of 1 to 9, which is much more similar to the rate estimated from the Thai study. Table 1 shows the best-fit estimates of the models (1), (2) and (4) to the both Thai and PNG data sets.
A second difference between the Thai and PNG studies is the age of the cohorts, which was young in PNG, but included all ages in Thailand. Therefore we asked whether age may affect the proportion of infections arising from hypnozoite reactivation. Primary infection arises soon after infectious inoculation, and thus should be proportional to the current rate of infectious inoculation. By contrast, reactivation from hypnozoites requires first the establishment of a ‘reservoir’ of hypnozoites from previous infections, and then their later reactivation. Thus, for example, after the first exposure in life, only hynozoites laid down by the first inoculation can reactivate. However, after many exposures, reactivation can occur from hynozoites laid down at different times in the past. This is evident from previous studies of P. vivax clonotypes in infection. These studies have analysed the relationship between P. vivax clonotypes from baseline infection, and in subsequent infectious episodes. In adults, these clonotypes are very often unrelated, consistent with reactivation of hypnozoites laid down prior to the most recent infection event [13, 14, 22]. However, in children it is more likely that baseline infection and subsequent reactivation will be due to the same clonotype [23].
We modeled the rates of primary infection and reactivation with age, to investigate how this might impact our analysis (using Equations (5), (6) and (7). The results suggest that for a given infection rate, the rate of primary infection is constant over time, but the rate of reactivation takes some time to reach its long-term level. Therefore at young ages we would expect an overall lower rate of infection (due to a lower rate of hypnozoite reactivation), as well as a higher ratio of primary infection to reactivation. How long this effect would be observed is directly related to the average time between laying down of hypnozoites and their subsequent reactivation (as shown in Fig. 5A). However, since the average time for hypnozoites to reactivate is thought to be of the order of months in tropical regions, this effect should only be present very early after initial exposure, and seems unlikely to have been the cause of the observed differences between the PNG and Thai cohorts.
The analysis above suggest that if age plays an important role in the ratio of primary infection to reactivation, then younger children treated with CQ in the Thai cohort should also exhibit an overall lower infection rate than older individuals (because of the reduced rate of hypnozoite reactivation). To test this we analysed the rate of infection of the children aged <5 years in the CQ treated cohort from the Thai study. Our fitting using Equation (2) showed no evidence for a significant difference in the rate of infection between those aged < 5 years or >5 years old in the Thai cohort (p = 0.2493, F-test, Fig. 5B). Thus, it seems unlikely that age alone would explain the difference between the Thai and PNG studies.
The model above considers how the infection rate from new inoculation and reactivation of P. vivax hypnozoites evolves with age in children exposed to infection. The same dynamics of accumulation of hypnozoites will also occur after successful primaquine therapy, when again the individual starts with no hypnozoite reservoir. After successful primaquine therapy both the number of hypnozoites and the rate of infection from hypnozoites will similarly increase over time.
Other mechanisms for the differences between the Thai and PNG studies include the overall rate of inoculation, as well as seasonal fluctuations in this. Changes in the inoculation rate per se should not affect the ratio of primary infection to reactivation. That is, the average ratio of primary infection to reactivation is determined by the proportion of sporozoites progressing immediately to infection versus becoming hypnozoites, and is relatively independent of the rate of inoculation. Moreover, comparing the groups with treatment for blood stage infection only (CQ in Thailand and AS in PNG), the rate of infection in these groups was similar (0.043 vs. 0.034 infections per day, respectively). Once we accounted for treatment resistance, our estimated rate of new primary infection was also similar (0.0017 vs. 0.0032 per day, respectively). Thus, it seems unlikely that differences in the rate of infection were a major factor.
Seasonal fluctuations in the rate of new infections will have a direct effect on the rate of primary infection over time. However, if the reactivation of hypnozoites happens on a longer timescale, it may be less susceptible to seasonal variation, and thus alter the ratio of primary infection to hypnozoite reactivation over the seasons. Since the PNG study may have a higher degree of seasonality compared with the Thai study [8, 9], we explored the predicted impact of seasonality. We modeled a sinusoidal variation in the rate of overall infection (Equation (9)), and a constant fraction of infections becoming primary infection (Equation (5) or being laid down as hypnozoites (Equation (6)) and later reactivating (Equation (7)). The rate of infection from primary infection and hypnozoite reactivation thus evolve over time as described in Equations (5) and (8) respectively. When the average time to hypnozoite reactivation was short (one month, Fig. 6A), the number of hypnozoites and rate of hypnozoite reactivation fluctuates a lot over time, mirroring (with slight delay) the fluctuations in primary infection. However, as the average time to hypnozoite reactivation gets longer (Fig. 6B and 6C)), the number of hypnozoites and rate of hypnozoite reactivation becomes less variable with season. For our purposes, we are most interested in how seasonal fluctuation in infection rate might affect the ratio of primary infection to hypnozoite reactivation. Somewhat counter-intuitively, the shorter the time to hypnozoite reactivation, the less seasonal fluctuation in this ratio is seen (Fig. 6D). Thus, seasonality of infection may significantly alter the ratio, but this is least likely to have an effect with tropical strains of P. vivax, where the time to hypnozoite reactivation is thought to be relatively short. In our modeling, the rate of hypnozoite reactivation was assumed to be independent of season. However some have suggested that there may be a seasonality in presentation from UK residents returning from endemic regions[24], supporting a varying reactivation rate with season, which could further affect the ratio of primary infections to reactivation.
In the study we have estimated the relative contribution of new primary infection and the reactivation of dormant liver-stage hypnozoites to the rate of blood stage infection in individuals living in the P. vivax endemic areas. Our modeling results showed that the vast majority of infections (96%) in Thailand were due to hypnozoite reactivation. The proportion of infection due to hypnozoite reactivation in PNG was less clear. Considering only the early phase of infection and 100% efficacy of primaquine, we would estimate that only 70% of infections in PNG were due to hypnozoite reactivation. However, we found that when the full time course of infection was considered, we had a significantly better fit to a model in which up to 40% of PNG individuals were resistant to primaquine therapy. If this level of primaquine resistance in the population is correct, then 90% of the infections were due to hypnozoite reactivation in the PNG cohort study.
A number of factors differed between the Thai and PNG studies that might affect our results. Firstly, the age of the cohorts was different, with the PNG cohort focused on children 1–5 years of age, whereas the Thai cohort included individuals of all ages. Age may play a role in susceptibility to P. vivax infection [25, 26], and in the ratio of primary infection to reactivation. However, if we restricted our analysis to the 1–5 years age group in the Thai study, we found a similar overall infection rate to the rest of the cohort. Secondly, there may have been differences in the infection rate or seasonality of infection between the studies. However, again we found that these were unlikely to have a profound effect on the ratio of primary infection to reactivation. Since the drug dosage differed between PNG and Thailand, this may have contributed to treatment failure. However, since the dose of primaquine given was higher in the PNG study (0.5 mg/kg in PNG vs 0.25 mg/kg in Thailand), this would not appear to support less effective treatment in PNG. Finally, it has been suggested that P. falciparum infection may precipitate P. vivax hypnozoite reactivation [27, 28]. In the Thai cohort, patients who had mixed infection at enrolment were excluded. However, in the PNG cohort these patients were included. Thus, if recent infection with P. falciparum precipitates P. vivax reactivation, this may have driven earlier reactivation in some patients the PNG study. As we do not have data on which patients were co-infected at enrolment in this cohort, we are unable to exclude this as a factor in our study.
In our analysis of primary infection and reactivation, we first assumed a constant ‘force of infection’. That is, we assumed that the rate of infectious inoculation, new primary infection, and reactivation of hypnozoites was relatively constant over the period of study. While the assumption of stable infection rate is easy to understand, the assumption of a constant rate of reactivation is less clear. That is, one might expect that following a single infection event, hypnozoites may be more likely to activate early than late, and therefore we might see a decrease in hypnozoite reactivation rate over time [19]. However, we show that since individuals are presumed exposed to the same rate of inoculation before and after therapy, they will be in a ‘steady state’ in which, regardless of the distribution of times for hypnozoites to reactivate, the observed rate of infection from hypnozoite reactivation is effectively constant with time. Exploring the impact of seasonal variations in infection rate, we find that these can significantly affect the ratio of primary infection to reactivation over time.
Our studies indicated that the infection dynamics of the AS + PQ group in PNG could be best fit by assuming that primaquine was ineffective in a proportion of individuals in the PNG study. This not only provided a significantly better fit to the shape of the AS+PQ reinfection curves, but also led to a ratio of primary infection too reactivation that was much more similar to the Thai study. Here, it is important to differentiate parasite resistance to primaquine from ineffectiveness of primaquine in a given host. The former would result in the clearance of only a proportion of [susceptible] hypnozoites, whereas the latter would lead to clearance of all hypnozoites in a proportion of individuals. The possibility of true primaquine resistance is questioned by some, who suggest that variation in host genetics may lead to ineffectiveness of primaquine in a subset of individuals [29]. Primaquine is not itself active, and requires metabolism by cytochrome P450 isoenzymes to its active form [30]. Human populations vary in the proportion of functional and non-functional CYP2D6 alleles, with PNG having a high proportion of novel and uncharacterized alleles [31]. Although these novel alleles have not been biochemically characterized, many are assumed to be active [32]. In addition, there appears evidence that parasites acquired in PNG are less susceptible to primaquine. Thus, it seems highly likely that treatment failure is indeed higher in PNG.
Understanding the relative contribution of primary infection and relapse in P. vivax infection is important to driving future treatment strategies. We have developed a novel analytical framework that allows estimation of the primary infection to reactivation ratio. This work indicates that reactivation of hypnozoites contributes 90–96% of observed P. vivax infections. In addition, our work suggests approaches for understanding the level of primaquine resistance in different populations. Understanding the mechanism of hypnozoite reactivation and identifying optimal approaches for targeting the hypnozoite reservoir will greatly reduce the burden of P. vivax infection.
|
10.1371/journal.pbio.0060123 | Notch-Deficient Skin Induces a Lethal Systemic B-Lymphoproliferative Disorder by Secreting TSLP, a Sentinel for Epidermal Integrity | Epidermal keratinocytes form a highly organized stratified epithelium and sustain a competent barrier function together with dermal and hematopoietic cells. The Notch signaling pathway is a critical regulator of epidermal integrity. Here, we show that keratinocyte-specific deletion of total Notch signaling triggered a severe systemic B-lymphoproliferative disorder, causing death. RBP-j is the DNA binding partner of Notch, but both RBP-j–dependent and independent Notch signaling were necessary for proper epidermal differentiation and lipid deposition. Loss of both pathways caused a persistent defect in skin differentiation/barrier formation. In response, high levels of thymic stromal lymphopoietin (TSLP) were released into systemic circulation by Notch-deficient keratinocytes that failed to differentiate, starting in utero. Exposure to high TSLP levels during neonatal hematopoiesis resulted in drastic expansion of peripheral pre- and immature B-lymphocytes, causing B-lymphoproliferative disorder associated with major organ infiltration and subsequent death, a previously unappreciated systemic effect of TSLP. These observations demonstrate that local skin perturbations can drive a lethal systemic disease and have important implications for a wide range of humoral and autoimmune diseases with skin manifestations.
| Skin is the largest organ of the body, forming an elaborate barrier that prevents water loss and protects the internal environment from outside invaders. When this barrier is compromised, keratinocytes, keratin-producing epidermal cells, alert and recruit the immune cells to the site of the breach as part of an adaptive defense mechanism. However, chronic activation of such an “alarm” could have undesired consequences. Using genetic engineering to progressively remove components of Notch signaling from mouse skin in utero resulted in chronic skin-barrier defects, mimicking a form of human skin disease called atopic dermatitis. Surprisingly, we discovered that a persistent alarm signal in newborns triggered a systemic B-lymphoproliferative disorder, which precisely mirrored the degree of skin defect and was lethal in its extreme form. This alarm signal, in the form of a cytokine called thymic stromal lymphopoietin, was produced by Notch-deficient keratinocytes that failed to form a competent skin barrier. Therefore, we uncovered a long-range proliferative effect on fetal pre-B cells in vivo that is induced by injured skin and mediated by thymic stromal lymphopoietin. These findings highlight the central role that skin-derived factors can play in initiating systemic diseases with skin involvement.
| The vertebrate skin is an organ in which keratinocytes, underlying mesenchymal cells, and circulating hematopoietic cells engage in reciprocal communication as they monitor organ integrity [1]. Therefore, skin is an ideal system in which to study how complex, multicompartmental networks function. Epidermal keratinocytes are organized in several distinct layers with the innermost (basal) layer containing the stem cells and transiently amplifying cells [2]. The next layer contains spinous cells that, under normal conditions, begin a terminal differentiation program, giving rise to granular and cornified layers [3,4]. However, after injury, spinous cells proliferate to contribute to the restoration of an intact integument and produce cytokines that trigger an inflammatory response [1]. Defects in execution of this terminal differentiation program can lead to what are collectively known as skin-barrier defects [5].
Notch activation contributes to spinous cell differentiation. Loss of canonical Notch signaling, induced by deletion of the RBPSUH gene (coding for RBP-j, the DNA binding partner of Notch), causes severe epidermal barrier and differentiation defects highlighted by reduced spinous, granular, and cornified layer cells [6]. However, overexpression of activated Notch1 exclusively in basal cells under the K14 promoter triggers their premature differentiation into spinous cells, concomitant with loss of proliferative capacity [6]. The pattern of Notch1 activation in epidermal keratinocytes is consistent with its proposed role in suppressing basal cell proliferation and promoting spinous cell differentiation via cell autonomous modulation of targets [6–11]. However, overexpression of activated Notch1 in the differentiated spinous cells (where it is normally present) triggers basal cell hyperproliferation and formation of acanthotic epidermis with a thickened granular layer [12], indicating that the Notch signaling pathway has a complex role by not only promoting differentiation and exit from the basal layer but also by contributing in a non-cell autonomous fashion to skin homeostasis [12,13]. How Notch performs its functions within spinous cells is a matter of some controversy, and multiple autonomous mechanisms have been proposed, including both canonical and noncanonical pathways [6–11].
To study the role of Notch signaling in skin homeostasis and barrier formation, we used the Msx2-Cre line to delete components of the Notch signaling pathway in skin keratinocytes [14]. A burst of Msx2-Cre expression on embryonic day 9.5 (E9.5) creates chimeric skin with dorsal and ventral patches (clones) lacking floxed alleles, confining the consequences of gene loss only to a fraction of the surface area (for a detailed analysis of Msx2-Cre expression in skin, see [14,15]). This allows animals lacking total Notch signaling to survive through birth (Figure 1A). With this system, we have shown previously that Notch loss also involves non-cell autonomous alteration in transforming growth factor ß and insulin-like growth factor signaling [15]. Removal of both Notch1 and Notch2 proteins or both presenilin-1 (PS1) and presenilin-2 (PS2) proteins (the catalytic subunits of γ-secretase) within epidermal clones is sufficient to cause early postnatal lethality [14]. In the current study, we investigated the mechanistic basis for this early demise.
The progressive loss of Notch alleles in skin keratinocytes generated a dose-dependent increase in thymic stromal lymphopoietin (TSLP) expression by suprabasal keratinocytes in direct response to defective skin differentiation/barrier formation in utero. TSLP is a recently discovered, epidermally derived cytokine implicated in the pathogenesis of atopic dermatitis and asthma [16]. Like interleukin 7 (IL-7), TSLP can support B cell development; however, the in vivo effects of high TSLP levels on B-lymphopoiesis are not fully understood [17]. A perinatal increase in TSLP levels produced a dose-dependent expansion of pre- and immature B cells in the periphery, causing a B-lymphoproliferative disorder (B-LPD), a previously unappreciated effect of TSLP. In its extreme form, B-LPD complications, including B cell infiltration into vital organs, culminated in death. Therefore, this study provides the first physiological confirmation that skin perturbation can cause a lethal systemic disease.
Removal of total Notch signaling from the skin using Msx2-Cre led to death at weaning (Figure 1B; see [14]). To identify the cause of death in Msx2-Cre/+; N1flox/flox; N2flox/flox (N1N2CKO) and Msx2-Cre/+; PS1flox/flox; PS2–/– (PSDCKO) mice, a comprehensive necropsy was performed on the moribund animals. Surprisingly, we found that mice from both genotypes had extremely high white blood cell (WBC) counts (>150,000 cells/μl) at around the time of death (Figure 1C). Importantly, analysis of an extensive Notch allelic series revealed that the severity of the leukocytosis during the first few weeks of life was correlated inversely with Notch dosage (Figure 1C).
To characterize the cells causing leukocytosis, we analyzed various hematopoietic parameters in Notch-deficient mice. Peripheral blood analysis showed that the leukocytosis in both N1N2CKO and PSDCKO animals was of a lymphoblastic/lymphocytic nature; hence the mice displayed a severe case of lymphoproliferative disorder (LPD; Figure 2A). This LPD was accompanied by the failure of bone marrow (BM) to participate in normal hematopoiesis, as demonstrated by reduced red blood cells and platelets (normocytic anemia and thrombocytopenia; Figure 2A). Because the blood phenotypes of N1N2CKO and PSDCKO mice were similar, we use the term “mutant” to describe the common LPD features of both genotypes and “wild type” for all genetic combinations that either lack Cre or carry Msx2-Cre and a wild-type allele of PS1. Macroscopic examination showed that mutant mice were smaller than their wild-type littermates and had enlarged spleen and lymph nodes but smaller than normal thymus, suggesting that the expanding lymphocytes are of B cell origin (Figure 2B and Table S1). Flow cytometry (FC) analysis on blood from mutant animals confirmed that the expanding population contributing to LPD was of the B cell lineage (B-LPD; Figure 2C). These observations suggested that extreme B-LPD could be the cause of death in PSDCKO and N1N2CKO animals.
Unexpectedly, Msx2-Cre/+; RBP-jflox/flox (RBP-jCKO) mice lived significantly longer (95 days on average) and showed lower WBC counts (∼70,000 cells/μl) compared to mice lacking Notch receptors or γ-secretase (Figure 1B and 1C). Although we could not exclude the possibility that Cre-mediated deletion of RBPSUH is marginally less efficient than that of PS1 or Notch1 and Notch2, examination of RBP-j protein in RBP-jCKO animals confirmed its loss in keratinocytes (Figure S1). If RBP-j could actively repress some target genes in the absence of Notch signaling and this repression would be lost upon RBP-j deletion, then the milder RBP-j phenotype could be explained by de-repression of some targets that ameliorated the phenotype [18]. Alternatively, the Notch pathway may have bifurcated; RBP-j-independent (yet γ-secretase-dependent) targets of Notch in skin [10,19] may contribute to the severity of the phenotypes reported here. To distinguish between these possibilities, mice that lack both γ-secretase and RBP-j in the skin were generated (Msx2-Cre/+; PS1flox/flox; PS2–/–; RBP-jflox/flox; PSDRBP-jCKO). In this genetic experiment, de-repressed targets should suppress the PSDCKO phenotype (i.e., PSDRBP-jCKO would display the RBP-jCKO phenotype [18]). However, if RBP-j-independent targets of Notch contribute to the life span and leukocytosis, then the γ-secretase mutation should be epistatic to RBP-j (i.e., the animal will have the PSDCKO phenotype). A combination of both would be expected to produce an intermediate phenotype.
The PSDRBP-jCKO mice lost RBP-j yet had the same life expectancy as PSDCKO mice (Figure 1B) and had WBC counts comparable to those of PSDCKO (Figure 1C), a result inconsistent with target de-repression. To confirm that these outcomes were Notch-dependent but RBP-j-independent, we also created the triple mutant Msx2-Cre/+; N1flox/flox; N2flox/flox; RBP-jflox/flox (N1N2RBP-jCKO). As with PSDRBP-jCKO animals, the life span of N1N2RBP-jCKO animals was not prolonged by removal of RBP-j (Figure 1B and 1C), indicating that a Notch-dependent activity contributes to the phenotype. This suggested that canonical Notch signaling could not be the sole determinant of the phenotypes described and instead supports the alternative assertion that RBP-j-independent effects of Notch in the skin may contribute to leukocytosis [10,20]. This result provides the first genetic evidence in a vertebrate to support the existence of a bifurcation in Notch signaling downstream of γ-secretase.
To understand B-LPD progression better, the WBC count from mutant animals was measured over their life span (Figure 2D). The WBC counts increased exponentially over the first 2 weeks starting at around P4 but plateaued during the third week of life within the same time window when most mutant animals expired. However, to our surprise, the longest surviving mice showed a trend toward normalization of their WBC counts (Figure 2D). A reduction in peripheral B cell number (and thus normalization of WBC count) was more evident in animals with partial loss of Notch signaling and in the longer living RBP-jCKO mice (unpublished data). Accompanying the surge in peripheral WBC count, B220+ B cells expanded the spleen (splenomegaly) and infiltrated several vital organs including lung and liver (Figure 2E, 2F, and Figure S2A). To characterize the identity of the expanding cells in B-LPD further, two additional markers, CD43 and IgM, were applied to segregate pro-, pre-, and immature B cell populations by FC (Figure 3A). The FC analysis showed that pre-B cells (B220+CD43–IgM–) constituted the majority of the expanding population in both BM and periphery (Figure 3B, red ovals). Immature B cells also expanded, as would be expected if differentiation of pre- to immature B cells persisted in the mutant animals (Figure 3B). The ability of pre-B cells to differentiate and the high WBC counts reached at an early age are consistent with a polyclonal origin of this B-LPD [21]. Large expansion of immature B cells was associated with elevated IgM levels, which precipitated at low temperature, a condition known as cryoglobulinemia (Figure S2B). Overall, despite the short duration of B-LPD in Notch-deficient mice, we hypothesized that extremely high WBC count, leukostasis, cryoglobulinemia, anemia, and infiltration of B cells into vital organs conspired with the skin phenotype to produce cachexic mice that failed to thrive during the early stages of postnatal development.
To determine whether B-LPD was an important contributor to early death, we performed an allogeneic bone marrow transplantation (BMT) experiment with mutant animals as the recipients. Lethally irradiated mutant mice transplanted with BM derived from their wild-type littermates around postnatal day 10 (P10) lived significantly longer than their untransplanted counterparts (Figure 4A). However, the transplanted mutants still died within a few weeks after transplantation, this time because of severe skin phenotypes including exfoliation, bleeding, inflammation, and infection (Figure S2C). When treated with systemic antibiotics, the life span of transplanted N1N2CKO animals was extended further and became comparable to that of RBP-jCKO mice (Figures 4A and 1B). However, antibiotic treatment of PSDCKO mice did not extend further their life span due to the greater severity of skin disease in these animals (Figure S2C). The WBC counts and FC analyses showed no reoccurrence of B-LPD in the transplanted mutants (Figure 4B and 4C). However, a significant expansion of granulocytes/monocytes was observed in the peripheral blood of transplanted mutant animals during the final days of life with WBC counts reaching ∼20,000 cells/μl (Figure 4B, 4C, and Figure S3). Similar granulocytosis accompanied by elevated peripheral T-lymphocyte percentage was observed in RBP-jCKO mice of the same age (Figure S3).
The observations detailed above suggested that if B-LPD was controlled, then life expectancy of the mutant animals could be increased significantly. Indeed, either a sublethal dose of total body irradiation (∼450 cGy) or focal irradiation of the mutant animals extended their life span. In both cases, life expansion correlated with a delay in B-LPD surge (Figure S4). Taken together, these experiments strongly demonstrated that B-LPD, acquired by animals with Notch- or γ-secretase-deficient skin, was a critical mediator of early lethality. However, the short life expectancy of transplanted mutants and RBP-jCKO was related to their skin disease, infection, and granulocyte/monocyte expansion, which again occurred in a Notch dose-dependent manner and in all adult animals lacking canonical Notch signaling (unpublished data and Dr. Freddy Radtke, personal communication).
Considering the importance of Notch signaling at various stages of lymphopoiesis [22,23] and the potential for ectopic expression of Msx2-Cre in a hematopoietic organ, we asked whether deletion of Notch in BM or any hematopoietic organ might have caused B-LPD. To map the sites of Msx2-Cre activity thoroughly, we applied the following approaches.
First, using a PCR protocol designed to detect the deleted allele of PS1 (PS1Δ), we confirmed that neither hematopoietic cells nor any hematopoiesis-related organ experienced Cre-mediated deletion of PS1 in PSDCKO mice (Figure 4D). Of note, we collected peripheral blood from the mutant animals at peak WBC counts, which thus was composed mostly (>90%) of expanding B cells, but still found no evidence of PS1 deletion. To rule out the possibility that loss of Notch signaling in a small subset of BM stromal cells could drive B cell expansion in the mutant animals, we re-analyzed Prx1-Cre; PS1flox/flox; PS2–/– mice in which Cre is active in BM stroma, the osteoblasts [24,25], and the osteoclasts [26]. The WBC analysis in all these mice showed no sign of B cell expansion (unpublished data), indicating that another organ provided the trigger for B-LPD. Next, we analyzed two reporter lines, Msx2-Cre; ZEG/+ and Msx2-Cre; Rosa26R/+. Only the skin was marked by these reporters (unpublished data) with no detectable reporter staining in any hematopoietic lineage or organ. Together, these findings led us to hypothesize that B-LPD was driven non-autonomously in wild-type B cells as a consequence of reduced Notch signaling in the skin.
To test this hypothesis, we applied the allogeneic BMT paradigm to ask whether hematopoietic stem cells isolated from the mutant animals could propagate B-LPD in normal recipients. The BM derived from mutant or wild-type littermates was equally competent in its ability to engraft in lethally irradiated wild-type littermate hosts and reconstitute a complete hematopoietic system that sustained a normal WBC count in the recipient animals over several months of follow-up (Figure 4E). The PCR analysis confirmed the complete repopulation of the recipients' hematopoietic system by donor-derived BM (Figure S5B). In addition, BM transplanted from mutant or wild-type animals into sublethally irradiated nonobese diabetic/severe combined immunodeficiency (NOD/SCID) mice were indistinguishable in their ability to rescue the recipients (Figure S6). Collectively, these findings confirmed the non-autonomous nature of B-LPD and implicated the skin as the primary organ responsible for the disease in Notch/γ-secretase-deficient animals.
Having identified skin keratinocytes as the only cells in which Cre-mediated deletion of Notch signaling was occurring, we hypothesized the existence of (an) epidermally derived cytokine(s) capable of driving B cell expansion that accumulated to high systemic levels in inverse correlation with Notch dose. To search for such factor(s), we performed microarray analysis of mutant and wild-type total-skin RNA samples collected at P9. Given the dose–response observed with life expectancy and B-LPD, we performed a modified trend analysis asking for transcripts that were modestly elevated in Notch1-deficient (N1CKO) skin but substantially elevated in N1N2CKO or PSDCKO skin. A small subset of altered transcripts, highly enriched for chemokines and chemoattractants, displayed the desired trend (Figures 5A and S7 and Table S2). Among them, TSLP was the second most abundant transcript and the only epidermally derived cytokine capable of driving fetal B cell proliferation in mouse [17,27].
Quantitative reverse transcription PCR (qRT-PCR) on epidermal mRNA samples confirmed a ∼20-fold increase of TSLP mRNA in mutants (Figure 5A and Table S2), and immunohistochemical analysis on skin sections identified suprabasal keratinocytes as the source of TSLP (Figure 5B and Figure S8). ELISA measurements detected TSLP levels >5000-fold above the normal levels in sera from mutant mice (∼50 ng/ml versus <10 pg/ml) that were already detectable at birth (unpublished data). A comprehensive serum analysis failed to detect differences in any other cytokine or autoimmune signature that could provide an alternative mechanism for B-LPD in the mutant mice (Table S3), including IL-7, the main cytokine implicated in B cell development. Interleukin 6 (IL-6), the only other cytokine implicated in B-lymphopoiesis, which showed moderate yet significant up-regulation in the skin microarray trend analysis, was elevated <2-fold in serum (Figure 5C). Examination of the sera from entire allelic series of Notch-deficient mice by ELISA confirmed and extended our trend analysis: The TSLP levels showed a strong inverse correlation with the dose of Notch signaling and life expectancy (Figure 5D) and a direct correlation with WBC counts of the mutant animals (Figures 1B, 1C, 5D, and Figure S9).
Collectively, RNA and protein analyses created a consistent picture pointing to TSLP as the likely candidate satisfying most of the criteria for the B-LPD-inducing agent: sensitivity to reduction in Notch dose, systemic availability, and the ability to promote B cell development [17,27]. However, such a proliferative role for TSLP has not been demonstrated previously. To satisfy Koch's postulate for disease causation, we injected recombinant mouse TSLP into wild-type mice. Daily injection of the animals with TSLP starting at birth and continuing for 7 days led to a dose-dependent elevation of WBC count (Figure 6A). The FC analysis on peripheral blood from the mice injected with TSLP identified the expanding cell population as B220+ B-lymphocytes, not seen in mice receiving carrier alone (Figure 6B). Further analysis of peripheral blood from wild-type mice receiving 1 μg of TSLP identified the expanding B cell population as pre- and immature B cells (Figure 6C). Injecting wild-type animals with 1 μg of TSLP daily resulted in steady-state serum TSLP levels of ∼250 pg/ml, comparable to that in N1CKO animals (Figures 6A, 5D, and Figure S9). Likewise, WBC counts in these animals were also indistinguishable from those seen in N1CKO mice at P8 (Figures 6A, 1C, and Figure S9). Thus, elevated TSLP levels were sufficient to cause mild B-LPD in an otherwise normal newborn mouse. Significantly and in agreement with published reports, no surge in WBC count was detected when the treatment regimen was initiated at P14 or later (Figure 6A; [28,29]). To demonstrate a correlation between endogenous levels of TSLP and high WBC counts, we analyzed K14-TSLPtg transgenic mice [30]. This analysis revealed high serum TSLP levels of ∼3 ng/ml in K14-TSLPtg newborns and, as predicted by our hypothesis, neonatal B-LPD similar to that observed in RBP-j-deficient animals (Figure 6D–G and S9). Importantly, B-LPD in K14-TSLPtg newborns developed in the absence of any overt skin morphology (unpublished data), indicating that elevated TSLP is not impeding epidermal differentiation. Consistent with the findings above, K14-TSLPtg animals also developed B-LPD only during the neonatal period, which disappeared later in life despite elevated serum TSLP levels (Figure 6D and 6E). This experiment confirmed that TSLP overexpression after the neonatal period did not sustain B-LPD in mice.
To ask if B-LPD represented the confluence of high TSLP with a responding, fetal pre-B cell population, we performed fetal liver transplantation (FLT) into lethally irradiated mutant animals. Surprisingly, FLT reconstituted normal adult hematopoiesis in the recipients, suggesting that fetal pre-B cells in an adult niche microenvironment lost their ability to respond to high levels of TSLP (Figure S10). To ask if continuous exposure to high TSLP levels sustained a responding pre-B cell population, we performed a third allogeneic BMT experiment in which BM from the mutant animals was transplanted into their lethally irradiated mutant littermates (which continued to produce high TSLP levels, unpublished data). Again, the donor-derived BM reconstituted normal adult hematopoiesis in the recipients, curing their B-LPD, consistent with the limited temporal window in which B-LPD can develop (Figure 4A). These results confirmed that B-LPD developed as a result of exposure to high TSLP levels in the perinatal period.
Loss of Notch signaling in skin keratinocytes leads to elevated epidermal TSLP production and high serum TSLP levels, reaching a maximum only when all Notch proteins or γ-secretase are removed (Figure 5). The TSLP overexpression thus may be a general response of epidermal keratinocytes to differentiation (and skin-barrier) defect [5]. Notch-deficient epidermis was defective in epidermal differentiation, leading to incomplete formation of upper spinous and granular layers, as reported in RBP-j-deficient animals (Figure 7A; [6]). This was accompanied by global down-regulation of skin lipid biosynthetic enzymes and reduced epidermal lipid content (Figure 7B and 7C and Table S4). Consequently, a defect in skin-barrier function was detectable directly by using the dye exclusion assay; this procedure stains areas with defective barrier, and indeed, only the stereotypical pattern of Cre-expression was stained in the mutant animals, consistent with a barrier defect in γ-secretase-deficient keratinocytes (Figure 7D; [6]).
To further ascertain if TSLP overexpression is a consequence of Notch loss or a consequence of defective differentiation/barrier formation, we isolated keratinocytes from mutant and wild-type pups and measured TSLP in the culture medium of cells grown on plastic. Both mutant and wild-type keratinocytes fail to form a fully differentiated epidermis under these conditions, and both secreted similar and significant levels of TSLP, as would be expected if TSLP overproduction was a general consequence of abnormal keratinocyte differentiation and not a specific consequence for loss of Notch signaling (Figure 7E). Because NFκB signaling, a potent inducer of differentiation, can activate TSLP [31,32], we asked if this pathway was responsible for TSLP expression in cultured keratinocytes. Addition of an inhibitor of NFκB signaling, BAY 11–7082 [33], abrogated TSLP production in a dose-dependent manner in both wild-type and γ-secretase-deficient keratinocytes (Figure 7E). To solidify the conclusion that TSLP overproduction is a readout of failed differentiation and not a specific repressed target of Notch signaling, we analyzed wrfr–/– mice that lack fatty acid transport protein 4 (FATP4) and die at birth because of severe epidermal differentiation and skin-barrier defects [34]. As in Notch-deficient mice, B-LPD was not yet detectable in wrfr–/– mice at birth. These mice, however, did show a substantial surge in skin TSLP transcript levels around the time of stratification and barrier formation (E15.5–17.5), which led to elevated serum TSLP levels at birth (Figure 7F). Because FATP4 expression is not altered in Notch-deficient skin (unpublished data) and Notch pathway targets are not altered in wrfr–/– skin (Figure 7F), this observation provided an independent confirmation that up-regulation of TSLP expression is a common readout of differentiation/barrier formation defects and not a repressed target of Notch signaling.
This report details the mechanism by which a local perturbation to the skin induced a lethal systemic disease. We demonstrate that progressive loss of Notch signaling in the embryonic ectoderm caused a dose-dependent impairment of epidermal differentiation and reduced lipid biogenesis, stimulating keratinocytes to secrete excess TSLP into systemic circulation. The pathophysiological consequence of persistent differentiation/barrier defects and the subsequent accumulation of TSLP had a potent proliferative effect on fetal/newborn B-lymphopoiesis, causing an exceptional expansion of pre- and immature B cell populations in newborn animals (i.e., neonatal B-LPD; Figure 8). A severe B-LPD with its systemic complications including infiltration of B cells into various vital organs, anemia, and cryoglobulinemia explained the early death of the animals lacking total (canonical and noncanonical) Notch signaling in the skin. Indeed, we were able to extend the life span of the mutant newborn animals simply by suppressing their B-LPD through either BMT or sublethal or focal irradiation administered before the onset of full-blown B-LPD in the second week of life. The ameliorating effects of these treatments also reflect the fact that pre-B cells emerging in the adult BM niche are refractory to high TSLP levels (see below).
A growing body of work has provided extensive evidence that TSLP overexpression in skin or lung epithelia can cause local allergic inflammation and subsequent development of atopic dermatitis and asthma, respectively [16,35,36]. However, the systemic consequences of TSLP overexpression were not clear [29,37]. In this report, we describe the mice in which defects in skin differentiation drive endogenous TSLP expression, beginning before birth. We find that TSLP levels are correlated directly with the degree of disruption in differentiation. This enabled us to recognize a novel consequence for TSLP elevation unique to fetal/newborn hematopoiesis and to ascribe TSLP production to failure in skin differentiation rather than its cause, as high levels of TSLP did not alter skin differentiation in newborn K14-TSLPtg mice.
It has been recognized that TSLP has a differential effect on fetal versus adult B-lymphopoiesis [27,38], despite the presence of fully functional TSLP receptors on both fetal and adult B cells [27,39]. In addition, in vitro results show that a liver-derived fetal pre-B cell line, NAG8/7, but not a BM-derived adult pre-B cell line, IxN/2B, proliferate in response to TSLP [40]. Consistent with these findings, we demonstrate that endogenously overexpressed or exogenously supplied TSLP, delivered during the perinatal period, lead to B-LPD and the appearance of a substantial number of pre-B cells (B220+CD43–IgM–) in the periphery in a dose-dependent manner (Figure S9). Transition from fetal/newborn to adult hematopoiesis occurs during the second week of life in the mouse BM [41]. We find that persistently high levels of TSLP do not produce B-LPD after this transition, explaining why B-LPD has not been detected in previous studies [29,37]. The narrow window in which TSLP can drive pre-B cell expansion explains why B-LPD in Notch-deficient and K14-TSLPtg animals is (1) confined to the first few weeks of life, (2) disappears in animals that experience a less severe disease, and (3) does not recur in the mutant mice after BMT, because adult-type pre-B cells that are reconstituted from the donor BM do not proliferate in response to high TSLP levels [38]. In addition, the fact that B-LPD does not recur in the mutant mice after FLT is consistent with the hypothesis that the niche is instrumental in regulating the response of pre-B cells to TSLP. This is reminiscent of the critical role the hematopoietic stem cell niche plays in defining the fetal versus adult characteristics of hematopoietic stem cells [42].
N1N2CKO or PSDCKO mutant animals transplanted with either wild-type or mutant BM do not enjoy a normal life span, dying before 3 months of age. Mice lacking some Notch alleles or lacking RBP-j do not develop lethal B-LPD but instead die prematurely from compromised skin integrity, exfoliation, inflammation, and systemic infection (unpublished data). Although antibiotic treatment delays the death of the mutant animals by controlling their infection, we find persistent granulocytosis even in antibiotic-treated mice, suggesting that systemic infection and the resulting inflammatory response are severe and incurable. Infection is the most likely cause of granulocytosis in the mutant animals; however, it is intriguing to speculate that persistently high levels of TSLP also contribute to this blood disease ([37] and Dr. Freddy Radtke, personal communication).
Notably, we observe a tight correlation between systemic TSLP levels, WBC counts, and degree of Notch loss in the neonatal skin. All Notch paralogs contribute to skin homeostasis, and stepwise reduction of Notch dosage leads to progressive skin perturbation resembling atopic dermatitis (unpublished data). However, we find no evidence arguing for a specific molecular connection between Notch loss and TSLP. Instead, the tight reverse correlation between serum TSLP levels and Notch dosage in the skin highlights the fact that TSLP levels reflect, with great precision, the magnitude of the differentiation defects. Thus, TSLP may be a direct readout of defects in differentiation, a systemic signal that many insults, including Notch loss, activate. To confirm this, we demonstrate that wild-type and mutant keratinocytes release significant, but similar, levels of TSLP when placed in culture in the absence of any exogenous stimulus (e.g., tumor necrosis factor α). In addition, wrfr–/– embryos show a substantial increase in epidermal TSLP transcripts in utero around the time of barrier formation. Therefore, TSLP is a keratinocyte “quality control” response to defective differentiation/barrier formation and not a secondary product of functional barrier failure after birth. One model for how the TSLP “sensor” works may be through the compensatory activation of NFκB. However, precisely how Notch loss in vivo leads to NFκB activation remains an important unanswered question that falls beyond the scope of this paper.
A striking finding is that RBP-jCKO mice show lower TSLP levels, a sublethal B-LPD, and improved epidermal morphology when compared to mice lacking total Notch signaling (Figure S11). To differentiate between de-repression of canonical targets and loss of noncanonical target expression, we performed a genetic analysis demonstrating unequivocally that RBP-j-independent Notch activity is of significance in skin homeostasis. This points to the importance of RBP-j-independent effects of Notch, including non-cell autonomous effects on ligand (Delta and Jagged) presenting cells, and only when all these arms of Notch signaling are lost, extreme TSLP levels are reached, causing lethal B-LPD (Figure 8). Identification of the relevant noncanonical targets also falls beyond the scope of this paper.
In this report, we identified canonical and noncanonical Notch signaling as essential regulators of epidermal differentiation/barrier formation, TSLP as a faithful reporter of keratinocyte differentiation/barrier defects, and B-LPD as the pathological consequence of chronic, perinatal TSLP elevation. Our analysis was done in a physiologically relevant setting: chronic skin-barrier formation defect caused by localized reduction in Notch signaling in a temporal and dose-dependent manner. How keratinocytes directly sense the degree of differentiation/barrier defect and translate such a stimulus into TSLP output remains an important, unsolved question. Nonetheless, the demonstration that defective skin differentiation can drive a lethal systemic B-LPD in mice through TSLP overexpression and the observation that human B cells also respond to TSLP [29] bring up a therapeutically important possibility that chronic high levels of TSLP may be an initiating factor in loss of B cell tolerance and/or B-lymphocytosis, a leukemia-like disease in humans. More important, this raises the possibility that skin has a central role in driving a wide variety of inflammatory and humoral diseases in which skin complications are also present.
Compound strains of mice were engineered as described [14]. All animals were maintained in mixed genetic backgrounds; however, littermates were compared whenever possible. All mice were kept in the animal facility under Washington University animal care regulations. In studies related to longevity, mice were monitored regularly for sign of cachexia and failure to thrive; care was taken to reduce competition for affected pups by removal of wild-type littermates unnecessary for the study. Severely affected individuals were left with their dams for their entire life span. Morphological details of the cutaneous phenotypes will be published elsewhere. The following cohort of animals was analyzed: Msx2-Cre/+; N1flox/flox (N1CKO), Msx2-Cre/+; N1flox/flox; N2flox/+ (N1N2hCKO), Msx2-Cre/+; N1flox/flox; N2flox/+; N3–/– (N1N2hN3CKO), N1N2CKO, PSDCKO, RBP-jCKO, PSDRBP-jCKO, N1N2RBP-jCKO, Prx1-Cre/+; PS1flox/flox; PS2–/–, Msx2-Cre/+; Rosa26R/+, Msx2-Cre/+; ZEG/+, CD4-Cre/+; ZEG/+, K14-TSLPtg. Wild-type cohorts in this study included Cre-negative littermates, Msx2-Cre/+; PS1flox/+; PS2–/–, Msx2-Cre/+; RBP-jflox/+, and Msx2-Cre/+; Notch1flox/+.
For hematoxylin-and-eosin staining, tissue samples were fixed in 4% paraformaldehyde in phosphate-buffered saline (PBS), dehydrated in ethanol, and paraffin-embedded. Tissues blocks were sectioned at 5 μm; Lac-Z staining on E15.5 embryo was performed as previously described [14]. Immunostaining on paraffin-embedded tissue samples was performed with the following biotinylated antibodies: anti-B220 (clone RA3-6B2, BD Pharmingen), anti-F4/80 (Abcam), anti-CD3e (clone 145–2C11, BD Pharmingen), and anti-TSLP (R&D Systems). Horseradish-peroxidase-conjugated streptavidin and DAB substrate kit (Pierce) were used to visualize the signal. For RBP-j staining, anti-RBP-j antibody (clone T6709, Institute of Immunology) was used together with biotinylated anti-rat secondary antibody. Sections were counterstained with hematoxylin.
Serum TSLP, IL-6, and IL-7 levels were measured according to the manufacturer's instructions in the Quantikine mouse TSLP, IL-6, and IL-7 ELISA kits (R&D Systems).
Single cell suspensions from peripheral blood, BM, and spleen were prepared for FC analysis as described [43]. The following antibodies were used: anti-B220 (RA3-6B2) conjugated to fluorescein (FITC), phycoerythrin (PE), or peridinin chlorophyll-a protein-cyanin 5.5 (PerCP-Cy5.5), anti-CD45 (30-F11) conjugated to PerCP-Cy5.5, anti-Ly-6G (1A8), anti-TER-119, anti-Thy-1 (5E10), anti-CD3e (500A2), and anti-CD43 (S7) conjugated to PE, anti-IgM (R6–60.2) conjugated to FITC and PE (all from BD Pharmingen). Stained cells were studied using a BD FACScan Flowcytometer (Cytek Development), and the data were analyzed using FlowJo software.
For allogeneic BMT, the immunocompetent recipient mice were lethally irradiated with 950 cGy at ∼P10 and transplanted with freshly harvested, unfractionated BM cells from their littermates as previously described [43]. However, NOD/SCID mice received BMT after a sublethal dose of irradiation (300 cGy). In each case, 2 × 106 cells in 100 μl of PBS + 2% fetal bovine serum were injected into the retro-orbital sinus of the irradiated recipient animal. A cohort of transplanted N1N2CKO mice were treated orally with 50 μl of 200 mg/ml cephalexin (Ranbaxy Pharmaceuticaks) twice daily.
To study disease progression and monitor disease occurrence/recurrence, all mutant and irradiated/transplanted mice were monitored closely over their life spans for signs of weakness, weight loss, and morbidity. Blood samples were collected from the mandibular vein. Hematological analysis (Hemavet 950 analyzer, Drew Scientific) comprised complete blood count including WBCs, platelets, red blood cells, white cell differential counts, and hemoglobin measurements [44]. White cell differential counts were confirmed on blood smears. In addition, blood samples were collected for FC analysis. Moribund mice were euthanized, and peripheral blood, BM, lymph nodes, thymus, lung, liver, and spleen were collected for a comprehensive pathological analysis as described previously [45].
Wild-type mice were injected intravascularly with carrier alone (50 μl PBS) or with carrier containing 0.5, 1, or 1.5 μg of recombinant mouse TSLP (R&D Systems) daily for 7 d starting on P0, P7, or P14. Mice were euthanized 12 h after the last injection, and tissues were collected for analysis. An ELISA assay of serum measured the systemic TSLP level. Complete blood count and FC analyses on BM, blood, and spleen were performed to check for any sign of B-LPD.
To test barrier function defect, a dye penetration assay was performed as previously outlined [46]. Briefly, intact E18.5 embryos were stained in X-gal (pH 4.5) for 12 h at 37 °C. After X-gal staining and three rounds of PBS wash, the embryos were photographed with a digital camera.
Primary keratinocytes were isolated from the dorsal midline skin of newborn mutant and wild-type littermates. The cells were maintained in 60-mm2 dishes with a medium of low calcium concentration as previously described [47]. Keratinocytes were plated at 40% confluence and allowed to double. Confluent plates (80%) were treated with 2.5, 5, or 10 μM NFκB inhibitor (BAY 11–7082) or carrier alone (DMSO). After 24 h, cells were re-fed with inhibitor/DMSO containing fresh medium with or without 2 mM CaCl2 to induce differentiation. After 24 h, cells and their media were harvested for analysis.
Keratinocyte lysates in SDS were immunoblotted to check for PS1 (H-70, Santa Cruz Biotechnology) and α-tubulin (B-5–2, Sigma-Aldrich) as described [15].
Frozen skin sections, 7 μm, were stained with 0.15 mg/ml Nile Red in 75% glycerol for 2 min and counterstained with DAPI [48].
Conventional PCR for PS1 alleles was done on genomic DNA of most tissues. For blood, fresh or frozen blood samples were used directly as template with KlenTaq10 (DNA Polymerase Technology) supplemented with 1.3 M final concentration of betaine. qRT-PCR was performed as described [15]. The primer sequences are provided in the Text S1.
Detailed description of microarray analyses on skin or epidermal mRNA samples from P9 Notch-deficient animals and skin mRNA samples from wrfr–/– embryos are provided in the Supporting Information (Text S1).
The bar graphs present the mean and standard deviation of each measured parameter. Student's t-test is applied as the test of significance unless otherwise specified.
Accession numbers for genes mentioned in this paper from the National Center for Biotechnology Information (http://www.ncbi.nlm.nih.gov) are FATP4 (AJ276492), NFκB (AY521463), Notch1 (NM_008714), Notch2 (NM_010928), Notch3 (NM_008716), PS1 (NM_008943), PS2 (NM_011183), RBP-j, (NM_009035), and TSLP (NM_021367).
|
10.1371/journal.ppat.1005961 | The Drosophila CD36 Homologue croquemort Is Required to Maintain Immune and Gut Homeostasis during Development and Aging | Phagocytosis is an ancient mechanism central to both tissue homeostasis and immune defense. Both the identity of the receptors that mediate bacterial phagocytosis and the nature of the interactions between phagocytosis and other defense mechanisms remain elusive. Here, we report that Croquemort (Crq), a Drosophila member of the CD36 family of scavenger receptors, is required for microbial phagocytosis and efficient bacterial clearance. Flies mutant for crq are susceptible to environmental microbes during development and succumb to a variety of microbial infections as adults. Crq acts parallel to the Toll and Imd pathways to eliminate bacteria via phagocytosis. crq mutant flies exhibit enhanced and prolonged immune and cytokine induction accompanied by premature gut dysplasia and decreased lifespan. The chronic state of immune activation in crq mutant flies is further regulated by negative regulators of the Imd pathway. Altogether, our data demonstrate that Crq plays a key role in maintaining immune and organismal homeostasis.
| Phagocytosis is a first-line host defense mechanism against microbes. Interactions between phagocytosis and other immune mechanisms, such as the humoral response, however, remain elusive. Defective phagocytosis can lead to immune deficiencies and chronic auto-inflammation. Here, we show that Croquemort (Crq), a Drosophila member of the CD36 family of scavenger receptors, plays a role in microbial phagocytosis. Crq is required in phagocytes for efficient uptake of bacteria and fungi, and mutants for crq succumb to both environmental microbes and infections. Crq is also required for phagosome maturation, and crq mutants lack the ability to fully clear bacterial infection. As a result, crq mutant flies enter a state of chronic immune activation. Notably, they show increased production of the cytokine Upd3 that induces intestinal stem cell proliferation. Consequently, crq mutant flies show early signs of gut dysplasia, as well as a shortened lifespan. Altogether, our study demonstrates a new link between phagocytosis and tissue homeostasis, and illustrates how the chronic induction of cytokine production secondary to defective phagocytosis can alter gut homeostasis and shorten lifespan.
| Mounting appropriate immune responses against pathogens is critical for the survival of all animals. Mechanisms to both eliminate microbes and resolve infection by returning the immune system to basal activity are necessary to maintain an adequate and balanced immune response [1,2]. Alterations in these responses can lead to immune deficiency or auto-inflammation [3–5]. Yet, to date, how these mechanisms are coordinated upon infection remains unclear.
Drosophila is a prime model to genetically dissect humoral and cellular innate immune responses to a variety of pathogens [6–8]. Humoral responses include the pro-phenoloxidase (PO) cascade, which leads to the generation of reactive oxygen species and melanization, and the rapid production of antimicrobial peptides (AMPs) regulated by the Toll and Imd pathways [7]. Upon recognition of microbial lysine (Lys)-type peptidoglycan (PGN), damage-associated molecular patterns (DAMPs), or exogenous protease activity, the Toll pathway promotes the nuclear translocation of the NF-κB-like transcription factor Dorsal-related Immune Factor (Dif) to induce AMP genes, such as Drosomycin [6,9]. In contrast, detection of bacterial meso-diaminopimelic acid (DAP)-type peptidoglycan activates the Imd pathway and leads to the nuclear translocation of the NF-κB-like transcription factor Relish (Rel) to induce transcription of AMP genes, such as Diptericin [10,11]. It has also been shown that proteases, such as Elastase and Mmp2, can activate the Imd pathway through cleavage of the receptor PGRP-LC [12]. As in mammals, chronic activation of immune responses is deleterious to the fly, and negative regulators are required to maintain immune homeostasis [13–15]. For instance, amidase PGN recognition proteins (PGRPs), such as PGRP-LB and PGRP-SC, negatively regulate the Imd pathway by enzymatically degrading PGN [14–16].
Phagocytosis and encapsulation are key cellular innate immune responses [7]. Phagocytosis allows for the uptake and digestion of microbes and apoptotic cells by phagocytes, including specialized immune cells called plasmatocytes [7,17]. Encapsulation results in the isolation and melanization of large materials, such as wasp eggs or damaged tissues, by dedicated immune cells named lamellocytes [18]. Both phagocytosis and humoral responses are required to fight infection. Indeed, decreasing the phagocytic ability of plasmatocytes by pre-injecting latex beads, which they take up, impairs fly survival upon infection with Gram-positive bacteria [19]. Similarly, inhibiting phagocytosis increases the susceptibility of Imd pathway-deficient flies to Escherichia coli (E. coli) infection, arguing that phagocytosis and the humoral response act in parallel [20]. Plasmatocytes were proposed to activate the production of AMPs by releasing immunostimulatory pathogen-associated molecular patterns (PAMPs) following phagocytosis [21]. They also express cytokines such as Unpaired 3 (Upd3), a ligand of the JAK-STAT pathway, which regulates immune-related genes [22]. Yet, ablation of the majority of plasmatocytes by targeted apoptosis has only a moderate effect on the fly’s ability to fight infection [23,24]. Therefore, the role of phagocytosis in the regulation of the humoral response and the resolution of infection remains unclear.
Several plasmatocyte receptors promote the recognition and engulfment of bacteria [25]. The scavenger receptor dSR-CI and a transmembrane protein, Eater, bind to both Gram-negative and -positive bacteria [26,27]. The membrane receptor PGRP-LC binds to and engulfs Gram-negative but not Gram-positive bacteria, and its membrane localization is dependent on the nonaspanin TM9SF4 [28,29]. Draper (Drpr) promotes clearance and degradation of neuronal debris and apoptotic cells via phagosome maturation, as well as phagocytosis of Staphylococcus aureus (S. aureus) together with the integrin βv and PGRP-SC1 [30–32]. Nimrod C1, which is related to Eater and Drpr, promotes phagocytosis of both S. aureus and E. coli by Drosophila S2 cells, and suppression of its expression in plasmatocytes inhibits phagocytosis of S. aureus [33]. Peste, a member of the CD36 family of scavenger receptors plays a role in the recognition and uptake of Mycobacterium by S2 cells [34]. Finally, croquemort (crq), another CD36 family member, promotes apoptotic cell clearance by embryonic plasmatocytes [35] and phagosome maturation of neuronal debris by epithelial cells [36].
In mammalian immunity, CD36 promotes the uptake of oxidized low density lipoproteins (oxLDLs) [37,38] and also regulates the host inflammatory response [39,40]. In addition, it is required to fight Mycobacteria and S. aureus infections in mice [41] and to induce pro-inflammatory cytokines in response to Plasmodium falciparum infection [42]. Using two lethal deficiencies that delete crq (as well as other genes), we previously proposed that Crq was specific to apoptotic cell clearance, as crq-deficient embryonic plasmatocytes retained some ability to engulf both E. coli and S. aureus in vivo [35]. However, Crq was subsequently implicated in phagocytosis of S. aureus by S2 cells, a heterogeneous cell line with phagocytic abilities derived from late embryonic stages [41]. Thus, we generated a knock-out of crq and further investigated its role in microbial phagocytosis and its relationship with the humoral response at larval and adult stages in vivo.
Drosophila plasmatocytes derive from pro-hemocytes originating either in the procephalic mesoderm of the embryo, with some further expanding by self-renewal in larval hematopoietic pockets, or from a second hematopoietic organ, the larval lymph glands, that persist to adulthood, or finally from adult hematopoietic hubs [43–51]. Here, we show that Crq is a major marker of plasmatocytes that is not required for hematopoiesis. The survival to adulthood of crq knock-out (crqko) mutants allowed us to quantitatively demonstrate that crq is required for pupae to survive environmental microbe infections and for adults to resist infection against Gram-negative and Gram-positive bacteria and fungi. crqko flies tolerate infections as well as control flies, but are unable to efficiently eliminate microbes. Indeed, crqko plasmatocytes are poorly phagocytic and defective in phagosome maturation. Crq acts parallel to the Imd and Toll pathways in eliminating pathogens, and crqko flies display elevated and persistent Dpt and upd3 expression, demonstrating that mutating crq promotes a state of chronic immune activation. As a consequence, crqko flies die prematurely with early signs of gut dysplasia and premature intestinal stem cell hyperproliferation. Therefore, we propose a model wherein crq is central to immune and organismal homeostasis. Overall, our results shed new light on the links between phagocytes, commensal microbes, gut homeostasis, and host lifespan.
In Drosophila adults, plasmatocytes (the phagocytic hemocyte lineage) originate from both embryonic and larval hematopoiesis [52]. crq is expressed in embryonic and larval plasmatocytes, as well as in S2 cells [53]. To test whether crq is expressed in adult plasmatocytes, we performed dual staining with combinations of GFP or dsRed and Crq antibodies of hemocytes bled from previously characterized transgenic plasmatocyte-reporter lines: eater-nls::GFP, eater-dsRed, and Hml-Gal4>UAS-GFP (Hemolectin-positive hemocytes) [54,55] (Fig 1A and 1B). We found that 83.3±4.4% of hemocytes of Hml-Gal4>UAS-GFP and eater-dsRed carrying flies were positive for both markers, while 16.7±4.4% were positive for eater-dsRed alone (Fig 1C). Crq immunostaining of hemocytes bled from eater-nls::GFP flies revealed that 85.2±2.6% of them were Crq and eater-dsRed positive, while 14.8±2.6% were Crq-positive but did not express eater-dsRed (Fig 1C). From this, we extrapolated that about 72.4% of circulating hemocytes are positive for all three markers, 12.8% are double positive for Crq and Eater, and 14.8% solely express Crq (Fig 1C). Therefore, crq is expressed in all Eater and Hml-positive hemocytes and marks the majority, if not all, adult plasmatocytes.
To study its role in vivo, we generated a knock-out allele of crq (crqko) by homologous recombination [36]. This mutant deletes the entire crq open reading frame (S1A Fig), and thus abolishes its expression [36]. As previously reported [35], crq was not required for embryonic hematopoiesis. As for crq deletion mutants, crqko embryonic plasmatocytes were less efficient at clearing apoptotic cells, having a phagocytic index of 1.6±0.2 versus 2.45±0.3 apoptotic cells/plasmatocyte for wild-type embryos (p<0.05, S1B Fig). Homozygous crqko flies were viable and appeared morphologically normal. To ask whether crq is required for hematopoiesis at later developmental stages, we recombined an eater-nls::GFP transgene (i.e., the broadest plasmatocyte reporter after Crq) (Fig 1C) into the crqko mutants, bled larvae and adults, and semi-automatically scored their eater-nls::GFP positive plasmatocytes by microscopy (S1C Fig and Fig 1D). As previously reported for wild-type [56,57], adult crqko flies had about 5-fold less plasmatocytes than larvae, and their number of eater-nls::GFP-positive plasmatocytes at both larval and adult stages were similar to that of wild-type flies (Fig 1D). Pro-hemocytes that differentiate into plasmatocytes can also differentiate into crystal cells, which are involved in melanization [58]. Furthermore, self-renewing plasmatocytes of the embryonic lineage can also differentiate into crystal cells by trans-differentiation [59,60]. Thus, we tested whether crqko flies have differentiated crystal cells by scoring the melanotic dots formed following heat-induced crystal cell lysis. We found no significant difference between crqko and wild-type larvae (S1D Fig). Therefore, Crq is a major plasmatocyte marker that is not required for hematopoiesis or hemocyte differentiation.
While crqko homozygous flies were viable to adulthood, we could not maintain a homozygous stock on conventional fly food. We found that 36±3.2% homozygous crqko larvae arose from crosses between crqko heterozygous flies over GFP-marked CyO balancer chromosome, indicating full viability of the homozygous larvae (Fig 1E). However, only 18±1.7% of emerging adults were homozygous crqko flies, indicating that half of the crqko homozygous progeny died during pupariation. Because flies with decreased plasmatocyte counts undergo pupal death associated with the presence of otherwise innocuous environmental microbes [23], we asked whether supplementing the food with antibiotics could rescue crqko lethality. With this treatment, we recovered 29±3.6% of crqko homozygous adults (Fig 1E), indicating a partial rescue of pupal lethality (homozygous vs balanced adults, p = 0.021). These results suggest that crqko pupae are susceptible to environmental microbes.
No adult progeny could be recovered from crqko homozygous crosses on conventional fly food, but crqko adults emerged in the presence of antibiotics that gave rise to a second adult progeny (Fig 1E and 1F). Maintaining a homozygous viable stock with antibiotics, however, remained difficult. We next bleached homozygous crqko embryos and raised them on sterile food. Under these axenic conditions, we successfully cultured a homozygous crqko line (Fig 1F). Therefore, environmental microbes represent a health constraint for crqko homozygous flies.
The susceptibility of crqko pupae to environmental microbes suggested that crq is required to mount an appropriate immune response. We next asked whether crq was up-regulated in flies injected with the Gram-negative bacterium Pectinobacterium (previously known as Erwinia) carotovora 15 (Ecc15) or the Gram-positive Enterococcus faecalis (E. faecalis). As anticipated, there was no crq expression in unchallenged (UC) or infected crqko flies as detected by RT-qPCR (Fig 2A). While crq was expressed in both UC pXH87-crq transgenic (the parental transgenic strain used for the generation of crqko flies, hereafter referred to as PXH87) and Canton S (Cs) control flies, it was not up-regulated within the first 24hrs of infection with Ecc15 or E. faecalis (Fig 2A). However, we cannot exclude the possibility that crq may be up-regulated in plasmatocytes specifically at these early time points after infection. Its expression was also not altered in mutant flies for the NF-κB-like transcription factor Relish (RelE20) downstream of the Imd pathway, or in flies mutant for the Toll ligand spz (spzrm7), upstream of the Toll pathway during that time-frame [9]. Surprisingly, we did observe an increase in crq mRNA levels at 36 (p = 0.0076) and 132 hrs (p = 0.0213) post Ecc15 infection (S2A Fig), but did not detect any upregulation of crq mRNA levels at 36 and 132 hours post E. faecalis infection (S2B Fig) (p>0.05). Altogether, our data show that crq does not appear to be induced by infection in whole adult extracts during the first 24 hours post infection with Ecc15 and E. faecalis, and its expression appears independent of the Toll and Imd pathways. However, at later time-points after infection crq can be upregulated in a pathogen-specific manner, as seen with Ecc15 here.
To assess the susceptibility of crqko male (Fig 2B–2G) and female (S2B–S2G Fig) flies to a variety of pathogens, we monitored their survival to these infections over time. When challenged by septic injury with the Gram-negative bacterium Ecc15, male (Fig 2B) and female (S2C Fig) crqko flies were more susceptible than Cs and PXH87 control flies to this infection (p<0.0001). crqko flies all died within 336 hrs post-infection (hpi), while only 64±6.8% and 67±6.5% of PXH87 and Cs flies had died by that time-point. crqko flies were, however, less susceptible than RelE20 mutants (p<0.0001), which are defective in the production of AMPs downstream of Imd [61]. All RelE20 flies died within 72 hpi, while only 56±7.7% of crqko flies had succumbed by that same time-point (Fig 2B). To verify that the susceptibility to Ecc15 infection was due to the crqko mutation and not to a background mutation, we infected trans-heterozygous flies for crqko and Df(2L)BSC16, which deletes crq, with Ecc15 (S2D Fig). These flies were as susceptible to Ecc15 infection as the crqko homozygous flies; they all died within 288 hpi, indicating that the crq mutation is responsible for this phenotype (S2D Fig). crqko flies also succumbed to infection with E. coli (39±8.1% survival at 336 hpi), a Gram-negative bacterium that does not kill Cs (97±2.5% survival) or PXH87 (86±5.6% survival) flies. However, crqko flies were less susceptible to E. coli infection than RelE20 flies, which all died within 312 hpi (p<0.0001) (Fig 2C and S2E Fig). Therefore, crqko flies are susceptible to various Gram-negative bacterial infections.
Similarly, crqko flies were more susceptible to infection with the Gram-positive bacterium E. faecalis than controls (p = 0.0006) (Fig 2D and S2F Fig) and died in 312 hpi. However, they were less susceptible than spzrm7 flies (p<0.0001), which are defective in the production of AMPs downstream of Toll and died within 72 hpi (Fig 2D and S2F Fig). crqko flies also died with intermediate susceptibility between that of control and spzrm7 flies (p<0.0001 for both) after septic injury with the pathogenic yeast Candida albicans (Fig 2E and S2G Fig). Similarly, crqko flies were significantly more susceptible to exposure to spores of the entomopathogenic fungus Beauveria bassiana than Cs and PXH87 flies (p<0.0001), but less susceptible than spzrm7 flies (p<0.0001) (Fig 2F and S2H Fig). Finally, crqko flies were more susceptible to S. aureus infection than spzrm7 flies (p = 0.0073) (Fig 2G and S2I Fig), and spzrm7 flies were only slightly more susceptible than Cs and PXH87 flies (p = 0.0006 and p<0.0001 respectively). Therefore, crqko flies are susceptible to Gram-positive bacteria and fungal infections and strongly susceptible to infection with S. aureus, a bacterium specifically cleared by phagocytosis [19,62,63].
These results argue that crq is required to fight infection. To further confirm this, we drove the expression of a UAS-crq transgene under the control of a crq promoter-Gal4 driver in the crqko flies (crqko; crq-Gal4>UAS-crq). These rescue flies were no longer susceptible to Ecc15 (S3A Fig), E. faecalis (S3B Fig), and B. bassiana (S3C Fig) infections (non-significant (ns) compared to PXH87, and p<0.0001 when compared to crqko flies) (S3A–S3C Fig). To assess the possible requirement of crq in hemocytes, we drove the expression of a UAS-crq transgene under the control of a hemocyte-specific serpent promoter-Gal4 driver in the crqko flies (crqko; srp-Gal4>UAS-crq). These flies were significantly less susceptible to Ecc15, E. coli, E. faecalis and C. albicans infections than crqko flies (p<0.0001, p<0.0001, p = 0.0004 and p<0.0001, respectively) (S3D–S3G Fig). We did not observe any significant differences between rescue experiments with the crq-Gal4 or srp-Gal4 drivers after infection with Ecc15, E. coli, or E. faecalis (p>0.05). The hemocyte-specific rescue of crqko flies infected with C. albicans, however, was slightly less efficient than the rescue with the crq-Gal4 driver (p = 0.0269). Thus, crq appears to be required mostly in phagocytes to fight infection by both Gram negative and Gram positive bacteria, although it appears to also be required in other tissues to fight C. albicans infection.
Multi-cellular organisms use two complementary strategies to fight infection: resistance, to eliminate microbes, and tolerance, to allow them to endure the infection and/or its deleterious effects [64,65]. Compared to controls, crqko flies die prematurely at around 552 hours even in the absence of infection (S4A Fig), suggesting these flies could be generally unfit or susceptible to damage. To test their response to abiotic damage, we pricked crqko flies with sterile needles at two separate thoracic sites. These flies did not die any earlier than non-pricked crqko flies (S4A Fig). Thus, despite their decreased lifespan, crqko flies are not susceptible to aseptic wounds.
To date, few studies have quantified the tolerance of immune-deficient flies [66,67]. Tolerance can be measured as the dose response curve relating health to microbe load. This curve takes the shape of a sigmoid; life expectancy in unchallenged conditions is considered as vigor, and the slope of the response curve (the portion of the health/load curve which is linear) estimates the ability to tolerate infection (S4B Fig) [67]. crqko flies have shortened lifespan and therefore an altered vigor (S4A Fig). We further aimed to estimate whether crqko flies show a decrease in tolerance by measuring the relationship (statistical interaction) between microbial load and the corresponding health of the host [64,67]. We used three approximations to relate health to microbe load of crqko flies and focused on the linear part for each regression. First, we estimated the regression between the LT50 (time at which 50% of the flies are dead) of Ecc15 or E. faecalis-infected flies and the number of bacteria injected (measured as colony forming units or CFUs) (S4C and S4D Fig). We did not detect any significant LT50~Time interaction between PXH87 and crqko flies (p = 0.21782 for E. faecalis, p = 0.55800 for Ecc15) (S4C and S4D Fig). However, this measure of bacterial load does not take into account the growth of the pathogen within the host. We therefore also quantified the regression between LT50 and the number of bacteria in the flies at 24 hpi (Fig 3A and 3B). We detected significant LT50~Time interaction between PXH87 and crqko flies (p = 0.008486 for E. faecalis, p = 0.018965 for Ecc15), with PXH87 flies having lower tolerance than crqko flies (Fig 3A and 3B). Finally, to get another estimate of the health of the flies, we plotted the health/bacterial load curve using survival at 3 time-points post Ecc15 infection and their corresponding bacterial load (S4E Fig). We did not detect any significant survival-time interaction between PXH87 and crqko flies (p = 0.335111). Thus, while crqko flies die prematurely in the absence of infection, they do not show any decreased tolerance to infection when compared to control flies.
These data suggest that the increased susceptibility of crqko flies to infection is due to their inability to control bacterial growth. In order to test this hypothesis, we monitored bacterial load during the course of Ecc15 and E. faecalis infections. In PXH87 flies, Ecc15 is eliminated within the first 48hrs of infection to reach an apparent plateau of low number of CFUs that persist at 72hrs post-infection (Fig 3C). crqko flies were less able to clear Ecc15 than controls with higher bacterial loads throughout the infection (p<0.001 for 24, 48 and 72hrs) (Fig 3C). In contrast, despite an initial decline of CFUs at 48hrs, E. faecalis grew within control flies at 96 and 168hrs (Fig 3D). During the whole course of infection with E. faecalis, the bacterial loads were significantly lower in wild-type control flies than in the crqko flies (p<0.001 at 48, 96 and 168hrs) (Fig 3D). These data indicate that crq is required for efficient elimination of both Ecc15 and E. faecalis.
crq is required for efficient phagocytosis of apoptotic cells (also known as efferocytosis) in vivo, and phagocytosis of S. aureus by S2 cells (S1B Fig and [35,41]). In addition, rescue of crq expression in hemocytes improved survival to various infections (S3D–S3G Fig), suggesting that crq could alter microbial phagocytosis. To test this hypothesis, we first compared the susceptibility of crqko flies to infection with that of mutants for two phagocytic receptors, Eater and Drpr [26,31]. crqko flies succumbed to Ecc15 infection significantly faster than eater-deficient (p = 0.0002) and drprrec8Δ5 loss-of-function flies (p<0.0001). 90±3.58% of crqko flies died within 192 hpi, while only 60±6.77% of drpr and eater mutants died in that same time (Fig 4A). However, the crqko flies were significantly less susceptible to Ecc15 than RelE20 flies (p<0.0001), which all died within 48 hpi (Fig 4A). In contrast, crqko flies succumbed to E. faecalis infection at a similar pace to that of both eater-deficient and drpr rec8Δ5 flies with 80–90% of all strains dying within 240 hpi (Fig 4B). However, all mutants were significantly less susceptible than spzrm7 flies, which died within 48 hrs of E. faecalis infection (p<0.0001) (Fig 4B).
To examine the precise role of crq in phagocytosis, we compared the amount of bacteria engulfed within 45min of thoracic injections of dead, Alexa 480-labeled E. coli and S. aureus in Cs, PXH87, and crqko flies as previously described [20,26] (Fig 4C and 4D). The crqko flies engulfed both E. coli and S. aureus bacteria with on average 66% less efficiency than control flies (Fig 4C–4F, respectively). This phenotype was completely rescued in crqko flies expressing a UAS-crq transgene under a crq-Gal4 driver (crqko, crq-Gal4>UAS-crq), which appeared to engulf more efficiently than control PXH87 flies (S5A Fig and S5B Fig). We speculate that this difference was due to the overexpression of crq in those flies.
To further assess the phagocytosis phenotype, wild-type and crqko flies carrying the eater-nls::GFP plasmatocyte-reporter were injected with dead rhodamine-labeled E. coli, bled, and their plasmatocytes were analyzed by confocal microscopy for internalized bacteria (Fig 4G). The rhodamine-fluorescence per eater-nls::GFP plasmatocyte was quantified at 45min, 3hrs, and 5hrs post-injection and normalized to that of WT plasmatocytes at 3hrs post-injection (Fig 4H). In WT plasmatocytes, the relative rhodamine-fluorescence increased as early as 45min, peaked at 3hrs, and decreased after 5hrs, as bacteria were presumably digested in mature phagosomes (Fig 4H). In contrast, crqko plasmatocytes accumulated about 2-fold fewer bacteria than controls at 45min and 3hrs post-injection, but accumulated 1.7-fold more bacteria by 5hrs post-injection. In addition, at 45min post-injection, most bacteria were internalized within wild-type plasmatocytes (S5C Fig), whereas bacteria were often bound to the cell surface of crqko plasmatocytes without being internalized (S5D Fig). Thus, crqko plasmatocytes can engulf bacteria but are less efficient at it than controls at early time-points; they also appear to accumulate internalized bacteria over time. These results are consistent with a role for crq in promoting efficient uptake of bacteria. Moreover, the observed accumulation of bacteria in crqko plasmatocytes at 5hrs post-injection suggested that crq could also be required for phagosome maturation and digestion of bacteria. To test this, we injected control, crqko, and rescue flies with pH-sensitive pHrodo E. coli and S. aureus. pHrodo bacteria fluoresce when engulfed into a fully mature, acidified phagosome [68] (S5E and S5F Fig). After quantification, we observed about 50% less fluorescence in crqko when compared to controls at 1, 3, and 5hrs post-injection (S5G and S5H Fig, p<0.5 when comparing PXH87 and crqko flies). This phenotype was again completely rescued in crqko, crq-Gal4>UAS-crq flies (S5E and S5F Fig and S5G and S5H Fig, p>0.5 when comparing PXH87 and rescue flies). At the single cell level, crqko plasmatocytes had up to 63±5.66% and 55±7.46% less pHrodo E. coli than controls at 3 and 5hrs, respectively (Fig 4I).
Finally, to ask whether mutating crq resulted in persistence of pathogenic bacteria, we injected live GFP-labeled Ecc15 in control and crqko flies carrying the eater-dsred plasmatocyte reporter (Fig 4J and S5I Fig). Control PXH87 plasmatocytes had little to no GFP signal at 4 days post-infection, indicating that most bacteria had been engulfed and digested (Fig 4J and S5I Fig). In contrast, crqko plasmatocytes had a 6-fold higher GFP signal, demonstrating that live Ecc15 accumulate in crqko plasmatocytes (Fig 4J and S5I Fig). Taken together, these results show that crq is required for efficient microbial phagocytosis by playing a role in bacterial uptake and phagosome maturation.
Phagocytosis has been proposed as a key step to initiate AMP production [21]. To assess the effect of mutating crq on AMP production downstream of both the Imd and Toll pathways, we next quantified the expression of Diptericin (Dpt) and Drosomycin (Drs)-encoding genes by RT-qPCR after Ecc15 or E. faecalis infections (Fig 5A and 5B). As previously reported, septic injury of control flies with Ecc15 induced Dpt expression, which peaked at 10 hpi and returned to near-basal levels within 48 hpi (Fig 5A)[7]. In crqko flies, Dpt expression was 2-fold higher than in control flies at 10hrs post-infection and failed to return to basal levels within 48hrs (Fig 5A). In contrast, there was no significant difference in Drs induction between control and crqko flies after E. faecalis inoculation (Fig 5B). Survival curves indicated that crqko flies were less susceptible to a non-pathogenic E. coli infection than RelE20 flies, while double mutants for crqko and RelE20 were statistically more susceptible than RelE20 or crqko mutants alone (Fig 5C). The extreme sensitivity of RelE20 flies to infection with pathogenic bacteria prevented us from carrying out these experiments with Ecc15. Instead, we inoculated the flies with 20 times fewer E. coli than previously used in Fig 2. Similarly, crqko and spzrm7 double mutants were also statistically more susceptible to C. albicans infection than the spzrm7 or crqko mutants alone (Fig 5D). Therefore, crq is not required for the induction of AMPs and acts in parallel to the Toll and Imd pathways.
These results suggested that aberrant phagocytosis in crqko flies can result in enhanced and persistent Imd pathway activation. Multiple negative regulators of the Imd pathway help maintain immune homeostasis. For example, Peptidoglycan Recognition Proteins (PGRPs) with amidase activity, such as PGRP-LB, degrade immunostimulatory molecules [15]. Thus, we next assessed Dpt expression levels by RT-qPCR in single and double PGRP-LBΔ and crqko mutants upon Ecc15 infection. Single crqko and PGRP-LBΔ mutants expressed statistically higher levels of Dpt than Cs and PXH87 controls at 48hrs (Fig 5E). The Dpt expression resolved back to basal levels within 72hrs post infection in control flies, but remained high in single PGRP-LBΔ or crqko mutants despite a steady decline in its expression (Fig 5E). Moreover, double mutants for crqko and PGRP-LBΔ expressed Dpt at levels 5-fold higher than controls at 24hrs post-infection, and levels remained high at 48 and 72hrs (Fig 5E). These results demonstrate the critical interplay between phagocytosis and negative regulators of the immune system to achieve proper resolution of AMP expression upon systemic infection.
Plasmatocytes are also a major source of cytokine production upon systemic infection. Upd3, the Drosophila analogue of IL-6, can induce the JAK-STAT pathway, which regulates the systemic immune response and metabolic homeostasis in the fat body, as well as gut homeostasis [6,22,69,70]. Using RT-qPCR, we asked whether crq is required for upd3 expression upon Ecc15 infection. Control flies displayed a small and temporary induction of upd3 expression that resolved within 72hrs (Fig 5F). In contrast, UC and Ecc15-challenged crqko flies showed a 1.5-fold stronger induction of upd3 expression, which further increased over 72hrs (Fig 5F). Thus crq is not required to induce upd3 expression, but crq mutation results in enhanced and continuously increasing upd3 expression. Altogether, these results demonstrate that crq is required for bacterial clearance and mutation of crq alters the resolution of AMPs and Upd3 cytokine production.
PGRP-LBΔ and RelE20 mutants all die prematurely, within about 696 hrs (29 days) of age, when compared to wild-type (p<0.0001) and PXH87 (p<0.0001) control flies, which die after about 912 hrs (37 days) on conventional food at 29°C [15] (Fig 6A). crqko flies died on average within 552 hrs (23 days), considerably earlier than RelE20 and PGRP-LBΔ mutants (p<0.0001). Double crqko and PGRP-LBΔ or crqko and RelE20 mutants died within about 480 hrs (20 days) and 408 hrs (17 days) of age, respectively (Fig 6A). Antibiotic treatment partially rescued these phenotypes, as the lifespan of crqko flies and the double mutants increased significantly (p<0.0001) (Fig 6B). To ask whether the premature aging of crqko flies might correlate with a loss of immune cells or their function, we estimated the number of plasmatocytes present in control and crqko flies using the eater-nlsGFP reporter (Fig 6C). As previously reported [56,57], the number of plasmatocytes was decreased by about 40% in 16-day-old control flies (Fig 6C), while similarly aged crqko flies had lost 80% of their plasmatocytes (Fig 6C). Treatment with antibiotics rescued this crqko phenotype but had no effect on the plasmatocyte counts of control flies. crqko flies also lost about 40% of their plasmatocytes at 4 days post-E. faecalis infection when compared to similarly challenged wild-type controls (Fig 6D). This loss of crqko hemocytes may be a consequence of accumulation of undigested bacteria inside their phagosomes. Thus, crq is required for plasmatocytes to survive innocuous or pathogenic bacterial infection.
Dpt expression in wild-type and PXH87 flies is relatively low and stable over the first 8 days of their lives and increases as flies age [71] (Fig 6E). Strikingly, Dpt expression was 70-fold higher in 8-day-old crqko flies and nearly 1,100-fold higher in the double mutants for crqko and PGRP-LBΔ compared to controls (Fig 6E). Thus, the Imd pathway is strongly up-regulated early on in the life of these mutant flies, even in the absence of infection. This points to a role for Crq in phagocytosis and in maintaining immune homeostasis. Likewise, upd3 expression steadily increased as PXH87 flies aged, and it was further enhanced by nearly 10-fold in 8- and 16-day-old crqko flies (Fig 6F). Antibiotic treatment partially rescued the levels of Dpt expression in crqko flies (S6A Fig), arguing that the hyper-activation of the Imd pathway in these flies results from their inability to control environmental microbes. To address this, we plated fly extracts on both LB (on which most pathogens can grow) and MRS (on which most Drosophila microbiota can grow) agar plates and quantified the resulting CFUs (S6B Fig). In line with previous studies, the CFUs obtained from 2 week-old control flies were in the range of 2,000 per fly (S6B Fig) [72–74]. Significantly fewer CFUs were recovered from PGRP-LBΔ mutants, while both crqko and RelE20 extracts showed a 10-fold increase. Double mutants for crqko and RelE20 had 50-fold more CFUs than controls (S6B Fig). Altogether, these results demonstrate that Crq and the Imd pathway act in parallel and are required for the management of environmental microbes.
Elevated levels of Upd3 are associated with midgut hyperplasia in aging flies [72,75]. In addition, loss of gut barrier integrity leads to early death in a microbiota-dependent manner [76,77]. Because 8-day-old crqko flies expressed high levels of upd3, we asked whether they also displayed premature gut hyperplasia by looking at the number of mitotic PH3-positive intestinal stem cells of their midgut. While PXH87 and crqko flies did not show any signs of midgut hyperplasia at day 7, midguts of 16 day-old crqko flies had a 2-fold increase in PH3-positive cells compared to that of similarly aged controls (p = 0.0109) (Fig 6G). This phenotype was completely rescued in crqko; crq-Gal4>UAS-crq flies (Fig 6G). The double mutants for crqko and PGRP-LBΔ or for crqko and RelE20 showed even higher levels of intestinal stem cell proliferation than controls (p = 0.03) and did so more prematurely (already in 7-day-old flies) (Fig 6G). The premature increase in midgut stem cell proliferation was partially dependent on Upd3, as upd3;crqko double mutant flies had significantly less mitotic cells (p = 0.04) and lived longer than crqko flies (p<0.0001) (Fig 6H and 6I). However, the lifespan of upd3;crqko double mutants flies was still shorter than that of PXH87 flies (p<0.0001), suggesting that additional mechanisms play a role in the shortened lifespan of crqko flies. We further asked whether crq is required in hemocytes to maintain intestinal homeostasis. Hemocyte-specific re-expression of crq led to a strong rescue of lifespan compared to crqko flies (p<0.0001) but not to the levels of PXH87 flies (p = 0.0462) and to a partial rescue of midgut hyperplasia in 16-day-old flies (p = 0.003 for crqko vs rescue and p = 0.0123 for rescue vs PXH87) (Fig 6H and 6I). Altogether, these results indicate that flies lacking crq display chronically elevated expression of upd3 that triggers early midgut hyperplasia and promotes premature death.
Our study shows that Crq is required for the engulfment of microbes by plasmatocytes and their clearance, and that the mild immune deficiency due to crq mutation is associated with increased susceptibility to infection, defects in immune homeostasis, gut hyperplasia, and decreased lifespan (S7 Fig). We have also re-confirmed a role for crq in apoptotic cell clearance, although the phagocytosis defect of crqko plasmatocytes is less severe than what had been previously observed with two lethal crq deficiency mutants, Df(2L)al and Df(2L)XW88 [35]. A possible explanation is that these deficiencies may have deleted at least one other gene required for apoptotic cell clearance. Additionally, morphological defects associated with secondary mutations could have exacerbated the crq phagocytosis defect by preventing efficient plasmatocyte migration to apoptotic cells. These same deficiency mutants had been assessed qualitatively for phagocytosis of bacteria by injecting embryos with E. coli or S. aureus; their plasmatocytes had no obvious defect in their ability to engulf these bacteria [35]. However, a role for crq in phagocytosis of S. aureus, but not that of E. coli, was subsequently proposed based on S2 cell phagocytosis assays following knock-down of crq by RNAi [41]. Here, we show that crq is required in vivo for uptake and phagosome maturation of both S. aureus and E. coli. A simple explanation of this discrepancy with E. coli could be that knocking down crq by RNAi is not sufficient to affect its role in E. coli phagocytosis (but sufficient to affect its role in S. aureus phagocytosis), and that completely abrogating crq expression by in vivo knock-out leads to a stronger phenotype with both bacteria. Our in vivo data in crqko flies further demonstrate that crq is required to resist multiple microbial infections, such as Ecc15, E. faecalis, B. bassiana, and C. albicans. These data therefore argue that crq plays a more general role in microbial phagocytosis than was previously anticipated. Our previous experiments to test whether crq is required for bacterial phagocytosis in embryos were qualitative rather than quantitative, and did not allow us to identify a role for crq at that stage [53]. In contrast, the experiments we now report in adult crqko flies are quantitative and allowed us to identify a delay in phagocytosis, followed by a defect in bacterial clearance in crqko hemocytes. A possible explanation for this discrepancy would be that hemocytes may differ in their expression profile, behavior, and phagocytic ability at various developmental stages due to differences in their microenvironment and/or sensitivity to stimuli. Accordingly, it has recently been shown that the phagocytic activity of embryonic hemocytes acts as a priming mechanism, increasing the ability of primed cells to phagocytose bacteria at later stages [78]. It is therefore possible that embryonic, larval and adult hemocytes display very different levels of priming and bacterial phagocytic activity, and that crq is required mostly in larval/adult bacterial phagocytosis. Alternatively, a potential defect in phagocytosis of bacteria by embryonic hemocytes of the crq deficiencies may have been suppressed by the deletion of (an)other gene(s) in that genomic region.
Because the immune competence of hemocytes varies during development [50,79,80], we were prompted to re-examine the potential role for crq in innate immunity by knocking it out. Here, we show that Crq is a major plasmatocyte marker at all developmental stages of the fly. We have found that crqko flies are homozygous viable, but short-lived, and can hardly be maintained as a homozygous stock in a non-sterile environment; crqko pupae become susceptible to environmental bacteria and their microbiota during pupariation. In a recent study, Arefin and colleagues induced the pro-apoptotic genes hid or Grim in plamatocytes and crystal cells using the hml-gal4 driver (Hml-apo) and observed a similar pupal lethality, but also associated with an induction of lamellocyte differentiation, and the apparition of melanotic tumors of hemocyte origin [81]. The authors therefore concluded that the death of hemocytes triggered lamellocyte accumulation and melanotic tumor phenotypes [81]. In contrast, we did not observe any obvious melanotic tumors in crqko flies, despite observing a loss of hemocytes in aging crqko flies (Fig 6C) and crqko flies subjected to Ecc15 infection (Fig 6D). One possible explanation is that hemocytes do not die of apoptosis in crqko flies, but of a distinct mechanism. Alternatively, crq mutation could affect more hemocytes than Hml-apo flies, as crq is expressed in all plasmatocytes, while Hml is only expressed in 72.4% of all plasmatocytes expressing crq (from Fig 1C). Thus the 27.6% of non-Hml plasmatocytes (thus non induced for apoptosis, which is hml-Gal4 dependent [81]) may respond to the death of the other plasmatocytes by inducing a signal that triggers the induction of lamellocytes and the subsequent formation of melanotic tumors. Considering the role of crq in apoptotic cell clearance, this signal may require a functional crq, which could explain why crqko flies do not develop melanotic tumors. Strikingly, in the Arefin study, as well as in previous studies, targeted ablation of plasmatocytes also made resulting ‘hemoless’ pupae more susceptible to environmental microbes [23,24,81]. Extensive tissue remodeling takes place at pupariation, and plasmatocytes are essential to remove dying cells, debris, and bacteria. Thus, it was argued that this increased susceptibility was likely due to environmental bacteria invading the body cavity after disruption of the gut [82]. In addition, it was found that the gut microbiome of Hml-apo flies could influence pupal lethality, as the eclosure rate of Hml-apo flies varied depending on the quality of the food they were reared on [81]. Accordingly, our rescue of the crqko pupal lethality with antibiotics demonstrates that their premature aging and death are indeed due to infection by normally innocuous environmental bacteria. Altogether, these data suggest that phagocytes and crq are important actors regulating the interaction between a host and its microbiome.
Hosts use both resistance and tolerance mechanisms to withstand infection and survive a specific dose of microbes [65,83]. crqko flies exhibit a shorter lifespan when compared to control flies, but they are equally tolerant to aseptic wounds and infections. The crqko flies are less resistant to infection, as crq is required to promote efficient microbial phagocytosis. crqko plasmatocytes can still engulf bacteria, albeit at a lower efficiency than their controls. Our data also demonstrate that crq plays a major role in phagosome maturation during bacterial clearance. This is in agreement with a recent study showing that crq promotes phagosome maturation during the clearance of neuronal debris by epithelial cells [36]. Thus, crq is required at several stages of phagocytosis. Similar observations have been made for the C. elegans Ced-1 receptor and for Drpr, as both promote engulfment of apoptotic corpses and their degradation in mature phagosomes [84,85].
‘Hemoless’, Hml-apo and crqko flies are all more susceptible to environmental microbes and their microbiota. While it is not known whether mutants of eater, which encodes a phagocytic receptor for bacteria but does not play a role in phagosome maturation, are more susceptible to environmental microbes during pupariation, both eater mutants and ‘hemoless’ flies showed either decreased or unaffected systemic responses [23,24,26]. Hml-apo larvae however, showed an upregulation in Toll-dependent constitutive Drs mRNA levels whereas Dpt expression was suppressed [81]. In contrast, crqko flies showed no significant difference in constitutive or infection induced expression of Drs, but showed an increased expression of Dpt with age, and infection induced an increased and chronic expression of Dpt. Altogether our results argue that phagosome maturation defects in crqko flies lead to persistence of bacteria and thus to an increased and persistent systemic immune response via the Imd pathway. Such defects in phagosome maturation are not present in hemocyte ablation experiments, which could explain different outcomes for the host immunity and survival.
We have found that Crq acts in parallel to the Toll and Imd pathways. In the mealworm Tenebrio molitor, hemocytes and cytotoxic enzymatic cascades eliminate most bacteria early during infection, and AMPs are required to eliminate persisting bacteria [86]. These data suggest that AMPs act in parallel with hemocytes to fight infections. We have also found that crqko flies are more susceptible to infection with S. aureus than wild-type and Toll pathway-deficient flies. These results are consistent with S. aureus infection being mainly resolved via phagocytosis and Crq having a major role in this process. Surprisingly, we have observed the opposite for infection with other Gram-negative or positive bacteria and fungi. Drosophila mutants for AMP production were more susceptible to infection than crqko flies, arguing that AMPs are critical to eliminate the bulk of pathogens. Indeed, crq (thus phagocytosis) is not essential for Ecc15 elimination, but accelerates bacterial clearance. Our results also suggest that the defects in phagosome maturation may allow some bacteria to persist and grow within hemocytes, where they are hidden from systemic AMPs. Thus, depending on the microbe, humoral and cellular immune responses can act at distinct stages of infection. In this context, phagocytosis acts as a main defense mechanism against pathogens that may escape AMPs or modulate their production.
Chronic activation of immune pathways can be detrimental to organismal health [13–15]. In Drosophila, multiple negative regulators of the Imd pathway, including PGRP-LB, act in concert to maintain immune homeostasis [14–16]. We have observed that crqko flies sustain high production levels of the AMP Dpt and the cytokine Upd3, demonstrating that defects in phagocyte function can lead to chronic immune activation. Notably, the level of Dpt expression induced by activation of the Imd pathway in unchallenged conditions is stronger in crqko flies than was previously observed in mutants of three negative regulators of the Imd pathway, namely pirkEY, PGRP-SCΔ, and PGRP-LBΔ [15], and over 1,000-fold higher in PGRP-LBΔ, crqko double mutants. This is despite the persistence of only a few hundred bacteria in these mutants. This phenotype may be due solely to the accumulation of these persistent bacteria, or Crq may also function in plasmatocytes to remove immunostimulatory molecules from the hemolymph. Nonetheless, our study shows that plasmatocytes, Crq, and phagocytosis are all key factors in the immune response, and that losing crq induces a state of chronic immune induction.
The ability of a host to control microbes decreases with age, a phenomenon called immune senescence [71]. The causes of immune senescence remain elusive, but the loss of immune cells with age and a decline in their ability to phagocytose have been suggested [56,57]. Recent studies have argued that microbial dysbiosis and disruption in gut homeostasis contribute to early aging [76,77,87]. In addition, persistent activation of the JAK-STAT pathway in the gut has been linked to age-related decline in gut structure and function [88]. Aging crqko flies lose a greater number of hemocytes than wild-type flies after infection, which may be the result of accumulating bacteria in these hemocytes in which phagosomes fail to mature. The premature death of crqko flies could be partially rescued by the presence of antibiotics. This demonstrates that phagocytosis, and phagosome maturation in particular, plays a crucial role in managing the response to environmental microbes and potentially, the gut microbiota directly to promote normal aging. We have also found that chronic upd3 expression in crqko flies triggers premature midgut hyperplasia, which is known to alter host physiology and promote premature aging [72,76,89]. It has recently been proposed that plasmatocytes can influence gut homeostasis by secreting dpp ligands and modulating stem cell activity [90]. Our results reinforce the possibility of an interaction between plasmatocyte function and gut homeostasis, and suggests that cytokines derived from hemocytes can trigger cell responses in the gut. These results are also in agreement with a recent publication showing that Upd3 from hemocytes can trigger intestinal stem cell proliferation [69]. Altogether, these results demonstrate that the interaction between hemocytes and the gut tissue are central to host health, and our data demonstrate that phagocytic defects can be associated with chronic gut inflammation and aberrant intestinal stem cell turn-over. As gut aging and barrier integrity are in turn important to maintain bodily immune homeostasis [76], we propose the following model: in crqko flies, plasmatocyte-derived cytokines accelerate gut aging promoting loss of gut homeostasis and microbial dysbiosis, with immune and plasmatocyte activation acting in a positive feedback loop (S7 Fig).
Collectively, our data show that Crq is essential in development and aging to protect against environmental microbes. Interestingly, the impact of mutating crq on host physiology is strikingly different from previously reported phagocytic receptor mutations. We speculate that this could be due to its dual role in uptake and phagosome maturation during phagocytosis. Crq is required for microbial elimination in parallel to the Toll and Imd pathways and acts to maintain immune homeostasis. This situation is surprisingly reminiscent of inflammatory disorders, such as Crohn’s disease, that result from primary defects in bacterial elimination and trigger chronic immune activation and disruption of gut homeostasis. Further characterization of the crq mutation in Drosophila will provide an interesting conceptual framework to understand auto-inflammatory diseases and their repercussions on immune homeostasis and host health.
All stocks were raised at 22°C on standard medium, unless otherwise specified. RelE20, spzrm7, and PGRP-LBΔ stocks were described in [15,61,91]. The crqko stock was generated by homologous recombination, which removed the majority of the crq open reading frame [36] and (S1B Fig).
For bacterial infections, males or females were pricked in the thorax with a needle previously dipped in a concentrated pellet of the tested pathogen. The following bacterial or yeast strains were used at the indicated optical density (OD) taken at 600 nm: Ecc15 (OD = 200), E. coli (OD = 200 and OD = 10), E. faecalis (OD = 5), S. aureus (OD = 0.5), C. albicans (OD = 200). For B. bassiana infection, flies were shaken in a petri dish with mature germinating Beauveria for spore coating. All infections and aging experiments were performed at 29°C. In antibiotic treatments, a cocktail of kanamycin, ampicillin, rifampicin, streptomycin, and spectinomycin (5mg/mL each) was added to the fly medium. Axenic stocks were generated as described in [72,73]. Survival experiments represent at least 3 independent repeats with 20 flies (60–100 flies tested). Survival was analyzed by a Log-rank test using the statistical programs R and Prism.
Flies were individually homogenized in 500 μl of sterile PBS using bead beating with a tissue homogenizer (OPS Diagnostics). Dilutions of the homogenate were plated onto LB agar or MRS agar with a WASP II autoplate spiral plater (Microbiology International), incubated at 29°C, and the CFUs counted. Results were analyzed using a Krustal-Wallis test in R.
Flies were injected in their thorax with 69nl of pHrodo red or Alexa 488 bacteria (Life Technologies Inc.) using a nanoject injector (Drummond). The fluorescence within the abdomen of the flies was then imaged at 45min, 3hrs, and 5hrs post-injection with a Leica MZFLIII fluorescent microscope and DFC300 FX camera and quantified using Image J 2.0.0-rc-30/1.49s (NIH).
For ex vivo imaging, flies were injected with 46nl of PBS at 45min, 3hrs and 5hrs after infection to release all hemocytes, and 10 flies were bled on a lysine-coated slide by mechanically scraping their hemocytes onto a drop of PBS. Once settled for 10min on the slide, hemocytes were quickly dried and mounted with AF1 mounting solution (Citifluor Ltd). Slides were automatically scanned using a Zeiss LSM 700 confocal microscope, and the number of plasmatocytes and average fluorescence signal per plasmatocyte quantified.
For immunostaining, flies were bled as described above and the hemocytes fixed in a solution of PBS, Tween 0.1%, PFA 4% for 30min. The samples were incubated in PBT with 1% normal goat serum and Crq [53] and GFP antibodies (Roche) at 1:500 overnight at 4°C. Samples were washed at RT three times for 5min in PBS, incubated with the appropriate secondary antibodies at 1:1000 in PBT for 2hrs at RT, and washed three additional times in PBT. Samples were imaged with a Zeiss LSM 700 confocal microscope.
Total RNA was extracted from pools of 20 flies per time point using TRIzol (Invitrogen). RNA was reverse-transcribed using Superscript II (Invitrogen), and the qPCR was performed using SYBR green (Quanta) in a Biorad instrument. Data represent the ratio or relative ratio (in %) of mRNA levels of the target gene (crq, Dpt, Drs or upd3) and that of a reference gene (RpL32 also known as rp49). The primer sequences used in this study are provided in the supplementary material. All experiments were performed at least 3 times.
|
10.1371/journal.pgen.1003205 | Next-Generation Sequencing Identifies the Danforth's Short Tail Mouse Mutation as a Retrotransposon Insertion Affecting Ptf1a Expression | The semidominant Danforth's short tail (Sd) mutation arose spontaneously in the 1920s. The homozygous Sd phenotype includes severe malformations of the axial skeleton with an absent tail, kidney agenesis, anal atresia, and persistent cloaca. The Sd mutant phenotype mirrors features seen in human caudal malformation syndromes including urorectal septum malformation, caudal regression, VACTERL association, and persistent cloaca. The Sd mutation was previously mapped to a 0.9 cM region on mouse chromosome 2qA3. We performed Sanger sequencing of exons and intron/exon boundaries mapping to the Sd critical region and did not identify any mutations. We then performed DNA enrichment/capture followed by next-generation sequencing (NGS) of the critical genomic region. Standard bioinformatic analysis of paired-end sequence data did not reveal any causative mutations. Interrogation of reads that had been discarded because only a single end mapped correctly to the Sd locus identified an early transposon (ETn) retroviral insertion at the Sd locus, located 12.5 kb upstream of the Ptf1a gene. We show that Ptf1a expression is significantly upregulated in Sd mutant embryos at E9.5. The identification of the Sd mutation will lead to improved understanding of the developmental pathways that are misregulated in human caudal malformation syndromes.
| Birth defects are the leading cause of infant mortality in the United States, accounting for 1 in 5 infant deaths annually. Birth defects that affect development of the caudal portion of the embryo can include malformations of the spine, such as spina bifida, and malformations of the kidneys and lower gastrointestinal tract. Little is known regarding the genetic causes of human caudal birth defects. The Danforth's short tail (Sd) mouse shares many similarities with these caudal birth defects that occur in humans. In this manuscript, we used next-generation sequencing to identify the genetic cause of the Sd mouse phenotype. We found that the Sd mutation is a retrotransposon insertion that inappropriately turns on a nearby gene that is normally important for pancreas development. Future studies of Sd mice will help us understand the pathogenesis of caudal birth defects in humans.
| The Danforth's short tail (Sd) mouse mutation arose spontaneously in the early 1920s in an inbred mouse colony maintained in the laboratory of C.H. Danforth [1], [2]. The inbred line in which the Sd mutation arose was being maintained for study of a dominant but incompletely penetrant posterior duplication phenotype. Danforth shared four mice with shortened tails (2 males and 2 females) with L.C. Dunn and S. Gluecksohn-Schoenheimer. None of the offspring of these shared mice displayed the posterior duplication phenotype, indicating that Danforth's original line was segregating two different mutations [1], [2]. The new short tailed line was named short-Danforth or Sd. Dunn et al. determined that the Sd mutation was inherited in a semi-dominant manner with complete penetrance and was not allelic to Brachyury (T) [2].
The Sd mutation causes severe defects in development of the axial skeleton, urogenital, and gastrointestinal systems [3]. Homozygous mutant mice (Sdsd/sd, to be denoted herein as Sd/Sd) are born live with viability equal to their littermates, although death occurs within 24 hours of birth [2], [3], [4]. The phenotype of homozygous mutants includes complete lack of tail development with a shortened spinal column due to absence of caudal and sacral vertebrae. The kidneys are completely absent, though occasionally a single small kidney rudiment can be identified medially. The bladder is present, but there is no urethral opening. The cloaca persists due to lack of proper growth of the urorectal septum to divide the cloaca into the primary urogenital sinus (ventrally) and the rectum (dorsally). This septation defect also results in a blind ending intestine and anal atresia.
Heterozygous animals (Sdsd/+, denoted herein as Sd/+) are less severely affected than homozygotes, and survive into adulthood [2], [3]. Adult heterozygotes are fertile, but fecundity is reduced. The heterozygous phenotype is 100% penetrant and includes caudal and sacral skeletal malformations and the characteristic extremely shortened tail. Both urogenital and anal openings are present. The kidneys are variably affected and are generally smaller in size (unilaterally or bilaterally), and unilateral kidney agenesis is not uncommon [4]. However, histologically the kidney tissue is normal. Despite this knowledge, it has been hypothesized that the kidney defects are the cause of reduced lifespan in Sd/+ mice [2].
During embryogenesis Sd/Sd mutants develop normally through approximately embryonic day (E)10.0–10.5, when the first external manifestation of the phenotype is shortening of the tail [4], [5], [6]. Histologically the phenotype first observed in Sd/Sd embryos at E9.5 is disintegration of the notochord [4], [6], [7], and no new notochord is formed caudally from this point [6], [7]. Additionally, the chordal cells remain abnormally close to the neural tube. By E14 the notochord has completely disintegrated except for a few fragments remaining in the sacral region. However, any floor plate that was properly formed remains. The notochord degeneration is equally severe in both homozygous mutant and heterozygous embryos, but it starts slightly earlier in Sd/Sd animals (E9.5) than in Sd/+ embryos (E10.5) [6], [7].
The phenotype of Sd mice resembles several human caudal developmental disorders characterized by malformations of the spine, lower gastrointestinal, and urogenital systems. These include caudal regression syndrome (CRS, OMIM #600145), urorectal septal malformation sequence (URSMS, also known as persistent cloaca), Currarino syndrome (OMIM #176450), and VACTERL (Vertebral-Anal-Cardiac-Tracheo-Esophageal fistula-Renal-Limb anomalies) association (OMIM #192350). Currarino syndrome, characterized by the triad of partial sacral agenesis, presacral mass, and anorectal malformation, is caused by mutations in the MNX1 (HLXB9) gene in 50% of sporadic and 90% of familial cases with the classic phenotype [8], [9], [10]. In addition, private mutations in VANGL1 [11], ZIC3 [12], HOXD13 [13], and PTEN [14] have been described in single cases of caudal dysgenesis and/or VACTERL phenotypes. However, the developmental mechanisms that lead to caudal malformations in humans are still largely unknown. Because of the significant overlap in phenotype, Sd mice are an ideal model to improve our understanding of the genetic and pathophysiologic mechanisms that lead to human caudal malformations.
Although the Sd mutation arose over 90 years ago and the mutation has been genetically mapped to a small region on mouse chromosome 2 [15], the specific genetic lesion has not yet been identified. Here, we report the identification of the Danforth's short tail mutation. We used the flanking markers from the previously published genetic map to delineate the corresponding physical map of the Sd critical region. Direct DNA sequencing of all the exons and exon/intron boundaries of the positional candidate genes and expressed sequence tags (ESTs) did not reveal any mutations. Since direct sequencing of the exonic DNA only provided ∼1% coverage of the Sd critical region, we performed next-generation sequencing (NGS) of the entire Sd genomic interval. Standard bioinformatic analysis of our NGS data did not reveal any causative mutations; therefore, we interrogated reads for which only a single end of a paired-end sequence mapped correctly to the Sd locus. Using this innovative technique, we identified an early transposon (ETn) retroviral-like insertion at the Sd locus. Expression analysis at E9.5 revealed that the Sd ETn causes inappropriate expression of the Ptf1a gene at this developmental timepoint.
We examined the Sd mutation on the recombinant inbred RSV/LeJ mouse line, and have confirmed the previously published phenotype of Sd/+ and Sd/Sd mice (Figure 1). The Danforth's short tail mouse phenotype is characterized by significant anomalies of the urogenital, digestive, and skeletal systems, and is 100% penetrant. Using backcrossing methods, Alfred et al. previously mapped the Sd critical region to an ∼0.9 cM region on mouse chromosome 2qA3 [15]. Two groups of flanking markers defined the proximal and distal borders; however, the mapping was not able to distinguish between the individual markers in each group. We used the flanking marker groups from the published genetic map to identify the corresponding Sd critical region on the mouse physical map. Using the UCSC Genome browser (July 2007; NCBI37/mm9) we defined the Sd critical region as 1.5 megabase pairs (Mb) spanning Chr2:18,901,614–20,456,182 flanked by the genetic markers D2Mit362 and D2Mit364 (Figure 2). The critical region contains 9 annotated RefSeq genes, representing both coding and non-coding genes. In addition, we identified 3 candidate expressed sequence tags (ESTs) not represented by RefSeq genes in the critical region (Figure 2; Table 1).
Using template Sd/Sd DNA from the RSV/LeJ line we individually sequenced a total of 87 annotated exons and intron/exon boundaries mapping to the Sd region. We compared sequences from the PCR products to the C57BL/6J (B6) mouse reference genome sequence and identified only 6 DNA changes. Among these changes, we identified a novel microsatellite in intron 9 of the 4921504E06Rik gene (IVS9+76_77DupTATC). Homozygous Sd mice have 16 copies of a TATC repeat, while Sd/+ mice have one allele with 16 copies and one with 15 copies, and wildtype mice from the RSV/LeJ line and the B6 reference genome have 15 copies. This microsatellite segregated with the mutant phenotype in all mice/embryos tested (26/26). Although linked to the Sd mutation and serving as a genetic marker for the Sd mutation, the microsatellite itself is unlikely to be mutagenic since it does not change any coding sequence or uncover a cryptic splice site. In addition, we tested 11 mouse lines and found that the A/J, CBA, CD-1, and DBA mouse lines also carry 16 copies of the TATC repeat, and the CAST/EiJ line has 18 copies, indicating that this TATC repeat is a polymorphic microsatellite (data not shown).
Our exon sequence analysis of homozygous mutant Sd DNA from the RSV/LeJ recombinant inbred line also identified 5 SNPs in exon 2 of the Etl4 gene. Three of these changes were represented in the SNP database (dbSNP: rs33065171, rs33061041, rs32925809), and two changes were novel (c.207A>G and c.210G>A). These 5 SNPs were homozygous in all mice from the RSV/LeJ recombinant inbred line regardless of phenotype (Sd/Sd, Sd/+, and +/+). Thus, no mutagenic DNA changes were identified in direct sequencing of the annotated exons mapping to the Sd critical region.
Since PCR and direct DNA sequencing of all known exons mapping to the critical region failed to identify the Sd mutation, we employed next-generation sequencing (NGS) technology for mutation detection. DNA capture/enrichment allowed us to focus only on the Sd genomic interval on mouse chromosome 2. An oligonucleotide based array was designed to capture the DNA mapping between bases 18,883,257 and 20,481,194 on mouse chromosome 2. Repeat masking software was employed to ensure only unique DNA sequences would be present on the array. Total genomic DNA isolated from a phenotypically heterozygous (Sd/+) neonatal mouse was hybridized to the capture array.
Prior to performing the NGS reaction we confirmed success of the DNA capture by quantitative real time PCR (qRT-PCR) experiments (Figure S1). Using two sets of primer pairs, one mapping within the captured region (to Etl4, exon 2), and the other mapping outside the captured region (to mouse β-Actin) we demonstrated an ∼20 fold increase in captured DNA, indicating a successful enrichment of the Sd interval genomic DNA sequences.
We performed 36 base pair paired-end sequencing using one lane on an Illumina Genome Analyzer IIx. The sequencing reaction produced 1.8 Gb of DNA sequence. Assembly and alignment of our NGS data was performed using the Bowtie software program [16]. We used the B6 genome as a template for the alignment and analysis of the NGS data. 93% of the obtained sequence reads mapped to the captured region on mouse chromosome 2 with both paired-ends mapping to the proper position and in the proper orientation. The sequence reads obtained cover 811,821 bases (51%) of the 1.5 Mb Sd critical interval as defined by Alfred et al. [15], and the average coverage of sequence reads across the interval was 443×. The Sd critical region is highly repetitive, and ∼49% of the region is made up of LINE, SINE and other highly conserved repetitive DNA sequences. In order for optimal capture and sequence alignment these DNA sequences were excluded from our capture array and are not fully represented in our data. Since we used heterozygous DNA as a template for sequencing, we set an initial variant threshold for our mutation analysis to be 40% of reads (meaning any change from reference showing up at least 40% of the time would trigger a heterozygous call). In the initial data analysis we did not identify any point mutations or small insertions or deletions. We increased the heterozygous call ratio to 25%/75% (where if 25% of reads differed from the reference a heterozygous call would be denoted) and still did not identify any mutations.
We hypothesized that the Sd mutation might be due to insertion of a retrotransposon. Retrotransposon insertions are frequently mutagenic in mice and are estimated to be the cause of 10% of spontaneous mouse mutations [17], [18]. In order to search for novel retrotransposon insertions, we sought to align separately individual ends of each pair from our NGS data. Reads for which one of the paired-ends mapped to the Sd critical region and the other end failed to align to unique (non-repeat masked) DNA were filtered to remove failed reads and low complexity sequences and screened for repetitive elements. Although our capture array was designed to exclude repetitive DNA, we expected some carry over, especially of paired-end reads where one end of the fragment mapped to unique DNA. New Perl scripts were written (available from the authors) to accomplish the analysis. Using this method we discovered a region in which one end of 166 independent paired-end reads mapped in the proper orientation to a consistent chromosome 2 position, and the other end of the pair mapped to the 5′ end of an ETn specific retrotransposon long terminal repeat (LTR) sequence. In addition, since the DNA sample sequenced came from a heterozygous mouse, we were also able to identify 1127 reads from the wildtype allele at the same location showing no variation from the reference genome. There are a larger proportion of reads representing the wildtype allele likely due to reduced capture efficiency from the mutant allele since repetitive sequences were excluded from the capture array. The method presented here provides a way to utilize “unmapped” reads which would normally be discarded. These unmapped data are important to understand the complexity of the genome, and the variation in individuals or strains from a reference sequence.
To confirm the presence of an ETn insertion at the Sd locus we performed both Southern analysis and long range PCR (Figure 3A). EcoRI and BamHI digested DNA isolated from Sd/Sd, Sd/+, and +/+ mice was hybridized with an ∼1000 bp probe from non-repetitive DNA flanking the location of the ETn insertion. The probe hybridized to a 7,300 bp fragment in wildtype DNA digested with EcoRI, which was predicted from the mouse genome reference sequence. However, hybridization of the probe to EcoRI digested Sd/Sd DNA resulted in detection of two bands of ∼4,000 bp and ∼6,000 bp, and hybridization to heterozygous DNA showed all three bands (Figure 3A). Similarly in BamHI digested DNA from a +/+ mouse, an expected band of 3,500 bp was detected. In an Sd/Sd mutant a single fragment of ∼5,000 bp was identified, and both BamHI fragments were detected in DNA from the Sd/+ mouse (Figure 3A).
To further confirm the presence of an ETn insertion at the Sd locus, long range PCR was utilized with PCR primers designed flanking the insertion site. Using wildtype genomic DNA as a template, the primers amplified a product of 253 bp as expected (Figure 3A). When Sd/Sd DNA was used as a template, the 253 bp product was absent, and instead a product of ∼9,000 bp was amplified. In heterozygous DNA, both fragments were amplified (Figure 3A). These data, in conjunction with the Southern analysis, confirmed insertion of a large piece of DNA at the Sd locus. The long range PCR product was cloned and sequenced using primer walking. The Sd insertion was identified as an early transposon (ETn) at mouse chromosome 2 position 19,355,026. There was no loss of wildtype DNA at the insertion site. However, the insertion resulted in a 6 base pair terminal duplication sequence (TSD) flanking the ETn. TSD sequences are a hallmark of retroviral insertion. Sequencing data showed that the long terminal repeats (LTRs) of the ETn were 847 base pairs in length and 100% identical to each other. The internal sequence of the ETn is 6,834 bp in length and does not contain any open reading frames. The total length of the Sd ETn is 8,528 bp (GenBank Accession JX863104).
Rigorous comparison of our next-generation sequencing data across the Sd critical region to the reference B6 genome, and available sequence from the CBA genome (CBA is the last known outcross of the RSV/LeJ line), did not reveal any significant sequence changes. Thus, we were not able to identify the mouse strain on which the Sd mutation originally arose in the 1920s. In order to rule out the possibility that the Sd ETn is a common polymorphism, we assembled a panel of DNA from 10 inbred and 1 outbred mouse strains for analysis. These strains were chosen to include a diverse number of strains spanning many arms of the inbred mouse genealogy chart [19]. We specifically included the A/J line since ETn sequences are reported to be highly active in this strain [17]. We designed a locus specific multiplex PCR and screened the 11 additional mouse strains (Figure 3B). The Sd ETn was not identified in any of the strains, indicating that it is not a common strain polymorphism.
ETn sequences are known to affect gene expression when they insert within a gene and are predicted to affect expression of genes when they land upstream [17]. The Sd ETn insertion is located 12,463 bases upstream of the Ptf1a start codon and within the gene's previously defined 15.6 kb promoter/enhancer region and in the same orientation as the Ptf1a gene (Figure 3C; Figure S2) [20]. Thus, we hypothesized that Ptf1a expression is affected by the ETn insertion. qRT-PCR on RNA isolated from whole embryos at E9.5, when the first manifestation of the Sd phenotype becomes apparent, indicated the Ptf1a gene is upregulated at ∼9 times normal levels in Sd/Sd embryos at this timepoint (Figure 4). This over expression was also seen in Sd/+ embryos where the Ptf1a is expressed ∼5 fold that of wildtype (Figure 4). These data demonstrate a dose-dependent up regulation of Ptf1a expression that correlates with the number of mutant alleles and the severity of the Sd phenotype.
RNA expression levels of other genes mapping to the Sd critical interval, including Pip4k2a, Armc3, Otud1, and Etl4 were equivalent in Sd/Sd, Sd/+, and +/+ embryos (Figure 4). The Msrb2 gene showed an ∼2 fold change that was not statistically significant (Figure 4). This increased expression change in Msrb2 is likely due to our small sample size (n = 3/genotype). We were unable to amplify message from the 4921504E06Rik gene at this timepoint. In addition, we identified an EST (BQ085140) that is not represented in the RefSeq data located 2 kb from the 5′ end of the ETn. BQ085140 is the closest expressed sequence to the ETn, however, its orientation is opposite that of the retrotransposon. We were unable to amplify BQ085140 from RNA isolated from E9.5 embryos of any genotype. EST AK053418 maps between the ETn insertion site and the Ptf1a gene, 8 kb from the 3′ end of the ETn and also in the opposite orientation. We were unable to amplify AK053418 from RNA isolated at this timepoint in either mutant or wildtype RNA. These data indicate that the two closest expressed sequences are not expressed at E9.5, and are not affected by insertion of the ETn at this timepoint even though Ptf1a is overexpressed. We did not test the expression of AK076779, 4930436L09Rik, and Gm3230 based on their distance from the ETn and lack of open reading frames.
To determine whether the ETn was acting as a promoter/enhancer or resulting in a dominant negative fusion RNA we performed 5′ RACE (rapid amplification of cDNA ends). Using RNA isolated from E9.5 Sd/Sd embryos, our 5′ RACE reactions only amplified the 5′ end of the Ptf1a gene (represented by GenBank Accession 007922). Thus, there is no incorporation of ETn sequence in the RACE products from homozygous mutant RNA. These data indicate that the Sd ETn is acting as an enhancer to strongly augment expression of normal Ptf1a transcripts.
We generated two transgene constructs in an attempt to recapitulate the Sd phenotype (Text S1). In the first experiment the Ptf1a coding sequence was cloned into the pCAGGS vector [21]. The pCAGGS vector contains the cytomegalovirus immediate early enhancer and the chicken β-actin promoter, resulting in ubiquitous expression of cloned sequences in mammalian cells [21], [22]. Among 116 harvested embryos only 6 carried the pCAGGS-mPtf1a transgene, compared with a typical yield of 10 to 20% transgenic offspring [23] (Table S1). Three of the six transgenic embryos did not express Ptf1a ectopically and were phenotypically normal at E16.5. The remaining three transgenic embryos had arrested growth and were dead and in the process of resorbing (Table S1). These data are consistent with a negative effect of Ptf1a overexpression on development.
In the second experiment we used recombineering [24] to create a BAC-based genomic clone containing both the Ptf1a gene and the Sd ETn (Text S1). Using gap repair we sub-cloned a 31 kb DNA fragment containing the Ptf1a gene and the previously reported known promoter, 5′ enhancer, and 3′ control region sequences (Figure S2) [20]. We inserted the ETn into the precise Sd location 12 kb upstream of Ptf1a to create an ETn/Ptf1a-containing genomic transgene. Both the ETn-containing transgene and the control BAC-based transgene lacking the ETn were injected into fertilized mouse eggs. Founder (G0) embryos were dissected between E14.5 and E15.5. No abnormal phenotype was apparent in the transgenic embryos with or without the Sd ETn insertion. However, there were significantly fewer transgenic embryos (3 of 23; p = 0.0086, Fishers exact test) carrying the ETn-containing transgene than the associated control transgene (16 of 32) (Table S1).
Our data reveal that the Sd phenotype is caused by an insertion of an early transposon (ETn) at the Sd locus. The Sd ETn maps upstream of the promoter of the Ptf1a gene causing overexpression of normal Ptf1a transcripts at E9.5. No fusion transcript of the ETn and the Ptf1a gene was detected, indicating that the ETn is functioning as a strong enhancer. Ptf1a (originally termed p48) is a basic-helix-loop helix transcription factor that was first cloned from a rat exocrine-specific pancreatic cell line [25]. Northern blot analysis in adult rat tissues revealed expression limited to the exocrine pancreas [25]. Thus, p48 was given the name “pancreas specific transcription factor 1a.” Northern and PCR-based analysis in developing mouse pancreas initially showed onset of Ptf1a expression at embryonic day 12 (E12) [25]. However, further examination of whole mouse embryo by RT-PCR indicated that Ptf1a expression is highest at E10.5 [26]. Analysis of the expression pattern of Ptf1a using whole mount in situ hybridization indicated that expression of the gene is not restricted to the developing pancreas. High levels of expression were found at E10.5 in the cerebellar primordium and throughout the length of the developing neural tube in addition to the developing pancreatic buds [26]. In humans, recessive mutations of PTF1A cause cerebellar agenesis and permanent neonatal diabetes due to pancreatic agenesis [27].
Ptf1a null mice are born in Mendelian ratios; however, they lack a pancreas and die shortly after birth [28]. The exocrine cells of the developing pancreas fail to form in null embryos, and exocrine specific genes (amylase, etc) are never expressed. However, some endocrine cells do form in null embryos and express endocrine specific genes (insulin, etc.); interestingly, these endocrine cells reside in the spleen. Based on the null phenotype it was thought that Ptf1a was essential for development of exocrine-specific pancreas cells and that the exocrine cells were secondarily crucial for the proper spatial organization of the endocrine pancreas [28]. Fate mapping using a Ptf1a-Cre locus specific knock-in and cre-mediated LacZ expressing mice showed that Ptf1a expression was present in all early pancreatic precursor cells as nearly all acinar, ductal, and islet cells showed a history of Ptf1a expression [29]. Sd mice do not have a reported pancreatic phenotype. Our analysis of Ptf1a overexpression in Sd mice only focused on total RNA isolated from E9.5 embryos, which is ∼1 day prior to pancreatic bud formation. It is unknown whether the ETn insertion influences Ptf1a expression occurring in the pancreatic primordia, or older pancreas.
Sd mice exhibit neuronal patterning abnormalities, and there is a significant decrease in the number of motor neurons in the most caudal portions of the developing embryo [30]. This could be secondary to failure of development of multiple caudal tissues in Sd mice, or could result from up regulation of Ptf1a. Mouse Ptf1a mRNA was injected into one blastomere of two cell stage Xenopus embryos [26]. Presence of the mouse Ptf1a mRNA in the cells of the developing Xenopus embryos suppressed growth in the cells which become interneurons and primary sensory neurons [26]. Injected embryos were not analyzed beyond neural development at early stages, so it is unknown whether overexpression of mouse Ptf1a in Xenopus would have any caudal phenotypic characteristics similar to Sd mice. Overexpression of Ptf1a has not yet been studied in other model organisms. Thus, the Sd mouse is the first report of a phenotype caused by ectopic Ptf1a expression.
Ptf1a is a transcription factor required in several tissue lineages and which interacts with tissue specific co-factors [31]. Ptf1a is part of a heterotrimeric complex called PTF1 [25], [26], [32]. The PTF1 complex includes Ptf1a, RPB-J or RPB-L, and a class A bHLH protein (HEB, E2, or E47) [31]. The RPB genes act as both repressors and co-activators [33]. RPB-J is ubiquitously expressed and responsive to the Notch intracellular domain, while RPB-L is tissue-specific in the developing pancreas, neural tube, and cerebellum, and is Notch-independent [33]. Ectopically increased Ptf1a could act to affect downstream gene expression through heterodimeric binding to RBP-J and/or heterotrimeric binding with RPB-J and a class A bHLH protein, as both heterotrimeric and heterodimeric complexes have the ability to bind DNA [31]. It is also possible that excessive Ptf1a may lead to de-repression of RPB-J regulated genes. However, it is unclear whether this would occur without specific signals from the Notch intracellular domain. It will be interesting to determine whether ectopic overexpression of Ptf1a causes misexpression of downstream genes, possibly through interaction with RBP-J.
Endogenous retrovirus (ERV) sequences containing long terminal repeats (LTRs) make up ∼8–10% of the human and mouse genomes [17]. ERV elements transpose in a copy and paste method via transcribed RNA intermediates. In humans, LTR mediated ERV retrotransposition has not been described and these elements have never been implicated in human disease. However, LTR mediated ERVs are mobile in mice and account for ∼10–15% of all spontaneous mouse mutations [17]. Early transposons (ETn) are a subfamily of the ERVs which also include intracisternal A particles (IAP) and MusD elements [34]. ETn elements range in size from 2 kb to 8 kb, do not contain open reading frames, and are non-autonomous [34], [35]. Rather, retrotransposition of the ETn elements require proteins encoded by the related MusD ERV [34], [36].
The ETn elements are known to be very highly expressed between E3.5 and E7.5 in all cells [37]. This high level of ETn expression likely accounts for the high rate of transposition of these elements in mice. After E7.5, global ETn transcription decreases to less than 5% of that seen in the undifferentiated cells of the early embryo. The expression pattern post-implantation (between E7.5–13.5) becomes more limited to mostly non-differentiated cells [38]. However, it is interesting to note that at E9.5 there is intense expression of ETn sequences in the caudal portion of the neural tube and caudal somites. [38]. It is tempting to hypothesize that this caudally localized ETn driven overexpression of Ptf1a at E9.5 results in the severe caudal defects observed in Sd mutant mice.
Luciferase assays have been used in cultured cells to study the transcriptional activity of LTR sequences contained in ETn elements [39]. Promoter activity can be up to 200-fold higher from the same LTR cloned into undifferentiated compared to differentiated cells [39]. Mutagenic ETn elements published to date have all landed intronically within the genes they affect [17], [18]. These ETn elements cause aberrant splicing of the gene in which the ETn inserted, resulting in premature polyadenylation and protein truncation. The related IAP elements are known to have transposed upstream of genes and act as strong promoters, as in the agouti locus [40]. Similarly, the Dactylaplasia (Dac) mouse is characterized by an incompletely penetrant limb phenotype consisting of missing central digital rays. The Dac mutation has been identified as a MusD element that maps downstream of the Fgf8 gene, and Fgf8 expression is downregulated during limb development in homozygous mutants [41]. The Sd ETn landed within the well-studied promoter/enhancer region of the Ptf1a gene [20]. It is plausible that insertion of the ETn within Ptf1a could usurp endogenous promoter activity until the ETn is methylated, thus suppressing endogenous Ptf1a promoter activity at early times. Our data do suggest that the Sd ETn is acting as a strong enhancer by causing a dose-dependent increase in Ptf1a expression that correlates with the severity of the Sd phenotype.
We attempted to recapitulate the Sd phenotype using two different transgene constructs; one with the ubiquitously expressed strong pCAGGS promoter driving the Ptf1a cDNA, and one BAC-based transgene with insertion of the ETn element upstream of the Ptf1a gene. Neither of these transgenes recapitulated the Sd phenotype. We hypothesize that the pCAGGS-driven expression of Ptf1a was more deleterious than the Sd mutation because of more widespread, and possibly higher level, of Ptf1a expression. The BAC-based ETn-containing transgene could also be expressed in novel spatial and temporal patterns that do not recapitulate the Sd mutation. Our ETn-containing BAC-based transgene may also be lethal since there were significantly fewer embryos carrying the ETn-containing transgene than the control. Although we designed our BAC-based genomic transgene to contain the known upsteam and downstream regulatory elements of the Ptf1a gene [20], it is possible that additional sequences necessary for recapitulation of the Sd phenotype are required and were unknowingly excluded in the design of our experiment (Figure S2). Passage through the germline might also be required for proper methylation of the ETn element to mirror the mutagenic effects of the Sd ETn.
Development of the notochord is severely affected in Sd mutant embryos [4], [7]. The notochord and associated floor plate develop normally through E9.5 in Sd/Sd embroys and E10.5 in Sd/+ embryos. No new notochord is formed caudally from these timepoints forward. The notochord that has developed degenerates completely in both Sd/Sd and Sd/+ embryos by E14 [6], [7]. However, the floorplate that developed alongside the notochord prior to notochord degradation remains intact. Sonic hedgehog (Shh) expression is affected by the lack of notochord, though the floorplate that remains after notochord degeneration still expresses Shh normally [42]. The lack of notochord and floorplate in the most caudal portions of the embryo is a plausible explanation for some aspects of the characteristic Sd phenotype. However, the lack of notochord is the same in both homozygous and heterozygous embryos even though the phenotype is less severe in heterozygous animals, suggesting that simple lack of Shh in the caudal portion of the embryo is unlikely the sole cause of the Sd phenotype. It is more likely that a combination of lack of Shh from the absent notochord as well as misexpression of other genes due to the ectopic overexpression of Ptf1a play a significant role in the phenotypic etiology. Global gene expression analysis of various timepoints in developing Sd mutants will be helpful to characterize these changes.
Mouse models have been invaluable in the study of human development, and the Danforth's short tail mouse is a striking model of human caudal birth defects. The Sd mouse exhibits features seen in caudal regression syndrome (CRS), urorectal septum malformation sequence/persistent cloaca (URSMS), VACTERL association, and Currarino syndrome. Shared phenotypic characteristics include kidney dysgenesis (or agenesis), vertebral anomalies, and anorectal malformations. The overlap of phenotypic characteristics suggests that there are similar pathways or related pathways that are disrupted in these disorders. Interestingly, Currarino syndrome, characterized by the triad of hemisacrum, anorectal malformation and pre-sacral mass, is caused by mutations in the HLXB9 gene (also known as MNX1), which is known to have a role in pancreas and motor neuron development [8], [9]. Recently, it has been reported that Ptf1a is a strong regulator driving Mnx1 expression during pancreatic development [43]. Although we identified a significant increase in Ptf1a in mutant Sd embryos at E9.5, we did not observe a concurrent change in Mnx1 expression in mutant embryos (data not shown). Further analysis of expression patterns may be needed at later timepoints to identify downstream changes due to Ptf1a misexpression. Importantly, mice lacking Mnx1 do not mirror the caudal phenotype seen in human patients with Currarino syndrome [44]. Thus, the genetic pathways leading to caudal dysgenesis via misexpression of Ptf1a and/or Mnx1 may differ between these species.
It is hypothesized that caudal malformation disorders result from failure of induction and proliferation of the caudal mesoderm [45]. In URSMS it is specifically postulated that the phenotype is caused by failure of the meosdermally derived urorectal septum to properly grow and divide the cloaca into the primary urogenital sinus (ventrally) and the rectum (dorsally) [46]. Studies of chimeric Sd mice in which Sd mutant cells carrying the LacZ gene were injected into wildtype blastocysts show specific exclusion of all Sd cells from the dorsal side of the urorectal septum (which overlaps the ventral portion of the hindgut), indicating a failure of this structure to grow in mutant animals [47]. Whether Ptf1a is overexpressed due to the ETn insertion in the developing mesoderm of the urorectal septum in Sd embryos is unknown and will need further investigation. Now that the genetic lesion has been identified, the study of downstream gene expression changes in Sd mice will undoubtedly lead to an improved understanding of human caudal development disorders.
All experiments involving mice have been approved by The University of Michigan University Committee on Use and Care of Animals.
The Sd mutation has been maintained at the Jackson Laboratories on a recombinant inbred strain (RSV/LeJ, stock# 000268) for over 127 generations, and we have established an Sd breeding colony at the University of Michigan. Data presented herein are from mice from the RSV/LeJ line, or from outcrossing the Sd mutation to CD-1 mice (Charles River Laboratories, Wilmington, MA). An outbred/mixed background was maintained by crossing inbred RSV/LeJ mice to CD-1 and harvesting embryos only from intercrossed F1 mice. The Sd phenotype is consistent between the inbred and outbred mice. Mice were housed in environmentally controlled conditions with 14 hour light/10 hour dark cycles with food and water provided ad libitum.
Matings for timed embryo isolation were set up using standard animal husbandry techniques. Noon on the day of vaginal plug observation was considered embryonic day (E) 0.5. DNA for embryo genotyping was isolated from yolk sacs via the HotSHOT extraction method.
Genomic DNA for next-generation sequencing and Southern analysis was isolated using phenol∶chloroform extraction and ethanol precipitation, and DNA was quantified on a NanoDrop spectrophotometer (ThermoFisher, Asheville, NC).
5 µg (at 50 ng/µl) of heterozygous Sd DNA was sheared into 300 bp fragments on a Covaris S-series sonicator (Covaris, Woburn, MA) and was subsequently used to create a total genomic DNA library for next-generation sequencing using standard Illumina protocols. Prior to sequencing, the DNA library was enriched for the Sd critical region (between chromosome 2 bases 18,883,257–20,481,194 on build 37 of the mouse genome (July 2007; mm9) as visualized on the UCSC genome browser (http://www.genome.ucsc.edu) using a 244 K Agilent SureSelect DNA oligo microarray designed utilizing Agilent eArray software (https://earray.chem.agilent.com/). Capture was performed by following the Agilent SureSelect Array protocol version 1. The captured DNA was tested for proper enrichment via quantitative PCR with locus specific primers and then subject to 36 bp paired-end sequencing on an Illumina Genome Analyzer IIx at the University of Michigan DNA Sequencing Core.
Next-generation sequence data were aligned to the corresponding region of chromosome 2 represented on the NCBI37/mm9 build of the mouse genome. Sequences were aligned using the Bowtie and MAQ software applications [16]. Overall assembly of the region was done using the Mosaik software suite and assembled discrepancies were mined using Consed software [48]. Sequences were also aligned using the Bowtie software application. Output files were produced in SAM and BAM format and converted to BED format for visualization on the UCSC Genome Browser. Histogram graphs to analyze coverage (copy number variation and deletions and duplications) were created.
The unmapped reads contain a combination of poor quality sequence and reads which couldn't be mapped to the existing reference sequence. Reasons for not mapping include repetitive and low complexity sequence as well as novel elements. Using the paired-end reads, and a combination of existing tools, and custom methods, we identified read pairs where there were at least 5 sequence reads where one end mapped uniquely to Chromosome 2 and the other end did not map but contained a repetitive element. We compared these repetitive sequences (LTR) with known locations of repeat sequences in the mouse genome to identify novel insertions and their locations.
Multiplex PCR was used for genotyping with primers CNV577 (5′-TTTCCACGGCCATTCTTTAC-3′), CNV578 (5′-GCTCAACCAGAACAATACATTCAG-3′) and CNV580 (5′-GCCAATCAGGAGACTGAAGC-3′). PCR was performed in 20 µl reactions containing 1 µM of each primer, and 1× TaqProComplete (Denville Scientific Inc., Metuchen, NJ) in an Eppendorf Mastercycler (Eppendorf North America, Hauppauge, NY). Cycling conditions were 94°C for 5 minutes, followed by 30 cycles of 94°C for 30 seconds, 58°C for 30 seconds, and 72°C for 45 seconds, and a final extension of 72°C for 10 minutes. PCR products were resolved on a 2%TAE gel stained with ethidium bromide. The resulting wildtype amplimer is 512 bp and the Sd allele amplimer is 405 bp.
6 µg of genomic DNA from wildtype, Sd/+, and Sd/Sd was digested with 4 U/µg of either EcoRI or BamHI (New England Biolabs, Ipswich MA) supplemented with 2.5 mM spermidine overnight at 37°C. Digested DNA was electrophoresed on a 1% TAE gel and transferred to Hybond-N+ nylon membranes as previously described. Digoxigenin labeled probe was created using the Roche PCR DIG probe synthesis kit (Roche Applied Science, Indianapolis, IN) per manufacturer instructions with primers CNV551 (5′-AACCACAGGAAAGGTTGCAG-3′) and CNV552 (5′-TCTGGGTACCAGCTTCAGTG-3′) using mouse genomic DNA as a template. Southern analysis was performed using the Roche DIG Easy Hyb system per manufacturer instructions (Roche Applied Science, Indianapolis, IN).
PCR primers flanking the Sd ETn insertion were designed using Primer3 software (http://frodo.wi.mit.edu/primer3/). The ETn was amplified using the Roche Expand Long Template PCR system (Roche Applied Science, Indianapolis, IN) per manufacturer instructions with primers CNV559 (5′-GAAGCTCTGCAGGCTGAAAGCAAAG-3′) and CNV560 (5′-GAATGAGGACTCTGCCCTTGAGTGG-3′). Amplified PCR product was cloned using the TOPO XL PCR cloning kit (Invitrogen/Life Technologies, Grand Island, NY). Sequencing of the cloned ETn was performed by primer walking (see GenBank Accession JX863104)
Primers for qRT-PCR were identified via PrimerBank (http://pga.mgh.harvard.edu/primerbank/) for Ptf1a, Msrb2, Otud1, Etl4, Pip4k2a, and Armc3. Primer3 was used to design primer pairs for 4921504E06Rik (CNV496 5′-TACCTAGCATGGTGCCTGAAGA-3′/CNV497 5′-GTCATTGCATACTGCCGGTAAA-3′), BQ085140 (CNV788 5′-GTGCTGGACCCAAACATAGC-3′/CNV789 5′-TGGGGAATCAACGAACTCTG-3′), and AK053418 (CNV792 5′-TAAGGGGATGGGAAGGTGTC/CNV793 5′-AGGTGCATCATCATGGCTTC-3′). cDNA template for qRT-PCR was synthesized using the First Strand cDNA synthesis kit from Invitrogen/Life Technologies (Grand Island, NY) from 1 µg RNA isolated from E9.5 Sd/Sd, Sd/+, and +/+ embryos (3 from each genotype) using the RNeasy Micro Kit (Qiagen, Valencia, CA). qRT-PCR reactions were performed in triplicate using 1× Applied Biosystems Power SYBR mix and run on an Applied Biosystems StepOne Plus PCR system (Applied Biosystems/Life Technologies, Grand Island, NY). Cycling conditions were 50°C for 2 minutes, 95°C for 10 minutes, followed by 40 cycles of 95°C for 15 seconds, 57°C for 30 seconds, 72°C for 30 seconds. The cycle threshold for each reaction was automatically calculated using StepOne Software v2.2. Fold change of the test genes was calculated for each genotype by normalization to β-actin using the Pfaffl method [49]. Statistical analysis was performed via Student's T-test with significance p<0.05.
5′ RACE was performed using the FirstChoice RLM-RACE kit (Ambion/Life Technologies, Grand Island, NY) per manufacturer instructions with 1 µg template Sd/Sd RNA as input. RACE products were sequenced at the University of Michigan DNA Sequencing Core.
|
10.1371/journal.pntd.0007139 | Water-induced strong protection against acute exposure to low subzero temperature of adult Aedes albopictus | As an important vector of dengue and Zika, Aedes albopictus has been the fastest spreading invasive mosquitoes in the world over the last 3–4 decades. Cold tolerance is important for survival and expansion of insects. Ae. albopictus adults are generally considered to be cold-intolerant that cannot survive at subzero temperature. However, we found that Ae. albopictus could survive for several hours’ exposure to -9 to -19 oC so long as it was exposed with water. Median lethal time (LT50) of Ae. albopictus exposed to -15 and -19 oC with water increased by more than 100 times compared to those exposed to the same subzero temperature without water. This phenomenon also existed in adult Aedes aegypti and Culex quinquefasciatus. Ae. albopictus female adults which exposed to low subzero temperature at -9 oC with water had similar longevity and reproductive capacity to those of females without cold exposure. Cold exposure after a blood meal also have no detrimental impact on survival capacity of female adult Ae. albopictus compared with those cold exposed without a blood meal. Moreover, our results showed that rapid cold hardening (RCH) was induced in Ae. albopictus during exposing to low subzero temperature with water. Both the RCH and the relative high subzero temperature of water immediate after cold exposure might provide this strong protection against low subzero temperature. The molecular basis of water-induced protection for Ae. albopictus might refer to the increased glycerol during cold exposure, as well as the increased glucose and hsp70 during recovery from cold exposure. Our results suggested that the water-induced strong protection against acute decrease of air temperature for adult mosquitoes might be important for the survival and rapid expansion of Ae. albopictus.
| Aedes albopictus is one of two most important vectors for dengue and zika. During the last 3–4 decades, this mosquito has spread from native Asian area to all continents except Antarctica, becoming the most invasive mosquitoes which imposed extensive public health threat to human beings throughout the world. Cold tolerance is important for distribution and survival of insects. During the expansion of Ae. albopictus, especially a spatial expansion to cooler climate areas, it needs to cope with cold temperatures. Moreover, because of such widespread distribution adult Ae. albopictus will certainly often encounter sudden drops in air temperature even below subzero that often happens in early spring and winter, and late autumn. Thus far, adult Ae. albopictus are generally considered to be cold-intolerant that can not survive at subzero temperature. In this study, we found that water can provide strong protection against low subzero temperature even below -10 oC. Cold exposure of adult female Ae. albopictus to low subzero temperature with water either before or after a blood meal have no detrimental impact on fitness costs of these adult mosquitoes. Considering water is common in nature, our results indicated that during the expansion of Ae. albopictus especially when adult mosquitoes encounter a sudden drop in air temperature water could be a good shelter for cope with such cold temperature below subzero.
| Cold tolerance or cold hardiness, the ability of an insect to survive at low temperature, is important in defining the distribution and survival of insects. There are two different cold hardening in insects at present. One is accomplished by long term (weeks or months) cold acclimatization to overwinter that occurs in an inactive or diapausing stage; the other called rapid cold hardening (RCH) which is accomplished by a brief exposure (minutes or hours) to low temperature that occurs even in feeding and reproductive stages [1–4]. As an efficient ability utilized by insects to survival in environment with rapid and unexpected changes in temperature, RCH has been found in numerous insect species belonging to different orders including Diptera [2, 4–7].
Aedes albopictus is an epidemiologically important vector for several arboviruses such as dengue, yellow fever, zika, and chikungunya. During the last 3–4 decades, Ae. albopictus has spread from native Asian area to all continents except Antarctica, becoming the most invasive mosquitoes which imposed extensive public health threat to human beings throughout the world [8–12]. During the spread of this species, a spatial expansion to cooler climate areas has also been reported and the ability to rapidly produce low temperature phenotypes has been considered as an important factor for the successful establishment in these cooler habitats [13]. Therefore, cold hardiness is a key trait for the distribution of this species and strong resistance to cold temperature provides more chances for efficient invasiveness to colder zones.
Adult Ae. albopictus are generally considered to be cold-intolerant that can not survive at subzero temperature and the only life stage that can cope with this low temperature is its eggs [13, 14]. Thus, the studies on the cold hardiness of this mosquito have been focused on eggs [13–19]. Moreover, the RCH has not been reported in adult and eggs of this species yet. In our current study, we found that adult Ae. albopictus could survive for several hours’ exposure under -10 oC when transferred from room temperature to low subzero temperature with water, which was also found in adult Aedes aegypti and Culex quinquefasciatus. Median lethal time (LT50) of adult mosquitoes was increased by nearly 100 times compared to those transferred directly to subzero temperature lower than -10 oC without water. This indicated that RCH was also existed in adult mosquitoes and that water in nature might provide strong protection for adult mosquitoes against sudden drop in air temperature that often happens in early spring and winter, and late autumn [2, 4, 6, 7]. The aim of this study is to compare the cold hardiness of adult Ae. albopictus exposed to low subzero temperature with water with those exposed directly, analyze the possible molecular mechanism for this cold hardiness and determine the impact of exposure to low subzero temperature with water on fitness costs of adult female Ae. albopictus.
Mosquitoes used in this study including Ae. albopictus, Ae. aegypti and Cx. quinquefasciatus were all established for several years in our laboratory with Guangdong origin. All mosquitoes were reared in a climate-controlled room at 28 ± 1°C and 80 ± 5% relative humidity with a 12:12-hour (light: dark) photoperiod. Adult mosquitoes were provided with 10% glucose solution.
To evaluate the survivorship, 3 groups of 20 adult mosquitoes with 3- to 5-day-old were immobilized by CO2 and then transferred to a disposable 100ml-plastic cup with (treatment) or without (control) dechlorinated tap water (50 mL). A plastic lid was used to cover with cup for preventing escape of mosquitoes during experiment. These adults were then allowed to recover from anaesthetization at room temperature for 1 h. Subsequently, the cups of mosquitoes with water were exposed to low subzero temperatures at -9, -15 and -19 oC for 3 to 8 h, and the cups of mosquitoes without water were exposed to the same temperatures for 3 to 30 min. Twenty adults (one cup) were removed from each temperature at a 1 h interval for cups with water and 1- to 5-min interval for cups without water until 100% mortality were attained. The exposed mosquitoes were then transferred to 650ml-plastic cages with a piece of wet filter paper to keep humidity and maintained at normal climate-controlled room. The mosquito survival was recorded 24 h after cold exposure and survival was defined as the ability of righting themselves and flies [20].
To evaluate the impact of cold exposed to low subzero temperature on fitness costs of female adult Ae. albopictus, 3 groups of 30 adult mosquitoes with 3- to 5-day-old were cold exposed to -9 oC with water for 3 to 5 h as described above. Subsequently, the exposed mosquitoes were transferred to cages, maintained under normal climate-controlled room and provided with 10% glucose solution. Control groups of 90 female mosquitoes without cold exposure were also transferred to cages and maintained under normal condition. Three days after recovery from cold exposure dead mosquitoes were discarded and survived mosquitoes that have starved for 24 h were blood fed on mice for 30 min. The engorged mosquitoes were then counted and transferred to new cages, and were aspirated into individual 50 mL Corning tubes 2 days post-blood meal (PBM) with bottom lining of moist filter paper supported by water-soaked cotton [21]. Two days after oviposition, mosquitoes were aspirated out and killed by cold, eggs were removed out for maturation for 3 days and then counted. After maturation, eggs on filter paper were immersed in 50 mL water in a 100ml-plastic cup and egg hatch rates were determined by counting the number of hatched second instar larvae [22]. Adult lifespan of female Ae. albopictus after cold exposure to -9 oC with water for 3 h were also monitored and compared with control group without cold exposure. Cold treatments were conducted as above and dead mosquitoes were removed at 24 h after recovery. Survived mosquitoes were reared under normal condition and dead mosquitoes were counted and removed daily for a month.
To evaluate the impact of cold exposure on blood-fed Ae. albopictus, engorged mosquitoes were collected and transferred to a new cage. Three groups of 30 adult mosquitoes were selected at 24 and 48 h PBM and exposed to -9 oC with water for 3 h. Mosquitoes without a blood meal were also selected and cold exposed at the meanwhile. Cold exposed mosquitoes were maintained in cages under normal condition and survivals were recorded 24 h after recovery. Two days after blood meal, mosquitoes of cold exposure at 24 h PBM were reared individually, and mosquitoes of cold exposure at 48 h PBM were reared as a pool, egg numbers and egg hatch rates were determined as above.
After transferring water from room temperature into low subzero temperature, the dynamics of water temperature was monitored by a mini-thermometer (Testo 175 H1, Lenzkirch, Germany). The temperature probe was wrapped by plastic film and immersed into 50 mL water in a disposable 100ml-plastic cup and then transferred to different low subzero temperature for 12 h. Temperature was detected and recorded at a 5-min interval.
Adult mosquitoes were cold exposed as above and collected at 0, 1, 2 and 4 h after recovery from cold treatment. Six adults were pooled in a single replicate and five replicate biological assays were performed. These sampled mosquitoes were homogenated in 300 μL distilled deionized water. Two hundred microlitres of the homogenates were used for RNA extraction and another 100 μL were filtered at room temperature through a spin filter (Pall, nanosep 10k Omega, NY) at 12,000 rpm for 15 min. For glycerol analysis, filtered homogenates were diluted 1:100 v/v with distilled deionized water. Next, 10 μL of the diluted homogenates were incubated with 100 μL Master Reaction Mix (Sigma-Aldrich, MAK117, USA) for 20 min at room temperature. Absorbance was measured at 570 nm and glycerol contents were calculated from standard curve. Glucose levels were determined by using the Glucose (GO) Assay Kit (Sigma-Aldrich, GAGO-20, USA) according to the manufacturer’s protocol with minor modifications. The above filtered homogenates were diluted 1:5 v/v with distilled deionized water and 50 μL of these diluted samples were then incubated with 100 μL Assay Reagent for 30 min at 37 oC. After the incubation, 100 μL of 12N H2SO4 were added to stop the reaction. Then, the absorbance was measured at 540 nm and glucose contents were calculated from standard curve.
Total RNA was extracted from samples collected above using TRIzol Reagent (Invitrogen, Carlsbad, CA) and the first strand cDNA was synthesized using HiScript II Q SuperMix for qPCR (+dDNA wiper) (Vazyme, Nanjing, China) following the manufacturer's protocol. Relative expression level of Hsp70 mRNA was performed by quantitative real-time PCR (qPCR) on the LightCycler96 Detection System (Roche, Mannheim, Germany) using TB Green Premix Ex Taq II (Tli RNaseH Plus) (TaKaRa, Otsu Shiga, Japan). The primer sequences for Hsp70 were forward (5’-TACCAACGGCGACACTCAC-3’) and reverse (5’-TTGCGGATGTCCTTACCCT-3’). Each reaction consisted 0.5 μL of cDNA, forward and reverse primer (10 μM), 10 μL of TB Green Premix Ex Taq II (2×), and 8.5 μL of distilled deionized water to a final volume of 20 μL. The qPCR program was 95 oC for 30 sec, then 40 cycles of 95°C for 5 sec and 60°C for 30 sec followed by a melt-curve analysis. Ae. albopictus rpS7 was used as internal control and the relative Hsp70 expression of cold exposed samples were calibrated by samples collected at room temperature that without cold treatment. The relative expression levels of Hsp70 were determined by using the 2-△△CT calculation method [23].
All data were analyzed by using SPSS statistics 19. Comparison of LT50, egg numbers and egg hatch rates per female (except egg hatch rate of mosquitoes cold exposed at 48h PBM) between different groups were conducted using Student’s t-test. Comparison of the percent of mosquitoes imbided a blood meal between cold exposed and non-exposed groups, of the survival of mosquitoes cold exposed at 24 h PBM with those without blood meal and of the egg hatch rate of mosquitoes cold exposed at 48 h PBM with those maintained at room temperature were performed using chi-square test. Comparisons of glycerol, glucose and hsp70 mRNA levels between groups were assessed using ANOVA followed by Tukey’s multiple comparison.
The GenBank accession number of hsp70 mentioned in the text is JN132155.1.
We found that adult Ae. albopictus could survive for several hours’ exposure to subzero temperature even below -10 oC when it was transferred from room temperature to the low temperature with water. Subsequently, we analyzed the survival of adult female Ae. albopictus that transferred to different low subzero temperatures from -9 oC to nearly -20 oC with water and compared to the mosquitoes that directly transferred to the same temperatures without water. The results showed that when exposed adult mosquitoes to low subzero temperature with water the cold tolerance of these mosquitoes were strongly increased compared to those directly exposed without water (Fig 1). About 70 to 45% of these mosquitoes survived a 5 to 2 hours’ exposure to -9 to -19 oC when exposed with water. However, mosquitoes that directly transferred to the same low temperature without water were found to 100% mortality within 3 to 30 min. The LT50 of these adult Ae. albopictus were about 292, 184 and 106 min for those exposed to -9, -15 and -19 oC with water, respectively, and these were 13.6, 145.1 and 108.7 times longer than those exposed to the same low temperature without water (Table 1). Moreover, we did the same analysis on adult male Ae. albopictus, female Ae. aegypti and Cx. quinquefasciatus to see if the phenomena also existed in male and other mosquitoes. The results suggested that when transferred these adult mosquitoes from room temperature to low subzero temperature (-15 oC) with water the cold tolerance were also significantly increased compared to those transferred to the same low temperature without water (Fig 2 and Table 2).
When transferring from room temperature to low subzero temperature with water adult mosquitoes will fall on the surface of water after anaesthetized by cold, we detected the change in temperature of water during this process (Fig 3). We found that when transferring from room temperature to -9, -15 and -19 oC, the water temperature would reach subzero and iced within 55, 45 and 25 min, respectively. After that the temperature of ice kept at a relative high level of -2, -3 and -5°C for 6, 3 and 2 h, respectively, and then rapidly decreased to the level equal to the low ambient air temperature. To evaluate whether it is the relative high temperature of ice just after dropping under subzero that protected adult mosquitoes from low temperature, we analyzed the survival of adult female Ae. albopictus when exposed to these relative high subzero temperature without water (Fig 4 and Table 1 in parentheses). The results indicated that when exposed to -2 to -5°C, 100% mortality of these adult mosquitoes were reached within 60 to 100 min and that the LT50 were about 85, 57 and 38 min for those exposed to -2, -3 and -5 oC, respectively. Even subtracting the time needed to cool water from room temperature to the relative high subzero temperature, the LT50 of adult female Ae. albopictus transferred to -9 to -19 oC with water still were 2.7 to 2.0 times longer than those directly exposed to -2 to -5°C without water (Table 1). These results indicated that when exposed adult mosquitoes to low subzero temperature with water a RCH response was induced during this process.
Since adult Ae. albopictus could survive several hours’ exposure to low subzero temperature with water, it is important to know the effects of cold exposure on the bite behavior and reproductive capacity of female adults. Our results showed that after exposure to -9°C for 3 and 5 h with water, there still were 70.2% and 56.7% of mosquitoes that successfully had a blood meal on mice, respectively, and there was no significant difference between cold exposed and non-exposed mosquitoes (Fig 5A). The fecundity (egg numbers per female) and egg hatch rates between cold exposed and non-exposed mosquitoes also had no significant difference except for the fecundity of mosquitoes exposed to -9°C for 5 h (Fig 5B and 5C). Although the fecundity of these mosquitoes was significantly lower than the mosquitoes without cold exposure, they still could lay about 41 eggs per female after 5 hours’ exposure to -9°C with water. Meanwhile, we found that blood meal had no impact on cold tolerance of adult female mosquitoes compared to mosquitoes without blood meal. There were no significant difference on survival capacity between blood-fed and non blood-fed mosquitoes after 3 hours’ exposure to -9°C with water at both 24 and 48h PBM (Fig 6A), and these mosquitoes could still lay viable eggs (Fig 6B and 6C). Indeed, the egg hatch rate of mosquitoes cold exposed at 48h PBM was significantly higher than mosquitoes without cold exposure after blood meal (Fig 6C). Moreover, the lifespan of adult female Ae. albopictus was also compared between cold exposure and no exposure, and no significant difference was observed (S1 Fig).
Since the cold tolerance was significantly increased when exposed adult mosquitoes to low subzero temperature with water, we analyzed the levels of two important cryoprotectants glycerol and glucose in whole body of adult Ae. albopictus that exposed to -15 oC with water for 2.7 h and compared to those exposed to -3 and -15 oC directly for 1.5 h and 3 min, respectively. The duration of time were chose because at this duration mosquitoes exposed to -3 and -15 oC directly were 100% mortality but mosquitoes exposed to -15 oC with water were less than 30% mortality at the same duration of time as those exposed to -3 oC directly after subtracting the time needed to cool water from room temperature to -3 oC. The results showed that the glycerol level of mosquitoes exposed to low temperature with water was 6.2 μmol per mosquito just after recovery from cold exposure and this was significantly higher than those directly exposed without water and those maintained at room temperature (Fig 7A). Another, the glycerol level of mosquitoes exposed to low temperature with water was significantly decreased at 1 h after recovery from cold exposure. However, the glycerol levels of mosquitoes exposed to subzero temperature without water had no difference with time after recovery from cold exposure. The glucose levels of mosquitoes exposed to -15 oC without water were significantly increased with time after recovery from cold exposure (Fig 7B). While those exposed to -3 oC directly and to -15 oC with water had no difference in glucose levels with time after recovery. However, the glucose level of mosquitoes exposed to -15 oC with water was significantly higher and lower than those exposed to -3 and -15 oC directly at 4 h after recovery from cold exposure, respectively. These results implied that glycerol and glucose might play important role in water-induced RCH of adult mosquitoes but at different stage of cold exposure.
Because the upregulation of heat shock protein 70 (Hsp70) have been reported in several insect species in response to cold temperature [24–27], we wondered if this protein was also involved in mosquitoes to cope with cold temperature. We found that Hsp70 expression gradually increased with time after recovery from cold exposure of mosquitoes exposed to -3 oC directly and to -15 oC with water and the expression levels were the highest at 4 h after recovery (Fig 7C). The Hsp70 expression of mosquitoes exposed to -15 oC without water was significantly increased at 1 h after recovery from cold exposure and maintained at these high levels till 4 h after recovery and the expression levels from 1 to 4 h after recovery of these mosquitoes were significantly higher than that exposed to -3 oC directly and to -15 oC with water. The results demonstrated that the degree of cold shock of mosquitoes exposed to -15 oC with water was similar to those exposed to -3 oC directly and both of them were significantly lesser than those exposed to -15 oC without water.
Adult Ae. albopictus has long been recognized as freeze-intolerant that can’t cope with subzero temperature. Most of the studies on cold hardiness of this species was focused mainly on eggs, which were the only life stage that can survive under subzero temperature as know so far [13–19]. However, in our current study, we found that when exposed adult Ae. albopictus to low subzero temperature (-9 to -19 oC) with water it could survive several hours. Moreover, LT50 of these adult mosquitoes were increased by 13.6 to more than 100 fold changes when compared with the counterpart exposed without water. In consistent, this phenomenon also existed in Ae. aegypti and Cx. quinquefasciatus.
Cold tolerance or cold hardiness is important for the distribution and survival of insects. Over the last 3–4 decades, Ae. albopictus has spread from its native Asian area to all continents except Antarctica [8, 9, 12]. Such widespread distribution of Ae .albopictus implied the strong cold hardiness of this species and that they might have more chance to experience sharp decrease of air temperature than other local mosquitoes. Thus far, studies about the cold hardiness of Ae .albopictus has been focused on egg stage and previous studies indicated that they overwintered predominantly through diapause eggs [16, 17, 28–30]. A previous study showed that diapause eggs from Ae .albopictus could only survive for 1 hour under -12 oC while non-diapause eggs could survive for 4 hours. In addition, neither Ae. albopictus nor Ae. aegypti eggs could be hatched after exposure to -15 oC [19]. However, our results showed that both the adult of Ae. albopictus and Ae. aegypti could survive under -15 oC for more than 3h-exposure (Figs 1 and 2). This indicated that the cold hardiness of adult mosquitoes was even stronger than eggs when exposed with water. The cold hardiness of Ae .albopictus eggs was also highly correlated with the origin. Eggs from northern were more cold-hardy than those from southern, while eggs from tropical Ae .albopictus are much more susceptible to low temperature than those from temperate counterpart [15, 16, 18, 19]. Moreover, similar comparisons of Ae .albopictus larvae from different regions have been conducted at low temperature above zero [31]. So far as we knew, only adult Culex pipiens could tolerate for several to tens of hours’ exposure to subzero temperature (-5 oC) [20]. Interestingly, our results showed that the cold hardiness of adult Ae. albopictus, Ae. aegypti and Cx. quinquefasciatus could be significantly enhanced in the presence of water. When exposed to -15 oC with water these adult mosquitoes can survive for several hours more. This is the first report that adult Aedes mosquitoes cold also cope with low subzero temperature even below -10 oC so long as there are waters when these mosquitoes exposed to this low temperature. Meanwhile, our results suggested that water-induced enhancement of cold hardiness might be a universal phenomenon in adult mosquitoes.
This study showed that the relative high subzero temperature of water immediate after transferring to low subzero temperature just provided partial protection for adult Ae. albopictus against low subzero temperature (Table 1). Previous studies showed that RCH of arthropods could be induced through gradual cooling from 0.1 to 1°C/min [4, 6]. We found that when transferred from room temperature to -9 to -19 oC, the cooling rate of water were about 0.42 to 0.95 oC/min till reaching the relatively high subzero temperature (Fig 3). These results demonstrated that RCH might be induced in these adult mosquitoes and provided partial protection during the process of cold exposure. Considering water is common in nature, our results suggested that RCH of adult mosquitoes induced by waters in nature might provide strong protection against acute decrease of air temperature to low subzero that would be lethal, which often happens in early spring and winter, and late autumn [2, 4, 7].
Ae. albopictus overwintered predominantly through diapause eggs, nevertheless, adult Ae. albopictus was also found occasionally during winter season [29, 32, 33]. This indicates these adults might experience the low subzero temperature sometimes during their lifetime like eggs do. Ae. albopictus has been considered to be the most invasive mosquito species worldwide and to be passively spread over long distance principally through the transportation of eggs by global shipments of used tires and other artificial containers [34–38]. In recent years, however, studies reported that, over a long distance, adult mosquitoes including Aedes species could be transported by aircraft [39–44], and, at a more regional level, adult Ae. albopictus are frequently transported by ground vehicles like cars [10, 37, 38]. These also pose a threat of experiencing low subzero temperature to the adult mosquitoes, especially those be transported to cooler climate areas. In the light of situations mentioned above, it is not known until now how the adult mosquitoes can cope with low subzero temperature in nature. In this study, we found that adult Ae. albopictus could cope with low subzero temperature even below -10°C in the presence of water and that after exposure to low subzero temperature for several hours the adult mosquitoes could still bite, lay eggs, and eggs could hatch to larvae (Fig 5). In addition, after a blood meal female mosquitoes must find a micro-habitat with water to lay eggs. We found that, after cold exposure to low subzero temperature with water, blood-fed Ae. albopictus could still lay viable eggs (Fig 6). This implies that water in nature not only provide a micro-habitat for mosquitoes’ egg-laying but also can act as a shelter against acute decrease of air temperature. In a word, our results indicated that the water-conferred strong protection against low subzero temperature might be an important means for adult Ae. albopictus to survive the lethal low temperature and therefore might be important for the expansion of this species to cooler areas.
Glycerol is the most commonly produced cryoprotectant for insects to cope with freeze damage [45, 46]. High accumulation in haemolymph and tissues was important for overwintering survival of insects while lower glycerol content often resulted in higher overwinter mortality [46–49]. Furthermore, the accumulation of glycerol was also highly correlated with survival of some insects in RCH [1, 3, 50–52]. In this study, we found that the glycerol levels of the adult Ae. albopictus transferred from room temperature to -15 oC with water was significantly higher than those exposed to -3 and -15 oC without water and those maintained at room temperature at 0h after recovery from cold exposure and then significantly decreased to normal level (Fig 7A). Because freeze damage might happen when adult mosquitoes anaesthetizing on -3 oC ice and underneath -15 oC air for a while, our study indicated that accumulated glycerol in adult Ae. albopictus during being exposed to -15 oC with water may confer strong protection to freeze damage and contribute to RCH induced by water.
Sugars are also important cryoprotectants in insects to eliminate or minimize freeze damage [45]. There were studies that the levels of glucose in some insects were increased in response to RCH or cold stress [50, 53–55]. Our results showed that the glucose level of adult Ae. albopictus exposed to -15 oC with water was significantly higher than those exposed to -3°C and lower than those exposed to -15 oC without water at 4 hours after recovery from cold exposure (Fig 7B). This results indicated that the accumulation of glucose in adult Ae. albopictus was important during recovery from cold exposure but not in the process of cold exposure. Moreover, we found that glucose levels went through significant changes with time-dependent manner during recovery from cold exposure to -15 oC without water. Heat shock protein 70 (Hsp70) also played an important role in cold hardiness of overwintering and cold stress of insects [27, 56, 57]. Our results showed that the expression of Hsp70 in adults Ae. albopictus, exposed to -3 or -15 oC with or without water, were all significantly up-regulated during recovery from cold exposure (Fig 7C). This is consistent with previous studies of different insects (including Culex pipiens) that Hsp70 expression were up-regulated during recovery from exposure to subzero temperature [20, 57–59]. Our results and others indicated that the up-regulation of Hsp70 might be required for the repair of cold injury caused by cold exposure [27]. Hsp70 protein went though obvious changes when exposing adult Ae. albopictus to -15 oC without water, which was similar with glucose. It increased dramatically by 6 fold from 1 to 4 h after recovery compared with those maintained under room temperature. This implied that severe acute cold shock might be happened in adult Ae. albopictus exposed to -15 oC without water and caused the serious disorders of glucose metabolism that eventually lead to the death of these adult mosquitoes.
In conclusion, adult mosquitoes especially Ae. albopictus and Ae. aegypti, which are the most important vectors for dengue and Zika virus, could survive at low subzero temperature even below -10 oC for several hours’ exposure in the presence of water and this cold exposure have no detriment impact on fitness costs of adult Ae. albopictus. Both the relative high subzero temperature of water immediate after cold exposure and RCH induced by gradual cooling of water provided this strong protection against low subzero temperature. The cold tolerance might be conferred by accumulation of glycerol during cold exposure stage, and contributed by both glucose accumulation and Hsp70 up-regulation during the recovery stage from cold exposure. The RCH of adult mosquitoes induced by waters in nature might provide strong protection against acute decrease of air temperature to low subzero temperature, which often happens in early spring and winter, and late autumn, and this might be important for the survival and rapid expansion of Ae. albopictus to cooler areas. Our subsequent studies would be performed further to identify whether water-induced protection could be eliminated by down-regulation of glycerol or Hsp70, and cold hardiness of eggs when exposed to low subzero temperature with water.
|
10.1371/journal.pcbi.1005680 | Interactions between the tumor and the blood systemic response of breast cancer patients | Although systemic immunity is critical to the process of tumor rejection, cancer research has largely focused on immune cells in the tumor microenvironment. To understand molecular changes in the patient systemic response (SR) to the presence of BC, we profiled RNA in blood and matched tumor from 173 patients. We designed a system (MIxT, Matched Interactions Across Tissues) to systematically explore and link molecular processes expressed in each tissue. MIxT confirmed that processes active in the patient SR are especially relevant to BC immunogenicity. The nature of interactions across tissues (i.e. which biological processes are associated and their patterns of expression) varies highly with tumor subtype. For example, aspects of the immune SR are underexpressed proportionally to the level of expression of defined molecular processes specific to basal tumors. The catalog of subtype-specific interactions across tissues from BC patients provides promising new ways to tackle or monitor the disease by exploiting the patient SR.
| We present a novel system (MIxT) to identify genes and pathways in the primary tumor that are tightly linked to genes and pathways in the patient systemic response (SR). These results suggest new ways to tackle and monitor the disease by looking outside the tumor and exploiting the patient SR.
| Breast cancer (BC) research has largely focused on understanding the intrinsic properties of the primary tumor in order to therapeutically target key molecular components that drive progression within the tumor epithelial cells [1]. For example, tamoxifen and trastuzumab target the estrogen and human epidermal growth factor receptors (ER, HER2) whose expression levels in tumors define the traditional clinical subtypes of BC. The vast majority of BC-related genomic studies have focused on bulk tumor samples that are expected to be enriched for neoplastic epithelial cells [2]. These efforts have produced subtyping schemes that classify patients into groups based on the similarity of expression of diverse molecular markers and processes [3–9] and generated gene signatures that can predict patient prognosis and benefit from therapy [10–13].
Cancers however are much more than an autonomous mass of epithelial cells. They constitute multicellular systems capable of bidirectional interactions with neighboring non-malignant cells and extracellular components i.e. the tumor microenvironment [14–16]. Tumor-microenvironmental interactions are necessary for tumor progression and drug sensitivity [16, 17] and are becoming better understood [18–21]. In fact, several genomics studies of the BC microenvironment, including our efforts, show that the microenvironment reflects its tumor and harbors prognostic information [22–24]. However, we also recently established that the primary tumor and its microenvironment does not harbor accurate prognostic signals in approximately 20% of BC patients [9]. Specifically, these patients are consistently misclassified by all hallmarks of breast tumors defining tumor epithelial cells (such as proliferation and ER status) and their microenvironment (such as the infiltration of immune cells, angiogenesis and fibroblast activation).
The systemic response (SR) in cancer patients refers here to the perturbations that occur in peripheral blood cells, which include immune effector cells and circulate throughout the body. The fact that a tumor exerts systemic effects (via eg soluble or exosomal factors) may provide an explanation for the clinical observation that patients with one tumor have an increased risk of developing several independent tumors, and that removal of primary cancer improves the survival of patients with distant metastases at the time of diagnosis [25]. In addition, since ER positive (ER+) BC tends to recur as long as 10–15 years after surgical removal of the tumor, it is important to understand systemic factors governing late recurrence and therapeutic approaches that target beyond the tumor site. In fact, there is a rapidly increasing understanding of the various means a tumor employs to favor metastasis in distant organs [26, 27]. For example, an “instigating” BC can exploit the patient SR so that otherwise-indolent disseminated tumor cells become activated [27–32]. The SR has also been investigated in BC at time of diagnosis. Specifically, our recent comparison of blood profiles of BC patients and matched controls yielded a gene signature that reports the presence of BC [33]. This diagnostic signature is specific to BC (i.e. the test classifies women with carcinoma other than breast as negative), and the composition of genes and enriched pathways in the signature suggest that a cytostatic immune-related signal in the SR of patients is associated with the presence of a tumor. Finally, recent evidence demonstrates that engagement of systemic immunity is critical to the process of tumor rejection in genetically engineered mouse models [34].
This study is the first large-scale genomics effort to study the molecular relationships between patient SR and primary tumor. We generated and analyzed expression profiles from peripheral blood and matched tumor cells in 173 BC patients. First, our results highlight how the patient SR is especially relevant to BC immunogenicity. Second, we present a novel tool entitled Matched Interactions across Tissues (MIxT) that starts by identifying sets of genes tightly co-expressed across all patients in each tissue. Then, MIxT identifies which of these gene sets and pathways expressed in one tissue are associated with gene sets and pathways in the second tissue by determining if their expression patterns in tumor and in the patient SR are tightly correlated. We find that there are very few such associations when all BC are considered. However, we do identify biological processes with significant associations between tumor and patient SR when we stratify our analysis by BC subtype. That is, we identify molecular processes in the tumor that are tightly co-expressed with (different) molecular processes in the SR across patients of a specific subtype. In particular, we detail how several tumor-permissive signals are associated between the tumor and SR of basal BC patients.
The Norwegian Women and Cancer (NOWAC) is a prospective population-based cohort that tracks 34% of all Norwegian women born between 1943–57. In collaboration with all major hospitals in Norway, we collected blood samples and matched tumor from women with an abnormal lesion, at the time of the diagnostic biopsy or at surgery, before surgery and any treatment (N ~ 300, S1 Text). RNA preservation for blood samples obtained followed our methodology previously described [33, 35] and detailed in S1 Text. RNA profiles from blood and tumor cells were generated using Illumina Beadarrays and data were processed following careful procedures (S1 Text, S1A Fig). After quality control, our study retained matched blood (SR) and tumor profiles of 173 BC patients diagnosed with invasive ductal carcinoma, and blood profiles of 282 control women (ie. women with no history of cancer with the exception of basal-cell and cervical carcinoma, which are both very common; Fig 1A). The controls are used to determine what constitutes a “normal” SR. BC patients and controls are comparable in terms of age, weight and menopausal status (Fig 1B). Several groups including ours have defined intra- and inter- individual variability of blood gene expression in healthy individuals [35–38]. All together, these studies demonstrate that intra-individual changes that can occur between blood draws are strikingly smaller than the variation observed among samples collected from different individuals. In this study, most women were 50 year-old or older and postmenopausal at time of sampling. Each profile measures the expression of 16,782 unique genes (S1 Text, S1A Fig). Almost all BC (95.4%) are early-stage disease (stage I or II).
Several tumor RNA-based subtyping tools were applied including PAM50 [5] that defines the intrinsic subtypes including luminal A (lumA), luminal B (lumB), normal-like (normalL), basal-like (basalL), and her2-enriched (her2E). The hybrid subtyping scheme partitions ER+ tumors according to their intrinsic subtype and partitions ER- tumors according to their HER2 status [9] (S1 Text, S1B and S1C Fig). In our dataset, all intrinsic luminals (lumA and lumB) and most normalL tumors (85.2%) are ER+; however, ~40% of basalL and ~50% of her2E BC are ER+ (Fig 1C, S1 Table). We also applied the Cartes d’Identité des Tumeurs (CIT) [8] subtyping scheme, which includes a ‘molecular-\ apocrine’ (mApo) subtype enriched for ER-/HER2+ tumors (78.6%) and the highly immunogenic ER+ luminal C (lumC) subtype enriched for ER+/basalL (39.1%). Fig 1C and S1 Table depict the relationships between these three schemes.
Although the IntClust (IC) subtyping scheme [6] is based on gene expression and DNA copy number profiles simultaneously, subtypes can be inferred using a reported RNA-based surrogate algorithm [7, 39]. S1 Table reports when subtypes from other schemes are enriched in each IC subtype. Most notably, IC1 and IC9 are enriched for CIT lumB; IC3, IC7 and IC8 are enriched for lumA; IC4+ is enriched for normalL and at lesser extent CIT lumC, IC5 enriched for mApo-her2E-HER2+, and IC10 enriched for basalL and ER-/HER2-. IC2, IC4-, and IC6 include very few patients (n < 10) and were therefore not further considered in our downstream analyses.
Restricting our attention to tumor profiles, we performed sparse hierarchical clustering with complete linkage using a permutation approach to select the tuning parameter that weights each gene to compute the dissimilarity matrix [40]. The resulting clusters were strongly associated with BC subtypes for all three RNA-based schemes (Fig 1D upper), which confirms that the transcriptional fingerprint of BC subtypes are also ubiquitous in our tumor samples. When restricting our attention to SR profiles, this unsupervised analysis does not identify patient clusters enriched for any given subtype across the three schemes (Fig 1D lower), suggesting that the transcriptional fingerprint of BC subtypes is not the predominant signal in the patient SR.
We then asked if there are genes in the patient SR whose expression covaries with the state of the pathological variables ER and HER2 measured in the primary tumor. Although both are key drivers in BC, neither was found to be associated with individual gene expression changes in the patient SR (limma, linear models for microarray data, false discovery rate, fdr ≤ 0.2, Fig 1E; S1 Text). Similarly, we asked if there are genes in the SR that are markers of tumor subtype (n patients > 10). For the intrinsic, hybrid, and IntClust subtypes, only the ubiquitin ligase RFWD3 is highly expressed uniquely in the SR of lumA patients, and TIMP3, an inhibitor of matrix metalloproteinases, is highly expressed uniquely in ER+/her2E patients (Fig 1E, S2 Fig). For the CIT subtypes [8], we found 70 univariate gene markers in the SR of patients of the lumC subtype. The genes are primarily involved in general cellular processes such as protein processing or transcription in blood cells (fdr ≤ 0.2, Fig 1E, S3 Fig). The lumC subtype is defined by strong activation of several immune pathways at the site of ER+ tumor (i.e. antigen presentation and processing pathway, hematopoietic cell lineage, NK cell mediated cytotoxicity, T-cell receptor signaling and Toll-like receptor signaling) [8], suggesting that the SR is informative in cases where the primary tumor exhibits strong immune properties.
To compare genome-wide molecular changes in tumor and SR across patients, we used WGCNA-based clustering to define sets of tightly co-expressed genes (termed modules) in tumor and blood, respectively [41] (S1 Text). Briefly, we opted for a distance measure based on topological overlap, which considers the correlation between two genes and their respective correlations with neighboring genes [42] (S1 Text). The WGCNA cut and merge routine [43] after clustering identified 19 and 23 modules in the patient tumor and SR, respectively (S4 Fig; S1 Text). Each of these modules can be considered as a unique and stable pattern of expression shared by a significant number of genes.
Modules of the primary tumor are enriched for genes from a broad range of BC hallmarks including angiogenesis (salmon module), extracellular matrix reorganization (greenyellow), proliferation (green), and immune response (brown and darkturquoise) (S2 and S3 Tables, S1 Text). For example, the proliferation tumor module is enriched for mitotic cell cycle-related genes (green, n = 1064 genes; weight01 Fisher test [44], p-value < 2e-17) including the well-known marker of proliferation MKI67, 12 serine/threonine kinases that are used in the calculation of the mitotic kinase score (MKS) [45], and several components of the Minichromosome Maintenance Complex (MCM).
Modules of the patient SR are often enriched for genes involved in either general cellular processes such as translation (black) and transcription (grey60), or immune-related processes such as inflammatory response (brown, green), B-cell response (saddlebrown), innate immune response (greenyellow) (S4 and S5 Tables). Thus, seven SR modules are enriched in genes that are specifically expressed in immune cells [46] (“iris” signature set in S5 Table; Fisher’s Exact Test FET fdr < 0.05).
We constructed a web-based system to visualize gene expression networks, heatmaps and pathway analyses of the modules in each tissue at http://mixt-blood-tumor.bci.mcgill.ca. In a network, genes are represented by nodes (colored by their module membership) that are connected by edges whose length corresponds to their level of co-expression across patients [47]. When selecting only strong gene-gene correlations (topological overlap > 0.1) and removing isolated nodes, the SR network has ~20% more genes than the tumor network (Fig 2A and 2B). Moreover, the SR network has approximately twice as many edges (89,465 connections between genes) than the tumor network (50,617 connections between genes). Thus, the underlying patterns of expression of the tumor genes (and modules) are more dissimilar from each other than the patterns of expression of the SR genes (and modules). In both tissues, the edges that span between modules reflect natural overlaps between cellular process (Fig 2A and 2B). For example in tumors, angiogenesis-related genes of the salmon module are strongly co-expressed with genes of the greenyellow module involved in extracellular matrix remodeling. In blood, modules enriched for genes involved in general cellular processes such as translation (black), RNA processing (violet), and RNA splicing (darkred) are also heavily connected to each other.
We first investigated the relationships between the expression pattern of each module and patient clinicopathological attributes. Towards this end, each gene of a module is used to rank the patient samples (S1 Text). In particular, the sum of gene ranks (ranksum) for each patient provides a linear ordering of the patient samples. Association tests then compare the ranksum values of patients with the attribute of interest eg tumor subtype (S1 Text).
When we consider tumor modules, the expression pattern of the green module (S5A Fig), previously established to be enriched for proliferation-related genes (S2 Table), ranks basalL, her2E and lumB tumors significantly higher than lumA and normalL tumors (ANOVA p-value < 1e-34, S5B Fig). In fact, we observe that the expression pattern of nearly every module is associated with BC subtype (15 of 19 modules, Fig 2C, fdr ≤ 0.15). Moreover, many tumor modules are associated with the proliferative state of the tumor encoded into the MKS score [45] (Pearson correlation, fdr ≤ 0.15) or with ER status (ER+ vs ER-, t-test, fdr ≤ 0.15), two variables that are strongly embedded in the definition of BC subtypes (Fig 2C). These results are consistent with our previous claim that patient subtype is a predominant signal in the primary tumor. Several tumor modules are associated with HER2 status of the tumor, however there are fewer such modules (n = 6) when compared with the proliferative state or ER status of the tumor (Fig 2C), suggesting that transcriptional fingerprint of HER2 is not as ubiquitous in tumor samples. A small number of modules are associated with the lumC subtype, including the brown module enriched for T-cell and inflammatory response genes (S2 Table). This is again consistent with the fact that this is a highly immunogenic subtype [8] (lumC versus not lumC, t-test, fdr ≤ 0.15, Fig 2C).
HER2 status, the lumC subtype and tumor size are all associated with modules of the patient SR (Fig 2D, t-test fdr ≤ 0.15). Although we did not find univariate gene markers in blood associated with HER2 status, the saddlebrown SR module is significantly underexpressed in patients with HER2+ tumors compared to other BC subtypes and controls (fdr = 0.07, S6A Fig) and is enriched for genes involved in B-cell receptor signaling and proliferation (including BLK, CXCR5, CD19, CD79A, CD79B and FCRL5; S4 and S5 Tables). Four SR modules are associated with the immunogenic lumC subtype; one of these modules are also associated with tumor size (Fig 2D, S6B and S6C Fig). Among the 70 univariate gene markers in blood of lumC tumors identified earlier, 31 are included in the darkgreen SR module predominantly underexpressed in lumC patients in comparison to other BC subtypes (fdr = 0.02, S6B Fig). In fact, all four SR modules associated with the lumC subtype are underexpressed compared to other BC subtypes and control samples (S6B and S6C Fig). This includes the purple module highly enriched for genes involved in T-cell (thymus) homing (CCR7, LTA, LTB, VEGFB, HAPLN3, SLC7A6, SIRPG, BCL11B0) and activation (CD47, TNFRSF25, MAL, LDLRAP1, CD40LG) which are underexpressed in lumC patients (fdr = 0.04, S6B Fig). Genes in the cyan modules are also found underexpressed in patients with large (> 2cm) tumors compared to other BC patients and controls (Fig 2D, S6C Fig). Finally, specifically for patients with large tumors, both the darkgrey module, which is enriched for MYC target genes, and the greenyellow module, which is enriched for genes involved in the lymphoid cell-mediated immunity (including GZMH, GZMB, GZMM, KLRD1, PRF1, KLRG1, and GNLY; S4 and S5 Tables), are underexpressed compared to the remaining BC patients and controls.
Together these results indicate that distinct SR are detected in BC patients with HER2+, lumC and/or large tumors, and that overall the patient immune response is underexpressed compared to patients of other subtypes and controls. These results also highlight the importance of distinct immune components for each of these disease groups. In particular, patients with HER2+ tumors exhibit low expression of genes specifically expressed in B-cell compared to patients with other BC subtypes. Patients with lumC tumors exhibit low expression of genes involved in T-cell homing and function compared to patients with other BC subtypes. Patients with large tumors (>2cm) exhibit low expression of genes involved in lymphoid cell-mediated immunity compared to patients with smaller tumors.
Our analysis to this point identified modules within each tissue independently. Our focus here is on the relationships between tissues by asking if specific biologies in one tissue are correlated with (possibly distinct) biologies in the second tissue. To do this, we constructed a software entitled MIxT (Matched Interactions across Tissues) that contains the computational and statistical methods for identifying and exploring associations between modules across tissues (http://mixt-blood-tumor.bci.mcgill.ca).
Using MIxT, we first ask if genes that are tightly co-expressed in the primary tumor are also tightly co-expressed in the SR, and vice versa (Fig 3A, S1 Text) by investigating the gene overlap between tumor and SR modules (Fisher’s Exact Test FET, fdr < 0.01). Genes that retain strong co-expression across patients regardless of tissue type are likely to be involved in the same biological functions in both tissues as a “system-wide” response to the presence of the disease (even if patterns of gene expression across tissues might differ).
Most modules, regardless of tissue, have significant overlap with three to five modules in the other tissue (Fig 3A). In some cases, it appears that a single (large) module in one tissue is in large part the union of several smaller modules from the other tissue. For example, the brown tumor module has 2765 genes including many involved in immune-related processes (T-cell costimulation, the IFN-gamma pathway and inflammation, S2 and S3 Tables). All of these genes/processes show very strong co-expression in the tumor however, in the SR, these genes divide into four distinct patterns of co-expression (Fig 3A), captured by four different modules: brown (inflammation), greenyellow (cytolysis and innate immune response), saddlebrown (B-cell) and pink (TNFA inflammatory response) (S4 and S5 Tables).
Of note, MIxT identifies three modules in each tissue (SR and tumor) that do not have significant overlap with any module in the other tissue (Fig 3A). For tumors, this includes the purple module enriched for genes involved in estrogen response, the lightcyan module enriched for genes involved in hemidesmosome assembly and cytoarchitecture, and the greenyellow module enriched for genes involved in ECM organization (Fig 3A, S2 and S3 Tables). For the SR, this includes the turquoise module enriched for genes expressed in erythrocytes and involved in hemoglobin production, the purple module enriched for genes in translational termination, and the green module enriched for genes involved in inflammation and specifically expressed in myeloid cells (Fig 3A, S4 Table). This suggests that these processes and responses are either specific to a tissue type (eg ECM organization specific to tumor, and hemoglobin production specific to blood cells) or that the co-expression of genes involved in a defined process is unique to a particular tissue (eg genes specifically co-expressed in peripheral myeloid cells).
There is only one instance where a single tumor module has significant overlap with only a single SR module: darkturquoise modules of size = 86 and 97 genes in SR and tumor, respectively with 50 common genes, including 20 involved in the type 1 IFN signaling pathway (S2 and S4 Tables). Although these two “mirrored” modules share many genes, their patterns of expression are significantly different between the two matched tissues (Fig 3B, correlation between ranksums p-value > 0.05; S1 Text), hinting at a non-concordant expression of the local (in tumor) and systemic (in blood) IFN-1 mediated signals.
Whereas the previous section considers interactions defined by a large number of shared genes between a tumor and a SR module, we also examined more general notions of interactions in MIxT. Here we identify tumor and SR modules that have similar expression patterns (ie both modules linearly order the patients in very similar manner in both tissues) but do not necessarily share any genes in common. More specifically, MIxT derives estimates of significance for interactions using a random permutation approach based on the Pearson correlation between ranksums of gene expression in modules across tissues (S1 Text). This type of interaction detects a biological process or response in the primary tumor that is tightly correlated (or anti-correlated) with a (possibly distinct) biological process or response in the SR, and vice versa. The specific expression pattern in the tissues allows us to then postulate the functional nature of the interaction across tissues.
MIxT identified only one tumor module (of 19) that interacts with only a single SR module (of 23) across all patients (MIxT statistic; p-value < 0.005). The paucity of pan-BC interactions across tissues suggest the need to stratify by patient subtype. After stratification for each of the five subtyping schemes (clinical, PAM50, hybrid, CIT, and Intclust) (Fig 4A), we identified 53 interactions involving 15 tumor modules and 19 SR modules (MIxT statistic; p-value < 0.005; Fig 4B, S7 Fig). Tumor and SR modules are indicated in columns and rows of Fig 4B, respectively. A non-empty cell corresponds to a significant interaction with color used to indicate in which subtype the association is found, grouping together similar subtypes across schemes (eg basalL tumors of the pam50 and CIT schemes). Nearly all interactions are significant in only a single subtype (four exceptions indicated by orange arrows, Fig 4B). For some subtypes, a single stimulus in the tumor affects several biological processes in the patient SR. For example, within the ER+/HER2- subtype and only within this subtype, the pink tumor module, enriched for genes involved in alternative splicing, is associated with three SR modules, enriched for a diverse range of biological processes (orange rectangle in Fig 4B).
The brown tumor module, which is enriched for genes involved in immune processes (S2 Table), has several interactions with SR modules across several subtypes (orange rectangle in Fig 4B). This includes interactions specific to normalL, lumB and IC9 but also several distinct interactions within the ER-/HER2- and basal subtypes. This suggests that immune signals expressed in tumor are associated with changes in expression of different molecular processes in the patient SR for a broad range of subtypes.
As alluded to earlier, only a few interactions are significant in two distinct subtypes simultaneously. For example, the brown tumor module is associated with green SR module in both ER-/HER2- and lumB although the directionality of the association differs between the two cases. More specifically, patients with high ranksums in the brown tumor module have low ranksums according to the green SR module, if the patient is of the ER-/HER2- subtype (Fig 5A, 5C and 5E, MIxT statistic, p-value = 0.004). At the same time, patients with high ranksums in the brown tumor module have high ranksum with respect to the green SR module, if the patient is of the lumB subtype (Fig 5B, 5D and 5F MIxT statistic, p-value < 0.004). In this manner the direction of correlation between the biological processes of the brown tumor module and of the green SR module is determined by the subtype of the patient.
For the brown tumor module in both subtypes, patients with a high ranksum (on the left of the ordering in Fig 5B or 5C for both subtypes) have the strongest immune signals in the tumors. This is because most of the immune-related genes in this brown module (within the red sidebar in Fig 5B and 5C, S3 Table) have highest expression in these patients. This includes genes involved in T-cell stimulation (incl. CD3, CD4, CD5, ICOS, several HLA-DR, -DP, -DQ), IFNɣ signaling (IFNG, IRF1-5, ICAM1, IFI30, HLA-A -B -C) and inflammation (incl. several interleukins, chemokines). For the green SR module in both subtypes, a high ranksum indicates an inflammatory SR (patients on the right in Fig 5E for ER-/HER2-, and patients on the left in Fig 5F for lumB). This is because almost every inflammation-related genes (incl. IFNAR1, IL15, TLR2, IL18RAP, RNF144B), and B-cell proliferation genes (incl. BCL6, IL13RA1, MIF, IRS2) (within the red sidebar in Fig 5E and 5F, S5 Table) have highest expression in these patients.
Thus, ER-/HER2- patients with low immune activity at the tumor site have a high inflammatory SR (right side of Fig 5C and 5E). In fact, the level of the inflammatory response in these BC patients is higher than healthy controls (Fig 5I, t-test, p < 0.001). However, for the lumB subtype, the relationship between tumor and SR is reversed. Here, it is the patients that have high immune activity at the tumor site that have a high inflammatory SR (left side Fig 5D and 5F). In fact, the CIT subtyping scheme calls these patients on the left side as belonging to the lumC subtype (Fig 5H), the highly immunogenic ER+ subtype. In these lumB patients the inflammatory response is also higher than in healthy controls (t-test, p-value < 0.01; Fig 5J).
Altogether these results indicate that a high inflammatory SR is observed in both ER-/HER2- and ER+/lumB patients but increase in systemic inflammation is associated with distinct immune activity at the tumor site depending on subtype.
Three tumor modules are enriched for genes within amplicons prevalent in BC [48] (highlighted in orange in Fig 4B, S3 Table). Two modules, the darkgrey and turquoise tumor modules, contain 68 genes (of 110) and 48 genes (on 71) located within the 16p11-13 amplicon highly prevalent in luminal tumors [48], respectively (S3 Table). The darkgrey module interacts with two distinct SR modules for the lumA and ER+/HER2+ subtype, respectively (S8A and S8B Fig). Tumors of both subtypes that over-express genes in the darkgrey module (left hand side S8C and S8D Fig) are likely amplified in 16p13. In these patients, the presence of this amplification is correlated with changes in expression of specific processes within the patient SR and these processes are distinct depending on subtype (S8E and S8F Fig, p < 0.005 in both cases). S8G and S8H Fig depicts associations between the presence of this amplification and patient clinico-pathological attributes. For example, in ER+/HER2+ patients (S8H Fig), the presence of 16p13 amplification is correlated with the luminal score of the tumor. In the lumA subtype, patients with the highest expression of the lightyellow SR module are significantly different than healthy controls (S8I Fig), and in the ER+/HER2+ subtype, patients with the lowest expression of the salmon module are significantly different than healthy controls (S8J Fig).
The third module enriched for genes involved in BC amplifications is the darkgreen tumor module. This module contains 43 (of 99) genes within the 8q23-24 amplicon prevalent in basal and her2E tumors [48] (S3 Table). Most associations with patient SR modules are specific to the basalL subtype (Fig 4B) and again suggest that basalL tumors that harbor this amplification have concomitant changes in expression of specific molecular processes in patient SR.
Approximately one-fourth of the interactions identified by MIxT are specific to ER-/HER2-, IC10 and basalL subtypes, indicating that the tumor and SR interact strongly in this family of BCs (Fig 4B). We study two tumor modules in greater depth here: the brown immune-enriched module and the darkgreen 8q-enriched module, and their interactions with SR modules in basalL patients (Fig 6A–6C). Here the brown tumor module interacts with one (tan) SR module enriched for genes involved in TOR signaling and cell proliferation (Fig 6A and 6B). BasalL patients with low immune activity at their tumor site (right side of brown tumor module) have low expression of the tan SR module, and this expression is significantly lower than healthy controls (boxplots in Fig 6B, t-test p < 0.0005).
The darkgreen tumor module interacts with four SR modules in basalL patients (Fig 6A and 6C). High expression of genes in 8q is associated with high expression of the green SR module. This module is enriched for genes involved in inflammation. For the remaining three SR modules associated with the 8q-enriched tumor module, almost all genes in these modules are underexpressed when 8q genes are highly expressed (ie. the patient orderings are reversed compared to the darkgreen tumor module). These SR modules contain genes involved in general cellular processes of blood cells (RNA/protein processing, cell proliferation; darkgreen module), genes involved in cytolysis and lymphoid cell-mediated immunity (greenyellow module), and MYC and CD5 target genes (darkgrey module) (Fig 6A–6C, S5 Table). The increase in inflammatory SR and the decrease in the three other molecular processes in the SR of basalL patients whose tumor is amplified on 8q are all significantly different from how these processes are expressed in healthy controls (boxplots in Fig 6C). Overall, we identified one distinct signature in the SR of basalL patients with low immune activity at their tumor site and several immuno-suppressive signals in the SR of basalL patients whose tumor is amplified on 8q.
Molecular profiles of peripheral blood cells and matched tumors were generated and compared for a large cohort of BC patients part of the NOWAC study. The NOWAC consortium provides a highly curated population-based study with extensive gene expression profiling across several tissues from BC patients and controls [35, 49]. A careful design and our extensive experience in blood-based expression profiles enable a detailed molecular description of the patient SR to the presence of BC where blood molecular profiles represent effectively an “averaging” over the transcriptional programs of the different types of cells in blood.
We first asked if the SR could provide accurate univariate markers of tumoral properties such as ER status or subtype. Although thousands of transcripts are differentially expressed in tumors between ER+ and ER- BC [9, 50], there is no gene in SR that can reliably predict ER status of the primary tumor. Moreover, the SR does not inform on the intrinsic BC subtype of the tumor such as lumA, lumB or basalL subtype or on IntClust subtypes. Interestingly, univariate markers in the patient SR were only identified for the CIT lumC subtype defined as particularly immunogenic ER+ tumors [8], suggesting that the SR is informative in cases where the primary tumor exhibits strong immune properties. This is consistent with previous reports that uses blood transcriptomics as a gateway into the patient immune system [51–53] and which is extensively used in the context of autoimmune and infectious diseases [54–56]. This result suggests that it is also applicable in cancer such as BC.
To further investigate the molecular changes in the patient SR, we extended our analyses to multivariate approaches where genes are combined into sets or “modules”. In particular, we performed cluster analysis to partition the genes of both tumor and SR profiles into modules with each module representing a distinct pattern of expression across patients. Our user-friendly website (www.mixt-blood-tumor.bci.mcgill.ca) provides access to these modules built in each tissue, enables investigation of their expression profiles in each tissue and allow user-defined queries of gene, gene sets, and pathway of interest. Further, our MIxT approach estimates gene module expression in both tissues and find significant associations between modules across tissues in a representative cohort of BC patients.
In our dataset, the primary tumor and SR have approximately the same number of modules (19 and 23, respectively) but their gene composition is qualitatively different. Not surprisingly, many modules in tumors were enriched for genes involved in hallmarks of cancer, while SR modules were enriched for either general cellular processes or specific immune responses. Only one module involved in the IFN-I pathway is highly conserved in both tumor and SR, although the common genes had markedly different expression patterns between the two tissues. This is important as it establishes that genes, whose expression patterns may act as good markers in the primary tumor, are not necessarily expressed in the same manner within blood cells.
Our multivariate approach was able to identify modules from the patient SR that could reliably identify not only lumC but also HER2+ and large (> 2cm) tumors. These three cases are among the most immunogenic subtypes of BC and are of relatively poor prognosis. For these patients, gene expression in blood cells is mostly decreased compared to other BC and controls. This result also highlights the importance of distinct immune components of the SR for each of these disease groups: B-cells for HER2+ tumors, T-cells for lumC, and aspects of the cellular immune response for large tumors. Interestingly, a previous study showed that her2E tumors have the highest B-cell infiltration and expression of B-cell receptor gene segments, although this was not predictive of improved patient survival [57]. Our study finds an impaired systemic B-cell response specifically in HER2+ patients, consistent with an inefficient anti-tumoral response in these patients, potentially due to a dysfunctional antigen receptor response and cell development. We could also speculate that the dysfunctional thymic T-cell homing signature in lumC patients reflects the well-documented effect of estrogen on thymic T lymphopoiesis [58–61] in patients diagnosed with a highly immunogenic ER+ tumor. These associations would certainly require validation in follow-up studies.
Finally, MIxT focuses on molecular associations between tissues and provides a holistic view of molecular changes in BC patients. Although the focus here is towards gene expression of blood and matched tumor, our approach could be extended to multiple tissues (eg. blood-microenvironment-tumor) or other levels of molecular data (eg. DNA level somatic aberrations, gene and miRNA expression, epigenetic profiles).
Interestingly, associations between BC tumor and patient SR are heavily dependent on subtype. Only one interaction between tumor and patient SR is identified when all BC patients are considered in the analysis but many are identified when we first stratify patients by BC subtype. This is perhaps not surprising given that there is a great deal of molecular heterogeneity between BC subtypes making “one SR fitting all” highly unlikely. We identified molecular stimuli in tumors that change patient SR in multiple ways only for patients within a particular subtype. For example, expression of genes involved in alternative splicing in ER+/HER2- tumors is associated with changes in expression of multiple processes in SR of patients and those associations are observed only within this specific subtype.
Of note, immune signals measured at the tumor site are associated with distinct SR across a broad range of subtypes. Immune-related processes are known to be more or less expressed within every subtypes and have prognostic capacity in almost all subtypes [9]. Here we show that a change in immune activity at the tumor site is not associated with equal SR across subtypes. Furthermore, high immune signals in tumor is associated with the patient inflammatory SR in opposite ways depending if the patient is ER-/HER2- or lumB. The high inflammatory SR in ER-/HER2- patients (with low immune activity at the tumor site) and in lumB patients (with high immune activity at the tumor site) were both significantly different from how systemic inflammation is “normally” expressed in controls.
Finally, we identify other examples of interactions between tumor and patient SR that occur in subtype-specific fashions. In particular, three tumor modules were enriched for genes in known large-scale BC amplicons (16p11-13, 8q23-24). The expression of these genes changes in a coordinated manner from high to low, suggesting that these genes measure amplification of the corresponding region in BC tumors. In turn, these patterns of expression were associated with distinct SR depending on subtypes highlighting the significance of each amplicon in defining patient SR for particular BC subtypes (eg 16p13 in lumA and ER+/HER2+, and 8q23-24 in basalL and her2E). Of note, these patterns of expression also define patients with particular clinico-pathological characteristics. For example, ER+/HER2+ tumors that do not highly express the genes on 16p have a lower luminal score than ER+/HER2+ tumors that highly express the genes on 16p.
When we restrict our attention to basalL patients, we observe that both the immune-related module and the presence of a 8q23-24 amplification is associated with the patient SR. In fact, the subset of basal patients with 8q23-24 amplification exhibit high inflammatory SR and underexpress genes involved in general cellular proliferation of blood cells, in immune cytolysis, and in MYC and CD5 targets. Together, our matched profiles offer a detailed map of tumor-permissive SR particularly relevant for basalL tumors amplified on 8q and highlight a signature in the SR of basalL patients with low immune activity at their tumor site. This is especially interesting in the context of BC-immunotherapy combination or for monitoring response to these therapies. Overall, our study set the groundwork for further investigation of promising new ways to tackle and monitor the disease by looking outside the tumor and exploiting the patient SR.
Tumor and blood samples were obtained as part of the NOWAC study [49, 62] with approval from Regional Committees for Medical and Health Research Ethics in Norway. Between 2006–10, we collected blood and biopsy samples from BC cases at time of diagnosis, and blood samples from selected age-matched blood controls together with associated lifestyle and clinicopathologic data (S1 Text). In total, and after data preprocessing, profiles include 16,792 unique genes expressed in primary tumors and blood from 173 BC patients, and in blood from 290 controls (S1A Fig).
We used ER status as measured by IHC and HER2 status measured by FISH or IHC where available. When unavailable, ER and HER2 status was determined using gene expression of the ESR1 gene and 6 gene members of the HER2 amplicon, respectively [9, 63] (S1 Text, S1B and S1C Fig). In addition, we calculated the HER2 score (HER2S) and the luminal score (LUMS) as the average expression of the HER2 amplicon gene members and the pam50 luminal genes, respectively. A proliferation score was calculated similarly using 12 mitotic kinases to produce the Mitotic kinase gene expression score (MKS) [45]. Samples were labeled according to our subtyping schemes from the literature: PAM50 [5], hybrid [9], CIT [8], IntClust [7, 39] (S1 Text).
Lists of differentially expressed genes in SR according to subtypes were obtained using the R/Bioconductor package Limma [64]. Whenever p-values were adjusted for multiple testing, the false discovery rate [65] was controlled at the reported level (S1 Text).
An unsigned weighted co-expression network was constructed independently in each tissue (SR and tumor) using the R/Bioconductor package WGCNA [41] (S1 Text). First, a matrix of pairwise correlations between all pairs of genes is constructed across blood and tumor samples, respectively. Next, the adjacency matrix is obtained by raising the co-expression measure to the power β = 6 (default value). Based on the resulting adjacency matrix, we calculate the topological overlap, which is a robust and biologically meaningful measure of network interconnectedness [42] (that is, the strength of two genes’ co-expression relationship with respect to all other genes in the network). Genes with highly similar co-expression relationships are grouped together by performing average linkage hierarchical clustering on the topological overlap. The Dynamic Hybrid Tree Cut algorithm [43] cuts the hierarchal clustering tree, and modules are defined as branches from the tree cutting. Modules in each network were annotated based on Gene Ontology biological processes (weight01 Fisher test [44]), MSigDB [66] and other curated signatures relevant to immune and blood cell responses [33, 46, 52] (S1 Text)
Our approach maps samples to a linear ordering based on expression of genes within a given module or signature of interest (S1 Text). In an univariate fashion, each gene within a given module/signature is used to rank all patients based on their expression. For each patient, the ranks of all k genes from the signature are summed and patients are then linearly ordered from right to left according to this ranksum vector. To identify the left and right boundaries of the low and high regions within the observed linear ordering, we delimit the region of independance (ROI95) for each module. Briefly, we compute (n = 10K times) the position of an artificial patient within the observed linear ordering by summing the randomized ranks over all k genes in the module (S1 Text). The ROI95 is defined as the region that contains 95% of the randomly generated samples. The three defined categories of patients correspond to those patients that have high ranskums of the module/signature (high category), low ranksums of the module/signature (low category), and a set of patients where the expression of the genes within the module/signature lose their pattern of pairwise correlation (mid category).
Using gene ranksums to capture module expression, we asked how modules are associated with patients’ clinical attributes and how they are associated across tissues. Pearson correlation and Analysis of Variance (ANOVA) was used to test association between a given module and continuous patient attributes (eg. age, weight, MKS, LUMS) and between a given module and categorical patient attributes (eg. ER, HER2, subtypes, lymph node status), respectively (S1 Text). For each variable. we computed empirical p-values after permuting clinical labels 1000 times. For each variable, we perform a total of 42 association tests (23 blood modules + 19 tumor modules) and used false discovery rate [65] to correct for multiple testing for each variable independently or for each “family” of tests when dependent variables are very similar (S1 Text).
Interactions between modules across tissues are identified using a random permutation approach based on the Pearson correlation between ranksums of gene expression in modules across tissues (S1 Text). ANOVA was used to compare SR module expression between BC patients (assigned to a given tumor module ROI95 categories) and controls.
|
10.1371/journal.pcbi.1003581 | Wikipedia Usage Estimates Prevalence of Influenza-Like Illness in the United States in Near Real-Time | Circulating levels of both seasonal and pandemic influenza require constant surveillance to ensure the health and safety of the population. While up-to-date information is critical, traditional surveillance systems can have data availability lags of up to two weeks. We introduce a novel method of estimating, in near-real time, the level of influenza-like illness (ILI) in the United States (US) by monitoring the rate of particular Wikipedia article views on a daily basis. We calculated the number of times certain influenza- or health-related Wikipedia articles were accessed each day between December 2007 and August 2013 and compared these data to official ILI activity levels provided by the Centers for Disease Control and Prevention (CDC). We developed a Poisson model that accurately estimates the level of ILI activity in the American population, up to two weeks ahead of the CDC, with an absolute average difference between the two estimates of just 0.27% over 294 weeks of data. Wikipedia-derived ILI models performed well through both abnormally high media coverage events (such as during the 2009 H1N1 pandemic) as well as unusually severe influenza seasons (such as the 2012–2013 influenza season). Wikipedia usage accurately estimated the week of peak ILI activity 17% more often than Google Flu Trends data and was often more accurate in its measure of ILI intensity. With further study, this method could potentially be implemented for continuous monitoring of ILI activity in the US and to provide support for traditional influenza surveillance tools.
| Although influenza is largely avoidable through vaccination, between 3,000–50,000 deaths occur in the United States each year that are attributed to this disease. The Centers for Disease Control and Prevention continuously monitor the amount of influenza that is present in the American population and compiles this information in weekly reports. However, because it can take a long time to collect and analyze all of this information, the data that is being reported each week is typically between 1–2 weeks old at the time of publishing. For this reason, we are interested in developing new techniques to determine the amount of influenza in the population that are accurate, can return results in real-time, and can be used to supplement traditional monitoring. We have created a method of estimating the amount of influenza-like illness in the American population, at any time of year, by analyzing the amount of Internet traffic seen on certain influenza-related Wikipedia articles. This method is able to accurately estimate the percentage of Americans with influenza-like illness, in real-time, and is robust to influenza seasons that are more severe than normal and to events that promote much media attention, such as the H1N1 pandemic in 2009.
| Each year, there are an estimated 250,000–500,000 deaths worldwide that are attributed to seasonal influenza [1], with anywhere between 3,000–50,000 deaths occurring in the United States of America (US) [2]. In the US, the Centers for Disease Control and Prevention (CDC) continuously monitors the level of influenza-like illness (ILI) circulating in the population by gathering information from sentinel programs which include virologic data as well as clinical data, such as physicians who report on the percentage of patients seen who are exhibiting influenza-like illness [2]. While the CDC ILI data is considered to be a useful indicator of influenza activity, its availability has a known lag-time of between 7–14 days, meaning that by the time the data is available, the information is already 1–2 weeks old. To appropriately distribute vaccines, staff, and other healthcare commodities, it is critical to have up-to-date information about the prevalence of ILI in a population.
There have been several attempts at gathering non-traditional, digital information to be used to predict the current or future levels of ILI, and other diseases, in a population [3]–[11]. The most notable of these attempts to date has been Google Flu Trends (GFT), a proprietary system designed by Google, which uses Google search terms that are correlated with ILI activity in the US to make a estimation of the current level of ILI [12]. Google Flu Trends was initially quite successful in its estimation of ILI activity, but was shown to falter in the face of the 2009 H1N1 swine influenza pandemic (pH1N1) due to much-increased levels of media attention surrounding the pandemic [13]. Similarly, GFT greatly over-estimated ILI activity in the 2012–2013 influenza season, again likely due to that fact that it was a more severe influenza season than normally observed and therefore garnered much media attention [14]. In the face of these obstacles, Google has continued to update and re-evaluate its models [15]–[17].
Although GFT has performed well in the past, with the exception of two high ILI activity time periods, new methods of estimating current ILI activity that are less susceptible to error in the face of media coverage should be sought. Additionally, as the global community continues to become increasingly in favor of open-access data and methods [18], new methods of ILI estimation should be freely available for everyone to investigate and improve upon, unlike GFT, which does not share the search terms it uses in its algorithms (though results are public).
To this end, we have created a method of estimating current ILI activity in the US by gathering information on the number of times particular Wikipedia articles have been viewed. Wikipedia is a massive, user-regulated, online encyclopedia. Launched in 2001, Wikipedia harnesses the power of the online community to create, edit, and modify encyclopedia-like articles that are then freely available to the entire world. Currently operating in 232 languages, Wikipedia has ∼30 million articles available, expanding at approximately 17,800 articles per day, with nearly 506 million visitors per month, representing 27 billion total page views since its launch, and has approximately 31,000 active Wikipedia editors (http://stats.wikimedia.org) [19].
With a wealth of detailed information on an almost limitless range of topics, Wikipedia is ideally suited as a platform that could potentially be of use for legitimate scientific investigation in many different areas. Not only is the information held within Wikipedia articles very useful on its own, but statistics and trends surrounding the amount of usage of particular articles, frequency of article edits, region specific statistics, and countless other factors make the Wikipedia environment an area of interest for researchers. It has previously been shown that Wikipedia can be a useful tool to monitor the emergence of breaking news stories, to track what topics are “trending” in the public sphere, and to develop tools for natural language processing [20]–[23]. Furthermore, Wikipedia makes all of this information public and freely available, greatly increasing and expediting any potential research studies that aim to make use of their data.
The purpose of this study was to develop a statistical model to provide near real-time estimates of ILI activity in the US using freely available data gathered from the online encyclopedia, Wikipedia.
In an attempt to use Wikipedia data to estimate ILI activity in the US, we compiled a list of Wikipedia articles that were likely to be related to influenza, influenza-like activity, or to health in general. These articles were selected based on previous knowledge of the subject area, previously published materials, and expert opinion. In addition to articles that were potentially related to ILI activity, several articles were selected to act as markers for general background-level activity of normal usage of Wikipedia. For example, information was gathered on the number of times the Wikipedia main page (www.en.wikipedia.org/wiki/Main_page) was accessed per day, as a measure of normal website traffic. As well, the Wikipedia article for the European Centers for Disease Control was included in models in an attempt to control for non-American article views. Table 1 displays the Wikipedia articles that were considered for inclusion in our models.
Wikipedia article view information is made freely available by Wikipedia, under a project called Wikimedia Statistics (http://en.wikipedia.org/wiki/Wikipedia:Statistics), and is available as the number of article views per hour, which may include multiple views on the same article by the same user. A freely available, user-written tool was independently developed to more easily access the information that Wikipedia makes available (http://stats.grok.se), which aggregates article view data to the day-level, and this tool was used to gather total daily article view information. Daily Wikipedia article view data was retrospectively collected beginning at the earliest available date, December 10, 2007, through to August 19th, 2013, and then aggregated to the week level, with each week beginning on Sunday.
The CDC compiles data on the weekly level of ILI activity in the United States by collecting information from sentinel sites across the country where physicians report on the number of patients with influenza-like illness. CDC ILI data is freely available through ILInet, via the online FluView tool (www.cdc.gov/flu/weekly), and downloadable as week-level data. Google Flu Trends data is also freely available through the Google Flu Trends website (http://www.google.org/flutrends) and is provided weekly at the country and state level. GFT data is the result of Google's proprietary algorithm that uses Google search queries to estimate the level of ILI activity in a given region.
We gathered Wikipedia article view data beginning from the week of December 10th, 2007, the earliest records available, until August 19th, 2013. Accordingly, retrospective CDC ILI data and GFT data was obtained for the same period as the Wikipedia article view information, although both the CDC and GFT data extends much further back in time. When aggregated to week-level, all data sources accounted for 296 weeks of retrospective information, capturing five full influenza seasons as well as partial 2007–2008 data. Due to a lapse in the Wikipedia database, article view information is not available between July 13th and July 31st, 2008, inclusive. Therefore, the total set of data available accounts for 294 weeks.
Models to estimate ILI activity using Wikipedia article view information were developed using a generalized linear model framework. The outcome variable, age-weighted CDC ILI activity, is a proportion and is therefore appropriately modeled using a Poisson distribution, and so the Poisson family was used in the GLM framework, with a log-link function. In an attempt to adjust for potential over-fitting, models were run using jackknife resampling. Two principle models were created, which include Mf, a Poisson model that used the full set of collected Wikipedia article page view data, and Ml, a Poisson model that used Lasso (Least Absolute Shrinkage and Selection Operator) regression analysis. Lasso regression dynamically and automatically selects predictor variables for inclusion or exclusion by penalizing the absolute size of the regression coefficients toward zero, thereby selecting a subset of predictor variables which best describe the outcome data [24], [25]. To investigate the reliability of the models, we used a split-sample analysis on the Ml models to compare how well the Lasso selected predictors for a subset of the data (including years 2007, 2008, 2009, and 2010) accounted for the observed data in the remaining subset (years 2011, 2012, and 2013).
Additionally, each of these aforementioned models were also run while excluding data at key time periods which reflect higher than normal ILI activity or Wikipedia article view traffic (during the early weeks of the 2009 pandemic H1N1 swine influenza pandemic and the unusually severe influenza season of 2012–2013) as a means of investigating the models' ability to deal with large data spikes. By comparing the models with or without higher than normal Wikipedia usage, we can investigate what impact, if any, spikes in Wikipedia activity (potentially caused by increased media reporting of influenza-related events) have on the accuracy of the models, and whether or not these spikes in traffic need to be accounted for. In addition to a factor variable representing the year being included in the models, the month was also controlled for in an effort to adjust for the seasonal patterns that influenza outbreaks exhibit in the United States. All models were investigated for appropriate fit using the Pregibon's goodness-of-link test [26] and by examining Anscombe and deviance residuals. Models were compared to one another by comparing Akaike's Information Criteria, response statistics, and by performing likelihood-ratio tests on the maximum-likelihood values of each model. Goodness-of-fit (GOF) tests, both Pearson and deviance, were tested for; all presented models had GOFs≫0.05. All statistics and models were performed using Stata 12 (Statacorp., College Station, Texas, US).
Across the 294 weeks of data available, the number of views of each Wikipedia article under consideration showed large variability. As an example of this variation, the mean number of daily views of the “Influenza” article was 30,823, but the total number of views ranged from 3,001–334,016 per day. Some of the articles under investigation had relatively few views, such as “influenza-like illness” with a mean of 1,061 article views per day (range: 0–15,629 views per day), while others had very high numbers of views per day, such as the Wikipedia Main Page, which had a mean of 44 million views per day (range: 7–139 million views per day).
Herein, we will discuss the characteristics of several models in an attempt to use Wikipedia article view information to estimate nationwide ILI activity based on CDC data. We consider a full model (Mf) that includes all dependent variables that were investigated and a Lasso-selected model (Ml) that includes only dependent variables chosen as significant by the Lasso regression method.
The Mf model, containing all 35 predictor variables (including year, month, CDC page views, ECDC page views, and Wikipedia Main Page views) and 294 weeks of data, resulted in a Poisson model with an AIC value of 2.795. Deviance residuals for this model ranged from −0.971–1.062 (mean: −0.006) and were approximately normally distributed. Although many of the dependent variables showed spikes in page view activity around the beginning of the 2009 pH1N1 event, the Mf model was able to accurately estimate the rate of ILI activity, with a mean response value (difference between observed and estimated ILI values) of 0.48% in 2009 between weeks 17–20, inclusive. Overall, the absolute response values for the Mf model ranged from 0.00–2.38% (mean: 0.27%, median: 0.16%). In comparison, the absolute response values between CDC ILI data and GFT data ranged from 0.00–6.04% (mean: 0.42%, median: 0.21%). The Pearson correlation coefficient between the CDC ILI values and the estimated values from the Mf model was 0.946 (p<0.001). The actual observed range of ILI activity throughout the entire period for which data is available, as reported by the CDC, was from 0.47–7.72%, with a median value of 1.40%. In comparison, the Mf model estimated ILI activity for the same period ranged from 0.44–8.37%, with a median value of 1.50%, and the GFT ILI data ranged from 0.60–10.56%, with a median value of 1.72%.
The Ml model, which contained 26 variables (including year, month, and CDC page views) that were chosen as significant by the Lasso regression method, resulted in a model with an AIC of 2.764. Deviance residuals for this model ranged from −0.790 to 1.205 (mean: −0.007) and were approximately normally distributed, though less so than in Mf. The absolute response values for this Ml model ranged from 0.00–2.53% (mean: 0.29%, median: 0.18%). During weeks 17–20 of the 2009 pH1N1 event, the mean response value for this model was 0.45%, suggesting it was slightly less accurate over this unusually high article view activity time period than the Mf model for the same period. The Pearson correlation coefficient between CDC ILI data and the estimated mean value for the Ml model was 0.938 (p<0.001), and the range of estimated ILI values for this model was from 0.55–8.66%, with a median value of 1.48%.
Split-sample analysis was used to investigate the reliability of the Ml model. A Lasso regression model that was trained on data from years 2007–2010, inclusive, and the selected predictor variables were used to estimate the ILI activity for each week in the remainder of the dataset (years 2011–2013, inclusive). The cross-validation Pearson correlation between the actual observed CDC ILI data and the ILI estimates provided by the Ml model based on the first subset of data was 0.9854 (p<0.001).
Figure 1 shows the time series for CDC ILI data, GFT data, and the estimated ILI values from both the Mf and Ml models.
In the following models, data from the beginning weeks of the 2009 pH1N1 event (weeks 17–20, inclusive), which showed large spikes in Wikipedia article views due to increased media attention, were excluded from analyses. As well, because of the higher-than-normal influenza activity of the 2012–2013 influenza season, that data was also removed from analyses, beginning from week 40 of 2012 to week 13 of 2013, inclusive. By running the Poisson models without these high volume time-sections, comparisons can be made to the full models in order to investigate the estimating ability of models in the face of higher-than-normal levels of influenza activity or Wikipedia article views.
When removing the above-mentioned data, the Mf model produced an AIC value of 2.772, only marginally smaller than that of the complete Mf model, and was comprised of 263 weeks of data. The range of deviance residuals from this model, −0.650 to 0.891, is slightly narrower than the complete Mf model, suggesting a better fit. For the truncated Lasso model, the Poisson regression model was refit to only include the available data, and therefore produced a different set of 24 predictor variables. From this model, an AIC value of 2.727 was obtained, with a range of deviance residuals from −0.677 to 1.081, a marginal narrowing over the original Ml model. Pearson correlation coefficient values between CDC ILI data and estimated values by the Mf and Ml models, for peak-truncated data, were 0.958 (p<0.001) and 0.942 (p<0.001), respectively.
In the United States, seasonal influenza activity usually peaks during January or February. Using the maximum value of the CDC ILI data in a single influenza season as the true peak time and value, we compared the peak value and week for influenza activity as estimated by our two models, Mf and Ml, as well as the Google Flu Trends data. Results are summarized by model and by year in Table 2.
The Mf model was able to accurately estimate the ILI activity peak in 3 of 6 influenza seasons for which data is available (2009–2010, 2010–2011 and 2012–2013 seasons), and was within one week of an accurate estimation in another season (2007–2008). The Ml model accurately estimated the ILI peak activity week in 2 of 6 seasons (2007–2008 and 2010–2011), and estimated 2 others within a week (2009–2010 and 2012–2013). In comparison, Google Flu Trends data was able to accurately estimate peaks of seasonal ILI activity in 2 of 6 influenza seasons (2009–2010 and 2010–2011 season), and was accurate within one week in 2 other influenza season (2007–2008 and 2008–2009). It should be noted that in the 2010–2011 season, the CDC data peaked at the same ILI percentage at both week 4 and week 6 in 2011, and week 6 was taken to be the true peak, as it agreed with both Wikipedia models and the GFT data. In the 2011–2012 season, the Mf and Ml models were 3 weeks early in their estimation of peak ILI activity and the GFT data was 10 weeks early. Finally, in the 2012–2013 influenza season, the GFT model was 3 weeks late and grossly over-estimated the severity by greater than 2.3-times.
Weekly ILI values based on Wikipedia article view counts were able to estimate US ILI activity within a reasonable range of error, with CDC data as the gold standard. While the CDC ILI data is routinely used as a gold standard, and is most often the best available source of ILI information for the country, this data source has potential biases of its own. There are over 2,900 outpatient healthcare providers that are registered participants of the CDC's ILI surveillance program, but in any given week, only approximately 1,800 provide ILI surveillance data [27]. As well, the population size/density of the area served by each outpatient healthcare provider is not uniform across locations and may lead to a skew in reporting. Additionally, increased media coverage of influenza may prompt healthcare providers to submit more samples for analysis or to report more potential ILI cases than they may have otherwise. Several models were fit to estimate ILI activity, including a model containing all 32 health-related Wikipedia articles investigated, a Lasso regression model which selected 24 health-related Wikipedia articles of significance, and each of these models were run without high media-awareness time periods representing the beginning of the H1N1 pandemic in spring of 2009 and the higher-than-normal ILI rates of the 2012–2013 influenza season. These models were compared to official CDC ILI values as well as GFT data.
Comparing the Mf and Ml models, the AIC value was slightly smaller for the Ml model, as was its range of estimation residuals. With a highly non-significant likelihood ratio test between the two models, there is no evidence to suggest that the Mf model performs better than the Ml model, which may be preferred here. However, since there is no cost or energy associated with collecting additional variable information, the full model may warrant continued use to account for the potential event where more health-related Wikipedia articles become useful in ILI estimation. Mf and Ml models that did not include data for the 2009 spring pH1N1 season and the 2011–2012 peak season resulted in slightly smaller AIC and residual values compared to their full-data counterparts, but did not show large enough improvements in estimates to suggest that higher than normal Wikipedia page view traffic or ILI activity were major factors in the models' ability to estimate ILI activity. This result exemplifies the Wikipedia model's ability to perform well in the face of increased media attention and higher than normal levels of ILI activity, whereas GFT has been shown on several occasions to be highly susceptible to these types of perturbations.
In comparison to GFT data, there are some areas where the Wikipedia models were superior, but others where they were not. Full Wikipedia models were able to estimate the week of peak activity within a season more often than GFT data. Out of the 6 seasons for which data was available, GFT estimated a value of ILI that was more accurate (regardless of whether or not the peak timing was correct) than the Mf or Ml models in 4 seasons, while the Wikipedia models were more accurate in the remaining 2. These analyses and comparisons were carried out on GFT data that was retrospectively adjusted by Google after large discrepancies between its estimates and CDC ILI data were found after the 2012–2013 influenza season, which was more severe than normal. Even with this retrospective adjustment in GFT model parameters, the peak value estimated by GFT for the 2012–2013 is more than 2.3-times exaggerated (6.04%) compared to CDC data, and was also estimated to be 4 weeks later than it actually was. For this same period, the Mf model was able to accurately estimate the timing of the peak, and its estimation was within 0.76% compared to the CDC data.
This study is unique in that it is the first scientific investigation, to the authors' knowledge, into the harnessing of Wikipedia usage data over time to estimate the burden of disease in a population. While Google keeps GFT model parameters confidential, the Wikipedia article utilization data in these analyses are freely available and are open to be modified and improved upon by anyone. Although it has not been investigated here, there is potential for this method to be altered for the monitoring of other health-related issues such as heart disease, diabetes, sexually transmitted infections, and others. While the above mentioned conditions do not have the same time-varying component as influenza, overall burden of disease may potentially be estimated based on the number of people visiting Wikipedia articles of interest. This is an open method that can be further developed by researchers to investigate the relationship between Wikipedia article views and many factors of interest to public health.
Data regarding Wikipedia page views is updated and available each hour, though data in this study has been aggregated to the day level, and then further aggregated to the week level. This was done so that one week of Wikipedia data matched one week of CDC's ILI estimate. In practice, if this Wikipedia based ILI surveillance system were to be implemented on a more permanent basis, it is possible that updates to the Wikipedia-estimated proportion of ILI activity in the United States could be available on a daily or even hourly basis, although this application has not yet been explored. It is hypothesized that hourly updates may have trouble dealing with periods of low viewing activity, such as nighttime and normal sleeping hours, and that the benefit of an hourly update versus a daily update might not be worth the effort involved in its perpetuation. Daily estimates are likely to be of greater use than hourly and hold potential for use as a tool for detecting outbreaks in real-time, by creating an alert when the daily number of Wikipedia article views spikes over a set threshold.
As with any study using non-traditional sources of information to make estimations or predictions, there is always some measure of noise in the gathered information. For instance, the number of Wikipedia article views used in this study represent all instances of article views for the English language Wikipedia website. As such, while the largest proportion of these article views comes from the United States (41%, with the next largest location being the United Kingdom representing 11%), the remaining 59% of views come from other countries where English is used, including Australia, the United Kingdom, Canada, India, etc. Since Wikipedia does not make the location of each article visitor readily available, this makes the relationship between article views and ILI activity in the United States less reliable than if the article view data was from the United States alone. To investigate this bias, it may be of interest to replicate this study using data that is country and language specific. For instance, obtaining Wikipedia article view information for articles that exist only on the Italian language Wikipedia website and comparing that data to specific Italian ILI activity data. Alternatively, the timing and intensities of influenza seasons in English-Wikipedia-using countries apart from the United States could be investigated as potential explanations of model performance. Depending on the timing of influenza activity in other countries, their residents' Wikipedia usage could potentially bolster the presented Wikipedia-based model estimations (if their influenza seasons are similar to that of the United States), or it could negatively impact estimations (if their influenza seasons are not similar to those of the United States). This is an interesting method of comparison and may potentially be explored in future iterations of this method.
If these models continue to estimate real-time ILI activity accurately, there is potential for this method to be used to predict timing and intensity in upcoming weeks. While re-purposing these models could potentially be a significant undertaking, we are interested in pursing this avenue of investigation in future works.
There has been much discussion in popular media recently about the potential future directions of Wikipedia. It has been noted in several papers and reviews that the number of active Wikipedia editors has been slowly decreasing over the past 6 years, from its peak of more than 51,000 is 2007 to approximately 31,000 in the summer of 2013. [19], [28] It has been speculated that the efforts made by the Wikimedia Foundation and it's core group of dedicated volunteers to create a more reliable, trustworthy corpus of information has limited the ability of new editors to edit or create new articles, thereby decreasing the likelihood that a new contributor will become a trusted source of information. Compounding this decrease in active editors, it has become increasingly evident that the vast majority of articles on the English Wikipedia website are both male and Western and European-centric, with comparatively few articles dealing with highly female-oriented topics or other geographic areas. Despite these concerns, the articles relating to influenza that have been investigated in this study are within the scope of the type of Wikipedia articles that are routinely and adequately maintained by long-time editors. The authors hypothesize that any decreases in the number of editors in the Wikimedia domain are unlikely to create significant changes in viewership of the articles of interest for estimating or predicting influenza-like illness, and therefore should not contribute meaningfully to the pursuit of this type of surveillance.
Due to an error in Wikipedia data collection, there were no article view data available between July 13, 2008–July 31, 2008, inclusive, resulting in a time gap of just over 2.5 weeks. Fortunately, this time gap occurred in a traditionally low ILI prevalence time of year, and is not suspected to meaningfully impact analyses.
The application of Wikipedia article view data has been demonstrated to be effective at estimating the level of ILI activity in the US, when compared to CDC data. Wikipedia article view data is available daily (and hourly, if necessary), and can provide a reliable estimate of ILI activity up to 2 weeks in advance of traditional ILI reporting. This study exemplifies how non-traditional data sources may be tapped to provide valuable public health related insights and, with further improvement and validation, could potentially be implemented as an automatic sentinel surveillance system for any number of disease or conditions of interest as a supplement to more traditional surveillance systems.
|
10.1371/journal.pcbi.1003493 | A Novel Bayesian Method for Detection of APOBEC3-Mediated Hypermutation and Its Application to Zoonotic Transmission of Simian Foamy Viruses | Simian Foamy Virus (SFV) can be transmitted from non-human primates (NHP) to humans. However, there are no documented cases of human to human transmission, and significant differences exist between infection in NHP and human hosts. The mechanism for these between-host differences is not completely understood. In this paper we develop a new Bayesian approach to the detection of APOBEC3-mediated hypermutation, and use it to compare SFV sequences from human and NHP hosts living in close proximity in Bangladesh. We find that human APOBEC3G can induce genetic changes that may prevent SFV replication in infected humans in vivo.
| Simian Foamy Virus (SFV) is a very common retrovirus in monkeys. When an infected monkey bites a human it can transmit the virus to the human; however, there are no documented cases of human to human transmission. There also appear to be significant differences between infection in monkey and human hosts. The reason for these differences in the two hosts is not completely understood. In this paper we show that a family of host defense enzymes called APOBEC3 may prevent replication of SFV in humans. They do this by changing the genome of the virus so that it cannot replicate. Although this same process also happens in monkeys, it appears to happen less than in humans, and the changes that the monkey APOBEC3 enzymes make are less likely to prevent the virus from replicating. We are able to make these inferences by seeing characteristic types of mutations in a collection of virus DNA sequences sampled in Bangladesh. We develop new statistical methodology to do this analysis.
| Simian foamy viruses (SFV) comprise a subfamily of retroviruses that naturally infect all primates examined with the notable exception of humans. In non-human primates (NHP), they show strong evidence of co-evolution with their hosts [1]. Persistent infection with SFV is ubiquitous in populations of free-ranging NHP [2], [3] and is not thought to be pathogenic in the natural host. However, recent work shows increased morbidity and mortality for macaques infected with SFV and SIV (simian immunodeficiency virus) compared to those infected with SIV alone [4]. SFV has been zoonotically transmitted to humans on more independent occasions than any other simian-borne retrovirus [5], [6]. There are no documented cases of human to human SFV transmission, including between discordant couples [7], [8]. The factors underlying the apparent lack of human-to-human transmission are not well understood. However, the apparent lack of viral replication in humans is probably an important factor [7], [9]. In NHP, SFV is believed to be transmitted through saliva, primarily through biting. This conclusion is supported by studies that have shown high levels of viral RNA in the oral mucosa of NHP, indicative of replication at that site [10], [11]. The large number of NHP infected with SFV and relatively frequent zoonotic transmission allow study of the roles that viral strain variation and host immune response may play in preventing SFV from becoming an endemic human virus.
There have been no direct experimental infections of a susceptible host with SFV or any other foamy virus. However, blood transfusions from an SFV positive NHP to an SFV negative NHP have been reported [12], [13]. From these studies, a model for the events that occur after SFV infection has been proposed. Briefly, initial infection is of PBMCs. Viral DNA integrations are found in these cells, but replication is not detectable. When a latently infected PBMC migrates to the oral mucosa, an unknown process occurs that leads to infection of superficial epithelial cells, in which the virus can replicate [10], [11]. Infections are persistent, but the only cells that have been found to replicate virus are in the oral mucosa. However, almost all organs in an infected NHP contain latent proviruses at levels suggesting there are many other cell types other than PBMCs that can be latently infected.
Host-viral interactions are better understood for SIV, an NHP-borne lentivirus, than for SFV. In particular the innate immune system is known to play an important role in limiting lentiviral inter-species transmission. Host factors such as SAMHD1, tetherin, and APOBEC3 [14] are known to restrict lentiviruses, which in turn have evolved viral protein antagonists to counter these specific host factors. Cross-species transmission of lentiviruses can be limited by the specificity of these viral antagonists for the host species to which the virus has adapted [15]. The APOBEC3 family of proteins are cytidine deaminases that act on negative strand single-stranded DNA, which is created during reverse transcription. Deamination changes C to U, which then appears as G to A mutations on the positive strand [14]. The importance of APOBEC3G as a barrier to cross-species transmission of SIV has recently been highlighted by Etienne et al [16], who provide evidence that the ability of SIVcpz Vif to adapt to restrict chimpanzee APOBEC3G was more important than its ability to counter SAMHD1 with another viral gene, vpx.
Human APOBEC3 has also been shown to be a potent SFV restriction factor in tissue culture [17]. Some G to A mutations have also been observed in SFV sequences derived from human hosts [17]. These authors suggested that the observed mutations may have been due to APOBEC3 hypermutation, but they noted that strain-level polymorphisms, random retroviral mutations, or other processes could not be excluded as alternative explanations. Also, current methods for detecting and quantifying APOBEC3-mediated hypermutation have limited sensitivities at low rates of hypermutation. Thus, new methods are needed to resolve how APOBEC3 proteins might protect humans from zoonotic transmission of retroviruses.
APOBEC3 activity against retroviruses can be inferred via the local sequence specificity of these editing enzymes. In general, APOBEC3 activity is detectable as an overall excess of plus-strand G to A mutations, however, the various members of the APOBEC3 gene family each have their own local nucleotide context specificity [18]. Much of the work on this specificity has focused on the dinucleotide pair formed by a G and the nucleotide immediately following on the positive strand. For example, human APOBEC3G is known to induce mutation in a GG context. Thus the level of activity of a given APOBEC3 enzyme can be characterized using the counts of G to A mutations in and out of context for that enzyme. Continuing the APOBEC3G example, by comparing the number of GG dinucleotide context G to A mutations to the number of such mutations outside this context, one can detect APOBEC3G hypermutation. Similarly, hypermutation by other APOBEC3 proteins can be inferred by G to A mutations in other dinucleotide contexts.
Currently, the most popular approach, as implemented in the widely used HYPERMUT program [19], is to use a Fisher test to determine if the in context mutations statistically exceed the out of context mutations. This application of the Fisher test has three shortcomings: first, when testing the equality of two binomial distributions, the nominal p-value of the Fisher test does not correspond to the actual rejection rate under the null [20]–[23]. Indeed, by simulating under the null in parameter regimes relevant to hypermutation analysis we show that it does indeed deviate from the nominal p-value, and importantly that the level of deviation depends on the parameters and thus cannot be ameliorated by a simple global change of cut off. However, we also find that the “mid-P” variant [24] does show significantly better performance than the classical Fisher test in this respect. Second, the Fisher test does not provide an estimate of the relative probability of mutation (i.e. the effect size). Third, because the Fisher test requires a strict segregation of sites into “in context” and “out of context,” it does not provide a foundation for further generalization to incorporate subtleties such as varying “strengths” of hypermutation contexts.
In this paper, we employ a Bayesian method to detect and quantify hypermutation by estimating the relative probability, along with uncertainty estimates, of G to A mutation in a given APOBEC3-associated context versus a control context. In addition to providing a more sensitive test, the Bayesian methodology provides an integrated means to estimate effect size (i.e., hypermutation strength) and significance (to decide whether hypermutation is occurring). The risk ratio (described below) is a natural choice to report alongside the Fisher p-value for effect size estimation, as HYPERMUT does. Our approach does a better job of effect size estimation than the risk ratio for a range of parameter values spanning the data sets we have analyzed. Finally, the Bayesian approach can be directly generalized to situations such as different strengths of various hypermutation contexts.
Using this Bayesian approach, we examined the hypermutation patterns of 1097 blood proviral DNA sequences from 169 rhesus macaques, as well as 152 buccal swab RNA sequences from 30 of these animals, and compared them to the hypermutation patterns of 77 SFV proviral DNA sequences detected in blood obtained from 8 zoonotically infected humans sampled from the same geographic areas as the macaques [3], [25], [9]. The buccal swabs are important for our analysis as they represent SFV as it is actively replicating rather than latently present in blood.
For our studies of SFV variation, we have examined 1125 nucleotides of the gag gene [3]. This region of the genome was chosen for our studies because in FV, the gag sequence is the most variable of those encoding virion associated proteins [26]. This is unlike the case of orthoretroviruses, where the env gene is the most variable. The 1125 nucleotides were also chosen because this region contains only one short motif (PSAP) that is known to be required for FV replication. We reasoned that the relatively high variability in this region of gag would allow us to define viral strains. Since we had a large data set from this region of gag [3], [25], [9], we used these sequences to determine potential APOBEC3 mediated hypermutation of SFV.
Although we found evidence of hypermutation in SFV sequences from both humans and macaques, the relative frequency and intensity of SFV gag hypermutation differed significantly between macaques and humans, as did the dinucleotide contexts, suggestive of different host APOBEC3 activities. Moreover, by comparing macaque buccal swab RNA sequences to those obtained from human whole blood, we conclude that the signature of hypermutation in human host SFV sequences is not present in the viruses shed from monkey oral mucosal tissues, but likely arose after at least one round of replication in the human host. Taken together, our results indicate that human APOBEC3G is at least one mechanism that protects humans from extensive replication of some SFV strains.
To ameliorate the issues with applying the Fisher test described in the introduction, we developed a Bayesian approach to use the in-context versus out-of-context mutation counts to statistically identify hypermutation and quantify its strength (Figure 1). Our method uses the same data as the Fisher test to describe the ratio, with uncertainty estimates, of the probability of G to A mutation in a dinucleotide context of interest compared to the corresponding probability in a control context. We call this ratio the relative probability ratio. The uncertainty estimates associated with the relative probability ratio are crucial. For instance, if we see mutation in one out of four context X positions, and two mutations out of four context Y positions, then we can guess that the relative probability ratio is 1/2. However, one can make this statement with much higher certainty if we have 1000 out of 4000 X context mutations and 2000 out of 4000 Y context mutations.
This notion of an estimate with uncertainty can be formalized using Bayesian statistics as the posterior distribution of a model parameter given the data. In our setting, the model parameter of interest is the relative probability of G to A mutation in a dinucleotide context associated with a particular APOBEC activity, the focus context, to the probability of the same mutation elsewhere, the control context. This relative probability will be simply quantified as the ratio of the probabilities that we will call the relative probability ratio.
We use two summaries of the posterior distribution of the relative probability ratio. The first is the location of the 0.05 quantile, which we abbreviate Q05. Q05 signifies the level for which, with posterior probability 0.95, the analysis predicts that the true relative probability ratio is greater than or equal to Q05. In casual terms, if Q05 is equal to 2, then we are 95% sure mutations in the focus context occur at least twice as frequently as those in the control context. We call the sequence as hypermutated in a given context when the corresponding Q05 value of the posterior distribution for the probability ratio exceeds 1.
The other summary used is the Maximum A Posteriori (MAP) value for the relative probability. The MAP is the most likely value, or mode, of the posterior distribution. As such it represents our best estimate of the relative probability ratio. It is important to note that the MAP of this ratio, the object of interest to us, is not the same as the ratio of the MAP numerator and MAP denominator. The difference between the two is especially apparent when the distributions on the numerator and denominator have substantial skew, as is often the case in our setting where the bulk of the probability can be on one side of the MAP value for each distribution. Indeed, the difference between the MAP of the ratio of two Beta-distributed random variables and the corresponding ratio of the MAP values can get arbitrarily large (Figure S1).
Note that we will be testing “overlapping” contexts such as GG and GR (G followed by a G or an A). When GR is preferred over GG, for example, this means that the combination of mutation in the GG and GA contexts was more significant than considering GG sites alone. For each sequence identified as hypermutated in more than one context, the context with the highest Q05 value was identified as the call pattern. The call pattern thus represents the context in which evidence of hypermutation is strongest.
Validations were carried out on mutation counts simulated from a range of relative probability ratios and background mutation probabilities (see Materials and Methods). Ideally, according to the definition of the p-value, one would get a uniform distribution of p-values under the null. Although it is not possible to get an exactly uniform distribution under the null in a discrete setting such as the Fisher test, it is desirable to have this distribution as close to uniform as possible (e.g., [24]). Under a variety of simulation conditions, we find that the classical Fisher test is far from having a uniform distribution under the null in that the observed p-value is consistently smaller than the nominal p-value. Thus, we confirm in this parameter regime the observations of others that the Fisher test is consistently “conservative.” These simulations showed that our method is more sensitive than the Fisher exact test (Table 1), and that the sensitivity of the classical Fisher test cannot be improved by a simple predetermined change of cutoff (Supplementary Figures S2 & S3). We note that our method is slightly “liberal” for some parameter regimes (in particular for testing the range between 0.05 and 0.1) and conservative for others.
Additionally, the simulations allowed us to directly compare our MAP estimates to the true relative probability ratios used to generate the simulated data. Typically researchers have calculated effect size (hypermutation strength) by the risk ratio (RR, also known as relative risk), as is done on the HYPERMUT web site (see Materials and Methods). For most of the parameter domain, MAP estimates were consistently closer to the relative probability ratios used for simulation than were the RR estimates in terms of mean squared error (Figure 2). The simulation parameter regime for this figure was chosen to span the range observed in the SFV and HIV sequences used in this study.
The “mid-P” variant of the Fisher exact test (reviewed in [24]) splits the probability of the observed contingency table in half, and assigns one half of the probability to the “more extreme table” category and half to the “less extreme table” category. This variant performed significantly better than the classical Fisher test in generating an appropriate p-value distribution (Supplementary Figures S2 & S3). For the simulations performed in this paper, this effectively corrected the issues of p-value cutoff observed with the classical Fisher test. However, the current methodology for hypermutation detection uses the classical Fisher test, rather than the mid-P version. Furthermore, in terms of the Receiver Operating Characteristic (ROC) curve to judge the true positive rate as parameterized by the false positive rate, the Bayesian approach performs slightly better than the mid-P approach (Figure S4).
We also validated our method using sequence data from an in vitro study by Refsland et al. [27], which involved knocking out members of the APOBEC3 family from human cell lines and measuring the consequent levels of hypermutation. On the Refsland data set, our methodology detected significantly more positives when the corresponding APOBEC was present, and the two tests had equal false positive rates when it was not. (Table S1). Using simulations based on the Refsland sequences, with no context-specificity to their mutations (see Materials and Methods), we see that the median positive probability for our method is below the expected 5% (Table S2).
In addition, we validated our method by applying it to sequence data from a study by Land et al. [28] that found a significant correlation between CD4 count and presence of strongly hypermutated HIV virus. We performed a similar analysis as in the original paper but with a slightly different bioinformatics pipeline, (see Materials and Methods) and did not see a significant effect when applying the Mann-Whitney test to compare CD4 counts between hypermutation positive and negative calls made by either the Fisher test or our approach. However, when we added the requirement that sequences considered positive for hypermutation by Q05 also have a large effect size as measured by MAP (in the top 25%) we did find a significant elevation in CD4 count compared to the rest of the sequences (p = 0.026). However, we did not see a significant effect when taking sequences that were positive according to mid-P and in the top 25% of effect size according to risk ratio (p = 0.31). Additionally, when restricting to the sequences found to be hypermutated, we find a much more significant nonparametric positive correlation between effect size and CD4 count using our method (Kendall tau p = 0.0026) than using mid-P together with the risk ratio (p = 0.060). These findings emphasize the importance of accurate effect size estimation, which forms an important part of our analyses of SFV sequences below.
Thus, a Bayesian framework to directly estimate the relative probability of mutation in or out of a given APOBEC3 context avoids problems associated with applying the Fisher test and provides a more accurate means for quantifying the level of hypermutation than previously described. The corresponding code is already publicly available (http://github.com/fhcrc/hyperfreq; see Materials and Methods for details) and will be made available as a web tool in the near future.
In order to investigate whether APOBEC3 activities alter SFV in macaques and/or humans infected with the virus, and to compare the levels of APOBEC3 activities in humans and macaques, we analyzed SFV gag sequences from a diverse collection of human blood samples as well as macaque blood and buccal samples collected across multiple urban and forested locations in Bangladesh [3], [25], [9]. Overall, 50 out of 77 (∼65%) human host SFV sequences obtained were found to be affected by hypermutation (Table 2). SFV from all but one of the 8 humans showed evidence of APOBEC3G hypermutation in at least one sequence. The exception was one individual (BGH150), whose 6 SFV clones showed no evidence of G to A hypermutation in any context. We note that the BGH150 sequences were similar to those detected in the macaques from the same region, indicating that the sequences were not amplified from contaminating plasmid. In two of our human subjects, both of whom were infected by more than one SFV strain, we observed hypermutation in clones corresponding to only one of the viral strains. Although buccal swabs were taken from the humans sampled as part of this study, none of these tested positive for SFV.
In contrast, only 82 out of 1097 (∼8.1%) of SFV sequences from monkey blood were found to be hypermutated, and only 42 of the 169 monkeys sampled had at least one hypermutation-positive sequence. Hypermutation was more prevalent in human blood sequences than monkey blood sequences (Fisher p = 1.3×10−32). Defining a sample to be hypermutated if at least one sequence obtained from the sample was hypermutated, hypermutation was more prevalent in human blood samples compared to monkey blood samples (Fisher p = 1.7×10−4). Additionally, the distribution of relative probability ratio across all sequences, irrespective of inferred hypermutation status, was higher for human host SFV sequences than for monkey host sequences (Figure 3). Furthermore, sequences marked as hypermutated showed a higher relative probability ratio of hypermutation in human blood than in monkey blood (Bonferroni-corrected Wilcoxon p = 1.9×10−6). Different context patterns were observed between human and monkey sequences (Figure 4).
Of the 152 sequences obtained from the 30 macaque buccal swab samples, only 8 – from 5 samples – were found to be hypermutated. Thus, hypermutation was also more prevalent in human blood sequences than monkey buccal sequences (Fisher p = 2.3×10−22). Similarly, more human blood samples had evidence of some hypermutation than monkey buccal samples (Fisher p = 4.3×10−4). Furthermore, the MAP relative probability ratios of monkey buccal sequences were significantly lower than those of the GG positive human blood sequences (Figure 5; Bonferroni-corrected Wilcoxon p = 0.023). While the frequency of hypermutation observed in monkey blood samples is higher than that of monkey buccal samples, no statistical significance was found for this relationship.
Thus, overall, with a high degree of statistical significance, more human host SFV sequences were found to be hypermutated than monkey host SFV sequences, and human host SFV sequences had a higher level of hypermutation than the SFV sequences from the macaque host.
Hypermutation of human host sequences in these data was most frequently associated with the GG and GR (i.e. GG or GA) dinucleotide contexts (45 out of 50 sequences; 90%), consistent with APOBEC3G activity as well as combined APOBEC3G and APOBEC3F activity [27]. In contrast, monkeys exhibited a significant amount of GA and GM (i.e. GA or GC) context hypermutation (37 out of 82 sequences; 45%). GM context hypermutation was also observed in a study that examined hypermutation of the XMRV retrovirus in macaques [29]. Overall, hypermutation in human host sequences was more likely to be called in GG and GR contexts than for monkey host sequences (Fisher p = 1.3×10−5). Furthermore, human blood SFV sequences identified as hypermutated in GG and GR contexts exhibited higher MAP relative probabilities than macaque blood SFV sequences (Bonferroni-corrected Wilcoxon p = 4.8×10−8 and p = 3.7×10−4, respectively for the two contexts), corresponding to stronger action of APOBEC3G. The GM context, characteristic of macaque APOBEC3DE hypermutation [29], showed elevated levels in SFV from macaque samples (Figure 4). While the 8 monkey buccal sequences (out of 152) marked as hypermutated all exhibited the strongest hypermutation signal in a GG context, as mentioned above, the strength and abundance of this hypermutation signal was significantly lower in monkey buccal samples than human blood samples.
Of the 77 human blood sequences, 36 (46.8%) contained stop codons within the coding region when the sequences were translated. These stop codons were “in-frame” in that they were the result of a point mutation rather than insertion or deletion and a consequent frame shift. In contrast, only 63 of the 1097 (5.7%) monkey blood sequences had such stop codons. Thus, such stop codons are more likely in blood samples from humans than those from monkeys irrespective of whether the entire sequences were called hypermutated by any test (Fisher p = 2.2×10−16). When considering only sequences called hypermutation positive, this statistical relationship held (Fisher p = 6.5×10−13). The same was true when looking at only GG context positive sequences (Fisher p = 1.0×10−12). Stop codons were correlated with presence of hypermutation activity in humans: all human sequences with stop codons were classified as hypermutated, and only 15 human host sequences called hypermutation positive lacked stop codons. Thus we find that the number of stop codons in sequences from human host blood samples is statistically significantly higher than in monkey host blood sequences.
6 of the 152 (3.9%) monkey buccal swab sequences had in-frame stop codons. Thus, stop codons are also significantly more prevalent in human blood sequences than they are in monkey buccal sequences (Fisher p = 1.1×10−14). While the empirical frequency of stop codons is higher in monkey blood samples than in buccal samples, this relationship was not found to be statistically significant.
Overall, by applying Bayesian analysis we show that hypermutation is statistically more prevalent, stronger and in distinct dinucleotide contexts in the human host sequences, and correlates with the presence of stop codons in a coding region for gag that would preclude virus replication (Figure 6).
We have developed Bayesian methodology to test for and quantify the strength of hypermutation. Our motivation for doing so was to quantify the relative probability of mutation in various nucleotide contexts. This Bayesian method tidily formalizes this idea as estimation, with uncertainty, of the ratio of probability of mutation in two contexts as a ratio of beta-distributed random variables. This enables a unified approach to significance testing (hypermutation detection) and effect size (hypermutation strength) estimation. We show that the Bayesian effect size estimate performs better than the classically-used risk ratio (henceforth RR) over a range of parameter values (Figure 2). Additionally, it is recognized in the statistics community that the Fisher test is only appropriate when the “marginals”, i.e. the row (in this study the number of mutants versus not) and column (in this study the number of sites in dinucleotide context versus not) sums, are fixed in advance [21]. This is not the case for hypermutation detection. A number of statistical papers have highlighted problems with applying the Fisher test when this assumption is violated [20]–[23]. For example, by direct enumeration of tables, D'Agostino et al. [20] have shown that the Fisher test does not produce appropriate p-values when testing the equality of two binomial distributions. In our simulated data we also find that the classical Fisher test is less sensitive than our method (Tables 1 and S1), and that this lack of sensitivity cannot be easily remedied by considering alternate globally-applied cut-offs (Figures S2 & S3). However, the “mid-P” variant of the Fisher test does generate a null distribution that is significantly closer to the uniform than the classical Fisher test and consequently is more sensitive. This variant should be preferred to the classical Fisher test when sensitive detection of hypermutation is desired using a Fisher-type test.
Others have proposed alternate means of investigating hypermutation. One approach is to test ratios derived from k-mer motif frequencies in sequences with a Hotelling T2 test [30]. This method has the advantage of not needing to have every sequence paired with a putatively non-hypermutated sequence, however, it requires long sequences to get sufficient power (in that paper they used whole HIV genomes). Another group [31] has made a software package to investigate potential hypermutation using plots, but did not formalize a statistical methodology.
Using validation and an application to real data, we have shown that the Bayesian framework is an appropriate way to analyze hypermutation-by-context data and that it avoids issues associated with applying the Fisher exact test in this setting for significance testing. We also show that the effect size estimates, which follow naturally from our framework, are more accurate than the standard risk ratio estimator.
A further advantage of the Bayesian framework proposed here is that it can incorporate diverse sources of information as well as uncertainty of “hidden” variables in a principled way. We will take advantage of this feature in future work. Specifically, our next step will be to account for a variety of “strengths” of k-mer context specificities. We are motivated by observations that some contexts are more strongly associated with hypermutation than others [32], [18], [33]. Thus it is not possible to strictly segregate motifs into “hypermutation associated” versus not, making it impossible to apply tests such as the Fisher exact test.
This flexibility comes at the cost of some non-trivial computation. Indeed, although we are able to employ a closed form expression for the probability density function in a ratio of Beta distributions, this expression involves hypergeometric functions that take work to evaluate beyond standard implementations of these functions. This is in contrast with the FET and the RR estimators, which are easily implemented and computationally efficient.
The code used to evaluate sequences for hypermutation using our posterior estimation framework is available at http://github.com/fhcrc/hyperfreq. This program, as well as the routines to perform clustering to find representative non-hypermutated sequences, will be made into a more user-friendly form released within the next year and linked to from the same hyperfreq website.
Using this methodology we found that hypermutation in SFV latent proviral sequences from zoonotically infected humans is common, strong, and primarily in the GG dinucleotide context with some in GA and GR (i.e. GG and GA combined). This corresponds primarily to APOBEC3G activity, perhaps combined with activity of another APOBEC3. In contrast, the hypermutation signal observed in macaques is rare, generally much weaker, and in a distinct set of dinucleotide contexts. A relatively small number of these sequences exhibit very strong GM (i.e. G followed by A or C) and GA context hypermutation, suggestive of rhesus macaque APOBEC3DE activity [29].
By quantifying the strength, frequency, and context specificity of APOBEC3 acting on SFV, we show that it is likely an important restriction factor that acts in vivo to limit replication of some SFV strains in the human host (Figure 6). This is true not only when comparing hypermutation levels between proviruses present in human blood and monkey blood, but also when comparing SFV sequences present in human blood and monkey buccal swabs. This is important, as oral mucosal tissues are the apparent source of infectious virus. APOBEC3G-mediated inhibition of replication in humans could explain the lack of human to human transmission of these strains.
The differences in hypermutation context and strength suggest that the observed hypermutation in human host sequences could not have originated in macaques prior to transmission, and must instead be occurring within human hosts. Other researchers have shown human APOBEC3 to be a potent SFV restriction factor in vitro [17]. These researchers also observed G to A mutations in SFV sequences derived from four bushmeat hunters from Southern Cameroon [17]. These individuals were persistently infected with gorilla SFV from 10 to 30 year old bites, and viral loads in PBMCs were described as being low. Several G to A mutations were observed, some of which were in GG and GA contexts, which may be explained by APOBEC3G or APOBEC3F activity that targeted the viruses. However, the authors of that study did not take a statistical approach and stated that they could not rule out alternate causes for the observed mutations. Thus the present study is the first to clearly show human APOBEC3 activity against SFV in vivo.
There are conflicting data on whether or not there is an SFV viral antagonist to APOBEC3 analogous to lentiviral Vif. While some researchers [34]–[36] report that the nonstructural protein Bet can counteract APOBEC3 activity, others [17] have not been able to detect a difference between restriction of wild-type viruses and viruses lacking Bet. However, it is possible that viruses can evade APOBEC3 using other mechanisms. For example, murine leukemia virus does this via modification of the Gag protein rather than through a specific viral antagonist [37], [38]. In either case, our data support a model where some strains of SFV are sensitive to inactivation by human APOBEC3G.
APOBEC3 enzymes work on ssDNA during reverse transcription. Unlike HIV, SFV primarily undergoes reverse transcription prior to infection of new cells, and only the DNA already present in the virion gets incorporated into new cells [39], [40]. Thus, evidence of human APOBEC activity acting on SFV implies at least one round of replication within the human host. This study provides the first evidence, although indirect, supporting SFV replication in humans. However, this conclusion is in contrast to other work failing to detect SFV replication in human oral or blood cells using other methods [7]. Indeed, in a companion study [9] we were unable to detect SFV RNA in buccal swab samples from the same seropositive humans. This suggests that the level of replication in humans may be below the limit of detection, which is consistent with the overall low proviral titers observed in human blood.
Almost half of the human host SFV gag sequences in this study contained in-frame stop codons within the coding region, which would prevent further replication. Although there are likely to be replication competent proviruses in humans, our studies have failed to detect any SFV transcripts. We cannot say there are no transcripts, only that our RT-PCR methods have failed to detect these.
We also could not exclude the possibility that there is a strain- or host-level effect on hypermutation frequency. In Feeroz et al. [3] we demonstrated that SFV gag sequences from free-ranging rhesus macaques in Bangladesh primarily cluster into six strains, and that these strains have a strong correspondence with sampling location and/or origin of the animal. Here we observe that some of these SFV strains show more evidence of hypermutation than others (Table 2). Two humans and 10 monkeys were infected with the karamjal strain, a strain characteristically found in animals that originate from the Karamjal region of Bangladesh. Only one out of the 73 sequences of the karamjal strain was found to be hypermutated, and that one hypermutated sequence was from a macaque. Additionally, no hypermutated sequences were found in a human infected with the charmaguria strain, a strain detected in the macaques in the town of Charmaguria. On the other hand, 22 of the 31 sequences in bormi2 sequenced from human hosts (see [25] for terminology) were positive for hypermutation, and every human of the four infected with bormi2 had at least one hypermutated sequence. This contrasts with only one sequence of the 102 bormi2 sequences obtained from monkey hosts being positive for hypermutation. Additional data are required to understand how viral strain and host response influence hypermutation.
The data set is completely described in [3], [25]. The human study population consisted of eight human subjects who were found to be positive for SFV by PCR as part of a larger study, as well as 169 free-ranging macaques (M. mulatta). The macaques and humans were sampled in regions of Bangladesh where they come into close contact in the context of daily life. RT-PCR was performed to clone partial gag sequences (1125 bp) from buccal swab RNA of 30 macaques [9], while gag proviral sequences were PCR amplified and sequenced from blood of macaques and humans. An average of six clones per sample were sequenced.
|
10.1371/journal.pntd.0006012 | Composition of the Schistosoma mansoni worm secretome: Identification of immune modulatory Cyclophilin A | The helminth Schistosoma mansoni modulates the infected host’s immune system to facilitate its own survival, by producing excretory/secretory molecules that interact with a variety of the host’s cell types including those of the immune system. Herein, we characterise the S. mansoni adult male worm secretome and identify 111 proteins, including 7 vaccine candidates and several molecules with potential immunomodulatory activity. Amongst the molecules present in the secretome, a 17-19kDa protein analogous to human cyclophilin A was identified. Given the ability of cyclophilin A to modulate the immune system by regulating antigen presenting cell activity, we sought to determine whether recombinant S. mansoni Cyclophilin A (rSmCypA) is capable of modulating bone-marrow derived dendritic cell (BMDC) and T cell responses under in vitro conditions. rSmCypA was enzymatically active and able to alter the pro-inflammatory cytokine profile of LPS-activated dendritic cells. rSmCypA also modulated DC function in the induction of CD4+ T cell proliferation with a preferential expansion of Treg cells. This work demonstrates the unique protein composition of the S. mansoni male worm secretome and immunomodulatory activity of S. mansoni Cyclophilin A.
| Helminths are known for their ability to alter the host’s immune response in order to promote their survival. One such mechanism is the propensity of helminths to secrete molecules with immunomodulatory activity; such molecules alter various aspects of host immunity to the benefit of the parasite. Following detailed characterisation of the secretome, we have identified that Cyclophilin A is secreted from adult Schistosoma mansoni and that the protein has the capacity to alter dendritic cell and T cell function in vitro by inducing a T cell regulatory phenotype.
| Schistosomiasis is one of the most prevalent parasitic diseases, with approximately 230 million people being infected globally [1]. To develop improved schistosomiasis control strategies, including new drugs and vaccines, it is important to advance our understanding of how the parasite manipulates the host’s immune system to achieve the chronic infection state that is characteristic of helminth infections in man. The trematode Schistosoma mansoni is the second most common schistosome species responsible for cases of schistosomiasis in the tropics and subtropics [2, 3].
The co-evolution of parasitic helminths and mammals has led to the development of a spectrum of mechanisms whereby the immune system of the infected host is bypassed or modulated to facilitate the completion of the parasite’s life cycle [4, 5]. The mechanisms that govern the host’s immune system modulation during helminth infection include the release of excretory/secretory (ES) products from the helminth, that for an example drive a ‘modified’ T cell response as a mechanism of suppression of the immune system, to regulate protective host responses and in addition, prevent immunopathology [6, 7]. S. mansoni infection evokes a spectrum of cytokines, such as IL-4, IL-5, IL-10, IL-13, IL-25, IL-33 and TGF-β, as well as modulating the function of various immune cells, including regulatory T (Treg) and B cells, eosinophils, alternatively activated macrophages and tolerogenic dendritic cells (DC) [8, 9]. Indeed, helminth ES contains potent immunomodulatory molecules (IM) that have activities beyond their function in helminth immunity, and have been explored as potential therapeutic molecules [10–13].
Helminth infection modify the functions of T cells with the generation of Th2 CD4+ T cell responses and expansion of Treg cells to evoke a state of helminth-induced T cell hypo-responsiveness [9, 14–17]. These effects are not only caused by the direct activity of ES molecules on T cells, but also indirectly, through modulation of antigen presenting cell (APC) activity [18]. DCs form a heterogenic network of cells comprised of several subsets that are capable of responding to a variety of danger signals [19]. Interactions with certain pathogen pattern recognition receptors, including toll-like receptors (TLRs), drive DC maturation and antigen presentation, with elevated MHCII and co-stimulatory molecules (CD80, CD86) expressed on their surface. Additionally, TLR ligands can differentially influence DC cytokine production with LPS stimulation enabling pro-inflammatory cytokine (IFN-γ, IL-6 and IL-12) secretion [20]. In the course of a helminth infection, and despite the availability of TLR ligands, which include helminth-secreted components, DC subtypes demonstrate altered activity, with impaired cytokine release, decreased expression of co-stimulatory molecules and altered antigen presentation capacity that induces the development of a hypo-responsive T cell response [21].
S. mansoni male worm infections of mice induce a state of modified immunity in vivo, rendering mice refractory to anaphylaxis, allergic lung inflammation and experimental colitis [22–25]. Therefore, we sought to characterise the worm excretory/secretory (WES) proteome of adult male S. mansoni worms, and analyse the immunomodulatory potential of WES to identify novel anti-inflammatory IM and vaccine candidates. In the secretome, seven known vaccine candidates were present, and several potential IM were detected, including a secreted non-classical cyclophilin A (SmCypA). Enzymatic activity of the generated recombinant SmCypA (rSmCypA) was assessed, and the effect of SmCypA on bone marrow (BM) derived DC was investigated with regards to BMDC activation in response to TLR stimulation. We also determined the effect of rSmCypA on DC antigen presentation and subsequent T cell proliferation in vitro. Collectively, we demonstrate that the S. mansoni secretome has a unique composition that includes potential vaccine candidates and IM as well as a human CypA analogue. SmCypA modulates DC leading to the preferential expansion of Treg cells in vitro that may contribute to T cell hypo-responsiveness during S. mansoni infection.
C57BL/6J and OT-II (TCROVA) transgenic mice were from Jackson Laboratory (Maine, USA) and bred in-house. All mice were bred in a specific pathogen-free barrier facility with male mice used at 8–10 weeks of age. A Puerto Rican strain of S. mansoni was maintained by passage in mice and albino Biomphalaria glabrata snails, as previously described [26].
All animal care and experimental procedures were performed under an Irish Department of Health and Children Licence (holder Padraic Fallon, Licence Number B100/3250) in compliance with Irish Medicine Board regulations. Animal experiments received ethical approval from the Trinity College Dublin Bioresources Ethical Review Board (Reference: 121108).
Mice were infected with 200–300 S. mansoni cercariae and portally perfused, to recover worms, six to seven weeks after infection, as described previously [26]. Mice were perfused in Minimum Essential Media (MEM) supplemented with Earle’s Salts (Gibco) and 26.2 mM sodium bicarbonate and further washed several times; male worms were then selected via microscopic examination and any damaged, stunted or dead male worms were discarded. Male worms were subsequently washed thoroughly with 10 mL RPMI-1640 supplemented with 2 mM L-glutamine (Gibco), 50 IU Penicillin/Streptomycin (Gibco) and 5 μg/mL Gentamicin (Gibco) at 37°C under sterile conditions. Two hundred male worms were transferred to cellulose membrane dialysis tubing with a 10 kDa molecular weight cut-off (MWCO; Thermo Scientific) in incubation media; RPMI-1640 supplemented with 2 mM L-glutamine, 100 IU Penicillin/Streptomycin and 5 μg/mL Gentamicin. The tubing was placed in a T-75 cell culture flask containing nutrient media, incubation media with 10% Foetal Bovine Serum (FBS; Sigma) for up to 72 hrs (S1 Fig).
Worms were incubated at 37°C for seventy-two hours; the nutrient media was changed after 24 and 48 hours of incubation. Worm incubation media was harvested and concentrated using a stirred cell concentrator at 50 psi with a 10 kDa MWCO membrane (Amicon). The concentrated supernatant was dialyzed with Dulbecco’s Phosphate Buffered Saline (PBS; Biosera) at 4°C using a 10 kDa MWCO dialysis cassette. The WES preparation was then centrifuged and filtered through a 0.22 μm filter. All WES batches were subjected to a quality control analysis (QCA). A batch is defined as a specimen containing the concentrated WES products from 200 male worms over 72 hours, pooled together and stored at -80°C.
AW molecules were prepared as described previously [27]. In brief, live male worms were isolated, as above, and disrupted under liquid nitrogen with a percussion mortar. The resulting paste was sonicated and centrifuged (10,000 × g) for 1 h at 4°C. The supernatant was repeatedly clarified using a micro-centrifuge at 4°C followed by filtration through 0.45 μm and 0.22 μm filters. AW was stored at a stock concentration of 1 mg/ml at -80°C.
QCA entailed a sample from a batch being resolved on a 12% sodium dodecyl sulfate polyacrylamide gel (SDS-PAGE) and visualized by silver staining. Serum from WES immunised rabbit was used in comparative Western blot analysis. Protein and endotoxin levels were assessed by BCA (Thermo Scientific) and LAL assays (Thermo Scientific), respectively.
300 μg of WES or AW proteins were precipitated using 10% w/v trichloroacetic acid (TCA; C2HCl3O2) and resuspended in rehydration buffer (8M urea, 40 mM Tris, 4% CHAPS, 2% v/v IPG Buffer, 0.002% w/v Bromophenol blue, 60 mM dithiothreitol; DTT). For first dimension separation, WES/AW products were loaded onto 13 cm IPG strips, pH 3–11 (GE Life Sciences), followed by reduction with 1% w/v DTT and alkylation with iodoacetamide (IAM, 2.5% w/v). Second dimension separation was performed on a 138 (W) x 130 (H) mm 12% SDS-PAGE using the ATTO Corporation electrophoresis system AE-6220 (ATTO Bioscience and Biotechnology). Running buffer was added to the buffer chamber and the gel was run at 25 mA and stained by Coomassie Brilliant Blue (Thermo Scientific). All visible spots were manually collected and identified by mass spectrometry.
Mass spectrometry was performed by the BSRC Mass Spectrometry and Proteomics facility at St. Andrews University (http://www.st-andrews.ac.uk/~bmsmspf/). The samples were analysed by a quadruple-time-of-flight mass spectrometer, the Q-STAR Pulsar XL (Applied Biosystems). Briefly, samples were digested with trypsin and loaded onto a capillary liquid chromatography system (nanoLC system). The peptides were separated by reverse phase chromatography and directly eluted into the mass spectrometer. The peptides were then subjected to electrospray ionization and tandem mass spectrometry (ESI-MS/MS) generating a mass (m) -to-charge (z) ratios (m/z) spectrum. The mass spectrometry data was analysed using the Mascot software (http://www.matrixscience.com/) searching the S. mansoni GeneDB sequence database (http://www.genedb.org/Homepage) for protein hits. Protein hits were identified through peptide mass fingerprints. Protein hits that showed at least 2 matched peptides with expectation values lower than 0.05 were considered positive hits.
The analysis for protein homologues was performed using the Protein-Protein Basic Local Alignment Search Tool (BLASTp) accessed at the National Centre for Biotechnology Information (NCBI) website (http://blast.ncbi.nlm.nih.gov/Blast.cgi). Sequences alignments that showed a bit score higher than 30 and an expected value (E-value) lower than 1e-16 were considered homologues. Analysis of Gene Ontology (GO) was performed using the QuickGo web-based browser provided by the European Bioinformatics Institute (EBI, http://www.ebi.ac.uk/). GO terms associated with biological process and molecular function were addressed to each protein sequence by searching the UniProtKB-GOA database. Protein sequences were subjected to analysis by the web-based browser SignalP 4.0 Server (http://www.cbs.dtu.dk/services/SignalP/) for predictions of a secretory signal peptide. Protein sequences that did not contain a signal peptide were subjected to analysis by the web-based browser SecretomeP 2.0 Server for prediction of non-classical secretion i.e. not signal peptide triggered protein secretion. SecretomeP analysis was performed using predictions for mammalian sequences and proteins that showed a SecP score higher than 0.5 where considered non-classically secreted proteins. The TMHMM Server v. 2.0 (http://www.cbs.dtu.dk/services/TMHMM/) was used for the predictions of transmembrane helices in the protein sequences. The InterProScan (http://www.ebi.ac.uk/Tools/pfa/iprscan/) sequence search was used for assignment of protein signatures. This tool combines different protein signature recognition methods that search against specific databases. The protein signatures that describe the same protein family or domain are grouped into unique InterPro entries, with a unique accession number. Alignment of multiple sequences was performed using the ClustalW2 programme provided by the EBI (http://www.ebi.ac.uk/Tools/msa/clustalw2/).
For the production of rSmCypA a construct was designed to incorporate an N-terminal Honeybee melittin signal peptide and a C-terminal polyhistidine (His) affinity tag. Recombinant protein was expressed via baculovirus infection of SF9 insect cells and purified by nickel-NTA and size exclusion chromatography. Endotoxin was removed by Triton-X washing, with rSmCypA having <0.01 EU/mg as assessed by LAL assay.
For confirmation of purity, 5 μg of protein was loaded onto SDS-PAGE gels followed by western transfer for detection of the His-tag. 15% SDS-PAGE gels were visualized using Coomassie Brilliant Blue. For western transfer, proteins were immobilized onto a PVDF membrane (Millipore) at 30V for 18 hours at 4°C. Protein transfer was confirmed via Ponceau S (Sigma) staining. Membranes were blocked and probed with HisProbe-HRP (1:5000 dilution; Thermo Scientific) in Odessey Blocking Solution (Li-cor) for 1 hour. Membranes were developed using ECL Western Blotting Substrate (Thermo Scientific Pierce) and visualized via the ChemiDoc MP system (Bio-Rad). Gel images were acquired via HP Scanjet G4050.
The peptidyl-prolyl cis-trans isomerase (PPIase) activity of rSmCypA was measured using protease-coupled spectrophotometric assay [28], which follows the cleavage of the trans form of the chromogenic peptide substrate, N-succinyl-Ala-Ala-Pro-Phe-4-nitroanilide (Suc-AAPF-pNA; Sigma) by chymotrypsin. The serine protease inhibitor PMSF was added to reactions to block proteolytic activity. Reactions were carried out at 23°C in 200 μl of assay buffer (35 mM HEPES buffer, pH 7.9, 86 mM NaCl, and 0.015% Triton-X-100). Ice-cold chymotrypsin solution (added from a 2 mM stock prepared in 10 mM HCl; Sigma) was added immediately followed by rSmCypA and 4mM of substrate to initiate the reaction. The cis-trans isomerization of the Pro-Phe bond was measured by following the absorbance increases at 405 nm over 0–600s using a microplate reader (VersaMax tunable microplate reader; Molecular Devices, Sunnyvale, CA).
Following lysis of 200 adult male or female S. mansoni worms using Trizol reagent (Invitrogen), total RNA isolation was performed using the RNeasy kit (Qiagen) and was followed by reverse transcription with the Quantitect reverse transcription kit incorporating a genomic DNA elimination step (Qiagen), as per the manufacturer’s instructions. PCR for the detection of SmCypa and housekeeping gene (Pai1, Gapdh) expression was performed using the following primers, SmCypa (forward primer: ACGTCTATGCCACTGACGAC, reverse primer: ATGTGTCAGGGTGGCGATTT), Pai1 (forward primer: TAGCTCCGACAGAAGCACCT, reverse ACGACCTCG ACCAAACATTC), Gapdh (forward primer: ATCCCAGCCTTCGCATCAAA, reverse primer: CATCCCGTGGGATAAGGACG). Fold expression change was calculated by performing densitometry on images of agarose gel electrophoresis for SmCypa and housekeeping genes PCR product bands using ImageJ v1.51.
Femur bones were isolated from male C57BL/6 mice and BM was then flushed out using a 27G needle with RPM1-1640 (Gibco). Erythrocytes were lysed using PharmLyse solution according to manufacturer’s instructions (BD Biosciences), and the remaining cells were washed and counted. Cells were subsequently seeded at 2 x 106 cells/ml in untreated petri dishes in complete media (CM; RPMI-1640 supplemented with 10% heat-inactivated Fetal Bovine Serum (Sigma), 2 mmol L-glutamine, 50 IU/ml penicillin and 50 μg/ml streptomycin) supplemented with 20 ng/ml GM-CSF (R&D Systems). BM cells were cultured as described in detail by Lutz et al., [29]. Following 9 days of culture, 85% of the non-adherent cells expressed the DC marker CD11c (Clone #HL3; BD Biosciences) as assessed by flow cytometric analysis. For functional assays, 1 x 106 BMDCs per ml were treated or not with various concentrations of rSmCypA followed by stimulation for 24 hours with 100 ng/ml ultra-pure LPS (Invivogen) with or without rSmCypA.
Spleen and lymph nodes (LN) were harvested from TCROVA mice and processed to generate single cell suspensions as described previously [30]. CD4+ T cells were isolated by magnetic bead negative selection (CD4+ T cell Isolation Kit II; Miltenyi Biotec) via AutoMACS. Cell purity was determined to be > 90% via flow cytometry. Cells were then labelled with Cell proliferation dye (eBioscience) as per manufacturer’s instructions before being utilized downstream. For T cell culture, 1 x 105 TCROVA CD4+ T cells were cultured with 2 x 104 BMDC previously treated or not with LPS and/or rSmCypA in the presence of 10μM OVA.
To determine OVA323-339 uptake in BMDCs, OVA peptide (Cambridge Research Biochemicals) was labeled with the AlexaFluor647 microscale protein labeling kit (Life Technologies) as per manufacturer’s instructions. Isolated BMDCs, as described above, were treated with labeled peptide at 1μM with or without the presence of 100 ng/ml LPS. Cells were analyzed via flow cytometric analysis after 24 hours of incubation with labeled OVA for AlexaFluor647 positive cells.
Concentrations of IL-12p70, TNF-α, IL-10, IL-4, IL-17A, IFN-γ (R&D Systems) from cell culture supernatants were determined via ELISA as per manufacturer’s instructions. Following Streptavidin-HRP treatment, ELISAs were developed using TMB substrate solution (eBioscience). Absorbance at wavelength 450 nm was read using a microplate reader (VersaMax tunable microplate reader; Molecular Devices).
Single-cell suspensions from in vitro cultures or harvested LNs were analysed by flow cytometry. Cells were washed in flow cytometry staining buffer (PBS with 2% FCS and 0.02% sodium azide) followed by blocking with anti-mouse CD16/32 (2.4G2; BD Bioscience). The following mAbs from BD Biosciences, CD4-PE/V450 (RM4-5), CD80-PE-CF594 (16-10A1), CD40-PE (3/23), eBioscience; CD3-PE-eFluor-610 (145-2C11), CD11b-PerCP-Cy5.5/PE-Cy7 (M1/70), CD11c-PE/PE-Cy7 (N418), MHC-II-FITC/eFluor-450 (M5/114.15.2), Biolegend; PD-L1-PE/APC (10F.9G2), DEC-205-APC (NLDC-145), and Miltenyi Biotec; CD86-FITC/PE (PO3.3) were used at optimally titrated concentrations. For transcription factor staining, cells were fixed and permeabilized with a commercial transcription factor staining kit (eBioscience) according to the manufacturer's instructions, and the cells were stained with the following transcription factors from BD Bioscience—FoxP3-PE/ PE-CF594 (MF23), T-bet-FITC/APC (O4-46), RORγt-BV421 (Q31-378) and eBioscience GATA3-PE (TWAJ). Viable cells were distinguished using LIVE DEAD Aqua (Life Technologies). Populations of interest were gated according to appropriate “fluorescence minus one” controls. Samples were acquired on a CyAn ADP flow cytometer (Beckman Coulter) and were analyzed with FlowJo software (Tree Star).
Data are expressed as mean ± SEM and were analyzed by two-way analysis of variance (ANOVA) test or unpaired Student's t-tests (Prism 6; GraphPad Software). Significance for all statistical tests was shown in figures as P < 0.05 (*), P < 0.01 (**), P < 0.001 (***) and P < 0.0001 (****).
A method for the collection of WES from live adult male S. mansoni worms in vitro [26] was adopted for bulk collection of WES (S1A Fig). Only adult male worms were utilized in WES preparations, to exclude any IM secretions from eggs that would contaminate the WES if female worms were included. Several independent batches of concentrated WES were analysed using one-dimensional SDS-PAGE to determine the spectrum of proteins being produced. The protein content of the WES batch was measured, with protein degradation and overall protein profile assessed. There was consistency, with little identifiable variation between the independent batches of WES (S1B Fig). WES Batches that showed no protein degradation, a similar protein profile of positive bands using rabbit polyclonal anti-WES antibody (S1C Fig) and a maximum of 0.5 endotoxin units per mg (EU/mg) were approved for further analysis. Protein yields were estimated to be equivalent to 110 ng of protein per worm within a 72-hour incubation period. Despite attempts to increase yield through elongated culture times there was little effect on the final protein yield.
To address the question of whether male adult S. mansoni worms selectively secrete a unique proteome, we compared WES to a soluble adult male worm homogenate (AW) by two-dimensional (2D) SDS-PAGE. AW and WES preparations were resolved in parallel. The overall distribution and intensity of the spots revealed a differing protein composition between the WES and somatic preparations (Fig 1A). The AW preparation appears markedly more complex, supporting the hypothesis that the parasite actively secretes/excretes a specific fraction of molecules. To confirm this visual demarcation, one spot unique to the AW preparation (spot 1), two spots common to both preparations, with similar isoelectric points (pI) and molecular weights (spots 2 and 3), as well as two spots unique to the WES preparation (spots 4 and 5) were selected for identification by mass spectrometry (Fig 1B). Spot 1, which is unique to AW preparation, was identified as S. mansoni Paramyosin (Smp_021920.1), a major structural protein of schistosomes and other invertebrates. Spots 2 and 3, which are common to WES and AW preparations, were identified as S. mansoni Fatty acid-binding protein (Smp_095360.1). S. mansoni fatty acid protein, also known as Sm14, was previously identified in the worm tegument and the S. japonicum homologue had previously been detected in the adult worm excretory-secretory proteome [31]. Spots 4 and 5, which are unique to the WES preparation, were identified as S. mansoni Cyclophilin (Smp_040130), also known as Smp17.7. These results demonstrate that WES and AW preparations have common and unique molecules therefore demonstrating a clear demarcation of the protein profile between these preparations.
To further identify the proteins present in WES, selected spots were picked from 2D gels and subjected to electrospray ionization and tandem mass spectrometry (ESI-MS/MS). A total of 170 spots were subjected to analysis by mass spectrometry, of which 136 (80%) showed positive hits when analysed against the S. mansoni GeneDB sequence database, resulting in the identification of 111 proteins (S1 Table). Homologues of some of the S. mansoni WES proteins identified were also present in two other human-infecting schistosome species, S. japonicum and S. haematobium as determined by BLAST analysis. S. japonicum homologues were detected for all S. mansoni WES proteins, with more than 70% of the homologues (80 proteins) detected having a similarity to their respective S. mansoni homologues higher than 80% (S2 Table). Furthermore, several schistosome antigens have been previously tested as vaccine candidates in different animal models [32]. A number of vaccine candidates were identified in our analysis of the S. mansoni WES proteins (Table 1). Three out of six potential candidates selected by the WHO in a study in the 1990’s were identified; TPI [33], Sm28GST [34] and Sm14 [35]. The ECL (200 kDa protein) [36], Sm21.7 [37], Sm-p80 [38] and Cu-Zn superoxide dismutase [39] were also detected. Two promising candidates, more recently tested, Sm29 [40] and Sm-TSP-2 [41] were not identified among WES molecules. Sm29 [40] and Sm-TSP-2 [41] have been characterized as membrane proteins highly expressed in the schistosome tegument and therefore are not expected to be excretory-secretory proteins [40].
WES molecules were analysed for predicted signal peptides, both classical N-terminal and non-classical internal. A small number of proteins had a classical signal peptide (6.3%; 8/111) while the majority of signal peptide containing proteins were non-classical (41.4%; 46/111). These proteins were classified as secretory. Three WES molecules were predicted to be trans-membrane (2.7%; 3/111) and the remaining contained no signal peptide (47.7%; 53/111), and were classified as excretory. The ratio however, of secreted to non-secreted proteins may have been influenced by such predictive software optimised using mammalian, rather than parasitic proteins [42].
S. mansoni WES proteins were screened for the presence of immunomodulatory candidates through homology analysis with other helminth proteins with known immunomodulatory activity (Table 2). Based on this analysis, five S. mansoni WES proteins were considered as potential immunomodulatory candidates: Calreticulin auto-antigen homologue precursor, Serpin, Peroxiredoxin 1 and two Cyclophilin proteins. S. mansoni Calreticulin auto-antigen homologue precursor contains a signal peptide and Serpin and Peroxiredoxin are predicted to be non-classically secreted with a SecP score of 0.601 and 0.597, respectively. The S. mansoni Cyclophilin A protein is predicted to be secreted via a non-classical pathway, while Cyclophilin B contains a signal peptide. Based on strong sequence homology we selected SmCypA to examine further for immunomodulatory activity.
rSmCypA was expressed in insect cells and after Nickel and size chromatography purification, two bands, with a molecular weight of 18–22 kDa, which corresponds to predicted characteristics of the cyclophilin family [49, 50] were detected, following coomassie staining of SDS-PAGE resolved proteins. Western blot, using an anti-His tag, confirmed the predicted size of the recombinant protein (Fig 2A). To determine if rSmCypA was enzymatically active, the cis to trans isomerization of succinyl-Ala-Ala-Pro-Phe-4-nitroanilide was measured using the standard protease-coupled assay. rSmCypA was found to have PPIase activity, 30 μg/min, with increased Suc-AAPF-pNA cleavage to levels above that of the control reaction of chymotrypsin-α alone (Fig 2B). Furthermore, by inhibition of the signal using the chymotrypsin-α inhibitor phenylmethanesulfonylfluoride (PMSF), the PPIase activity was determined to be chymotrypsin-α dependent (Fig 2B). Therefore, the recombinant SmCypA protein generated purified as an enzymatically active protein.
Given that cyclophilins from other species have been reported to modulate BMDC function [49], we sought to demonstrate whether rSmCypA has a similar activity. TLRs have been shown to play a key role in the recognition of helminth products by the host, and certain helminth products have been shown to modulate the activation of DC in response to LPS activation [51, 52]. Therefore, BMDC were treated with rSmCypA during simultaneous activation with LPS for 24 hours. Representative example of gating strategy for flow cytometry analysis of BMDC and the effect of LPS treatment on DC activation is shown (S2 Fig). Treatment with rSmCypA during LPS induced activation did not have an effect on cell surface expression levels of several molecules (CD80, CD86, MHCII, CD40) involved in DC antigen presentation (Fig 3A).
We then assessed the capacity of rSmCypA treated BMDC to uptake antigen. For that purpose, OVA323-339 was labeled with Alexa Fluor-647 and antigen uptake by LPS-only treated or LPS activated and rSmCypA co-treated BMDC was assessed by flow cytometric analysis. The uptake, by DC, of antigen and presentation capacity of labeled OVA323-339 was not altered following co-treatment with LPS and rSmCypA (Fig 3B). Interestingly, despite the inability of rSmCypA to alter BMDC cell surface expression of several co-stimulatory molecules, rSmCypA altered BMDC pro-inflammatory cytokine production, resulting in significantly (P<0.05) reduced expression of TNF-a and IL-12 (Fig 3C).
We next assessed the capacity of BMDC co-treated with LPS and rSmCypA to process and present antigens and subsequently drive T cell activation. CD4+ T cells isolated from TCROVA mice were labeled with CFSE and cultured in the presence of OVA with LPS + rSmCypA co-treated BMDC. rSmCypA treatment of BMDC did not alter the viability of TCROVA CD4+ T cells, but reduced their ability to evoke antigen-specific proliferation of T cells (Fig 4A). These results indicate that the reduced CD4+ T cell proliferation in response to rSmCypA treated BMDC, could potentially be the effect of altered T cell skewing with a preferential expansion of suppressive Treg cells. BMDC with LPS stimulation and rSmCypA treatment were cultured with TCROVA CD4+ T cells ± antigen (Fig 4B). Interestingly, rSmCypA-treated BMDC induced significantly (P<0.05) increased Foxp3+ Treg cell expansion but reduced Gata3+ Th2, while Rorc+ Th17 cell expansion had a modest increase (Fig 4B). The rSmCypA expansion of Treg cells was both dose and antigen dependent and it occurs irrespectively of LPS-induced stimulation (Fig 4B). In accordance with a preferential expansion of Foxp3+ Treg cells, when BMDC co-treated with rSmCypA and LPS were used as APC for the activation of TCROVA CD4+ T cells there is a significant (P<0.05) increase in secreted IL-10 and a decrease (P<0.05) in IFN-γ, marked elevated IL-17A, while IL-4 levels remained unchanged (Fig 4C).
S. mansoni infection can modulate host immunity with the molecules that are excreted or secreted from adult worms acting at the interface of engagement with the immune system of the infected host. The characterisation of WES will potentially be of great significance in the identification of novel therapeutic targets for S. mansoni and other Schistosoma species’ infections of man [53, 54]. Several previously proposed vaccine candidates were found to be located in extracellular vesicles of adult S. mansoni worms [55]. In addition to the identification of vaccine candidates, adult S. mansoni worm proteome has previously been serologically screened for the successful detection of biomarkers and infection susceptibility markers [56]. In our study, 111 proteins of the S. mansoni male WES were identified by mass-spectometric analysis. This led to the identification of 7 previously proposed vaccine candidates and 5 molecules with potential immunomodulatory activity. Over 70% of the identified S. mansoni WES proteins share high homology with proteins described in S. japonicum. Our data show a clear demarcation of the WES and the AW of S. mansoni, revealing that only a fraction of the S. mansoni worm proteins are excreted/secreted. It should be noted that WES and AW used in this study were prepared from adult sexually mature male worms, with female worms not included in antigen preparations. This was to remove female worms and therefore presence of eggs, and release of egg secretions during in vitro culture, and prevent the potential confounding effects of the potent IM in eggs [57]. Further, infections of mice with male worms only have been shown to induce marked modulation of the immune system of the host [8]. Amongst the WES proteins, an enzymatically active homologue of human CypA was identified. SmCypA showed immunomodulatory activity in in vitro cell culture assays, with alterations in DC function and cytokine production leading to a DC mediated preferential expansion of CD4+ Treg cells.
Treg cells are important for limiting immunopathology during helminth infections, with significant expansion of these cells during helminth infections [9]. Evidence suggests that this effect is restricted to live helminths, indicating that Treg cell expansion is a result of released components found in the parasite’s WES, with TLR ligands playing a key role in DC modulation and subsequent Treg cell expansion [17, 58]. In this study, we identified SmCypA as a WES component that modulates DC to induce Treg cell expansion in vitro. It remains to fully characterise the FoxP3+ Treg cells that are expanded by SmCypA treated BMDC in vitro and further address the functionality of such cells in mediating the suppression of bystander T cells. Further, in the context of in vivo functions of SmCypA, it would be important to determine if modulation of DC by SmCypA induced Treg cells in vivo.
Cyclophilins are a group of proteins that have peptidyl-prolyl cis-trans isomerase activity, which have been identified in both prokaryotes and eukaryotes. These proteins are widely expressed and are found in several subcellular compartments at micromolar levels [59]. Cyclophilins have broad functionality, including roles as chaperones and cell signalling molecules [60, 61]. Initially identified as intracellular proteins, later studies showed that CypA and CypB could be secreted under conditions of stress. However, the exact mechanisms and degree of secretion of CypA and CypB are not fully understood [62, 63]. Elucidating these mechanisms would provide a greater understanding of the roles of cyclophilins during inflammation. The first member of the cyclophilins to be identified in mammals, cyclophilin A, is the major cellular target for the immunosuppressive drug cyclosporin A (CsA) as well as newly developed small molecule analogs of CsA that have been shown to limit inflammation and injury by inhibiting neutrophilia in a model of LPS induced acute lung injury [63].
Identification of SmCypA in the adult male worm secretome, highlights the different means by which species can produce similar proteins as part of their immune-suppressive stratagem (as demonstrated by S2 Table). Interestingly, a CypA homologue has been identified and shown to be present in all developmental stages of S. japonicum and in both male and female worms [64]. In S.mansoni CypA was detected by PCR in both adult male and females worms (S3 Fig). Furthermore, S.mansoni-infected patients have IgE responses against the protein [65], supporting that SmCypA is recognized by the infected host during infection. While these studies highlight the potential importance of schistosome CypA in modulation of the host’s immune system, in the context of human schistosomiasis we have not explored if SmCypA can modulate human DC and T cells.
Several recent studies show that helminths have the ability to alter DC function leading to the generation of DC that can significantly dampen immune responses. Adoptive transfer of BMDC exposed to helminth homogenates, resulted in CD4+ T cell IL-10 mediated disease suppression in an experimental model of colitis [66]. Chronic helminth infections have additionally been shown to evoke a preferential expansion of CD11clo DC that are less capable at inducing T cell proliferation and Th2 cytokine secretion in comparison to CD11chi DC [67]. Despite recent advances, the mechanisms that lead to alterations in DC function during S. mansoni infection are not fully understood. While previous studies show that predominantly helminth WES and not AW components are able to modulate DC function and result in the generation of a Th2/Treg skewed response [68, 69], limited individual WES components that induce Treg cells have been described, including a H. polygyrus TGF-β homologue [70]
We have not explored how rSmCypA interacts with BMDC and thereby modulates their function. Studies by Zhu et al., have highlighted that recombinant human CypA binds the CD147 receptor with ligation inducing elevated pERK, pStat3 and pAkt [71]. Whereas, CypA derived from Histoplasma capsulatum has been described as binding to the VLA-5 receptor on DC that modifies its adhesion properties [49]. While these studies showed extracellular CypA to have a preferentially pro-inflammatory effect [49, 71], in this study we show that rSmCypA has a DC immune-regulating capacity that attenuates the veracity of DC mediated T cell activation by specifically inducing a Treg cell response. It is possible therefore that rSmCypA competes with the abundant endogenous host CypA during infection to modulate the infected host’s immune system. This highlights that further work is required to elucidate the exact mechanism(s) of action for rSmCypA modulation of DC.
Whilst it is clear that Schistosoma species may utilize CypA as an immune modulator, it must be emphasized that this is just one of the potential IM that are produced by the parasite. A number of the other molecules in WES may also have immunomodulatory activity. In addition to this, understanding their composition, relative to each other, is key to elucidate how elegantly helminths modulate immune responses. Further work will be required to carefully delineate the vast array of molecules which helminths generate and to understand how this unique composition of IM act together for immune modulation.
|
10.1371/journal.pcbi.1003446 | Leadership and Path Characteristics during Walks Are Linked to Dominance Order and Individual Traits in Dogs | Movement interactions and the underlying social structure in groups have relevance across many social-living species. Collective motion of groups could be based on an “egalitarian” decision system, but in practice it is often influenced by underlying social network structures and by individual characteristics. We investigated whether dominance rank and personality traits are linked to leader and follower roles during joint motion of family dogs. We obtained high-resolution spatio-temporal GPS trajectory data (823,148 data points) from six dogs belonging to the same household and their owner during 14 30–40 min unleashed walks. We identified several features of the dogs' paths (e.g., running speed or distance from the owner) which are characteristic of a given dog. A directional correlation analysis quantifies interactions between pairs of dogs that run loops jointly. We found that dogs play the role of the leader about 50–85% of the time, i.e. the leader and follower roles in a given pair are dynamically interchangable. However, on a longer timescale tendencies to lead differ consistently. The network constructed from these loose leader–follower relations is hierarchical, and the dogs' positions in the network correlates with the age, dominance rank, trainability, controllability, and aggression measures derived from personality questionnaires. We demonstrated the possibility of determining dominance rank and personality traits of an individual based only on its logged movement data. The collective motion of dogs is influenced by underlying social network structures and by characteristics such as personality differences. Our findings could pave the way for automated animal personality and human social interaction measurements.
| How does a group of family dogs decide the direction of their collective movements? Is there a leader, or is decision-making based on an egalitarian system? Is leadership related to social dominance status? We collected GPS trajectory data from an owner and her six dogs during several walks. We found that dogs adjusted their trajectories to that of the owner, that they periodically run away, then turn back and return to her in a loop. Tracks have unique features characterising individual dogs. Leading roles among the dogs are frequently interchanged, but leadership is consistent on a long timescale. Decisions about running away and turning back to the owner are not based on an egalitarian system; instead, leader dogs exert a disproportionate influence on the movement of the group. Leadership during walks is related to the dominance rank assessed in everyday agonistic situations; thus, the collective motion of a dog group is influenced by the underlying hierarchical social network. Leader/dominant dogs have a unique personality: they are more trainable, controllable, and aggressive, additionally they are older than follower/subordinate dogs. Dogs are an ideal model for understanding human social behaviour. Therefore, we address the possibility of conducting similar studies in humans, e.g. walking with children and detecting interactions between individuals.
| Groups that are not able to coordinate their actions and cannot reach a consensus on important events, such as where to go, will destabilise, and individuals will lose the benefits associated with being part of a group [1], [2]. Decision-making usually involves some form of leadership, i.e. ‘the initiation of new directions of locomotion by one or more individuals, which are then readily followed by other group members’ ([3] p83).
Several factors may give rise to the emergence of leadership. In some species or populations, leaders are socially dominant individuals (consistent winners of agonistic interactions [4]) and have more power to enforce their will [5]. For example, in rhesus macaques (Macaca mulatta) the decision to move is the result of the actions of dominant and old females [6]. Similarly, dominant beef cows (Bos taurus) have the most influence on where the herd moves. They go where they wish while subordinates either avoid or follow them [7].
Leaders could appear in species or populations without any dominant individuals, or independently from social dominance. Leaders may have the highest physiological need to impose their choice of action [1], [3], [8]–[10], or they may possess special information or skill [11], [12].
Finally, an individual of a personality type that is more inclined to lead or does not prefer following others may also initiate collective movements [13], [14]. For example, leadership is associated with boldness in sticklebacks (Gasterosteus aculeatus) [15], [16]. The investigation of the relationship between leadership and personality might reveal which personality types occupy particular positions in the leadership network, and conversely, network metrics could identify potential personality traits.
With this study our aim was to reveal potential links between leadership in collective movements, motion patterns, social dominance, and personality traits in domestic dogs (Canis familiaris). It is often assumed that domestic dogs inherited complex behaviours from their wolf ancestors (Canis lupus). The typical wolf pack is a nuclear or extended family, where the dominant/breeding male initiates activities associated with foraging and travel [17]. However, family dog groups may consist of several unrelated individuals with multiple potential breeders. In large wolf packs with several breeders, leadership varies among packs, and dominance status has generally no direct bearing on leadership, but breeders tend to lead more often than non-breeders [18]. Similarly, leadership in Italian free-ranging dogs interchanged between a small number of old and high-ranking habitual leaders. Interestingly, affiliative relationships had more influence on leadership than agonistic interactions [19].
Family dogs are often kept in groups (for instance, 33% of owners in Germany [20] and 26% of owners in Australia [21] have 2 or more dogs), however interactions within freely moving dog groups and their relationship with social dominance are still unexplored. The capacity of dogs to form robust dominance hierarchies is highly debated [22], [23]. However, the reason for the inability to detect hierarchies might be due to methodological issues in certain cases, as instead of aggression patterns, submissive behaviours appear to be better indicators of dominance relationships in dogs [24].
To describe what characterises the collective movement of a group of dogs, and to investigate links between leadership, social dominance, personality [25], and characteristics of individual motion trajectories, we collected high-resolution spatio-temporal (1–2 m, 0.2 s) GPS trajectory data from a group of dogs and their owner during everyday walks. Directional choice dynamics and potential leading activity were assessed by quantitative methods inspired by statistical physics [26], [27]. Personality and dominance rank of the dogs were measured by questionnaires completed by the owner. Because the capacity to form dominance hierarchies is likely to vary from breed to breed [28], we chose a group that contains multiple individuals of the same breed, the Hungarian Vizsla. The studied group is composed of five Vizslas (with two dam-offspring pairs) and one small-sized, mixed-breed dog.
A general overview of the GPS-logged trajectories (see Figure 1 and Video S1: our animation showing a 3-minute-long part of a walk) shows that the dogs run away from the owner periodically, then turn back and return to her, in a loop. Figure S5 shows a typical trajectory of dog V1. It can also be seen that they prefer running these loops or a part of them with one or more group members (see details in the Data Analysis). Given that the dogs' speed was significantly higher than that of the owner (1.5–3.7 times), this motion pattern allows dogs to cover a greater distance than the owner while also keeping the group together. We calculated several simple characteristics of the trajectories and performed an analysis concerning the returning events (Table 1 and Text S1).
The preferred running speeds of the dogs, the relative distances covered, and the distances from the owner were unique and consistent characteristics of an individual dog's path, while other characteristics (e.g, distance from dogs) were less consistent and/or distinctive (for details see Text S1).
To extract information about the interactions between group members, we used a directional correlation analysis [26] with a time window to quantify the fast, joint direction changes for all possible pairings of the dogs (Figure 2A; Table S1; for more details see Data Analysis and Text S1).
We detected frequent short-term interactions and leading tendency differences between dog pairs within the group. The leading and following roles between interacting pairs were often changed during walks and between walks. To check the robustness of the interactions, the directional delay times were calculated for the first 7 and the second 7 walks separately for all pairs. High correlation was found (two-tailed Pearson correlation: r = 0.635, n = 15, p = 0.011), i.e. significant differences in leading tendency were detected over longer timescales. Calculated from a Gaussian fit to the peak of the relevant distributions (Figure S8, Table S1) we found that dogs play the role of leader in a given pair about 50–85% of the time (57% to 85% when directed leader-follower relationships were found).
Based on the directional delay time values, we created a summarised leadership network (Figure 2B). In the network each directed link points from the individual, which played the role of the leader more often in that given relationship toward the follower. We used this network to calculate leading tendency, which is the number of followers that can be reached travelling through directed links.
We also calculated ‘active connections’, which shows the number of how many interactions a dog has (with the number of edges a dog is connected with in the network).
Correlations between trajectory-based variables, leading tendency, personality traits (Jones, 2008, Table 2) and dominance rank (Pongrácz et al., 2008, Table 2) were calculated using two-tailed Pearson correlation for the Vizslas only (n = 5) (Figure 3) and also for all subjects (n = 6). We tested our data for normality using a Shapiro-Wilk test (p<0.05), and where a significant deviation from a normal distribution was found, we used Spearman correlations (indicated as rS).
Our main aim was to investigate whether the leadership we defined based on the motion patterns had any connection with the social dominance.We found that the leading tendencies calculated from the GPS data significantly correlated with the dominance ranks gained from the dominance questionnaire [29] (r = 0.92, n = 5, p = 0.026). To support this result, we performed a comparison with a randomisation using all possible permutations, and this correlation value proved to be significantly higher than it was for the randomised cases. For more details see Text S1 and Figures SI11–13.
To find more correlations in our dataset of trajectory variables and personality traits, all 300 possible pairings were analysed. Note that due to the large number of variable pairs and the small number of dogs involved in the study, none of the p-values remain significant after correction for multiple comparisons (Bonferroni, Sidak or Benjamini–Hochberg procedure). But the correlations mentioned here were all significantly higher than the corresponding values of the randomly permuted cases.
The distance from other dogs correlated with the fear of dogs facet (rS = 0.92, n = 5, p = 0.028) and the excitability facet (rS = 0.92, n = 5, p = 0.026). Dogs that, according to the owner, avoid other dogs and seek constant activity maintained a longer distance from their group mates during the walks.
The time period of the returns (the average time duration between returning events) was found to be inversely correlated with the controllability facet (r = −0.82, n = 6, p = 0.046), and the dominance rank measure (r = −0.84, n = 6, p = 0.036). Dominant dogs who were more responsive to training returned to the owner more often.
The far-from-owner ratio (the time ratio of being relatively far from the owner, for more details see Text S1) correlated negatively with companionability (r = −0.87, n = 6, p = 0.024). Dogs that, according to the owner, seek companionship from people also like staying in the owners' proximity.
The preferred running speed correlated with the general aggression facet of the aggression toward people factor (r = 0.95, n = 5, p = 0.015). More aggressive dogs ran faster during the walks.
In addition to being correlated with dominance rank (mentioned earlier), leading tendency was positively correlated with: age (r = 0.91, n = 5, p = 0.032), responsiveness to training (rS = 0.92, n = 5, p = 0.028), controllability (r = 0.98, n = 5, p = 0.003), and aggression towards people (r = 0.95, n = 5, p = 0.013). These relations indicate that those dogs that have a tendency to take the leading role during walks are more aggressive and dominant, and they are also more controllable by the owner, based on the personality questionnaires (Figure 3).
By analyzing the GPS trajectories of freely moving dogs and their owner during walks, we found significant differences in simple path characteristics of the individual dogs. The preferred running speed of Vizslas ranged from 1.5 to 4.0 m/s (5.4–14.4 km/h), they covered a 1.8–3.7× longer distance than the owner during a walk, and the usual distances from the owner ranged from 16 to 20 m. These results might be useful for conservation managers in establishing areas where dog walking is prohibited [30] and may also help in designing parks, as dog-walking is a popular method for increasing human physical activity (for a review, see [31]).
A directional correlational analysis [26], [27] revealed leader-follower interactions between the group members. We detected a loose but consistent hierarchical leadership structure. Due to the dynamic nature of the pairwise interactions, role reversals did occur during walks and an individual took the role of the leader in a given pair in about 73% (ranging from 57% to 85%) of their interactions, where directed leader-follower relationships were found. This ratio is of similar magnitude to the case of wild wolf packs with several breeding individuals, where leaders led for 78% of the recorded time, ranging from 58% to 90% [18]. The role of initiating common actions is also frequently interchanged between guide dogs and the owner [32] and between dogs during play [33]. But over a longer timescale, differences in leading tendency remained consistent; thus decision-making during the collective motion was not based on an egalitarian system in our sample.
Although the existence of an overall dominance hierarchy in dogs is debated [23], and the Vizsla is a “peaceful” breed, which, compared to other breeds, rarely fights with conspecifics [34], we detected a dominance hierarchy via a questionnaire assessing agonistic and affiliative situations [29]. We found that dominance rank and leadership were strongly connected. Dogs who tend to win in everyday fighting situations, eat first, bark more or first, and receive more submissive displays from the others, and have more influence over the decisions made during collective motion.
The correlation between leadership and dominance is consistent with a trend in ‘despotic’ social mammals [5], but probably not characteristic in wolves with several breeding individuals [18]. In large wolf packs (with 7–23 individuals), breeding individuals lead during travels, independently from dominance status. But this situation is relatively rare, as the typical wolf pack is a nuclear or extended family, where the only breeding male leads the pack during travel [17]. Unlike wolves, the dog is a promiscuous species, and in a group, there is usually no single pair of breeders [22]. In our family dog group, the highest ranking dog (V2) was neutered, which may suggest that both leadership and dominance have little or no relationship with reproductive behaviour in family dogs, consistent with observations in feral dogs in India [35]–[37].
We also investigated the relationship between leadership and personality to reveal which personality types occupy particular positions in the leadership network. We found that leaders/dominants were more responsive to training, more controllable, and more aggressive than followers/subordinates. Other data also suggest that dominance cuts across different contexts and is correlated with boldness, extraversion, and exploratory tendencies in several taxa [38], and assertiveness in wolves [18], but reported links between personality and leadership are rare [14].
Age was a reliable indicator of leadership and dominance. Several studies have reported a positive correlation between age and dominance [39]. Age-related dominance might be due to greater fighting skills (e.g. [40]) or enhanced possibility of forming alliances with other individuals, among other factors [41]. If rank acquisition is learnt at an early age with regular reassessments of dominance, younger dogs may remain subordinate, long after initial body weight differences have disappeared. In our group, both dams were dominant over their adult offsprings, and each adult Vizsla dominated the juvenile Vizsla, which supports the hypothesis that the acceptance of subordinate status within a dog group is probably mediated by conditioning.
Not only leadership and dominance, but movement characteristics were also related to personality. Fearful and excitable dogs maintained a longer distance from other dogs. More controllable and dogs returned to the owner more often, while less companionable dogs spent more time far from the owner. Surprisingly, more aggressive dogs ran faster during the walks. As male dogs harvest more game than females in preindustrial societies [42], and experimental evidence on mice suggests that testosterone increases persistence of food searching in rodents [43], higher speed might be related to testosterone levels. Note, however, that even the most “aggressive” score was relatively low in our sample (2.67 out of the maximum 8).
Social organization and social structure vary among populations [44], and in the case of dogs, they vary among breeds and groups [45], thus group decision-making processes are expected to vary accordingly [46]. The main limitation of our study is the low sample size. Observing other groups and breeds may provide different results. For example, the hierarchical network of sled dogs which work as a team with a lead dog [47] is more robust than that of our sample. It would also be interesting to investigate what happens with the leadership network if the owner runs or rides a bike, and her speed is comparable to the dogs' speed.
To summarise, by using GPS devices we found that the leader and follower roles are dynamically interchanged during walks, but are consistent over a longer timescale. The leader-follower network was hierarchical, and the dogs' positions in the network correlated with dominance order derived from everyday life situations. Leadership also correlated with age and personality traits such as trainability and aggression.
Our findings on the connection between variables extracted from GPS trajectory data, dominance rank, and personality traits could pave the way for automated animal personality and dominance measurements. As dogs are ideal models of human social behaviour [48], [49] and social robots [50], the present study may also be applied to measure social interactions in humans, as in the case of parents walking with their children, or humans interacting with robots.
Non-invasive studies on dogs are currently allowed to be done without any special permission in Hungary by the University Institutional Animal Care and Use Committee (UIACUC, Eötvös Loránd University, Hungary). The currently operating Hungarian law “1998. évi XXVIII. Törvény” – the Animal Protection Act – defines experiments on animals in the 9th point of its 3rd paragraph (3. §/9.). According to the corresponding definition by law, our non-invasive observational study is not considered as an animal experiment. The owners volunteered to participate and gave written consent to the publication of the photos.
6 dogs (5 Hungarian Vizslas and one mixed breed; labelled V1 to V5 and M, respectively) and their owner took part in the experiments. Demographic characteristics are shown in Table 2. Photos of the subjects are presented in Figure S2, kinship is depicted in Figure S3.
GPS data were collected during 14 daily walking tours, each lasting about 30–40 minutes between 2 May 2010 and 25 November 2010. We analysed 823,148 data points. The high-resolution GPS devices were attached to the dogs with ordinary harnesses (Figures S1, S2), while the owner carried one device attached to her shoulder. The 5 Hz custom-designed GPS devices had a time resolution of 0.2 s and previous independent tests with the same devices showed a spatial accuracy of 1–2 m ([4] – Text S1). Weighing only 16 g, and with dimensions of 2.5 cm×4.5 cm, it is reasonable to suppose that the devices did not hinder the dogs' movements.
The group always walked on the same open grassy field, with the approximate dimensions of 500×1000 m, near Budapest, Hungary (located 47°25′17″N latitude, 19°8′45″E longitude).
The task of the owner was walk continuously and with a constant speed as far as possible during the walks. The dogs were allowed to walk and run freely, and the owner called the dogs back to herself only when she noticed some kind of danger, which happened on just a few occasions. Graphical summary of the Procedure is presented in Figure S1.
The personality of the dogs was quantified using two questionnaires that were completed by the owner at the end of the GPS measurements.
To extract information concerning the interactions between group members, we used a directional correlation analysis [26] with a time window to quantify the fast, joint direction changes of pairs. Highly correlated direction changes of pairs are usually found only when two dogs interact by running a part of a loop together. The timescale of the owner's direction changes was much larger than that of the dogs, and – due to the short time window and the typically small time delays – it was not covered in the calculations. Therefore interactions between the owner and the dogs were not detected with this method. However, we know that the owner was walking on a predetermined route, and clearly led the whole group on a longer time scale (Figure 1, Figure S5 and Video S1).
We calculated directional correlation values for all short trajectory segments that were in a 6 s time window (twin; in other details the method was identical to [26]), thus isolating short-term effects. We used twin = 6 s in the study, but the exact choice for the time window size has no substantial effect on the results (Figure S8). A local interaction event was defined to exist when corresponding trajectory segments had a higher correlation value than Cmin = 0.95 (Figure S7).
To extract leading tendency differences between members of pairs, the temporal directional correlation delay times (τij) were determined with the maximal correlation value. Positive τij values correspond to leading events when dog i leads dog j, as the direction of motion of i is ‘copied’ by j delayed in time. For each pair, leading-following events corresponding to different τij time delays were summed for each case in a walk, and for all 14 walks measured. For a detailed description of the applied method and a histogram of the found time delays between dog i and dog j, see Figure 2A and Figure S8.
If a clear maximum of the time delay histogram exists, it indicates frequent interaction between a dog pair at and near a well-defined time delay (see detailed description in Text S1 and Figures S8, S9). In many cases it can be seen from the histograms of those dog pairs where interaction was found (Figure 2A shows a typical example) that the leading and following roles (i.e. the sign of the time delay) are dynamically changing during a walk and also between walks. Significant deviation from zero in the location of the maximum value indicates that the dogs in the current pair have different leading propensities, suggesting a directed leader-follower interaction. The full width at half maximum of the histogram (see Text S1) characterises how stable the leader-follower relationship between a pair is.
We constructed an interaction network based on the detected interactions and leading tendency differences (Figure 2B, see also Figure S10). An edge (or link) indicates detected interaction between a dog pair. In those pairs where there is a significant difference in leading tendency we defined a directed edge (pointing from the dog who was found to lead more frequently to the one who more often assumes the role of follower).
The result of the method using the directed edges of the leadership network to characterise active connections was confirmed in an independent way. From the positional data we determined whether members of a pair spend more time in the close vicinity of each other compared to a randomized case (for more details see Text S1). This vicinity method does not require synchronised movement from interacting pairs. The resulting “social” network of the directional correlation and the vicinity method are in high correlation (two-tailed Pearson correlation, r = 0.600, n = 15 (number of possible pairs), p = 0.018).
|
10.1371/journal.ppat.1000898 | The Type III Effectors NleE and NleB from Enteropathogenic E. coli and OspZ from Shigella Block Nuclear Translocation of NF-κB p65 | Many bacterial pathogens utilize a type III secretion system to deliver multiple effector proteins into host cells. Here we found that the type III effectors, NleE from enteropathogenic E. coli (EPEC) and OspZ from Shigella, blocked translocation of the p65 subunit of the transcription factor, NF-κB, to the host cell nucleus. NF-κB inhibition by NleE was associated with decreased IL-8 expression in EPEC-infected intestinal epithelial cells. Ectopically expressed NleE also blocked nuclear translocation of p65 and c-Rel, but not p50 or STAT1/2. NleE homologues from other attaching and effacing pathogens as well OspZ from Shigella flexneri 6 and Shigella boydii, also inhibited NF-κB activation and p65 nuclear import; however, a truncated form of OspZ from S. flexneri 2a that carries a 36 amino acid deletion at the C-terminus had no inhibitory activity. We determined that the C-termini of NleE and full length OspZ were functionally interchangeable and identified a six amino acid motif, IDSY(M/I)K, that was important for both NleE- and OspZ-mediated inhibition of NF-κB activity. We also established that NleB, encoded directly upstream from NleE, suppressed NF-κB activation. Whereas NleE inhibited both TNFα and IL-1β stimulated p65 nuclear translocation and IκB degradation, NleB inhibited the TNFα pathway only. Neither NleE nor NleB inhibited AP-1 activation, suggesting that the modulatory activity of the effectors was specific for NF-κB signaling. Overall our data show that EPEC and Shigella have evolved similar T3SS-dependent means to manipulate host inflammatory pathways by interfering with the activation of selected host transcriptional regulators.
| Bacterial intestinal pathogens have evolved distinct ways of colonizing the gut and causing disease. Enteropathogenic E. coli (EPEC) and its close relative enterohemorrhagic E. coli O157:H7 (EHEC) are extracellular pathogens that cause a characteristic lesion on the intestinal mucosa known as an attaching and effacing lesion. In contrast, Shigella is an intracellular pathogen that invades the intestinal mucosa and spreads from cell to cell. Both pathogens utilize a bacterial type III secretion system that “injects” virulence effector proteins into the host cell upon contact. We have discovered that an effector shared by EPEC/EHEC and Shigella, known as NleE or OspZ, as well as another EPEC/EHEC effector, NleB, inhibit the host cell inflammatory response by preventing translocation of the immune regulator NF-κB to the cell nucleus. Thus, although EPEC/EHEC and Shigella have evolved different colonization strategies, they share a common virulence determinant that suppresses the inflammatory response of the host, and both pathogens mediate a multi-effector attack on NF-κB signaling.
| Many bacterial pathogens have the ability to “inject” virulence effector proteins into the host cell using a type III secretion system (T3SS). The effector proteins perform a variety of functions that allow the pathogen to persist in the host and cause disease [1]. Enteropathogenic Escherichia coli (EPEC) and enterohemorrhagic E. coli (EHEC) deliver T3SS effector proteins to the intestinal epithelium that mediate attaching and effacing lesion (A/E) lesion formation. A/E lesions are characterized by intimate bacterial attachment, effacement of the brush border microvilli and actin pedestal formation [2]. T3SS effectors from other pathogens such as Salmonella and Shigella have various roles in invasion, intracellular survival and the inhibition of innate immune responses through targeting host inflammatory signaling pathways [1]. Many of the T3SS effectors belong to conserved protein families that are found in a range of bacterial pathogens of plants and animals. For example, the OspF family of T3SS effectors from Shigella, Salmonella and Pseudomonas exhibit phosphothreonine lyase activity and induce irreversible dephosphorylation of mitogen-activated protein kinases (MAPKs) in the host cell nucleus [3], [4], [5]. In Shigella, this leads to gene-specific repression of a subset of NF-κB regulated genes, including IL8 [3]. Given the remarkable specificity of their biochemical function, the discovery of the mechanism of action of T3SS effectors remains an important step towards understanding the pathogenesis of many bacterial infections.
The activation of gene expression during inflammation is tightly regulated by transcription factors such as NF-κB. The NF-κB/Rel family comprises five members that share an N-terminal Rel homology domain that mediates DNA binding, dimerization and nuclear translocation [6]. The p65, c-Rel and RelB NF-κB subunits have an additional C-terminal transactivation domain, which strongly activates transcription from NF-κB-binding sites in target genes. The p50 and p52 subunits lack the transactivation domain but still bind to NF-κB consensus sites and act as transcriptional repressors [6]. The most abundant form of NF-κB in mammalian tissues is a p65/p50 dimer that activates the expression of multiple cytokine genes in response to inflammatory signals. In resting cells, NF-κB subunits are held in an inactive form in the cytoplasm by binding IκB proteins. Activation of NF-κB signaling stimulates the phosphorylation and proteosomal degradation of IκB, whereupon the NF-κB dimer is transported into the nucleus through the nuclear pore complex [6]. The canonical NF-κB pathway is stimulated by a range cell surface receptors such as the TNF receptor, IL-1 receptor, Toll-like receptors and T-cell receptor. Although the upstream components of these pathways vary, they converge at the point of IκB kinase complex (IKK)-mediated phosphorylation of and subsequent degradation of IκB [7].
Enteropathogenic Escherichia coli (EPEC) and enterohemorrhagic E. coli (EHEC) utilize a type III secretion system (T3SS) to deliver effector proteins to the intestinal epithelium that induce actin pedestal formation [2]. Multiple additional effectors are transported into the host cell where their targets and effects on host cell biology remain largely uncharacterized [8]. NleE is a highly conserved 27 kDa T3SS effector protein of A/E pathogens encoded in an operon with the 38 kDa effector, NleB. The NleE homologue in the invasive pathogen, Shigella, is called OspZ [9], [10]. While investigating the effect of EPEC infection on NF-κB activation, we observed that wild type EPEC prevented translocation of the p65 subunit of NF-κB to the host cell nucleus, whereas an nleE mutant was defective for this activity. Here we report that NleE inhibits p65 nuclear translocation, thereby reducing the IL-8 response during bacterial infection, and that OspZ shares this activity. In addition, we show that NleB suppresses NF-κB activation but appears to act in distinct manner to NleE.
Recent work has shown that EPEC and EHEC infection inhibits inflammatory cytokine production and NF-κB activation [11], [12], [13]. Previously, we found that translocated NleE localised to the host cell nucleus and we postulated that NleE had a role in subverting innate immune signaling [10]. Here we investigated the effect of NleE on NF-κB activation during EPEC infection. As actin accumulation beneath adherent EPEC depends on successful translocation of the T3SS effector, Tir [2], we used the fluorescent actin staining (FAS) test as a general marker for the translocation of T3SS effectors. HeLa cells were infected with wild type EPEC E2348/69, a T3SS (escF) mutant, an nleE deletion mutant of EPEC or an nleE mutant complemented with full length nleE. Cell monolayers were either infected for 4 h and left unstimulated or infected for 90 min and stimulated with tumour necrosis factor α (TNFα) or interleukin-1β (IL-1β) for 30 min. Nuclear translocation of the p65 NF-κB subunit was visualised by immunofluorescence microscopy of FAS-positive cells for EPEC E2348/69, ΔnleE and ΔnleE (pNleE) (Fig. 1A) and cells with adherent bacteria for ΔescF. In unstimulated cells, there was no significant difference in p65 nuclear exclusion between wild-type infected cells and the escF mutant (Fig. 1B). In contrast, in cells stimulated with TNFα or IL-1β, wild type EPEC E2348/69 inhibited p65 transport to the nucleus, whereas the escF mutant had little inhibitory effect on p65 nuclear translocation (Fig. 1B). The nleE mutant also showed greatly reduced inhibition of p65 nuclear transport in response to TNFα or IL-1β compared to wild type EPEC E2348/69 which was restored upon complementation of the nleE mutant with a copy of full length nleE. Similar results were obtained in response to IL-1β in Caco-2 intestinal epithelial cells (Fig. S1).
Caco-2 intestinal epithelial cells were then utilised to determine if the inhibition of p65 translocation resulted in the suppression of IL-8 production. Caco-2 cells were incubated with wild type EPEC E2348/69, a T3SS (espB) mutant, the nleE mutant or the nleE mutant complemented with nleE. Cells were then stimulated with IL-1β [14]. Compared to infection with the espB mutant, wild type EPEC inhibited IL-8 production from Caco-2 cells. The nleE mutant showed a diminished capacity to inhibit IL-8 production, which was complemented to wild type levels upon reintroduction of full length nleE (Fig. 2A). A similar trend was observed in CaCo-2 cells left unstimulated, although the differences were not as great as in IL-1β-stimulated cells (Fig. 2B). Real time PCR analysis of IL8 from Caco-2 cells infected with derivatives of EPEC and stimulated with IL-1β, showed that levels of IL8 mRNA were suppressed by NleE expression (Fig. 2C).
To determine if NleE was sufficient for the inhibition of p65 nuclear translocation, we expressed GFP-NleE or GFP transiently in HeLa cells. Upon stimulation with TNFα, the NF-κB p65 subunit was excluded from the nucleus in the presence of ectopically expressed GFP-NleE but not GFP alone (Fig. 3A). To ensure that this effect was not an artefact arising from the over expression of GFP-NleE, we examined a range of transfected cells exhibiting low, moderate and high GFP expression. Even in cells exhibiting low levels of GFP-NleE, p65 was excluded from the nucleus upon stimulation with TNFα (Fig. 3A and data not shown). We also investigated the effect of NleE on nuclear localization of other NF-κB proteins, c-Rel and p50. Similar to p65, NleE blocked nuclear translocation of c-Rel in response to TNFα (Fig. 3B), however p50 nuclear localization was unaffected by the presence of NleE (Fig. 3C). A dual-luciferase reporter system measuring the activation of NF-κB-dependent transcription confirmed the absence of NF-κB p65 nuclear activity in GFP-NleE transfected cells stimulated with TNFα (Fig. 3D).
To determine if NleE-mediated inhibition of signaling affected other transcription factors, we tested the ability of NleE to inhibit nuclear translocation and activation of STAT1 and STAT2. Following stimulation with interferon α, both STAT1 and STAT2 were translocated to the nucleus in the presence of GFP or GFP-NleE (Fig. 4A and B). NleE also had no impact on STAT1/2 activation using a ISRE-Luc luciferase reporter (Fig. 4C) [15]. This suggests NleE acts on a subset of signaling pathways that includes NF-κB but not STAT1/2.
To determine if the function of NleE and OspZ was conserved across A/E pathogens and Shigella, GFP-NleE and GFP-OspZ fusions generated from enterohemorrhagic E. coli O157:H7, Citrobacter rodentium, Shigella boydii and Shigella flexneri were expressed by transfection in HeLa cells. NleE from EHEC O157:H7 and C. rodentium, and full length OspZ from S. boydii and S. flexneri serogroup 6 inhibited NF-κB activation and p65 nuclear translocation (Fig. 5A and B). In contrast, OspZ from S. flexneri serogroup 2a which carries a 36 amino acid truncation at the C-terminus (Fig. S2), had no impact on NF-κB activation and did not block p65 nuclear translocation. Further screening of three clinical isolates of S. flexneri 2a revealed that all strains encoded a truncated OspZ protein. Similar to the truncated form of OspZ from S. flexneri 2a, a GFP-NleE truncation lacking the C-terminal 36 amino acid residues, GFP-NleE1-188 was unable to prevent NF-κB activation. However, the C-terminal region was not sufficient for this antagonism, as GFP-NleE188-224 did not inhibit NF-κB activation (Fig. 5C). Interestingly, in contrast to the other full length GFP-NleE/OspZ fusion proteins, GFP-OspZ from S. flexneri 6 and S. flexneri 2a was largely excluded from the nucleus (Fig. 5B). Although the molecular basis of this is unknown, it may indicate that the mechanism of action of NleE/OspZ is in the cytoplasm of the cell since GFP-OspZSF6 inhibited NF-κB activation to the same degree as GFP-NleE (Fig. 5A).
To examine further the C-terminal region of NleE and OspZ, we performed a deletion analysis of NleE. Whereas truncated NleE1-214, inhibited NF-κB activation to the same degree as full length NleE, truncated NleE1-208 was inactive (Fig. 6A). This suggested that the region between amino acid residues 208 and 214, with the motif IDSYMK was critical for NleE function. We then constructed a deleted form of NleE lacking amino acids I209DSYMK214 which was unable to inhibit NF-κB activation (Fig. 6A) as was a corresponding deleted form of OspZ from S. flexneri 6 lacking amino acids I209DSYIK214 (Fig. 6B). In fact GFP-OspZΔIDSYMK appeared to have a pro-inflammatory effect even in unstimulated cells (Fig. 6B).
The C-terminal 50 amino acids of NleE and OspZ are strongly predicted by Jpred [16] to form an alpha-helical region, therefore it was possible that the IDSY(M/I)K deletion had disrupted protein secondary structure (Fig. 6C). To account for this possibility and preserve the predicted alpha helix, we changed all six amino acids to alanine to generate GFP-NleE6A. NleE6A had the same alpha-helical prediction by Jpred as NleE (Fig. 6C). Similar to GFP-NleEΔIDSYMK, GFP-NleE6A was unable to inhibit NF-κB activation (Fig. 6D), and NleE6A delivered by the T3SS during infection was also unable to inhibit p65 nuclear translocation (Fig. 1B). Further mutation of individual amino acids within the IDSYMK motif of NleE to alanine did not make a significant difference to NF-κB inhibition compared to the fulllength GFP-NleE fusion (Fig. 6D). Expression of all GFP-NleE derivatives was tested by immunoblot using anti-GFP antibodies to ensure that differences in activity were not due to differences in the levels of GFP fusion proteins (Fig. S3).
To determine if variations in amino acid sequence at the C-termini of NleE and full length OspZ had any functional significance (Fig. 6C), we performed a domain swap by exchanging the last 40 amino acids of NleE with the last 46 amino acids of OspZ and vice versa. Both chimeric forms of NleE and OspZ (NleE-OspZcterm and OspZ-NleEcterm) were fully functional (Fig. 6E) and they retained their native localization (data not shown). Therefore these regions were functionally interchangeable, suggesting that NleE and OspZ use the same molecular mechanism to inhibit NF-κB activation.
To compare the activity of NleE with another T3SS effector, NleH, reported to interfere with NF- κB activation [17], we generated GFP-NleH1 and GFP-NleH2 fusions from EPEC E2348/69 and tested these for their ability to inhibit NF-κB activity following stimulation with TNFα. In this system, NleE showed greater inhibition of NF-κB activation than either NleH1 or NleH2 (Fig. 7A). To test for possible non-specific effects on NF-κB activation of effector over-expression by transfection, we also tested the effect of NleD and NleB on NF-κB activation following stimulation with TNFα. While NleD had no effect on luciferase induction in response to TNFα, NleB inhibited NF-κB activation as effectively as NleE (Fig. 7A). NleB is encoded directly upstream from NleE and this organization is highly conserved among A/E pathogens [18]. Fluorescence microscopy of GFP-NleB transfected cells stimulated with TNFα confirmed that GFP-NleB inhibited p65 translocation (Fig. 7B).
NleE inhibited NF-κB activation in response to both TNFα and IL-1β so we tested the ability of NleB to inhibit IL-1β signaling. Whereas, GFP-NleE was effective against both TNFα and IL-1β stimulation, GFP-NleB had no effect on NF-κB activation stimulated by IL-1β (Fig. 7C). This suggested that the two effectors act at different points in the NF-κB signaling cascade. To examine the effect of NleE and NleB on other signaling pathways, we used an AP-1 reporter to monitor JNK/MAPK signaling. Neither NleE nor NleB inhibited AP-1 activation by phorbol 12-myristate 13-acetate (PMA) (Fig. 7D), and NleB also had no effect on STAT1/2 activation (data not shown). This suggests that the effectors target only a subset of signaling pathways involving NF-κB.
To ensure that NleE and NleB translocated by the LEE-encoded T3SS conferred the same phenotype as ectopic expression of the effectors by transfection, we infected HeLa cells with wild type EPEC E2348/69 and a double island mutant that lacked the genomic regions, PP4 and IE6 [19]. The double island mutant was used to eliminate genes encoding NleE and NleB in IE6 as well as NleB2, a close homologue of NleB, encoded in PP4 [19]. The ΔPP4/IE6 island mutant was complemented with pNleE or pNleB to examine the contribution of each effector to the inhibition of p65 translocation. In unstimulated cells, there was no difference in p65 nuclear translocation between uninfected cells and those infected with wild type EPEC E2348/69 or the T3SS mutants (escN and escF) (Fig. 8A and B and Fig. 1B). This suggested that, over a 4 h infection, bacterial products such as flagellin and lipopolysaccharide were not sufficient to stimulate signaling. In contrast, the ΔPP4/IE6 island mutant induced substantial p65 nuclear translocation (Fig. 8B), which may indicate that the translocation and/or biochemical function of some effectors is proinflammatory. In infected cells, both NleB and NleE injected by the T3SS had the capacity to inhibit p65 nuclear translocation in response to TNFα but only NleE was effective in response to IL-1β (Fig. 8A and B).
Since EPEC infection has been reported to inhibit IκB degradation [12], a critical event in the activation of NF-κB, the effect of NleE and NleB on IκB degradation was examined here in TNFα and IL-1β stimulated cells. Ectopically expressed GFP-NleE and GFP-OspZ inhibited IκB degradation in response to both stimuli whereas GFP-NleB inhibited IκB degradation in response to TNFα only (Fig. 9A). GFP-NleE6A lacked the ability to inhibit IκB degradation as did GFP-NleEΔIDSYMK and GFP-OspZΔIDSYMK (Fig. 9B). In addition, we tested whether NleB and NleE delivered by the T3SS had the same effect as ectopically expressed protein on IκB degradation stimulated by TNFα and IL-1β. Wild type EPEC E2348/69, ΔnleE (pNleE), ΔPP4/IE6 (pNleE) or ΔPP4/IE6 (pNleB) inhibited IκB degradation in HeLa cells stimulated with TNFα but IκB degradation was not inhibited in cells infected with ΔPP4/IE6 (pNleB) and stimulated with IL-1β. This suggests that NleE and NleB act in different ways to interfere with NF-κB signaling (Fig. 9C).
The activation of NF-κB signaling is a critical host response to infection. In this study, we found that the T3SS effector NleE from EPEC prevented nuclear translocation of the p65 NF-κB subunit, leading to diminished IL8 expression and a compromised IL-8 response. The inhibition of p65 nuclear translocation occurred when NleE was expressed ectopically or when NleE was delivered through the T3SS by infection. We also observed that NleE inhibited nuclear translocation of c-Rel but not nuclear import of activated p50, STAT1 or STAT2. Both p65 and c-Rel are structurally similar and contain transcriptional activation domains that initiate gene expression [6]. In contrast, p50 lacks a transcriptional activation domain, so that p50/p50 homodimers act as transcriptional repressors. Thus, NleE appears to obstruct nuclear translocation of Rel family transcriptional activators while allowing nuclear import of a transcriptional repressor, resulting in the suppression of IL8 expression. The selectivity of NleE for p65 and c-Rel is not unprecedented as lack of nuclear translocation of p65 and c-Rel but not p50 was recently reported for oestrogen-induced inhibition of NF-κB activation, although the mechanism is unknown [20].
NleE is one of the conserved core type III effectors of A/E pathogens [19]. We observed that ectopically expressed NleE from EHEC O157:H7 and the murine pathogen, C. rodentium also inhibited NF-κB activation and p65 translocation. A close homologue of NleE, OspZ, is found in Shigella, however in S. flexneri 2a, OspZ is truncated to the length of NleE1-188 [10]. Both the truncated form of OspZ from S. flexneri 2a and a C-terminal 36 amino acid deletion mutant of NleE were inactive, suggesting that the C-terminus was critical for the immunosuppressive function of NleE. However, this region was not sufficient for inhibition of p65 nuclear translocation as a region encompassing the last 36 amino acids of NleE alone was unable to prevent NF-κB activation. A domain swap between NleE and OspZ of the last ∼40 amino acids showed that these regions were functionally interchangeable and we identified a 6-amino acid motif, IDSY(M/I)K, that was critical for both NleE and OspZ function.
Although A/E pathogens stimulate an inflammatory response in vivo and proteins such as flagellin are recognised by TLR5 [21], [22], previous work has suggested that A/E pathogens modulate that inflammatory response by inhibiting p65 nuclear translocation as well as IκB degradation [11], [12], [23]. Here, we found that NleE inhibited nuclear translocation of p65 by preventing IκB degradation in response to TNFα and IL-1β. In contrast, we found that NleB inhibited IκB degradation in response to TNFα only. Since TNFα and IL-1β signaling converges at the point of IKK phosphorylation (Fig. 9D) [24], NleE may act on IKK or IκB itself to prevent IκB degradation. TAK1 or other MAPK may also have involvement in IKK phosphorylation leading to JNK activation [24], however, JNK signaling, represented here by the AP-1 reporter, was not affected by NleE and so we predict that NleE interferes with IKK or IκB function directly. Indeed while this work was under review, Nadler et al reported that NleE inhibits IKK phosphorylation [25]. The authors also proposed that NleB assists the inhibition of IκB degradation by NleE [25]. Here we hypothesize that NleB acts upstream of IKK in the TNFα pathway since NleB did not inhibit IκB degradation in response to IL-1β (Fig. 9D). We therefore propose a model where NleE and NleB act at different points in the NF-κB signaling pathway and each plays a distinct role in the inhibition of p65 nuclear translocation. In the TNFα pathway, NleE and NleB have overlapping and somewhat redundant inhibitory roles as complementation of the ΔPP4/IE6 double island mutant with either NleE or NleB was sufficient to block p65 nuclear translocation. In the IL-1β pathway however, NleB was not able to compensate for the lack of NleE. Although we believe that NleB acts independently of NleE, these results do not exclude the possibility that in IL-1β stimulated cells, NleB acts in concert with NleE [25].
The fact that both NleE and NleB inhibit NF-κB activation raises the possibility that more effectors contribute to the suppression of innate signaling pathways. Although compromised compared to wild type EPEC, the nleE mutant showed significantly greater inhibition of IL-8 secretion than an T3SS mutant, which lacks the ability to translocate all T3SS effectors. While one of the additional effectors inhibiting p65 translocation is clearly NleB, a close homologue, NleB2 may also have anti-inflammatory activity and perhaps other effectors in the genomic islands, PP4 and IE6. In addition, NleH1 and NleH2 were recently reported to interfere with the activation of NF-κB by binding ribosomal protein S3 (RPS3), a co-factor of nuclear NF-κB complexes, and sequestering it in the cytoplasm [17]. We also found that ectopically expressed NleH1 and NleH2 inhibited NF-κB activation, but not to the same degree as NleE and NleB. Together these anti-inflammatory effectors may balance the action of other effectors that through their biochemical activity stimulate inflammatory signaling, as suggested by the ΔPP4/IE6 double island mutant, which showed increased p65 nuclear translocation in uninfected cells compared to a T3SS mutant. Therefore, despite the fact that EPEC and Shigella infection ultimately induces gut inflammation, we propose that NleE/OspZ and NleB contribute to pathogenesis by inhibiting an initial host inflammatory response to allow the bacteria to persist in the early stages of infection.
A multi-effector attack on NF-κB signaling occurs during Shigella infection, which modulates NF-κB activation through the effectors OspG and OspF [3], [26]. Our studies suggest that Shigella strains carrying full length OspZ have evolved a further distinct mechanism to modulate NF-κB signaling. This makes the absence of a functional OspZ protein in S. flexneri 2a curious and may also explain previous findings that OspZ from S. flexneri 2a potentially enhanced inflammation by inducing polymorphonucleocyte migration across a polarized epithelium [10]. The truncation rendering OspZ inactive in S. flexneri was serotype specific however, as S. flexneri 6 encoded functional full length OspZ, similar to S. boydii.
In this study, we have ascribed a function to the NleE/OspZ family of T3SS effectors shared by attaching and effacing pathogens and Shigella as well as the EPEC effector, NleB. Despite the remarkably different infection strategies of these two groups of pathogens, they appear to have a mutual need to inhibit the host inflammatory response during infection. NleE, NleB and OspZ are the latest T3SS effectors to target NF-κB activation and the expression of NF-κB-dependent genes. Neither NleE nor NleB inhibited STAT1/2 or AP-1 signaling, suggesting that the proteins target the NF-κB pathway specifically. The ongoing identification of T3SS effectors that act on this and other inflammatory pathways will continue to provide insight into the molecular mechanisms by which bacterial pathogens inhibit immune signaling and establish infection.
The bacterial strains and plasmids used in this study are listed in Table 1. The construction of vectors and culturing of bacterial strains for infection is described in detail in the supplementary methods (Protocol S1). EPEC strains were used to infect HeLa cells for 4 h without exogenous stimulation or for 90 min after which the media was replaced with DMEM supplemented with 20 ng/ml TNFα or 10 ng/ml IL-1β (eBioscience, San Diego, CA) and the infection was continued for a further 30 min. For Caco-2 cells, EPEC infection continued for 4 h after which the cells were washed, treated with 100 µg/ml gentamicin for 2 h. For mRNA analysis monolayers were incubated for 3 h in media supplemented with 50 ug/ml gentamicin with or without 5 ng/ml IL-1β. For analysis of IL-8 secretion, monolayers were infected for 4 h and incubated for 24 h in media supplemented with 50 ug/ml gentamicin with or without 5 ng/ml IL-1β. IL-8 secretion into cell culture supernatants was measured by ELISA (Peprotech EC). The expression of IL8 from total RNA was determined using the comparative quantification method included in Rotor-Gene 1.7 software (Qiagen) as described in the supplementary methods using gene specific primers (Protocol S1).
Plasmids were transfected into HeLa cells for ectopic expression of GFP fusion proteins using Lipofectamine 2000 in accordance with the manufacturer's recommendation (Invitrogen, Carlsbad CA, USA). Transfected HeLa cells were treated with 20 ng/ml TNFα, 10 ng/ml IL-1β or IFNα (500 U/ml; Calbiochem, La Jolla, CA,USA) for 30 min at 37°C and 5% CO2. Transfected or infected cells were fixed in 3.7% (wt/vol) formaldehyde (Sigma) in PBS for 10 min and permeablized with acetone-methanol (1∶1, vol/vol) at −20°C for 15 min. Following a 30 min blocking in PBS with 3% (wt/vol) bovine serum albumin (Amresco, Ohio, USA) samples were exposed to rabbit polyclonal anti-p65 (SC-109, Santa Cruz, Santa Cruz CA, USA), anti-c-Rel (#4727, Cell Signaling, Beverly MA, USA), anti-STAT1 (SC-345, Santa Cruz), anti-STAT2 (SC-476, Santa Cruz) or mouse monoclonal anti-p50 (2E6, Novus Biologicals, Littleton CO, USA). Antibodies were used at a 1∶100, or 1∶50 for anti-c-Rel, (vol/vol) diluted in blocking solution for 1 h at 20°C. Alexa Fluor 488 or Alexa Fluor 568 (Invitrogen) conjugated anti-mouse or anti-rabbit immunoglobulin G were used at 1∶2000. Coverslips were mounted onto microscope slides with Prolong Gold containing 4′,6-diamidino-2-phenylindole (DAPI; Invitrogen). For the fluorescence actin staining (FAS) test, HeLa and Caco-2 cells were infected with bacterial strains, fixed and permeabilized as described above and cells were incubated with 0.5 mg/ml phalloidin conjugated to rhodamine for 30 min. Images were acquired using a confocal laser scanning microscope (Leica LCS SP2 confocal imaging system) with a 100x/1.4 NA HCX PL APO CS oil immersion objective. Nuclear exclusion of NF-κB, STAT1 and STAT2 was quantified from at least 3 independent experiments for both transfection and infection studies.
To examine the activity of NF-κB, a dual luciferase reporter system was employed. HeLa cells were seeded into 24-well trays and co-transfected with derivatives of pEGFP-C2 (1.0 µg) together with 0.2 µg of pNF-κB-Luc (Clontech, Palo Alto CA, USA) and 0.04 µg of pRL-TK (Promega, Madison WI, USA). Approximately 24 h after transfection, cells were left untreated or stimulated with 20 ng/ml TNFα or 10 ng/ml IL-1β for 16 h. Firefly and Renilla luciferase levels were measured using the Dual-luciferase reporter assay system (Promega) in the Topcount NXT instrument. For each sample, the expression of firefly luciferase was normalized for Renilla luciferase measurements and NF-κB activity was expressed relative to unstimulated pEGFP-C2 transfected cells.
To measure the induction of STAT1 and 2, the IFN-α/β-responsive luciferase reporter plasmid p(9-27)4th(–39)Lucter (ISRE-Luc) [15] was used in combination with the Renilla luciferase plasmid pRL-TK. HeLa cells were transfected with both plasmids as described above and stimulated with IFNα (500 U/ml; Calbiochem) for 30 min. Luciferase activity was measured as described above. To measure the induction of AP-1, the cAMP response element (CRE)-dependent luciferase vector pAP-1-Luc was used in combination with the Renilla luciferase plasmid pRL-TK. HeLa cells were transfected with both plasmids as described above and stimulated with 25 ng/ml phorbol 12-myristate 13-acetate (PMA) for 30 min. Luciferase activity was measured as described above.
To test the effect of ectopically expressed NleE, NleB and OspZ and on IκB degradation, HeLa cells were mock transfected or transfected with pEGFP-C2 or pEGFP-NleE, pEGFP-NleB, pEGFP-OspZ and derivatives and incubated for 16 h before being left untreated or treated with TNF-α or IL-1β for 10, 20 or 30 min. Cell lysis was performed by incubating cells in cold lysis buffer (50 mM Tris-HCl pH 8.0, 150 mM NaCl, 5 mM EDTA, 1% NP-40) on ice for 5 min before collecting lysate and incubating on ice for a further 10 min. Cell debri was pelleted and equal volumes of supernatant were collected for SDS-PAGE. Proteins transferred to nitrocellulose membranes were probed with mouse monoclonal anti-IκBα (Cell signaling) diluted 1∶1000 or rabbit polyclonal anti-p65 (Santa Cruz) diluted 1∶1000. For infection studies, HeLa cells were infected with derivatives of EPEC E2348/69 for 90 min before stimulation with TNF-α or IL-1β for 30 min as described above.
|
10.1371/journal.pgen.1004831 | The Global Regulatory Architecture of Transcription during the Caulobacter Cell Cycle | Each Caulobacter cell cycle involves differentiation and an asymmetric cell division driven by a cyclical regulatory circuit comprised of four transcription factors (TFs) and a DNA methyltransferase. Using a modified global 5′ RACE protocol, we globally mapped transcription start sites (TSSs) at base-pair resolution, measured their transcription levels at multiple times in the cell cycle, and identified their transcription factor binding sites. Out of 2726 TSSs, 586 were shown to be cell cycle-regulated and we identified 529 binding sites for the cell cycle master regulators. Twenty-three percent of the cell cycle-regulated promoters were found to be under the combinatorial control of two or more of the global regulators. Previously unknown features of the core cell cycle circuit were identified, including 107 antisense TSSs which exhibit cell cycle-control, and 241 genes with multiple TSSs whose transcription levels often exhibited different cell cycle timing. Cumulatively, this study uncovered novel new layers of transcriptional regulation mediating the bacterial cell cycle.
| The generation of diverse cell types occurs through two fundamental processes; asymmetric cell division and cell differentiation. Cells progress through these developmental changes guided by complex and layered genetic programs that lead to differential expression of the genome. To explore how a genetic program directs cell cycle progression, we examined the global activity of promoters at distinct stages of the cell cycle of the bacterium Caulobacter crescentus, that undergoes cellular differentiation and divides asymmetrically at each cell division. We found that approximately 21% of transcription start sites are cell cycle-regulated, driving the transcription of both mRNAs and non-coding and antisense RNAs. In addition, 102 cell cycle-regulated genes are transcribed from multiple promoters, allowing multiple regulatory inputs to control the logic of gene activation. We found combinatorial control by the five master transcription regulators that provide the core regulation for the genetic circuitry controlling the cell cycle. Much of this combinatorial control appears to be directed at refinement of temporal expression of various genes over the cell cycle, and at tighter control of asymmetric gene expression between the swarmer and stalked daughter cells.
| The regulation of timing and ordered progression of cell cycle events is central to the survival of any organism and is one of the fundamental processes of life. The gram-negative α-proteobacterium Caulobacter crescentus (Caulobacter, hereafter) is an important model organism for the study of the regulation of cell cycle progression and asymmetric cell division, shown in Fig. 1A [1]–[3]. A hallmark of Caulobacter asymmetric cell division is that the daughter stalked cell immediately initiates DNA replication and the daughter swarmer cell has a period of motility before differentiating into a stalked cell and initiating chromosome replication. Control of cell cycle progression and asymmetric division occurs through coordinate regulation of transcription, protein phosphorylation, DNA methylation, protein localization, and protein degradation [1], [4]–[6]. A cyclical genetic circuit, comprised of five master regulator proteins, including DnaA, GcrA, CtrA, and SciP, and the DNA methyltransferase CcrM, drives the cell cycle [2], [4] (see Fig. 2B). The circuit regulates the transcription of more than 200 genes controlling sequential polar differentiation events including flagella biosynthesis, pili biosynthesis, chemotaxis complex formation, DNA-replication, and cell division [3], [7]–[12]. However, the mechanism of cell cycle control for only a subset of these has been described. To decipher the regulatory landscape that guides the cell cycle we need to identify transcription start sites (TSSs), measure their cell cycle stage-specific levels, and define the regulatory motifs within each cell cycle-regulated promoter.
Here, using a detailed map of the coding and non-coding features in the genome based upon ribosome profiling [13], we applied global 5′ RACE to map approximately 75 percent of the Caulobacter TSSs at single base-pair resolution and to measure the abundance of RNAs carrying a 5′ tri-phosphate (5′ PPP) group. This was done at multiple time points during the cell cycle to determine the timing of activation of TSSs. We also identified binding sites of the key cell cycle regulatory transcription factors directly upstream of the TSSs. When multiple TFs were predicted to bind within these TSS-proximate regions, we were able to provide an initial estimate of the combinational control logic. For example, the core cell cycle circuit regulators DnaA and GcrA often regulate gene expression in combination with other transcription factors. For genes controlled by the cell cycle circuit regulator CtrA, the location and presence of full palindromic or half CtrA binding motifs, and co-appearance of SciP binding motifs, dictates the cell cycle timing of their transcriptional regulation. We discovered that 107 antisense TSSs positioned within Coding DNA Sequences (CDSs) are temporally regulated and identified 241 genes transcribed from multiple promoters whose activation is independently controlled, yielding different timing of TSS activation. Furthermore, we found internal promoters in operons that were independently regulated to alter the expression profiles of encoded genes. Cumulatively, these observations suggest that the regulation of Caulobacter TSS levels during the cell cycle is much more complex than previously reported and this dataset provides a powerful resource for the elucidation of the cell cycle regulatory circuit.
We used a global 5′ RACE (rapid amplification of cDNA ends) method in combination with Illumina high-throughput sequencing to obtain a single-nucleotide resolution map of TSSs and their cell cycle-dependent activation level. Isolated swarmer cells (0 minutes) were grown in M2G minimal media for 140 minutes until cell division (Fig. 1A). We collected cell samples at 8 time points during the cell cycle and carried out total RNA extraction to prepare an Illumina high-throughput sequencing library for each time sample (S7 Fig. and Materials and Methods).
At each TSS the 5′ nucleotide contains a 5′ PPP group, whereas products of nuclease cleavage yield either a 5′ mono-phosphate (5′ P) group or a 5′ hydroxyl (5′ OH) group. As many processed RNAs such as mature ribosomal and transfer RNA (rRNA, tRNA) have a 5′ P, we prepared two additional libraries from unsynchronized culture in mid-exponential growth in minimal media to selectively distinguish between RNA segments with 5′ PPP ends and 5′ P ends. In one of these two libraries, the RNA was treated with tobacco acid pyrophosphatase (denoted +TAP) to hydrolyze the 5′ PPP to 5′ P. The other library (denoted -TAP) was prepared without TAP treatment (S7 Fig.). Libraries were ligated with a 5′ sequencing adapter followed by reverse transcription using a random-hexamer primer conjugated with a second Illumina sequencing adapter (S7 Fig.). Since T4 RNA ligase reacts only with a 5′ P, removal of the pyrophosphate group allows for the ligation of the 5′ sequencing adapter [14].
About 180 million 38 bp reads were obtained from the ten sequencing libraries (8 TAP treated cell cycle time point libraries and the +TAP and -TAP libraries from an unsynchronized mid-log phase culture). The reads were aligned onto the Caulobacter NA1000 genome DNA sequence NC_011916 [15] using Bowtie 0.12.7 software [16]. Only non-rRNA reads that mapped to a unique genomic location without mismatches were used for our analysis. Sequencing reads for all libraries were normalized to the total number of non-rRNA reads in each library.
To identify TSSs, we used 34 biochemically-characterized TSSs as a positive control (S2 Dataset) and 24 tRNA 5′ P-sites as a negative control (n = 24) (S1 Fig.). We compared the natural log of the ratio (θ) between the number of sequencing reads obtained at the 5′ ends of positive and negative controls in +TAP/-TAP libraries. The θ's obtained for positive and negative controls fall into two separate normal distributions with slight overlap (Welch two sample t-test, df = 41.2, p-value = 1.3 e−11, S1 Fig.). To minimize false-positive TSSs, we set the threshold value at θ>0.26 which corresponds to approximately two standard deviations (α = 2.3%) above the sample mean of the negative control. We set the minimum read threshold in the +TAP asynchronous library to be 25. Using RNA-seq data from [13] we selected only TSSs that had more than a 35% increase in the downstream RNA-seq coverage (S1 Dataset). In total, this procedure identified 2,726 Caulobacter TSSs (S1 Dataset). The parameters we set for TSS identification allowed us to identify the TSSs for approximately 75% of genes and operons, in keeping with similar estimates of the percentage of TSSs reported for Listeria monocytogenes [17] and Escherichia coli [18]. Two factors are likely to contribute to the null identification of TSSs: strict parameter cutoff based on TAP-enrichment (S1 Fig.) and 5′ RACE dependence on a 5′ ligation, which is inefficient for RNAs that are highly structured at the 5′ region [19]. To verify our approach, we tested 36 of the TSSs identified by 5′ global RACE using β-galactosidase reporter assays and found that all 36 exhibited significant expression activity (S3 Dataset).
Canonical bacterial promoters contain binding sites for σ factors upstream of the TSS. A motif search of 50 bp upstream of the identified TSSs revealed a−35 (TTG) and −10 (A/T) binding site (n = 1,666) consistent with σ73 (RpoD), the most abundant housekeeping sigma factor in Caulobacter [20] (S2 Fig., S4 Dataset). The 5′ nucleotide of 93% (1,542/1,666) of the identified RpoD binding motifs are positioned between −34 to −37 bp upstream of the TSSs.
Based on recent functional re-annotations of the Caulobacter genome (CP001340) using ribosome profiling and computational analysis [13], [21], we categorize TSSs into four categories with overlap (Fig. 1B). TSSs located upstream of CDS are denoted as primary (P, 1443); those located in intergenic regions or upstream of annotated RNAs are denoted as non-coding (N, 155); those initiated from within coding sequences and transcribed in the same direction are denoted as internal (I, 344), and those transcribed in the opposite direction of the CDS are denoted as antisense (A, 503). There is overlap between primary and antisense TSSs (A+P, 84) and between primary and internal TSSs (I+P, 197). Thirty-three of the 93 previously characterized non-disruptable intergenic gaps in the Caulobacter genome [22] were found to contain a TSS within the non-disruptable gap, suggesting these TSSs may play a role in cell viability.
We also observed a slight directional bias in the number of TSSs encoded in the same direction as DNA replisome movement (n = 1544 co-directional, n = 1182 opposing) (Fig. 1C) to minimize collisions between DNA polymerase and RNA polymerase during chromosome replication [23]. Comparison of global TSS levels of the swarmer and stalked cell stage of the cell cycle revealed differences in the pattern of global site-specific TSS levels (Fig. 1D). We analyzed the direction of cell cycle-regulated TSSs peaking in the swarmer cell and find no significant directional bias (n = 21 co-directional, n = 19 opposing). As the swarmer cell does not actively replicate its chromosome, it is likely that these promoters have no selective pressure to be encoded co-directional with DNA replication.
To distinguish cell cycle-regulated TSSs from constitutively active TSSs, we implemented a modified Fourier Transform algorithm, similar to [24], including both minimum sequence read and expression fold-change cutoffs on the corresponding normalized time-course sequencing data of the 2,726 identified TSSs. Using this approach, we identified cell cycle-regulated TSS levels for 586 TSSs (Fig. 2A). In general, the cell cycle TSS levels measured by 5′ global RACE yield similar timing as the steady state mRNA levels as determined previously by microarrays [25] (S3 Fig.).
To improve upon lower resolution and coverage studies of transcription factor binding motifs [25], we used our base pair resolution TSSs and a new genome annotation [13], [21] to search in DNA segments upstream of cell cycle-regulated TSSs for binding motifs of the core cell cycle circuit (Fig. 2B): CtrA (Fig 2C, S4 Fig., S5–S7 Dataset), SciP (Fig. 2C, S4 Fig., S8–S9 Dataset), DnaA (Fig. 2C, S4 Fig., S10 Dataset) and the CcrM DNA methyltransferase (Fig. 2C, S4 Fig., S11 Dataset). To identify TSSs regulated by the GcrA transcription factor, we searched upstream of TSSs for DNA segments enriched in GcrA ChIP-seq signal (S12 Dataset) [26]. About 57% of cell cycle-regulated TSSs had upstream binding motifs for one or more of the four transcription factors or CcrM DNA methlytransferse in the core cell cycle regulatory circuit (Fig. 2C, S4 Fig., S5–S12 Dataset). We identified 199 cell cycle regulated TSSs with a single upstream regulatory binding motif for one of the known master regulators (DnaA, 29; GcrA, 34; CtrA, 89; SciP, 7; CcrM, 40). In addition, we found another 135 TSSs that have multiple master regulatory factor binding sites, suggesting they are under combinatorial control (S13 Dataset). The TSSs that are preceded by multiple regulatory motifs are enriched for genes encoding critical cell cycle proteins, including the genes of the core regulatory circuit itself.
We also identified binding motifs for sigma factors SigT (S5 Fig., S14 Dataset) and RpoN (S5 Fig., S15 Dataset). The TSSs with binding motifs for SigT, a cell cycle-regulated ECF sigma factor [11] that is also induced under stress [27], exhibited peak levels during the swarmer to stalked cell transition coincident with the expression pattern of sigT (S5 Fig.). The previously identified Caulobacter SigT binding motif (GGAAC-N16-CGTT, e-value = 1.9 e−39, n = 26) [25], [27] is located at the −35 bp region relative to the TSS (S5 Fig.). RpoN, a sigma factor induced upon nitrogen limitation and required for flagella gene expression in Caulobacter [28], [29], controls two classes of genes in both the swarmer to stalk transition and in flagellar genes. The previously identified RpoN binding motif (GGCNC-N4-CTTGC, e-value = 1.5 e−27, n = 33) [25], [30] is located between −35 bp and −25 bp relative to the TSS (S5 Fig.). The RpoD, SigT, and RpoN binding motifs together account for 63% (1725/2726) of the observed upstream TSSs regions. The remaining TSSs are likely activated by the additional 13 known sigma factors encoded in the Caulobacter genome.
The CtrA response regulator is a master transcriptional regulator of the Caulobacter cell cycle that was shown to directly control 95 cell cycle-regulated genes using ChIP-chip [11], [12]. We identified 183 cell cycle-regulated TSSs with an upstream CtrA binding motif (Fig. 3A) that were also enriched in CtrA ChIP-seq data [31] (S5–S7 Dataset). Among these are promoter regions of the cell cycle master regulators sciP, ccrM, and ctrA, a regulator of CtrA degradation rcdA, cell division proteins ftsK, ftsQ, and mipZ, the response regulator divK, flagellar genes, and 6 non-coding RNAs (S5–S7 Dataset). Surprisingly, we observed two classes of CtrA binding motifs, a full palindromic TTAA-N7-TTAA (Fig. 3A) and a half motif TTAA (Fig. 3A). Based on hierarchical clustering of the cell-cycle profiles we identified 3 groups of CtrA binding motifs (CtrA full, CtrA half repressor, and CtrA half activator). Expression of genes controlled by CtrA full (n = 52, S5 Dataset) mirrored CtrA protein levels in the predivisional cell stage (60–120 min) (Fig. 3A). The 5′ nucleotide of these motifs is positioned near the −35 region, consistent with CtrA activity as a transcriptional activator in the predivisional cell. Conversely, CtrA half repressor containing promoters (n = 24, S6 Dataset) exhibited an anti-correlated temporal pattern of TSS activation with the CtrA protein levels (Fig. 3A). These half sites were positioned over the −10 site, consistent with CtrA functioning as a transcription repressor, similar to the observed repression of ctrA P1 by CtrA [32] (Fig. 2B). Fifty eight percent (14/24) of promoters with a CtrA half repressor motif also contain a DnaA binding motif or a GcrA binding site. CtrA half site-containing promoters that function as activators (n = 107, S7 Dataset) correlated with CtrA protein levels throughout the cell cycle, with transcription activity occurring in both the swarmer and predivisional cell stages. These CtrA half activator binding sites are also near the −35 region, consistent with transcriptional activation (Fig. 3A). Indeed, gene expression profiling studies using microarrays from strains with altered CtrA activity [28] show good agreement with CtrA full and CtrA half activator in transcription activation and CtrA half repressor in transcriptional repression (S5–S7 Dataset). These data show that the activity and timing of transcription is controlled by the precise position of CtrA DNA binding upstream of the TSS. Additionally, two transcription factors (MucR 1/2) that act to regulate the S-G1 phase transition were reported to bind to CtrA target promoters. We found that 76% of cell cycle-regulated promoters under MucR 1/2 ChIP-seq peaks, as determined by the Fumeaux et al. [33], contains a CtrA binding motif, and 10% of the cell cycle-regulated promoters under MucR 1/2 ChIP-seq peaks contain a SciP binding motif. Those CtrA regulated promoters under MucR 1/2 peaks were more highly repressed in the stalk/early-predivisional cells than CtrA-regulated promoters without MucR 1/2 (S10 Fig.). This is consistent with the proposal by Fumeaux et al. [33] that MucR 1/2 specifically represses CtrA activated genes in the S phase while SciP specifically represses CtrA activated genes in the swarmer cell.
There are a total of 61 TSSs with a SciP binding motif [8] in the upstream promoter region, and among these are the promoter regions of ctrA, the DNA methyltransferase ccrM, and polar development protein podJ. Previously, the promoter regions of 30 genes were identified as targets of SciP by expression arrays [7], 15 of which were shown to be direct by ChIP-PCR [8]. Seven of the 15 promoter regions shown to directly bind SciP, were identified by the TSS motif search. As with the CtrA binding motifs, we found that the SciP motif falls into two categories: a palindromic motif GCGNC-N5-GNCGC and a half motif GCGAC (Fig. 3B) that was identified previously (reverse complement in [8]). TSSs with the palindromic motifs exhibited peak levels at 120 min (n = 29, S8 Dataset) and those with half SciP motifs exhibited levels peaking at 100 min (n = 32, S9 Dataset) (Fig. 3B). Both groups exhibited an anti-correlated cell cycle profile with SciP protein levels indicating SciP acts as a repressor, in agreement with previous reports [7], [8]. To confirm this role as a repressor, we showed that the mutation of the SciP site in three CtrA activated promoters leads to an increase in the promoter activity (S9 Fig.). The half motif is associated with early TSS repression and the full motif with repression later in the cell cycle (Fig. 3B). SciP binding motifs are predominantly found between −60 bp and −90 bp relative to the TSS (Fig. 3B–C) and 80% (49/61) of TSSs with a SciP binding motif also have a CtrA binding motif (Fig. 3C). On average the SciP ChIP-seq signal peak occurs upstream of the CtrA ChIP-seq signal in agreement with the upstream position of SciP binding motifs relative to CtrA (S8 Fig.). When SciP sites are combined with CtrA sites, SciP represses genes in the late predivisional cell and the swarmer cell (Fig. 3C) where SciP protein levels peak (Fig. 3B).
The DnaA protein directs the initiation of chromosome replication in addition to functioning as a transcription factor [10]. We identified DnaA binding motifs in 77 promoter regions of cell cycle regulated TSSs (S10 Dataset). DnaA regulates transcription of GcrA, FtsZ, PodJ, and components of the replisome and nucleotide biosynthesis proteins [10]. The DnaA binding motif occurs as the sole master regulator site in 29 promoter regions, while it is commonly accompanied at promoter regions by 25 CtrA, 19 CcrM, and 19 GcrA sites (Fig. 2C, S4 Fig.).
The GcrA protein is a master transcription factor that is activated by DnaA (Fig. 2B) [34] and whose protein levels are anti-correlated with CtrA [9]. We searched for enrichment of the GcrA signal in promoter regions of cell cycle regulated TSSs in the ChIP-seq dataset reported by [26] and found GcrA binding sites in promoters of CtrA P1, podJ, and mipZ in addition to 91 other promoter region (S12 Dataset). GcrA binding occurs as the sole master regulator site in 34 promoter regions, while it is accompanied by 30 CtrA, 29 CcrM, and 19 DnaA binding motifs (Fig. 2C, S4 Fig.).
Hemi-methylated GANTC sites in the Caulobacter chromosome are recognized by the CcrM DNA methyltransferase yielding 6-methyl adenines [35], [36]. The transcription of CtrA and DnaA is affected by the cell cycle-dependent methylation state of their promoters, a link that helps synchronize the progression of the core cell cycle regulatory circuit with the progress of DNA replication [3], [37], [38]. We identified a total of 96 TSSs with GANTC sites located within 50 bp upstream of the TSS that were activated at specific stages in the cell cycle (Fig. 4A–B, S11 Dataset). Eleven of the 96 TSSs contain more than one upstream GANTC site yielding a total of 108 GANTC sites within 50 bp upstream of cell cycle regulated TSSs. These cell cycle-regulated TSSs fell into three distinct temporal clusters. For the TSSs with the GANTC site positioned between −10 and −35 of the promoter region, the time of TSS activation occured between 40 and 60 minutes (Fig. 4B–C red). If the GANTC motif is positioned outside this region the TSS levels is lowest between 40 and 60 minutes (Fig. 4B–C blue). A third cluster of cell cycle-regulated TSSs that have GANTC sites equally distributed within 50 bp upstream of the TSS exhibits maximal levels between 80 to 100 minutes (Fig. 4C green). Fifty-eight percent of all TSSs with upstream GANTC motifs contained other master regulator binding motifs (Fig. 2C, S4 Fig.).
Based on this TSS study and a recent RNA-seq study [13], we identified 587 (503 A, 84 A+P) antisense transcripts in the Caulobacter genome. Only eight antisense transcripts for genes encoding transposases have been reported previously [39]. We found an additional 179 putative antisense TSSs that have RNA-seq coverage [13] below our mapping threshold (S16 Dataset). Despite the low RNA-seq coverage, 7 (out of 7 assayed) antisense TSS with low RNA-seq coverage had significant β-galactosidase activity when 75 bp of the promoter were inserted in front of the β-galactosidase gene (S3 Dataset), suggesting they are indeed antisense TSSs. 583/3,885 (∼16%) of Caulobacter CDSs have at least one antisense TSS; as compared to Helicobacter pylori (27%) [40], Escherichia coli (20%) [41], [42], and Mycoplasma pneumoniae (12%) [43]. Seventy-four antisense TSSs reside within essential genes including those that encode DnaA, CtrA, an RNA polymerase beta chain, and MreB.
Of the 766 (587+179) antisense TSSs in the Caulobacter genome, 107 are cell cycle regulated (Fig. 5A). Of these 107, 42 are within genes that are constitutively expressed and 13 are within genes that are cell cycle controlled. We observed for the spmX gene, that promotes the localization and activation of the DivJ histidine kinase [44], the timing of antisense transcription is correlated with the timing of sense transcription (Fig. 5B). Perhaps the antisense transcript stabilizes the spmX mRNA, as reported for the gadX mRNA by the antisense GadY transcript in E. coli [45]. In 12 of the 13 cell cycle-regulated antisense TSSs residing within with a cell cycle controlled gene (S17 Dataset), the levels of the antisense TSS and the corresponding cell cycle-regulated primary TSS peak at different times over the course of the cell cycle, as shown for CCNA_01391 (Fig. 5C). The antisense TSS with a SigT binding site in its promoter is induced at the swarmer to stalked cell transition. Upon the decrease in levels of the antisense TSS, we observe an increase in the levels of the primary TSS (Fig. 5C). The coordinated transcriptional patterns of these genes and their antisense TSSs, suggest that the antisense RNA may control gene expression.
Bacterial intergenic small non-coding RNAs (ncRNAs) have been shown to enable cells to adapt to environmental and physiological challenges [46]. We have separately reported 199 ncRNAs in the Caulobacter genome [13], including four new ncRNAs that are encoded in nondisruptable regions of the genome [22]. We identified 155 TSSs within intergenic non-coding regions (category N, Fig. 1B) using 5′ global RACE in minimal medium; these included 50 TSSs for tRNA or rRNA genes. In 46 of these 155 intergenic TSSs, the TSS matches the 5′ nucleotide in the ncRNA identified by RNA-seq [13] (S18 Dataset). While only 5 Caulobacter ncRNAs were previously observed to be cell cycle-regulated [39], [47], we identified 33 cell cycle-regulated non-coding TSSs activated in different phases of the cell cycle (Fig. 6A, S1 Dataset). One of these cell cycle-regulated TSS drives a ncRNA of 182 nt in length (CCNA_R0094) transcribed from within the chromosomal origin of replication (Fig. 6B) that appears to be essential [22].
We identified a total of 241 CDSs with multiple upstream TSSs (S19 Dataset) that appear to be independently controlled. Fifty seven of these CDSs containing multiple promoters are essential for viability [22]. In Caulobacter, only 18 cell cycle-regulated genes, including ctrA, dnaN, clpX, and rcdA, have been shown previously to be transcribed from multiple cell cycle controlled promoters using either tiling microarrays, nuclease protection, or primer extension assays [25], [32], [48], [49]. We found that 102 CDSs (42% of those with multiple upstream TSSs) have at least one cell cycle-regulated TSS, and 25 CDSs, including ctrA (3 promoters, Fig. 7A) and podJ (2 promoters, Fig. 7B), have more than one cell cycle-regulated primary TSS that are independently regulated.
The ctrA gene was previously shown to be transcribed from two promoters (P1 and P2) where P1 is activated by GcrA after the ctrA locus is replicated and P1 becomes hemi-methylated (Fig. 2B) [9], [37]. The P1 promoter is thereafter repressed by CtrA which simultaneously strongly activates P2 [32]. We confirmed the previously reported temporal sequence of P1 and P2 activation, and identified an additional cell cycle-regulated promoter, P3, located between P1 and P2 (Fig. 7A). A CtrA half repressor binding motif TTAA is located −14 bp upstream of the P1 TSS, in addition to a SciP binding motif GCGAC located −78 bp upstream, and a CcrM methylation site GANTC (▾) located at −28 bp upstream of P1 (S6, S9, S11 Dataset). A full CtrA binding motif TTAA-N7-TTAA is located at −39 bp upstream of P2 and at −14 bp upstream of P3, and a half SciP binding site is located at −74 bp upstream of P3 (S5, S9 Dataset). Both SciP and CtrA have been shown previously to bind at both locations [8], [32]. The full CtrA motif likely functions to activate P2 and simultaneously repress P3 because the 5′ nucleotide of the CtrA full motif is at −14 bp upstream of P3 (repression) and at −39 bp upstream of P2 (activation).
PodJ is an essential protein that mediates polar organelle development by contributing to the synthesis of pili and the control of the polar localization of the PleC kinase/phosphatase [50], [51]. Two cell cycle-regulated primary TSSs (P1 and P2) are located 120 bp and 22 bp upstream, respectively, of podJ (Fig. 7B). The levels of P2, which contains a CcrM methylation site GANTC (▾) −24 bp upstream (S11 Dataset) and a DnaA binding motif (CTCCACA) (Hottes et al, 2005) at −82 bp upstream (S10 Dataset), peaks at 40 min into the cell cycle during the swarmer to stalked cell transition. P1 contains a CtrA full binding motif TTAA-N6-TTAA at −49 bp upstream (S5 Dataset) and a SciP half binding motif GCGAC at −73 bp upstream (S9 Dataset) with P1 levels peaking at 100 min into the cell cycle in the predivisional cell, contributing to a second wave of podJ transcription.
For 77 out of the 241 CDSs transcribed with multiple primary TSSs one promoter is constitutively expressed while the other is cell-cycle regulated. For example mipZ (Fig. 7C), which encodes the essential division plane positioning ATPase MipZ [52], has a constitutive P1 and a cell cycle-regulated P2 promoter containing a CtrA half repressor motif at −10 bp upstream of the TSS (Fig. 7C, S6 Dataset). Activation of the P2 promoter results in an increase in the transcription of MipZ at the same time as transcription of FtsZ, thereby coordinating the temporal control of MipZ regulation of FtsZ function [6].
In many instances, the transcription of individual genes within the 848 operons is differentially regulated. There are 115 operons that have a cell cycle-regulated TSS upstream of the leading CDS. There are 52 operons (S20 Dataset) with internal cell cycle-regulated TSSs for downstream CDSs enabling independent cell cycle regulation of downstream operon genes. One example is the operon consisting of CCNA_00875, CCNA_00876, and CCNA_00877, where CCNA_00875 encodes a 7,8-dihydro-8-oxoguanine-triphosphatase, CCNA_00876 encodes a Flp/Fap pilin component protein, and CCNA_00877 encodes a protein of unknown function (Fig. 7D). In this operon, the P1 start site is constitutively active, but the P2 start site has a putative CtrA full binding motif TTAC-N7-TTCA upstream of it (S5 Dataset) where the 5′ nucleotide of the motif resides at -39 bp and a CcrM methylation site GANTC (▾) is at −35 bp (S11 Dataset). The P2 TSS has a cell cycle-dependent profile suggesting cell cycle-regulated expression of just the downstream CDSs in the operon.
We used a global 5' RACE method to identify TSSs at single base pair resolution for approximately 75% of Caulobacter genes and identified their pattern of cell cycle regulation. Previous studies have identified Caulobacter cell cycle-regulated steady state mRNA levels using microarrays or RNA-seq [10], [11], [25], [53]. Here, we measured the abundance of 5′PPP ends corresponding to the relative activity of individual promoters at different time points during the cell cycle. We identified 586 cell cycle-regulated TSSs and assayed their time and level of activation as a function of cell cycle progression. This detailed map of TSSs and their temporal profiles of transcription activation have revealed new layers of gene regulation by the cell cycle control circuit.
Genome-wide assays have shown that antisense transcription occurs in many bacteria species [40], [41], [43]. Further, antisense transcripts have been shown to regulate genes involved in a wide variety of processes such as photosynthesis in Synechocystis PCC6803 [54], acid and SOS response in E. coli [45], [55], virulence in Salmonella enterica [56], and iron transport in Vibrio anguillarum [57]. The 107 cell cycle-regulated antisense TSSs suggest that antisense regulation is a significant element of cell cycle regulation. Supporting this, we found that 23 of the 107 cell cycle regulated antisense TSSs have binding sites in their promoter regions for the core cell cycle regulatory factors. Antisense TSSs are also found within the CDSs of essential cell cycle-regulated genes, including dnaA, ctrA, spmX and mreB. Most previously described mechanisms of antisense transcripts involve base-pairing with the corresponding sense mRNA to alter the RNA stability [45], [54], translation [58], or transcription termination [57]. Depending on the context, the antisense RNA has been observed to effect either down-regulation or up-regulation of genes. In Caulobacter, for 12 cell cycle regulated primary TSSs, the TSS levels is anti-correlated with the time of the time of activation of the antisense TSS (S17 Dataset), suggesting that antisense transcripts may down-regulate the levels of their target mRNAs. Additionally, the Caulobacter antisense TSSs show multiple cell cycle expression patterns, suggesting that this mechanism is active in regulation of gene expression during all stages of the cell cycle. Twenty four promoters upstream of the 107 cell cycle-regulated antisense TSSs contain master cell cycle regulator binding motifs allowing them to be controlled directly by these cell cycle regulating transcription factors. Antisense TSSs are also abundant in Sinorhizobium meliloti [59], another α-proteobacteria. However, as master regulator binding motifs residing within the protein coding sequences have generally not been included in global analyses of the α-proteobacterial cell cycle transcription control circuitry [25], [60], [61] and it is not known whether antisense TSSs controlled by the cell cycle master regulators are conserved within the α-proteobacteria.
In Caulobacter, 199 intergenic ncRNAs have been identified in the genome including rRNAs and tRNAs [13], [39]. Of these, we identified 155 TSSs for ncRNAs (S1 Dataset) and for 33 of these the TSS levels are cell cycle-regulated (Fig. 6A). The functions of only two ncRNAs (non rRNA/tRNA/RNaseP/4.5S RNA) have been characterized: the tmRNA which rescues stalled ribosomes and alters the timing of replication initiation, and crfA which controls the carbon starvation response [47], [62], [63]. While ncRNAs perform many functions, a common ncRNA function in bacteria involves ncRNA base-pairing to mRNAs to regulate gene expression, sometimes mediated by the RNA chaperone Hfq [46], [64]. Four of the ncRNAs lie within non-disruptable intergenic gaps [22] suggesting that these ncRNAs may be essential for the regulation of Caulobacter cell cycle progression.
Across the Caulobacter genome we identified 241 CDSs with multiple upstream TSSs (S19 Dataset) including 57 of essential [22] and 102 of cell cycle regulated genes. The genes encoding the CtrA global cell cycle regulator and the PodJ polar differentiation factor have multiple upstream promoters with different cell cycle timing that modulates the pattern of expression of the genes. In the case of ctrA transcription, GcrA activates the hemi-methylated P1 promoter to initiate ctrA transcription followed by a boost in CtrA production from the subsequent auto-regulated activation of the P2 promoter [9], [32] (Fig. 2B, Fig. 7A). Based on the identification of a third temporally controlled promoter, P3, whose temporal pattern of activation is similar to that of P1, we suggest that P1 and P3 accelerate initial production of CtrA. Both P1 and P3 have a CtrA binding site in the −10 region upstream of each TSS, which are then repressed by CtrA. We also know, as mentioned earlier, that the subsequent expression of SciP represses ctrA transcription. The net effect appears to be aimed at modulating the shape of the pulse of CtrA production to make it stronger, yet shorter in time.
In the case of podJ, the function of the additional promoters seems to be to extend of the duration of PodJ production over a longer interval of the cell cycle. In other cases, such as the cell division gene mipZ, a cell cycle-regulated promoter is activated at a specific time during the cell cycle, corresponding to the time in which the protein product is needed, but the gene is also transcribed at a low level from a constitutive promoter, presumably ensuring that a low level of MipZ is always present during the cell cycle (Fig. 7C).
We found that 209 operons contain internal TSSs, 55 of which are cell cycle-regulated (S20 Dataset). In some operons, downstream genes are activated at different times in the cell cycle. Internal TSSs that exhibit independent cell cycle-regulated expression have nearby upstream transcription factor binding motifs. Additionally, we report separately that some promoters internal to operons lead to production of alternative shortened forms of the encoded protein [13], presumably changing the protein's function. The spatial ordering of multiple upstream promoters, in conjunction with promoters internal to operons, could conceivably have a regulatory impact since each mRNA has a different 5′ UTR sequence that would enable differential post-transcriptional control. The exciting implications of the combinatorial promoter logic possible for genes and operons with multiple TSSs remain to be explored.
The largest class of cell cycle-regulated TSSs are those that have one or more CtrA binding motifs in the promoter region (183 of the 586 of the cell cycle-regulated TSSs) (Fig. 2C, 3A, S4 Fig.). We observe two CtrA binding motifs, a full palindromic binding site and a half palindromic binding site (Fig. 3A). We find that the 5′position of the CtrA full palindromic motif relative to the TSS is most commonly at the −35 region corresponding to activation of transcription in the predivisional cell. Conversely, the 5′position of the CtrA half-palindromic motifs can either function as a repressor by binding over the −10, or an activator by binding over the −35 region. In predivisional cells, CtrA full-palindromic TSSs maximally activate at 80 to 100 min while those with CtrA half activator, do so later at the 120 min time point, likely due to the tighter binding affinity of the CtrA full site to CtrA∼P [65]. TSSs with CtrA half motifs are also active in the swarmer cell, while CtrA full motif TSSs are only active in the predivisional cell. We interpret this switch in TSS levels to be primarily due to overlapping control by SciP which inhibits CtrA activated promoters as reported by [7], [8]. About 80% (49/61) of TSSs containing an upstream SciP motif also have a CtrA motif (Fig. 2C, 3C, S4 Fig.). The SciP binding motifs are positioned between −60 and −90 upstream of the TSSs (Fig. 3B, S8 Fig.). SciP binds directly to DNA as shown in [8] and confirmed here by mutation of SciP sites. We have mutated both the half and full SciP sites in 3 CtrA activated promoters and in each case we observed an increase in promoter activity as measured by β-galactosidase activity, providing evidence that the SciP motif does indeed function to bind SciP and acts that bound SciP is a repressor of these promoters (S9 Fig.). In the presence of DNA containing both CtrA and SciP binding motifs, both CtrA and SciP become resistant to proteolysis [66]. Since direct interaction between CtrA and SciP has been demonstrated [7], [8], it seems likely that the combined DNA binding energy provides additional regulatory capacity. Interestingly, SciP and CtrA binding motifs are primarily positioned together with at least one motif present as a full palindromic binding site (91.8% of co-regulated promoters) and the co-positioning of SciP and CtrA half motifs occurs only in four co-regulated promoters. It is possible that SciP and CtrA require sufficient binding to DNA to form a stable complex that is not accomplished with weaker half sites.
A recent report by Fumeaux et al. [33] analyzed the top 50 SciP ChIP-seq peaks and found them to contain TTAACAT motifs, similar to the CtrA half binding motif. We performed motif searches of the SciP ChIP-seq peaks reported by Fumeaux et al. [33] using the CtrA half- and SciP half-binding motif presented in this paper, and found a total of 63 CtrA half motifs and 143 SciP half motifs with a P value less than 1−3. Both Fumeaux et al. [33] and our own analysis of their data revealed that only a subpopulation of SciP ChIP-seq peaks (47%) contain the TTAACAT motif. However, we found that 88% of the SciP ChIP-seq peaks contain a SciP binding motif. Of the 47% that contain a CtrA binding motif, 33/35 of these also contain the SciP binding motif. The SciP motif used in this study is based on previously reported direct footprints to DNA, ChIP-chip analysis and microarray analysis reported in Tan et al. [8], which is in agreement with a motif previously identified for a cohort of genes expressed at the same time in the cell cycle by McGrath et al. [25]. The SciP binding motif correlates with a larger percentage of the SciP ChIP-seq peaks than the CtrA motif. Both SciP half and CtrA half (TTAACAT) motifs are present in the CtrA promoter foot-printed by Tan et al. [8]. However, only protection of the SciP sites was observed with purified SciP. We suggest that the Fumeaux et al. ChIP-seq data analysis [33] missed many of the SciP motifs in the presence of stronger CtrA motif signals from peaks including SciP and CtrA co-regulated promoters. Indeed, the peaks of the Fumeaux et al. [33] SciP and CtrA ChIP-seq signal match the positions of their respective binding motifs (S8 Fig.).
We found that 57% of the cell cycle-regulated TSSs have upstream binding sites for known cell cycle transcriptional regulators (Fig. 2B-C, S4 Fig., S5-S12 Dataset) whose activity and protein levels oscillate in time [1], [4]. While 199 cell cycle regulated TSSs contain a single regulatory factor binding site, 135 have binding sites for 2 or more of these factors (S13 Dataset). We have not yet identified the regulatory factors (and their binding sites) that control the other 43% of the cell cycle regulated TSSs but we expect to find that the activity of these promoters are controlled by a second layer of regulatory factors that are turned on by the core circuit (Fig. 2B). Co-regulation among the 5 cell cycle master regulators acts to tune the timing of cell cycle transcription profiles, and we expect to find equally complex timing regulation in every genetic regulatory sub-system.
DnaA directly controls the initiation of DNA replication by binding to the origin of replication [67] and it also serves as a transcription factor for a large complement of cell cycle regulated genes [10]. We identified 77 cell cycle-regulated TSSs with a DnaA binding motif. Additionally, from our analysis of the ChIP-seq data of [26] we found 94 cell cycle-regulated TSSs with enriched upstream binding of GcrA, a transcriptional activator [9]. In contrast to to the temporal expression of CtrA-controlled genes, DnaA and GcrA regulated TSSs have a multitude of different cell cycle profiles with lower levels of activation (S6 Fig.). This increase in profile diversity is likely due to an increased number of co-regulated promoters for genes controlled by both DnaA and GcrA. Additionally, when DnaA and GcrA binding sites are combined with a CtrA binding site, the transcriptional profile appears to be dominated by the activity of CtrA.
Finally, we identified CcrM methylation sites within 96 cell cycle-regulated TSSs (Fig. 4A–B). We found cell cycle-controlled TSSs with GANTC motifs in these promoter regions, such as the ctrA P1 promoter which is activated by hemi-methylation, in agreement with a previous report of the control of this gene by methylation state of its promoters [37]. A cluster of these TSSs, located between −35 and −10 of the promoter region (Fig. 4B–C Red), are activated between 40 and 60 minutes. Another cluster of TSSs, with GANTC sites positioned outside of the −10 and −35 region, exhibit minimal TSS levels between 40 and 60 minutes (Fig. 4B–C Blue). While the underlying mechanism of methylation dependent transcription regulation in Caulobacter is unknown, the GcrA transcription factor has been implicated in the control of some GANTC containing promoters [68]. However, we find that only 31% of cell cycle regulated TSSs enriched for GcrA binding (from the published ChIP-seq data from [26]) also contain GANTC methylation sites within 50 bp upstream of the TSS (Fig. 2C, S4 Fig.). Additionally, the promoters of only 11 genes found to bind GcrA by ChIP-seq are differentially expressed in GcrA or CcrM depletion arrays [69].
Cumulatively, the newly-revealed complexity of transcriptional regulation that drives the Caulobacter cell cycle, coupled to multiple modes of post transcriptional regulation [13] has opened up new avenues of bacterial systems control architecture. The known Caulobacter cell cycle regulatory circuit, composed of four transcription factors and a DNA methyltransferase, drives stage specific expression of a majority of cell cycle-regulated promoters [2]. We found multiple examples of co-regulation using these five master regulators of the cell cycle regulatory circuit, and many of these examples involve essential genes and essential regulatory regions of the genome [22] (S13 Dataset). The presence of each master regulator is restricted to specific times in the cell cycle [8], [70], so that there are multiple cell cycle stage-specific regulatory options to tune the timing of transcription. The core Caulobacter cell cycle transcription regulatory network is conserved in a majority of α-proteobacteria [2], [60] and has been shown to be integrated with other regulatory pathways such as quorum sensing [71], plant symbiosis [61], [72], gene transfer agent production [73], [74], host cell infection [75], [76], and motility [73]. These α-proteobacterial species have adapted their cell cycle control circuits to their specific biological niches. Negative feedback by antisense regulation and the existence of many promoters with multiple TSSs that appear to modulate the cell cycle stage of mRNA production identifies yet other levels of regulatory control.
All 5′ Global RACE experiments were performed with C. crescentus strain CB15N (NA1000) [77]. Cells were grown to OD600 0.4 in M2G [78] and synchronized using standard procedures [77]. Aliquots were taken at 20 min intervals over ∼1 cell cycle (140 min), pelleted and immediately frozen in liquid nitrogen. Total RNA was isolated using Trizol Plus kit (Invitrogen) according to the manufacturer protocol. Ten micrograms of total RNA isolated from each aliquot of cells were treated with MICROBExpress (Ambion) to remove ribosomal RNA following the manufacturer protocol. Ribosomal RNA depleted total RNA was treated with 10 U of Tobacco Acid Pyrophosphatase (TAP) (epicentre) for 1.5 hours at 37°C. In addition, two sequencing libraries were prepared from an unsynchronized cell population grown to OD600 0.4 in M2G, one of which was not treated with TAP. Purified RNA was ligated with 10 pmol of RNA adapter (5'-ACACUCUUUCCCUACACGACGCUCUUCCGAUCU-3') using 25 U of T4 RNA Ligase 1 (New England BioLabs) for 12 h at 16°C. cDNA was then generated from purified RNA using SuperScript II Reverse Transcriptase (Invitrogen) with 10 pmol of primer (5'-CTCGGCATTCCTGCTGAACCGCTCTTCCGATCTNNNNNN-3') following standard manufacturer protocol then treated with 1 Unit of RNase H (Invitrogen) for 20 min at 37°C. cDNA was PCR amplified for 12 cycles using primers (5'-AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCCGATCT-3', 5'-CAAGCAGAAGACGGCATACGAGATCGGTCTCGGCATTCCTGCTGAACCGCTCTTCCGATCT-3'). DNA was size selected from 100 bp to 300 bp using agarose gel-electrophoresis and purified using QIAquick Gel Extraction Kit (Qiagen). 100ng of DNA was then treated with duplex-specific nuclease (evrogen) to remove ribosomal cDNA following standard manufacturer protocol and PCR amplified for 10 cycles.
Sequencing reads were mapped using Bowtie version 0.12.7 [16] to the C. crescentus NA1000 reference sequence (Genbank: NC_011916). All valid alignments were made, and only reads that mapped to a unique location were used with no mismatches allowed. Sequencing reads were normalized by the total number of non-ribosomal RNA reads for comparison. +TAP and -TAP reads were normalized with relation to each other, and reads from synchronized cells were normalized with relation to each other.
To differentiate between reads initiating from TSS and RNA processing sites, thirty-four previously biochemically characterized TSS with a minimum normalized read value of at least 30 in the +TAP library were used as a positive control, and twenty-four tRNA processing sites (n = 24) with a minimum normalized read value of at least 30 in the -TAP library were used as a negative control. The natural log value of the ratio (θ) between the number of sequencing reads obtained at the 5′ ends of positive and negative controls in +TAP/-TAP libraries was calculated. The mean and standard deviation of θ obtained for positive and negative controls were obtained (Welch two sample t-test, df = 41.2, p-value = 1.3 e−11, S1 Fig.). To minimize false-positives or Type I errors in TSS prediction, we set α = 2.3%, corresponding to θ>0.26. An increase of >35% of RNA-seq coverage [13] in a 38 bp window upstream compared to downstream of the TSS was also required.
The distance of mapped 5′ read locations upstream to CDSs in the Caulobacter genome was then determined. Reads that are mapped within 300 bp upstream of annotated CDSs or on the first base of the start codon were categorized as upstream promoters (category “P”); those within 300 bp upstream that are inside the upstream CDS but reside within the last 30% of the CDS if the CDS is longer than 600 bp or within the last 30% of the CDS if the CDS is shorter than 600 bp are categorized as internal upstream promoters (category “IP”). If the upstream CDS is oriented in the opposite direction as the TSS, then the TSS is categorized as an antisense upstream promoters (category “AP”). Unless already categorized as IP or AP for another CDS, reads that are mapped inside CDSs are categorized as internal promoters (category “I”); if oriented in the opposite direction of the CDS, the TSS is categorized as an antisense promoter (category “A”). All other reads were categorized as non-coding TSS (category “N”).
The total number of reads corresponding to each CDS as obtained from the +TAP library was calculated for both the plus and minus strands. Upstream promoter locations categorized as P, AP, or IP with reads contributing more than 10% of the total reads in the same orientation for a single CDS were kept. Antisense or internal promoter locations (A or I) with reads contributing more than 25% were kept. If multiple reads were mapped within a distance of ≤5 bp, the reads were summed and placed at the position with the highest number of reads in the +TAP library. TSS with ≥2 normalized reads in at least one of the cell cycle time points were kept for categories (P, AP, or IP); TSS with ≥10 normalized reads in at least one of the cell cycle time points were kept for other categories. Some TSSs categorized as N were actually upstream promoters for genes with unusually long 5′ UTRs. All TSSs categorized as N were manually curated to ensure accuracy.
On the cell cycle expression profile of each TSS, Fourier coefficients (coeff 0-7) were calculated using normalized reads obtained from synchronized cells. The maximum and minimum normalized read values (defined as a and b respectively) from time 0 min to time 140 min were also determined. Cell cycle-regulated TSS are those that meet the following three criteria: 1.) coeff1/(coeff1+ coeff2 + coeff3 + coeff4) ≥ 0.35 2.) ln(a) - ln(b) ≥ 1.1 3.) a ≥ 20.
Regulatory binding sites were identified from groups of promoter sequences upstream of temporally clustered TSSs using MEME [79] and the position of the 5′ nucleotide of the conserved motifs are numbered relative to the TSS sites. The RpoD motif was obtained by searching within 50 bp upstream of constitutively active TSSs. Cell cycle-regulated TSSs were clustered (k-means, 15 clusters) by normalized cell cycle profile where the maximum of normalized read value within time 0–140 min is equal to 1. From the resulting clusters, up to 100 bp of upstream sequences were used to search for enriched CtrA, SciP, SigT, CcrM, and RpoN motifs using MEME. DnaA binding sites from [10] were used to generate a position weight matrix (PWM) to search for DnaA motifs within 100 bp upstream of cell cycle-regulated TSSs using FIMO (p-value setting: <0.001) [80]. To search for GcrA binding, we searched for ≥3 fold ChIP-seq enrichment within 50 bp upstream of cell cycle regulated TSSs, using ChIP-seq data from [26]. The enrichment was calculated relative to the average coverage across the genome.
75 bp of the upstream sequence of 36 TSSs including primary (P), anti-sense (A), and intergenic non-coding (N) TSSs (see S3 Dataset) were cloned into the Bgl II and Xho I restriction sites of vector pNJH185 [81], resulting in transcriptional fusions with the lacZ reporter gene. For the ctrA antisense promoter, 200 bp of upstream sequence was cloned between the Bgl II and XhoI sites. For promoter constructs mutating the SciP binding motif, 105 bp of upstream DNA were cloned between the Bgl II and Xho I sites (S9 Fig.). In each other case, the Bgl II site is upstream and the Xho I site is downstream of the +5 site of the RNA transcript. These 36 constructs and a pNJH185 empty vector control were introduced into Caulobacter crescentus NA1000 cells by electroporation. The LacZ activity of all constructs was measured using mid-log phase NA1000 cell cultures grown in minimal media and according to standard ONPG based β-galactosidase assays. Results in S3 Dataset represent the average of three independent measurements for each strain.
|
10.1371/journal.pgen.1005026 | Recombination between Homologous Chromosomes Induced by Unrepaired UV-Generated DNA Damage Requires Mus81p and Is Suppressed by Mms2p | DNA lesions caused by UV radiation are highly recombinogenic. In wild-type cells, the recombinogenic effect of UV partially reflects the processing of UV-induced pyrimidine dimers into DNA gaps or breaks by the enzymes of the nucleotide excision repair (NER) pathway. In this study, we show that unprocessed pyrimidine dimers also potently induce recombination between homologs. In NER-deficient rad14 diploid strains, we demonstrate that unexcised pyrimidine dimers stimulate crossovers, noncrossovers, and break-induced replication events. The same dose of UV is about six-fold more recombinogenic in a repair-deficient strain than in a repair-proficient strain. We also examined the roles of several genes involved in the processing of UV-induced damage in NER-deficient cells. We found that the resolvase Mus81p is required for most of the UV-induced inter-homolog recombination events. This requirement likely reflects the Mus81p-associated cleavage of dimer-blocked replication forks. The error-free post-replication repair pathway mediated by Mms2p suppresses dimer-induced recombination between homologs, possibly by channeling replication-blocking lesions into recombination between sister chromatids.
| Ultraviolet (UV) light is a ubiquitous agent of exogenous DNA damage. In normal cells, the nucleotide excision repair (NER) pathway is the primary mechanism for repair of UV-induced DNA lesions. Defects in the NER pathway are associated with the human disease xeroderma pigmentosum (XP), and XP patients are prone to skin cancer. Mitotic recombination is strongly stimulated by UV treatment. In this study, we examined whether such stimulation requires the NER pathway. We show that, in the absence of NER, UV is still able to greatly induce recombination. We then characterized a nuclease that is required to generate recombinogenic breaks. Finally, we examined a previously known recombinogenic pathway called the “post-replication repair (PRR) pathway.” Our results suggest that the PRR pathway mainly promotes recombination between sister chromatids, and suppresses recombination between chromosome homologs.
| The primary types of DNA lesions caused by UV radiation are pyrimidine dimers [1]. Although UV strongly stimulates recombination in wild-type yeast cells [2–5], it is unclear whether this stimulation in wild-type cells primarily reflects unexcised dimers, or single-stranded DNA gaps and double-stranded DNA breaks (DSBs) resulting from incomplete nucleotide excision repair (NER) of dimers. One approach to simplifying the nature of the recombinogenic lesion is to examine UV-induced recombination events in NER-deficient cells. Previously, Kadyk and Hartwell (1993, [6]) showed that NER-deficient rad1 strains had reduced levels of UV-induced inter-homolog recombination and elevated levels of sister chromatid recombination compared to UV-induced events in an NER-proficient strain. Based on these observations, they argued that the inter-homolog recombinogenic effects of UV in wild-type cells were likely a consequence of DNA lesions introduced during NER. Since Rad1p can process various types of secondary DNA structures [7,8], one caveat to this conclusion is that Rad1p could be involved in the downstream events of recombination in addition to its role in producing the recombinogenic lesion. Consequently, in the current study, we examined UV-induced recombination in a rad14 diploid. Strains that lack Rad14p cannot perform NER, but are not known to have any other recombination defect [9]. Our analysis demonstrates that inter-homolog recombination events of a variety of types (crossovers, gene conversions unassociated with crossovers, and break-induced recombination events) are greatly elevated by unexcised dimers. Our analysis is the first detailed study of the recombinogenic effects of this biologically-important DNA lesion.
Unexcised pyrimidine dimers block or slow the progression of replication forks [10]. In yeast, G1-synchronized rad14 mutants treated with UV have large (up to 3000 base) single-stranded regions on the leading strand of the replication fork, as well as smaller single-stranded gaps on both the leading and lagging strands [11]. In addition, these strains have asymmetric fork structures diagnostic of broken forks. Presumably, either the single-stranded regions or the DSBs resulting from broken forks could act as recombinogenic lesions. Regressed replication forks are not observed as a consequence of unrepaired UV damage in yeast [11], although evidence for such structures has been obtained in E. coli [10].
Broken DNA molecules can be repaired by a variety of homologous recombination (HR) pathways involving the intact homolog (Fig. 1). Gene conversion events unassociated with crossovers occur primarily through synthesis-dependent strand-annealing (SDSA), although a small fraction are a consequence of processing of a double-Holliday junction (dHJ) or dHJ dissolution [8,12]. In addition, a broken DNA end can invade a duplex, generating a replication structure that duplicates sequences extending to the terminus of the homolog. This process, termed “break-induced replication” (BIR), involves conservative DNA replication [13,14]. These pathways generate a variety of branched intermediates including nicked HJs, double HJs, and single HJs that must be resolved to produce mature recombinant products. Processing of intermediates may involve either cleavage or dissolution (Fig. 1).
For the cleavage of various types of HJs formed during mitotic HR in S. cerevisiae, the primary resolvase is the Mus81p/Mms4p complex with Yen1p serving as a backup role [15]. The Mus81p/Mms4p complex prefers nicked HJs as a substrate, whereas Yen1p more efficiently cleaves intact junctions [8]. The nucleases involved in cleavage of HJs are somewhat organism specific and/or specific to the type of recombination (meiotic versus mitotic). For example, Rad1p has no role in junction resolution during meiotic recombination in yeast [16], but its ortholog in Drosophila (MEI-9) is required for meiotic recombination [17]. We show below that recombination between homologs caused by unexcised dimers requires Mus81p but not Yen1.
In addition to DNA repair events that involve inter-homolog interactions, DNA lesions that stall replication forks can be bypassed by both error-free and error-prone pathways of post-replication repair (PRR; [18]). The error-free pathway utilizes HR, primarily operating within the context of the replication fork, to bypass DNA damage. One sub-pathway involves template switching between the two arms of the replication fork, whereas a second pathway involves fork regression. Since fork regression is not evident for UV-induced damage [11], we assume that the primary error-free pathway for UV-induced damage in yeast is template switching. Efficient template switching depends upon the poly-ubiquitination of PCNA by the Ubc13p/Mms2p complex [19]. In contrast, the error-prone pathway utilizes the error-prone translesion synthesis (TLS) DNA polymerases to replicate the damaged DNA strand [18]. The recruitment of TLS polymerases requires mono-ubiquitinated PCNA catalyzed by the Rad6 and Rad18 proteins and is not known to be recombinogenic [19]. In our analysis, we examined strains lacking Mms2p, since such strains should be defective specifically in the error-free component of PRR. We find that mms2 strains have elevated levels of inter-homolog recombination, indicating that the Mms2p suppresses inter-homolog recombination.
The system that we used to detect and map UV-induced mitotic recombination events has been described previously [3,4,20]. In brief, we constructed a diploid heterozygous for about 55,000 heterozygous single-nucleotide polymorphisms (SNPs) by crossing two haploids (W303–1A and YJM789) that had 0.5% sequence divergence. In addition, the diploid was heterozygous for an insertion of the ochre suppressor tRNA SUP4-o located near the left end of chromosome V and homozygous for the ade2–1 ochre suppressible mutation. The starting diploid with one copy of SUP4-o partially suppresses the red pigment that accumulates in ade2 strains. Following a mitotic crossover, if the recombined chromosomes are segregated into different daughter cells, one daughter cell would lack the suppressor and one cell would have two copies of SUP4-o (Fig. 2A). Continued divisions of each of these cells would result in a red/white sectored colony with red sector derived from the daughter cell with no copies of the suppressor and the white sector derived from the daughter cell with two SUP4-o genes.
The position of the crossover is the junction between heterozygous SNPs and homozygous SNPs. Mapping of the position of loss of heterozygosity (LOH) was done using oligonucleotide-based microarrays [3]. For about 13,000 SNPs distributed throughout the genome, these arrays contained short (25-base) oligonucleotides that were specific to the W303–1A form of the SNP or the YJM789-specific form. By comparing the level of hybridization of the genomic samples with the recombination event to a heterozygous control strain (details in Materials and Methods), we could determine the transition between heterozygous SNPs and homozygous SNPs.
Several different patterns of LOH were observed. In some red/white sectored colonies, the positions of the transition were identical in both sectors (Fig. 2A). The microarray analysis for the two homologs is depicted as a single line for each sector with heterozygous chromosomal regions shown in green, and regions homozygous for W303–1A-derived SNPs and YJM789-derived SNPs shown as red and black segments, respectively. There were two other types of red/white sector colonies that were commonly observed. As shown in Fig. 2B, some of the sectored colonies had a chromosome segment (boxed in blue), adjacent to the crossover that was not reciprocally transferred between chromosomes. These segments are gene conversion events and likely reflect DNA mismatch repair within a heteroduplex tract adjacent to the crossover (Fig. 1). About 80–90% of mitotic crossovers are accompanied by a gene conversion [4,20]. Considering both chromosomes in both sectors, the conversion event shown in Fig. 2B has three chromosomes with W303–1A-derived SNPs and only one chromosome with YJM789-derived SNPs. Such events will be called “3:1” conversions, and these events are a consequence of repair of a single chromatid break (SCB) as shown by the arrow in Fig. 2B. In previous studies of spontaneous and UV-induced mitotic recombination [4,20,21], we also found red/white sectors in which the conversion event produced a 4:0 conversion associated with the crossover. This pattern is a consequence of the repair of double sister chromatid breaks (DSCB) at the same position. Our favored interpretation of this result [22] is that a chromosome that received a break in G1 was replicated to produce two broken sister chromatids (Fig. 2C).
Other patterns of LOH can be detected as unselected events by microarrays. For example, BIR can produce one daughter cell with a terminal LOH event and a second daughter cell that has no LOH (Fig. 2D). In addition, unselected gene conversion events that are unassociated with crossovers can produce one daughter cell with an interstitial LOH event and a second daughter cell with no LOH (Fig. 2E). Examples of LOH patterns diagnostic of a conversion-associated crossover, and a conversion unassociated with a crossover are shown in Fig. 3.
In most of the experiments described below, yeast strains with the SUP4-o marker located near the left end of chromosome V or near the right end of chromosome IV were synchronized in G1 and treated with UV. All diploids used in our study had deletions of the MATα gene, allowing their synchronization with α pheromone [22]. The cells were plated on solid growth medium, immediately treated with UV, and subsequently screened for sectored colonies. Cultures derived from each side of the sector were examined by microarrays. For most of the experiments, we used arrays that allowed us to monitor LOH events on the selected chromosome (either V or IV) and unselected events throughout the genome.
The NER-deficient rad14 mutant is very sensitive to UV radiation [9]. While 15 J/m2 of UV in G1-arrested wild-type diploids did not significantly reduce cell viability in the wild-type (76% viability relative to unirradiated cells; [4]), less than 0.02% of the rad 14 cells survived 5 J/m2 of UV. Treatment of rad14 cells with 1 J/m2 of UV resulted in 61% cell viability, comparable to the viability of wild-type cells treated with 15 J/m2. Since formation of a sectored colony requires the survival of both daughter cells containing the products of recombination, it is important that the UV dose has a minimal effect on viability of the treated cells. Previously, we mapped UV-induced LOH in both 1 J/m2- and 15 J/m2-treated wild-type cells [4]. Below, we show sectoring frequency, as well as the frequency of unselected LOH events, in the rad14 mutant treated at 1 J/m2, in wild-type cells treated at 1 J/m2 (resulting in the same number of UV-induced lesions), and in wild-type cells treated with 15 J/m2 (an approximately equitoxic radiation dose).
The sectoring frequency data for diploids with the SUP4-o marker on either chromosome V or on IV are summarized in Table 1. Since only half of crossovers have the segregation pattern that produces reciprocal LOH in the daughter cells [23], the rate of mitotic crossovers is twice the frequency of red/white sectored colonies for all strains. For most strains treated with UV, the sector frequencies were determined non-selectively. For the untreated wild-type strain with the SUP4-o marker on chromosome V, crossovers were selected as described by Lee et al. [21]. This method, which involves selection of crossovers as canavanine-resistant sectored colonies, did not work well for strains with the rad14 mutation or for strains with the SUP4-o marker on chromosome IV [20]. In the rad14 mutant treated with 1J/ m2 of UV, the red/white sectoring frequency on the left arm of chromosome V is 0.5%; this frequency is similar to that of the wild-type strain treated with 15 J/m2, and is five-fold higher (p = 1.7×10-10) than the wild-type treated with 1 J/m2 [4] as shown in Table 1. For the same interval, the spontaneous sectoring frequency for the wild-type strain is 1.10.5×10-6. In the wild-type strain, we can select crossovers on the left arm of V as canavanine-resistant red/white sectors [21,24]. For unknown reasons, this selection does not work in the rad14 diploids and, therefore, we screened for sectored colonies. Since no such colonies were detected in 22797 total colonies (Table 1), we conclude that the frequency of spontaneous crossovers in rad14 strains is very low, less than 4.4 ×10-5.
In order to obtain a more accurate estimate of the frequency of sectored colonies in the unirradiated rad14 strain, we examined sectored colonies in strains in which the SUP4-o marker was located on the right arm of chromosome IV (Table 1). The distance between CEN4 and SUP4-o is 1.1 Mb, about eight times larger than the interval between CEN5 and SUP4-o in the other strains. Sectored colony formation in the rad14 mutant was stimulated about 200-fold by the 1 J/m2 dose of UV relative to the unirradiated rad14 strain. The frequency of red/white sectored colonies in rad14 strains treated with 1 J/m2 UV dose was about four-fold higher than the frequency of such colonies in the wild-type strain treated with the same dose. Interestingly, the rad14 mutation resulted in a modest (about five-fold, p = 1.8×10-9) increase in the frequency of sectors in the absence of UV irradiation.
In addition to the frequencies of red/white sectors, we also show in Table 1 the frequencies of sectors that were either red/white or pink/red/white. Pink/red/white sectors could arise in two different ways: 1) persistence of DNA damage beyond the first division cycle following UV treatment (S1 Fig.), or 2) irradiation of two adjacent cells resulting in a recombination event in one of them (the red/white portion of the sector) but no recombination event in the other (the pink portion of the sector). By micromanipulation of individual cells before irradiation, we showed that most (9 out of 10) of the selected crossovers occur in the first cell division in UV-treated wild-type strains [4], therefore, the pink/red/white sectored colonies we recovered from plating experiments are more consistent with the second explanation.
The recombinogenic effects of unexcised UV-induced DNA damage were further demonstrated by our whole-genome analysis of sectored colonies derived from UV-treated wild-type and rad14 strains (Table 2). This analysis allows the detection of unselected crossovers, gene conversions, and BIR events throughout the genome. The rad14 strain treated with 1 J/m2 had about seven unselected LOH events per sectored colony, whereas the wild-type strain treated with the same dose had only 0.4 LOH events per sectored colony. The unselected events from the wild-type strain treated with 15 J/m2 are described in S1 Table and S2 Table, whereas the unselected events from the rad14 strain are described in S3 Table and S4 Table. S1 Table and S3 Table depict the transitions between heterozygous and LOH regions within each sector schematically, and S2 Table and S4 Table show the SGD coordinates for each transition. Comparable data for unselected events for the wild-type strain irradiated with 1 J/m2 were described previously [4], as were the methods for depicting various classes of LOH events [3,4]. S5–S13 Tables summarize the transitions for the other UV-treated mutant strains, and S14 Table and S15 Table show the genome regions covered by the SNP arrays.
For statistical comparisons of the number of unselected recombination events for the wild-type and rad14 strains irradiated with 1 J/m2, we determined the number of LOH events per sectored colony for each strain (ten and twelve sectored colonies from the wild-type and rad14 strains, respectively; Table 2), and compared these distributions by the Mann-Whitney test. By this test, the number of unselected events per colony for the rad14 strain (average of 6.7) was significantly greater (p = 0.0003) than the number observed for the wild-type strain (average of 0.4).
As shown in Fig. 2, the patterns of LOH in a sectored colony can reveal whether a mitotic recombination event is a consequence of the repair of a DSB on a SCB or repair of a DSCB. 3:1 conversion tracts (Fig. 2B and Fig. 2E) are interpreted as SCB events, whereas 4:0 tracts (Fig. 2C) are indicative of DSCBs [4,21]; crossovers without a detectable conversion tract cannot be classified as SCB or DSCB. We found previously that high doses of UV (15 J/m2) in G1-synchronized wild-type strains result in more DSCBs than SCBs, but low doses (1 J/m2) produce primarily SCBs [4]. In the G1-synchronized rad14 strain treated with 1 J/m2, 78% of the selected events were SCBs, similar to the fraction observed for the wild-type strain treated with the same UV dose. The data from both the selected events on chromosome V and the unselected events on other chromosomes derived from sectored colonies are in Table 3. In summary, our results demonstrate that UV-induced recombination events in rad14 strains usually involve a single broken sister chromatid.
The rad14 strain irradiated with 1 J/m2 had significantly different proportions of recombination events relative to the wild-type strain irradiated with an equitoxic UV dose (15 J/m2) (Table 2). More specifically, the rad14 strain had significant elevations in the number of BIR events and the number of deletions relative to the wild-type strain (p = 0.0003 and 0.0004, respectively). Interestingly, we found presence of early/middle-firing ARS sequences [25] within the deleted sequences in all the 4 deletion events. Although no significant differences were observed between the rad14 and wild-type strains irradiated with 1 J/m2, there were very few unselected LOH events in the wild-type strain.
In considering the frequency of BIR events in the rad14 strain, we note an important caveat. BIR events in red/white sectored colonies are defined as unselected non-reciprocal terminal LOH events (S2 Fig.). If UV-irradiated rad14 strains retain recombinogenic lesions beyond the first division cycle, this same LOH pattern could be produced by having one UV-induced crossover on chromosome V in the first division cycle (resulting in the red/white sectored colony) and a second UV-induced LOH event on an unselected chromosome arm in the second division cycle (S3 Fig.). The second LOH event would produce two different types of granddaughters (shown as GD2–1 and GD2–2 in S3 Fig.). In our analysis of sectored colonies, we purify single colonies derived from each sector. If we examined one colony derived from the white sector (either GD1–1 or GD1–2) and a colony derived from GD2–1 from the red sector, we would conclude that there was a reciprocal crossover on chromosome V and an unselected BIR event on the other homolog. In conclusion, the relative frequency of unselected crossovers and BIR events in the UV-irradiated rad14 strain is unclear. This same caveat applies to all genetic analyses in which the experimental strain has high levels of on-going genetic instability.
To clarify whether unrepaired UV-induced lesions in rad14 strains were capable of generating recombination events after the first cell division, we micromanipulated single G1-arrested rad14 cells to specific positions on solid medium that we then irradiated. Of 1096 irradiated (1 J/m2) rad14 cells, 681 cells survived. Of four resulting sectored colonies, two were red/white as expected for a crossover on chromosome V in the first division, and two were pink/red/white as expected for a crossover in the second division (S1 Fig.). The same micromanipulation analysis for the wild-type strain treated with 15 J/m2 yielded nine colonies with red/white sectors and only one colony with pink/red/white sectors [4]. Although the numbers of these events were small and the difference in proportions of red/white and pink/red/white sectors in wild-type and rad14 cells is not statistically significant (p = 0.18 by Fisher exact test), these results suggest that recombinogenic DNA damage can persist in rad14 cells, as expected from unexcised pyrimidine dimers.
Mus81p is the main Holliday junction resolvase for mitotic crossovers in S.cerevisae [15] and for meiotic crossovers in S. pombe [26]. In addition, mus81 strains are UV sensitive [27]. To determine whether Mus81p had a role in the processing of UV-induced recombinogenic lesions, we examined the frequency and types of UV-induced recombination in a mus81 strain and in a mus81 rad14 double mutant strain.
For the mus81 strain, we used microarrays to perform a genome-wide analysis of seven red/white sectored colonies derived from G1-synchronized cells irradiated with 15 J/m2. There was no significant difference (p = 0.89 by Mann-Whitney test) in the average number of interstitial LOH events (primarily gene conversions) in the mus81 strain compared to wild-type (7.7 and 7.1, respectively). In contrast, the average number of unselected crossovers in the mus81 strain was reduced to 40% of the level observed in the wild-type strain (Table 2). This loss of crossovers is statistically significant (p = 0.004 by Mann-Whitney test). These results are similar to those observed for HO-induced DSBs by Ho et al. [15], demonstrating that Mus81p was involved in the resolution of HJs by crossovers, but not in the SDSA conversion pathway. In addition, the single mus81 mutant does not significantly (p = 0.72 by Mann-Whitney test) reduce the frequency of recombination events, since the average number of LOH events per sectored colony is very similar for the wild-type and mus81 strains (10.1 and 9.6, respectively).
Double mutant mus81 rad14 strains were very sensitive to UV. A dose of 1 J/m2, which reduced the viability of the rad14 strain to 61% that of the untreated strain, lowered the viability of the double mutant strain to 0.2%. Consequently, we monitored the effect of UV in the double mutant strain on UV-induced mitotic recombination by analyzing single colonies formed from cells treated with 1 J/m2 rather than by measuring the frequency of red/white sectored colonies. The results of this analysis are shown in Table 4, and the distributions of the number of LOH events per colony for rad14 and rad14 mus81 strains are shown in Fig. 4. We found that the total unselected LOH events were significantly reduced in the double mutant compared to the single rad14 mutant (p = 0.0005 by Mann-Whitney test). This result suggests that Mus81p is likely involved in generating recombinogenic DNA lesions at replication forks that are blocked at unexcised pyrimidine dimers; this interpretation will be further considered in the Discussion.
We also examined the effects of the Yen1p HJ resolvase on the UV-induced recombination events in a rad14 strain. After a dose of 15 J/m2 of UV, the single yen1 mutant has about the same number of LOH events per cell as wild-type (Table 2). The double mutant rad14 yen1 strain had approximately the same average number and types of LOH events per irradiated cells as the single rad14 mutant (Table 4); none of the observed differences with wild-type were statistically significant. In summary, our analysis suggests that the Mus81p, but not the Yen1p, has a role in generating recombinogenic lesions in UV-treated rad14 strains.
As discussed in the Introduction, unexcised pyrimidine dimers can be bypassed by either error-prone or error-free pathways [18]. The error-free pathway utilizes HR enzymes to catalyze either template switching between the two arms of the replication fork or fork regression. Since fork regression has not been observed for UV-treated strains, we assume that the primary pathway for error-free repair in our experiments is template switching. Template switching requires Ubc13p and Mms2p to poly-ubiquitinate PCNA. Although template switching is generally considered to be limited to the arms of the replication fork or to sister chromatids, most previous studies have been done in haploid strains. Consequently, we compared the frequency of UV-induced recombination in rad14 and rad14 mms2 diploids. If Mms2p is required for inter-homolog recombination, we would expect a reduction in recombination in the double mutant whereas, if Mms2p was required solely for inter-replication fork or inter-sister chromatid recombination, the double mutant should have elevated inter-homolog recombination.
The rad14 mms2 strains, like the rad14 mus81 strains were very UV sensitive. At a UV dose of 1 J/m2, the double mutant has 1% survival compared to 61% survival for rad14 strain, and 0.2% survival for the rad14 mus81 strain. Consequently, we compared the numbers and types of LOH events in single colonies derived from cells treated with 1J/m2 of UV (Table 4, Fig. 4). The average number of LOH events in a double mutant of rad14 mms2 increased by 62% compared to the rad14 strain; this difference is statistically significant (p = 0.008 by Mann-Whitney test). The main class of events that is elevated in the double mutant is the CO/BIR category (p = 0.001 by Mann-Whitney test). These results suggest that Mms2p promotes recombination between the arms of the replication fork and/or sister chromatids and, in the absence of Mms2p, the recombination intermediates are directed into an inter-homolog pathway of recombination.
In our previous analysis of the lengths of UV-induced mitotic gene conversion tracts in wild-type strains treated with 15 J/m2, we found a median length of 6.4 kb for all conversion tracts; median lengths of crossover-associated tracts (CO events) were longer than for conversions unassociated with crossovers (NCO events), 8.2 kb and 5.7 kb, respectively [4]. The median lengths of conversion events (CO plus NCO) in rad14, mus81, and yen1 are all between 5.2 and 7.9 kb (Table 5). These lengths are not significantly different from the wild-type strain. As observed for the wild-type strain, the median lengths for the CO conversions exceed those for the NCO conversions for all strains. When we compared the lengths of conversions in NCO and CO categories, the only significant difference was that the lengths of CO events in the mus81 strain significantly exceeded those of the wild-type strain irradiated with 15 J/m2 (p = 0.027 by Mann-Whitney test). One interpretation is that failure to cleave nicked HJs in the mus81 mutants results in their maturation to double Holliday junctions that are associated with longer heteroduplex regions.
For crossovers, the DNA lesions that initiate recombination should be located near the transitions between heterozygous and homozygous regions (the breakpoints of the LOH events). For gene conversions unassociated with crossovers, the initiating lesion should be located between the heterozygous sites flanking the conversion tract. Consequently, we determined whether these regions were significantly enriched for various types of chromosome elements such as replication origins, retrotransposons, palindromes, and regions associated with G4 motifs. The methods used for this analysis are described in the S1 Text, and the data are presented in S16 Table. We previously showed that spontaneous mitotic recombination events and LOH events induced by low levels of DNA polymerase α were enriched for replication-termination (TER) sequences [25] and other motifs associated with slow-moving or stalled replication forks [20,28]. In contrast, none of the chromosome elements tested were significantly over-represented among UV-induced LOH events in a wild-type strain [4].
We examined sixteen categories of chromosome elements for their representations at LOH breakpoints including: centromeres, tRNA genes, non-coding (nc) RNAs, solo long-terminal repeats (LTRs), early/middle ARS elements, late ARS elements, Rrm3p pause sites, palindromes, retrotransposons, G4 motifs, highly-transcribed and weakly-transcribed genes, TER sequences, regions with high levels of gamma-H2AX, snoRNAs and snRNAs, and genomic regions with inefficient TT dimer repair (references for the mapping of these elements are in S16 Table). In the rad14 single mutant, none of these elements were significantly over- or under-represented. In contrast, in both the mus81 and rad14 mus81 strains, G4 (quadruplex) motifs are over-represented at the LOH breakpoints. In addition, in the rad14 mus81 double mutant strain, we observed over-representation of retrotransposons (Ty elements), Rrm3p pause sites, regions with high levels of gamma H2AX, and regions associated with slow thymidine dimer repair (S16 Table).
Many of the over-represented motifs are associated with delayed or stalled replication forks in wild-type yeast cells, and are hotspots for mitotic recombination events induced by low levels of DNA polymerase α [28]. One explanation of our results is that the probability that a stalled replication fork will be broken is a function of at least three factors: certain DNA motifs (for example, quadruplex DNA) that cause replication forks to move slowly or stall, the presence of unexcised pyrimidine dimers near these motifs, and the presence or absence of Mus81p. Although Mus81p is required for about 60% of the LOH events induced by unexcised dimers, 40% are independent of Mus81p (Table 4). No significant associations were observed at the breakpoints of the rad14 yen1 strain, although the rad14 mms2 mutant had significant over-representations of ncRNAs and TER sequences.
In strains with the SUP4-o marker near the right end of chromosome IV, we did a limited analysis of spontaneous recombination events derived from red/white sectored colonies in the rad14, mus81, and yen1 strains. The primary purpose of this analysis was to ensure that none of the single mutants substantially elevated LOH in the absence of UV irradiation. The numbers of sectored colonies divided by the total number of colonies examined, the frequencies, and the strains names were: wild-type (55/1761664; 3.1x10-5; JSC25; [20]), rad14 (20/144771; 1.4x10-4; YYy37.6/YYy37.8), mus81 (10/277704; 3.6x10-5; YYy45.2/YYy46.1), and yen1 (12/276315; 4.3x10-5; YYy72.22/YYy72.41). Although there is a statistically significant elevation of spontaneous crossovers in the rad14 strain (as noted previously), the frequency of sectored colonies in unirradiated cells is two orders of magnitude less than in the irradiated cells. We also looked for sectored colonies derived from unirradiated cells of the rad14 mus81 and rad14 mms2 genotypes that had the SUP4-o marker on the left end of chromosome V. No sectored colonies were observed in colony totals of 22797 (rad14), 39466 (rad14 mus81), and 37493 (rad14 mms2). Although this analysis does not allow an accurate measurement of the sectoring frequency, we estimate that the frequency of spontaneous sector formation is less than 4x10-5 for all three strains. Thus, the spontaneous events do not contribute significantly to our estimates of the frequencies or types of UV-induced exchanges.
We performed a genome-wide analysis of inter-homolog recombination events induced by unexcised pyrimidine dimers, and we monitored the effect of various DNA repair enzymes on the frequency and types of these recombination events. The main conclusions from our study are: 1) unexcised pyrimidine dimers induced by irradiation of G1-synchonized cells strongly induce mitotic recombination between homologs, 2) most of the induced events reflect the repair of a single broken sister chromatid, 3) the recombinogenic effects of unexcised dimers are largely dependent on the Mus81p resolvase while in the NER-proficient strains, only crossovers are reduced by deletion of MUS81, and 4) rad14 mms2 strains have elevated levels of UV-induced recombination events relative to the rad14 single mutant, consistent with the hypothesis that the Mms2p-mediated PRR pathway channels DNA lesions for repair to the sister chromatid.
Although the recombinogenic effects of UV irradiation in yeast have been known for a long time [5], the nature of the recombinogenic DNA lesion and the pathways of lesion repair are still unclear. Based on the effects of UV irradiation in synchronized cells, Galli and Schiestl [2] suggested that NER-generated lesions (single-stranded nicks or gaps) could be converted into broken chromatids during DNA replication. Alternatively, the blocking of replication forks by unexcised dimers [10] could be converted into a recombinogenic DSB. One approach to distinguishing these possibilities is to compare UV-induced recombination events in wild-type and NER-deficient strains. Using this approach, Kadyk and Hartwell [6] found that wild-type diploid strains synchronized in G1 and treated with 30 J/m2 of UV had about a five-fold elevation in inter-homolog gene conversion and a two-fold induction in sister-chromatid recombination. In contrast, an NER-deficient rad1 diploid strain irradiated with 1 J/m2 (approximately equivalently genotoxic with a dose of 30 J/m2 in a wild-type strain) had a six-fold elevation in sister-chromatid recombination and a three-fold reduction in the frequency of inter-homolog conversion events. From this analysis, they concluded that inter-homolog recombination was likely a consequence of nicks and gaps formed by NER, whereas unexcised dimers primarily induced sister-chromatid recombination rather than inter-homolog recombination.
In our study, UV treatment very strongly (>100-fold) stimulated mitotic crossovers (Table 1), as well as other types of recombination events (Table 2), in an NER-deficient (rad14) strain. There are several explanations for the discrepancy between our results and those of Kadyk and Hartwell [6]. First, different mutations were used to inactivate NER, rad14 in our experiments and rad1 in those of Kadyk and Hartwell. Rad1p, unlike Rad14p, is needed to process certain types of mitotic recombination intermediates [29] in addition to its requirement for NER. However, since Rad1p is not required for I-SceI-induced recombination events between homologs [7], lack of Rad1p would not be expected to reduce the frequency of UV-induced inter-homolog exchange. In addition, in the rad14 (strain YYy23.4) and rad1 (strain YYy327.1) cells treated with 1 J/m2 of UV, the frequencies of survival (49% and 47%, respectively), and red/white sectored colonies (8.1 x 10-3 and 6 x 10-3, respectively) were similar in the two strains.
A more likely explanation is the nature of the genetic system used to detect inter-homolog exchange in the two studies. In the Kadyk and Hartwell study [6], inter-homolog conversion events were detected by measuring the frequency of leucine prototrophs derived from a diploid with leu1 heteroalleles. This system likely detects a very small fraction of the induced events. Assuming that mitotic conversion is a consequence of heteroduplex formation followed by DNA mismatch repair, a prototroph will be detected in only two situations: 1) the heteroduplex includes only one of the two leu1 mutations or 2) the heteroduplex includes both mutations but the resulting repair event using different strands as the repair template (“patchy” repair). However, mitotic conversion events are usually longer than one gene (> 4 kb) and multiple mismatches within a heteroduplex are usually repaired using the same strand as a template [4,20]. In our analysis, we examine inter-homolog recombination events by a sectoring assay for crossovers or by detecting unselected crossovers and conversions using SNP microarrays. These systems have the advantages of having fewer constraints on the detection of gene conversion events and of being less selective as to the location of the events.
Although our assay is designed primarily to detect LOH events occurring between homologs, we recovered four interstitial deletions in the UV-treated rad14 strains (Class R in S3 Table). Three of four deletions had directly-oriented Ty elements at the breakpoints of the deletion, and the fourth had one Ty element and a directly-oriented delta (LTR) (S4 Table). Thus, these deletions are likely generated by HR between the direct repeats rather than non-homologous end-joining. We have detected similar events previously in other genetically-unstable yeast strains [28,30]. The deletion events can be explained as intrachromatid “pop-outs”, unequal sister-chromatid crossovers, single-strand annealing, gene conversion, or off-set BIR events between sister chromatids. Some of these possibilities are shown in S4 Fig. Although deletions are observed in UV-treated rad14 strains, none were observed in UV-treated wild-type strains. The reason for this difference is not understood although it could be related to the nature of the DNA lesion, or the kinetics of the repair process.
It should be emphasized that our system (and almost all genetic systems) are capable of detecting inter-homolog and unequal sister chromatid recombination events but cannot detect sister chromatid exchanges between perfectly-aligned chromatids. A UV dose of 1 J/m2 is expected to introduce about 500 pyrimidine dimers/diploid genome [31]. Since we detect about seven LOH events/sectored colony in the rad14 strain (Table 2), most of the UV lesions must be bypassed by mechanisms that do not result in LOH, either recombination events that involve equal sister chromatid recombination or that utilize TLS polymerases.
As discussed above, one likely source of recombinogenic lesions in G1-synchronized wild-type cells treated with UV is nicks or gaps generated by NER; replication of such molecules would lead to DSBs [2]. If this lesion was solely responsible for initiating recombination in a wild-type cell, we would expect UV-irradiation of G1-synchronized cells to generate recombination events that indicate the break of a single chromatid (SCBs) (Fig. 2). In wild-type cells irradiated with 15 J/m2, however, we found that more than half of the events involved the repair of two sister chromatids broken at approximately the same position (DSCBs) [4]. Our interpretation of this result was that high doses of UV in G1-synchronized cells resulted in DSBs, and this interpretation was supported by physical evidence of a low frequency of UV-induced DSBs (5–10 DSBs in cells irradiated with 40 J/m2; [32]). In G1-synchronized cells irradiated with low levels of UV (1 J/m2), most of the recombination events were SCBs [4]. These results argue that the DSBs likely reflect the excision of closely-opposed pyrimidine dimers that arise independently [4]. An analysis of UV-induced recombination in exo1 cells treated with 15 J/m2 showed that Exo1p-expanded NER gaps were required for both SCB and DSCB events [33]. In addition, as described in [34], an analysis of UV-induced events in G2-synchronized haploid cells showed that single-stranded gaps stimulated recombination between sister chromatids. In summary, there are likely several types of recombinogenic lesions generated by UV in wild-type cells: G1-associated DSBs resulting from the repair of closely-opposed dimers, S/G2-associated DSBs resulting from replication of a DNA molecule with an NER-generated gap, and single-stranded gaps; Exo1p-expanded NER gaps contribute to the formation of DSBs in both G1 and S/G2.
Most (76%) of the recombination events induced by UV in G1-synchronized rad14 cells are SCBs (Table 3). These recombination events cannot be related to NER-generated DNA lesions, and we suggest that the relevant lesion is produced by breaks that occur at replication forks stalled by unexcised pyrimidine dimers. Two of the important HJ or nicked HJ resolvases in yeast are Mus81p and Yen1p. Consistent with the observations that Mus81p is required for DSB formation at fragile sites [35] and at camptothecin-stalled replication forks in mammalian cells [36], we found that loss of Mus81p reduced the frequency of UV-induced LOH events in the rad14 strain by about 60%. This reduction could be explained as a reduction in the level of recombinogenic DSBs or as less efficient processing of recombination intermediates to yield LOH events. Previously, Ho et al. [15] showed that mus81 mutants had reduced frequencies of I-SceI-induced mitotic crossovers but wild-type levels of induced conversions. Our observation that rad14 mus81 mutants have reduced UV-induced gene conversion unassociated with crossovers relative to rad14 (Table 3) argues that Mus81p is involved in generating DSBs, although it may also have a role in downstream recombination events. In comparisons of the frequency of gene conversion events in wild-type and mus81 strains treated with 15 J/m2 of UV, we found that Mus81p was required for crossovers, but not conversions unassociated with crossovers (Table 2), as expected from previous results.
Loss of Yen1p had no detectable effect on the frequency of LOH events in UV-treated rad14 strains. Based on the substrate preferences for Mus81p and Yen1p [8], this result argues that the relevant recombinogenic structure of the blocked replication fork is a nicked HJ rather a ligated HJ. It has been suggested that extensively-regressed replication forks resembled ligated HJs which should be processed by Yen1p [37]. Thus, the lack of effect of Yen1p on the frequency of UV-induced exchange is consistent with the observation of Lopes et al. that UV blocks replication forks without causing extensive fork regression. Our observations are also in agreement with other studies [15,38] that show that Mus81p is more important than Yen1p for the generation of crossovers in yeast. The role of Yen1p may be somewhat lesion-specific, however, since loss of Yen1p reduces the frequency of sister chromatid recombination of DSBs generated by replication of nicked template [39]. We also cannot rule out the possibility that the Yen1p is not present during the time of replication fork blockage [40].
In addition to Mus81p-mediated resolution of a blocked replication fork producing a DSB, several other pathways allowing bypass of unexcised dimers exist: an error-prone pathway utilizing TLS polymerases and an error-free pathway that likely involves template switching between the arms of the replication fork [18]. Lopes et al. [11] showed that long single-stranded regions accumulate at replication forks in UV-treated yeast cells, and suggested that these events reflecting an uncoupling of leading and lagging strand replication. Both nicks and gaps are recombinogenic [34,41], and Giannattasio et al. [42] suggested that most template switching events were initiated with single-stranded gaps. Error-free template switching requires Mms2p, which catalyzes the polyubiquitination of PCNA, and Rad5p, a protein that mediates D-loop extension [43,44]. The error-prone TLS pathway is in competition with the error-free pathway. Elimination of the TLS pathway results in elevated frequencies of UV-induced sister-chromatid exchange [34].
In order to determine whether mutations in the error-free pathway contributed to LOH events between homologs, we examined rad14 mms2 diploid strains. The double mutant strains had significantly elevated frequencies of inter-homolog recombination relative to the single mutant rad14 strain (Table 4). Our preferred interpretation of this result is that Mms2p primarily promotes recombination between the arms of the replication fork and/or between sister chromatids rather than between homologs. Thus, a reduction in the frequency of recombination between sisters would result in an elevation in the frequency of inter-homolog interactions. This observation, as well as those of others, are consistent with a model in which Mus81p is primarily involved in producing DSBs by resolution of blocked replication forks, whereas Mms2p is mainly concerned with promoting recombination intermediates that involve a single-stranded gap.
In conclusion, unexcised UV dimers are a potent inducer of inter-homolog recombination in yeast. About 60% of this type of recombination is dependent on the Mus81p resolvase. The error-free PRR pathway mediated by Mms2p suppresses recombination between homologs, likely by channeling replication-blocking lesions into recombination between sister chromatids instead of homologs. These conclusions are summarized in Fig. 5.
All strains are diploids formed by mating two sequence-diverged haploids W303–1A [45] and YJM789 [46]. The wild-type diploid (PG311) used in our analysis was described previously [21], and has the genotype: MATa/MATα::natMX4 ade2–1/ade2–1 can1–100/can1::SUP4-o GAL2/gal2 his3–11,15/HIS3 leu2–3,112/LEU2 RAD5/RAD5 trp1–1/TRP1 ura3–1/ura3 V9229::hphMX4/V9229 V261553::LEU2/V261553). The disruption of the MATα locus prevents sporulation of the diploid, and allows synchronization of the diploid in G1 using α pheromone. As described in the text, a crossover between the heterozygous SUP4-o marker and CEN5 results in a red/white sectored colonies. The constructions of isogenic strains with homozygous mutations in RAD14, MUS81, YEN1, MMS2, and various double mutant strains are described in S17 Table.
In experiments to look at the rate of spontaneous crossovers, we used strains isogenic with JSC25 [20] in which the SUP4-o marker is inserted near the right telomere of chromosome IV. The genotype of JSC25 is: MATa/MATα::hphMX4 ade2–1/ade2–1 can1–100::natMX4/CAN1::natMX4 GAL2/gal2 his3–11,15/HIS3 leu2–3,112/LEU2 RAD5/RAD5 trp1–1/TRP1 ura3–1/ura3 IV1510386::kanMX4-can1–100/IV1510386::SUP4-o). Mutant strains isogenic to JSC25 with homozygous mutations in RAD14, MU81, and YEN1 were constructed as described in S17 Table.
Standard media, genetic methods (mating, sporulation, transformation, and dissection), and DNA isolation procedures [47] were used unless otherwise indicated.
All the UV experiments were done with cells synchronized in G1 with α pheromone as described previously [22]. After two hours of treatment with α pheromone, the cells were diluted, and plated on MAB6 solid medium (omission medium lacking arginine and containing 10 μg/ml adenine). The cells were immediately treated with 1 J/m2 or 15 J/m2 of UV using the TL-2000 UV Translinker. The plates were then wrapped in foil to prevent photoreversal of pyrimidine dimers, and incubated at 30°C. for two days. The plates were then incubated overnight at 4°C. This incubation period results in an intensification of the red color in red/white sectored colonies. To examine spontaneous crossovers, we allowed cells to form colonies by growing them on rich growth medium (YPD) for two days. Colonies were then diluted into water, plated onto MAB6 solid medium, and incubated for three days at room temperature. The plates were incubated at 4°C overnight as described above.
The methods used for microarray analysis have been described in detail previously [3]. DNA from sectors or single colonies was isolated and sonicated using standard protocols. Experimental and control DNA samples were labeled with Cy5-dUTP and Cy3-dUTP, respectively, and hybridized in competition to Agilent-constructed SNP Microarrays. The whole-genome microarray was used to analyze all the UV-induced sectors and single colonies. The whole-genome microarray contains oligonucleotides that detect LOH for about 13,000 SNPs that differ between W303–1A and YJM789. Each SNP is represented by four 25-base oligonucleotides: two with sequences of the Watson and Crick strands of the W303–1A allele and two with sequences of the YJM789 allele. Following hybridization of the samples to the microarray, we scanned the array with GenePix scanner and GenePix Pro software. In-house R scripts were used for analyzing the microarrays. Based on the relative median hybridization ratio between experimental strain and control strain, which is heterozygous for all the SNPs, we determined which the SNPs were heterozygous or homozygous.
The data were analyzed using chi-square, Mann-Whitney, or Fisher exact tests. Tests were performed with the VassarStat Website (http://vassarstats.net/), Excel, and R functions. Corrections for multiple comparisons were done using the method of Hochberg and Benjamini [48].
|
10.1371/journal.pbio.1000204 | The Mechanochemistry of Endocytosis | Endocytic vesicle formation is a complex process that couples sequential protein recruitment and lipid modifications with dramatic shape transformations of the plasma membrane. Although individual molecular players have been studied intensively, how they all fit into a coherent picture of endocytosis remains unclear. That is, how the proper temporal and spatial coordination of endocytic events is achieved and what drives vesicle scission are not known. Drawing upon detailed knowledge from experiments in yeast, we develop the first integrated mechanochemical model that quantitatively recapitulates the temporal and spatial progression of endocytic events leading to vesicle scission. The central idea is that membrane curvature is coupled to the accompanying biochemical reactions. This coupling ensures that the process is robust and culminates in an interfacial force that pinches off the vesicle. Calculated phase diagrams reproduce endocytic mutant phenotypes observed in experiments and predict unique testable endocytic phenotypes in yeast and mammalian cells. The combination of experiments and theory in this work suggest a unified mechanism for endocytic vesicle formation across eukaryotes.
| Endocytosis is a complex and efficient process that cells utilize to take up nutrients and communicate with other cells. Eukaryotes have diverse endocytic pathways with two common features, mechanical and chemical. Proper mechanical forces are necessary to deform the plasma membrane and, eventually, pinch off the cargo-laden endocytic vesicles; and tightly regulated endocytic protein assembly and disassembly reactions drive the progression of endocytosis. Many experiments have yielded a lot of detailed information on the sub-processes of endocytosis, but how these sub-processes fit together into a coherent process in vivo is still not clear. To address this question, we constructed the first integrated theoretical model of endocytic vesicle formation, building on detailed knowledge from experiments in yeast. The key notion is that the mechanical force generation during endocytosis is both slave to, and master over, the accompanying endocytic reaction pathway, which is mediated by local membrane curvature. Our model can quantitatively recapitulate the endocytic events leading to vesicle scission in budding yeast and can explain key aspects of mammalian endocytosis. The phenotypes predicted from variations within the feedback components of our model reproduce observed mutant phenotypes, and we predict additional unique and testable endocytic phenotypes in yeast and mammalian cells. We further demonstrate that the functional significance of such mechanochemical feedback is to ensure the robustness of endocytic vesicle scission.
| During clathrin-mediated endocytosis, cells regulate plasma membrane molecular composition and internalize essential nutrients. This process involves coordination of biochemical activities with membrane shape changes [1],[2]. Multicolor real-time fluorescence microscopy studies in mammalian cells and yeast established that proteins are sequentially recruited to the endocytic site to drive membrane invagination and vesicle scission [3],[4],[5],[6],[7],[8],[9],[10],[11],[12]. Real-time movies and EM studies in yeast and mammals have demonstrated that the endocytic membrane is composed of different regions (bud and tubule/neck), each with a distinct protein composition and spatial profile [13],[14],[15]. Comparisons between yeast and mammalian endocytic systems have highlighted similarities and differences [2],[16]. The extent to which common principles underlie endocytosis in different eukaryotic cells is currently a matter of speculation and debate. Among the most obvious differences, clathrin-mediated endocytosis in mammalian cells involves formation of spherical clathrin-coated vesicle buds and recruitment of the GTPase dynamin to the vesicle neck, while endocytic structures in yeast are tubular invaginations lacking dynamin [15],[17]. Also, actin assembly is required for formation of the membrane invagination and for vesicle scission in yeast [8], while in mammalian cells these steps appear only to be assisted by actin assembly [18]. On the other hand, many endocytic proteins, including clathrin, adaptor proteins, and cytoskeletal proteins, are highly conserved from yeast to mammals. In both yeast and mammalian cells, dynamics of the key endocytic proteins are coordinated in space and time, and internalization and vesicle scission are accompanied by a transient burst of actin assembly [1],[2]. Despite intensive study in many laboratories, the mechanisms underlying coordination of protein recruitment, lipid modification, and membrane shape changes are not well understood in any organism.
From a mechanical standpoint, endocytosis appears to proceed in two stages: invagination of the membrane, followed by pinching off of the vesicle. The cell cortex is quite resistant to deformation, so the shape changes accompanying endocytosis incur a large energy penalty [19]. Consequently, the cell must generate a considerable mechanical force to deform the endocytic membrane. To do so, endocytosis must involve biochemical reactions at the endocytic site that control the pulling and pinching forces. In budding yeast, actin polymerization and myosin motor activity have been implicated in providing the pulling force for membrane invagination [10]. Pinching off of the membrane vesicle entails even larger membrane curvatures at the scission site than does generation of the invaginated membrane. In mammalian cells, dynamin GTPases have been proposed to act as “pinchases” that physically constrict membrane tubules [20],[21]. However, endocytic vesicles form in budding yeast despite the absence of dynamin at endocytic sites. In vitro studies have suggested a possible scission mechanism; an interfacial force arising at the boundary between two lipid phases can provide the driving force for vesicle scission [22],[23]. We previously proposed that such a mechanism might drive endocytic vesicle scission in vivo [11],[24].
Reciprocally, emerging experimental evidence suggests that membrane curvature created by mechanical force can modulate the local biochemical activities of several key endocytic proteins [25]. Experiments suggest that membrane curvature may act as a guiding signal to direct BAR (Bin/Amphiphysin/Rvs) domain proteins to the endocytic membrane invagination [26]. Conversely, BAR domain proteins (BDPs) are also capable of deforming the membrane into the preferred shape for their binding [27],[28],[29],[30],[31],[32]. However, in the context of the coherent process of endocytosis, the exact functional role of these physical properties of endocytic proteins remains elusive.
Here we attempt to combine detailed knowledge of endocytic protein dynamics and function in budding yeast with mechanochemical concepts to develop an integrated systems model for the endocytic internalization pathway. Our model stands in contrast to previous models [22],[23],[24]. Rather than focusing on one sub-process, our model seeks to reproduce the correct sequence of events in a coherent manner, including the local biochemical reactions and membrane shape changes. We propose a mechanochemical feedback mechanism that can generate successful endocytosis over a broad range of its parameter space. The model fits quantitatively the correct temporal and spatial profiles measured in budding yeast. Furthermore, when the parameters are varied to mimic endocytic mutants, the model accounts for many endocytic phenotypes in budding yeast and yields experimentally testable predictions. Finally, we argue that, despite some differences in molecular details, the underlying principles likely apply to mammalian endocytosis as well.
In this section, we will describe the qualitative features of our model. The quantitative mathematical formulations will be relegated to the Experimental Procedures.
Temporal control and spatial arrangement of proteins and the lipid PI(4,5)P2 at budding yeast endocytic sites are key features in the development of our model (Figure 1). First, each of the key endocytic proteins appears to localize along the membrane invagination with a distinct spatial profile, predicted by dynamic properties [8],[9],[11] and confirmed by EM [15]. Second, these proteins can be grouped into four “protein modules” based on their distinct dynamics and functions [9],[15]. Lastly, we previously obtained evidence for a PI(4,5)P2 “lipid module” that is dynamically regulated during endocytosis [11].
For this model, we describe clathrin-mediated endocytic dynamics on the level of functional modules, which allows us to look beyond roles of individual molecular players that may vary from one organism to another and to focus upon collective behaviors in membrane shape transformations and local biochemical pathways. Thus, our model can serve as a unified framework for endocytosis across diverse organisms. We propose that the five modules along with their functions are as follows (Figure 2 provides an overview of the model):
From a mechanical standpoint, the pulling forces generated by the actin/myosin functional module impinge on the bud and invaginate the membrane. The initial pinching force is generated as follows. Because of the protection afforded by BDPs on the tubule, more PIP2 is hydrolyzed at the bud region. This leads to lipid phase segregation—PIP2 levels along the membrane invagination differ, and the resulting interfacial force at the bud-tubule interface squeezes the neck. From a chemical perspective, the local chemical reactions (e.g., actin assembly, PIP2 hydrolysis) control pulling and pinching forces. Equally important, the resulting membrane curvature generated by the mechanical forces also influences the local reaction rates (Figure 2). In this way, endocytic dynamics are controlled by mechanochemical feedback between endocytic membrane shape changes (membrane curvature) and the local chemical reactions that control the mechanical forces (pulling and pinching forces). This key notion, as we will show below, is essential for the robustness of the sequential endocytic protein recruitment and timely vesicle scission.
This qualitative picture is captured by Equations 1–6 in the Experimental Procedures. The coupling between the mechanical and chemical processes of endocytosis is specified by the dependence of the reaction rates on membrane curvature and by the dependence of the local membrane curvature and the mechanical force on the local levels and activities of the functional modules. To calculate the dynamics of endocytic events, we numerically integrate Equations 1–6 over time starting from the initial condition: the endocytic membrane is flat and the initial coverage for all of the protein modules is set to zero. The initial PIP2 coverage is set to 2% corresponding to its normal average level [48]. At each step, the system is characterized by the instantaneous shape of the endocytic membrane and the local levels of the functional modules as represented in mole fraction. The values of the parameters used in the model are listed in Table 1 with references in Protocol S1. Below, we first study the endocytic dynamics of budding yeast by choosing the parameter set that quantitatively fits the time-lapse experimental data in Figure 1A. We then vary the parameters to mimic mutant experiments to predict and analyze the associated phenotypes.
As the model dynamics are controlled by many parameters in Equations 1–6, there could in principle be many outcomes depending on parameter choices. To circumvent this problem, 21 of the 25 parameters used in the model were taken from independent experiments (Table 1 in Protocol S1). The four unmeasured parameters all characterize BDP dynamics; they are the intrinsic BDP recruitment rate, actin-aided recruitment rate, turnover rate, and the relative timescale of BDP dynamics with regard to actin dynamics. With 21 measured parameters being fixed, we only vary the four free value parameters to fit the five time-lapse curves of endocytic dynamics observed in wild-type budding yeast (Figure 1A). The values of these four parameters are constrained because these kinetic rates must be comparable to those experimentally determined for each of the other functional modules. The dynamics of all of the modules are tightly coupled: one sub-process cannot be much faster/slower than the others. In what follows, we use specific proteins or lipids to represent the corresponding functional modules. We stress from outset that the goal of the paper is to illuminate the collective dynamics of endocytosis generated by the interactions among the functional modules, rather than identifying detailed molecular players.
Figure 3A shows that the endocytic dynamics predicted by our model (continuous lines) fit quantitatively with the experimental data (discontinuous lines) [8],[11, and the measurements in this paper]. Figure 3B shows snap shots of the corresponding computed membrane shape changes (a movie of the process derived from model calculation is provided in Video S1). Because the fitting parameters are constrained by measurements from independent studies, the agreement between our theoretical results and experimental observations lends validation to our model. An important feature of the process is that each functional module is activated sequentially in step with the membrane shape changes (Figure 3A and 3B). We next describe steps in the endocytic process in greater detail based on our model.
Early in the process (0–20 s, Figure 3A) coat proteins begin to accumulate. During this period, the membrane is deformed by the coat proteins, which generate a small dome (less than 50 nm in height and ∼50 nm in width, t∼20 s in Figure 3B). However, there is a delay before actin polymerization fully commences, because it takes a while for the nucleation factors to be recruited and activated and because actin assembly is autocatalytic due to Arp2/3 activation by actin filaments. Without the assistance of the actomyosin force, the dome-like membrane deformation would not progress further, which is consistent with observations from recent EM studies [15]. Indeed, this dome shape could be the prerequisite for further development of a deep invagination, because the local membrane shape may provide a suitable angle at which the F-actin pulling force can be exerted upon the bud region effectively.
At ∼20–25 s (Figure 3A), F-actin polymerization is promoted by nucleation factors recruited by the coat proteins, and the pulling force upon the bud region increases. This drives the endocytic membrane to invaginate further (t∼22 s in Figure 3B). As the membrane invaginates, actin monomers rapidly incorporate into the existing actin filaments with their barbed ends facing the cell cortex [8], while myosin pushes the actin network away from the plasma membrane into the cytoplasm. Meanwhile, the PIP2 phosphatase begins to accumulate all over the endocytic site. Concurrently, BDPs also start to accumulate along the tubule region rapidly, and they increase from 10% to the peak level in only 3 s (Figure 3A).
Now the question is: what drives the rapid BDP accumulation? We show that curvature-sensing and deforming activities of BDPs form an intrinsic positive feedback loop (see quantitative calculations in Figure S1). As schematized in Figure 4A, as they bind to the membrane, BDPs deform the adjacent membrane into the preferred curvature for their binding. This leads to a faster recruitment rate, which further promotes BDP recruitment and tubulation of the membrane. This positive feedback also explains and reconciles the two classes of experimental observations, which provided evidence for curvature-sensing and membrane-deforming activities [25],[26],[27],[28],[29],[30],[31],[32]. In our scenario, actin assembly and myosin contractile forces invaginate the membrane. The resulting membrane curvature fits relatively favorably to the preferred shape of BDPs and, hence, promotes rapid BDP binding at the right location and at the right time due to the curvature-sensing activity. In turn, BDP binding invaginates the membrane further and generates optimal curvature for BDP binding in the elongating tubule, which self-accelerates BDP accumulation. Thus, the initial membrane invagination generated by the actin/myosin force triggers the positive feedback between BDP binding and membrane tubulation.
During this same period, PIP2 hydrolysis rates are faster on the bud than on the tubule, as the BDPs protect the PIP2 on the underlying tubule from hydrolysis. Lipid-protein interactions involving BDPs could limit PIP2 diffusion in the membrane [45], allowing formation of a lipid-phase boundary. An interfacial force at the bud-tubule boundary thus starts to build up, constricting the neck.
Eventually (t∼29 s in Figure 3A and 3B), the interfacial force narrows the neck down to <5 nm, at which distance the opposed bilayers would fuse spontaneously [49], resulting in rapid vesicle scission. Upon vesicle scission, BDPs disassemble from the membrane tubule within 3 s as the tubule retracts due to loss of the actin pulling force. A second crucial effect of the PIP2 phosphatase activity on the vesicle bud is to trigger disassembly of the endocytic coat (t∼25–29 s, Figure 3A). As coat proteins disassemble, the F-actin attachments to the bud weaken, resulting in loss of pulling force on the invagination. We predict that this leads to a small retraction of the endocytic membrane tip concurrent with vesicle scission (see Figure 3A) and propose that loss of the pulling force on the membrane may be a prerequisite for vesicle scission.
Our description of the endocytic process (Figure 3) raises the following interesting questions: How is the interfacial scission force generated? How does vesicle scission occur so rapidly? And what turns off the positive feedback loop for BDP assembly and drives their extremely fast disassembly? In this section, we propose answers to these questions. Our proposal that an interfacial force can drive vesicle scission is supported by in vitro experiments [22],[23], in which lipid phase segregation is induced by lowering temperature. In vivo, however, cells always maintain constant temperature; instead, lipid-protein interactions could be utilized to yield effective lipid phase segregation. Here, we present two possible scenarios for how the interfacial force is developed in endocytosis (schematized in Figure 5A): (1) As PIP2 hydrolysis at the bud eliminates hydrogen bonds that had bridged the interfacial boundary, hydrogen bond shielding of the hydrophobic hydrocarbon chains is lost, and at the boundary these aliphatic tails are exposed to water, which is energetically unfavorable. The resulting line tension is proportional to the PIP2 difference across the interface, which contracts to minimize these unfavorable contacts, thus squeezing the neck. (2) The reduced hydrogen bond network at the bud lowers the membrane surface tension of the outer leaflet, which thus tends to expand. Effectively, this is a lateral surface pressure that propagates from the high-lateral pressure region towards the interfacial boundary. Due to the local concavity of the membrane created by the initial interfacial tension, this lateral pressure is directed inwards at the phase boundary and provides an additional pinching force. This additional lateral pressure also increases with the difference in PIP2 levels across the phase boundary (see the detailed derivations in Protocol S1).
Figure 5B shows the calculated time course for interfacial force development during endocytosis, while Figure 5C shows the calculated profiles for PIP2 levels around the bud-tubule boundary at different time points. Figure 5B and 5C show that the interfacial force undergoes rapid changes. During t∼0–21 s, PIP2 accumulates uniformly over the entire endocytic site, as promoted by kinase-mediated synthesis. From around t∼21 s (Figure 5C), PIP2 levels decline non-uniformly; consequently, the interfacial force starts to build up (Figure 5B). This spatial non-uniformity is because around the same time as the phosphatase is recruited, BDPs start to accumulate at the tubule region of the endocytic membrane (t∼21 s in Figure 3A). As a result of the BDP protection at the tubule, relatively more PIP2 is hydrolyzed on the bud, leading to lipid phase segregation at the BDP–coat protein boundary. This phase separation gives rise to the initial interfacial force at the phase boundary.
From t∼21–27 s (Figure 5B and 5C), the interfacial force grows sharply. Such rapid growth of the interfacial force is the result of another positive feedback loop involving curvature-enhanced PIP2 hydrolysis. We schematize the qualitative mechanism in Figure 4B. As the initial interfacial force squeezes the neck, it creates a higher mean curvature at the interface. The higher the mean curvature of the membrane, the more PIP2 is exposed and susceptible to phosphatase activity. Consequently, more PIP2 is depleted at the interface region along the membrane invagination. Thus, a larger difference in local PIP2 levels bounding this location is induced (∼21–27 s, Figure 5B and 5C), which in turn speeds up the growth of the interfacial force and, hence, further squeezes the interface. This is a self-accelerating process.
The sharp dip of the PIP2 levels around the bud-tubule interface compared to the smaller difference between those of tubule and bud (t∼23 s and 27 s in Figure 5C) suggests that curvature-dependent PIP2 hydrolysis is the predominant driving force for generating the interfacial force. Our model thus predicts that the pinching force arises as a result of differential phosphatase activity along the membrane invagination. This prediction is consistent with the observations that phosphatase activity is essential for endocytic vesicle scission in yeast, that the phosphatase concentrates at the endocytic site during the late stages of endocytic vesicle internalization, and that it moves into the cell with the forming vesicle, possibly suggesting enrichment at the vesicle tip [11].
During t∼27–29 s (Figure 5B and 5C), as the pinching force squeezes the neck, the membrane curvature in the radial direction of the tubule deviates from the optimal shape for BDP binding (t∼28 s and 29 s in Figure 3B). This deviation acts as a “disassembly signal” and invokes the intrinsic positive feedback loop between curvature sensing and curvature deforming of BDPs (Figure 4A), triggering the rapid BDPs turnover (∼27–29 s in Figure 3A). Meanwhile, PIP2 gets hydrolyzed not only at the bud but also on the tubule due to the lack of BDP protection (t∼29 s, Figure 5C). Although this leads to a fast decrease in the interfacial force (∼27–29 s, Figure 5B), the pinching force is still sufficient to drive rapid vesicle scission according to our calculations.
We need to point out that, while the in vitro systems on lipid phase segregation are crucial for identifying mechanical forces that might be involved in vesicle scission, the experimental conditions used are quite different from the in vivo conditions during endocytosis. Once the lipid phase segregation takes place in the in vitro systems, the resulting interfacial force persists and there is no time limit for the vesicle scission process. All that matters is that the interfacial force needs to be sufficiently large to overcome the membrane bending resistance [24],[50]. In cells, the timing of the lipid phase segregation is predicted to be critical for successful endocytosis. The threshold interfacial force value required for scission can be determined by force-balance calculations [24],[50]. A rapid nonlinear time course for interfacial force development in endocytosis means that successful scission in vivo can only occur within a short time window (the shaded region in Figure 5B).
In this section, we will explore in detail how mechanochemical feedback ensures the precise timing and sequence of endocytic events and guarantees rapid endocytic vesicle scission. In Figure 6A–6D phase diagrams for endocytosis are computed for different pairs of model parameters. These diagrams serve several purposes. First, they show that the model is robust: it can generate successful endocytosis over a large range of the parameters. Second, equipped with these phase diagrams, we can vary the parameters to mimic the conditions of mutant experiments. Third, they constitute an independent experimental test of the model. This is because the identities of the functional modules were in part derived from mutant experiments, but we did not explicitly take into account the mutant phenotypes in the model. That is, we used the five time-lapse curves and membrane shape changes to determine the four free parameters in the model, and then used these parameter values to predict mutant phenotypes. Thus, these predictions are independent of the parameter set, and consequently the agreement between predicted and observed phenotypes constitutes cross-validation of the model. Finally, based on the calculated phase diagrams, we can predict endocytic phenotypes for mutants that have not yet been made, thus guiding further experiments.
Figure 6A shows that endocytosis can only be successful when the curvature-dependent PIP2 hydrolysis rate is sufficiently fast. Otherwise, the PIP2 level difference across the interfacial boundary will not have had sufficient time to grow before the membrane bending energy resists squeezing and quickly balances the interfacial force without triggering the positive feedback loop (Figure 4B). Accordingly, the absence of positive feedback between the interfacial force and the local membrane curvature leads to a distinct phenotype (phenotype 1, wherein the PIP2 hydrolysis rate k2 is reduced from 20 (nm) per second to zero): F-actin associated forces could still drive membrane invagination; the interfacial force, however, would not squeeze the neck effectively, because the force cannot grow large enough. Thus, the whole system would eventually reach a mechanochemical equilibrium wherein a slightly curved membrane invagination could persist for a time without vesicle scission. This phenotype is consistent with the budding yeast mutant sjl1Δ sjl2Δ [11], wherein the PIP2 hydrolysis is dramatically reduced.
If the PIP2 hydrolysis rate is very fast but independent of the local membrane curvature, then the positive feedback between the interfacial force and the local membrane curvature is ablated (see Figure 4B and 45). Without this positive feedback, the interfacial force would always remain at its initial basal level, which is insufficient to pinch off the vesicle (see details in Figure S2). Successful endocytosis, therefore, requires the positive feedback between interfacial force and curvature-dependent PIP2 hydrolysis activity. This is further dictated by two conditions: first, the PIP2 hydrolysis rate must be faster than the typical response time scale of the membrane, and second, PIP2 hydrolysis must be curvature-dependent. The former can be tuned by the local concentration of phosphatases, and the latter is intrinsic to the mechanism of enzyme activity.
Figure 6A shows that, even with a sufficiently high curvature-dependent PIP2 hydrolysis rate, endocytosis may not be successful unless the protection of PIP2 at the tubule by BDPs is sufficiently effective (large K2). Otherwise (small K2), the resulting interfacial force would be too small to drive vesicle scission. On the other hand, if the protection is too effective, then PIP2 levels at the tubule would be maintained at a high level, which in turn would lead to persistent BDP accumulation. As BDPs tend to deform the membrane to a specific, preferred shape (diameter ∼30 nm), persistence of the BDPs would effectively hold the neck and prevent any further narrowing of the membrane tubule, hindering vesicle scission. This leads to prediction of a unique phenotype (phenotype 2, wherein the protection strength of PIP2 hydrolysis at the tubule region increases from 0.5 μM−1 to 2.5 μM−1), in which the absolute levels and the lifetimes of the BDPs would increase significantly as compared with the wild-type situation. Furthermore, a long and narrow membrane invagination could persist without vesicle scission. This is because BDPs have their own preferred shape (a tubule of ∼30 nm in diameter), and their persistence would tend to preserve the shape of membrane tubule, preventing any further squeezing in response to the interfacial force.
Our model predicts that within the successful endocytosis region in Figure 6A, increasing the curvature-dependent PIP2 hydrolysis rate k2 will speed up endocytosis and that this effect will saturate at large k2. This is because in this case endocytic dynamics are limited by the phosphatase recruitment rate, instead of by its activity. As shown in Figure 6B, positive feedback between interfacial force development and local membrane curvature will not develop if the phosphatase activity is not sufficient. Insufficient phosphatase results in a phenotype similar to those observed when PIP2 hydrolysis curvature dependence is insufficient, as shown in Figure 6A, and/or when PIP2 hydrolysis is independent of curvature, as shown in Figure S2.
On the other hand, endocytosis will also be impeded if the phosphatase is overexpressed or overactive at the endocytic site, which leads to phenotype 3 (where the curvature-dependent factor of phosphatase recruitment rate α increases from 100 nm to 500 nm). Here scission fails because the excessive phosphatase diminishes the initial PIP2 level difference across the bud-tubule boundary, thus preventing the development of the initial squeezing force. As a result, the membrane at the interface cannot be deformed sufficiently to invoke positive feedback between interfacial force development and the curvature-dependent PIP2 hydrolysis activity.
A surprising conclusion from our model is that coat proteins will still assemble at the endocytic site in the presence of excessive phosphatase and will disassemble slowly. This conclusion is based on the linear dependence of the PIP2 hydrolysis rate on the local membrane curvature, which is in accordance to experimental observations. PIP2 hydrolysis is relatively slow despite high phosphatase levels because the membrane is not highly curved (e.g., phenotype 3). Thus, even though the phosphatase recruitment is very fast in phenotype 3, its action is limited by the lack of membrane curvature, which is low because a pronounced phase boundary does not develop.
Figure 6C shows that successful endocytosis also critically depends on the coordinated dynamics of BDP recruitment and F-actin polymerization. Without actin polymerization, the endocytic membrane cannot become deeply invaginated. Failure to invaginate the membrane prevents BDP accumulation and the ensuing development of the interfacial force. Consequently, the membrane cannot deform into a deep invagination, nor proceed to vesicle scission. This situation is similar to having excessive phosphatase at the endocytic site, leading to phenotype 3 in Figure 6, consistent with actin-assembly inhibition phenotype in budding yeast [8].
When actin polymerizes normally, efficient endocytosis requires sufficiently fast BDP accumulation. Insufficient BDP recruitment would lead to phenotype 4 (wherein the BDP recruitment rate drops to zero): the endocytic membrane will be pulled out and will then retract without vesicle scission (a movie of the process is given in Video S2). This is because although the peak interfacial force is large enough to squeeze the neck in phenotype 4, the force declines so rapidly that the membrane does not have time to undergo deformation and, hence, the vesicle cannot be successfully pinched off. A large interfacial force can develop in the absence of the BDPs in phenotype 4 because the actin filaments contact actin-binding proteins associated with the coat so that the actin pulling force impinges on the entire bud region of the endocytic membrane, including the bud-tubule boundary. Although very small, the force from the actin module can still deform the membrane at the neck slightly, which activates the curvature-dependent PIP2 phosphatase activity. Hence, the positive feedback loop is triggered, leading to generation of a large interfacial force. However, without BDP protection, this large interfacial force is too short-lived and vesicle scission does not occur.
On the other hand, in the absence of sufficient numbers of BDPs, the high curvature of the membrane invagination generated by F-actin polymerization would still induce phosphatase recruitment, which would result in disassembly of the entire endocytic apparatus and retraction of the membrane invagination. This predicted phenotype is consistent with the phenotype of a budding yeast rvs167 (a BDP) knockout mutant [9] and a lipid-binding defective rvs167 point mutant (Kishimoto and Drubin, unpublished).
The lifetime of BDPs at endocytic sites is extremely short (∼10 s) in wild-type budding yeast [9],[11]. We have shown for phenotype 2 of Figure 6A that prolonged accumulation of BDPs could prevent endocytosis. A key message emerging from these two observations is that the interplay between the interfacial force and BDP turnover is critical for successful endocytosis. As the interfacial force squeezes the interface, it tends to narrow the adjacent membrane tubule, which deviates from the shape preferred by BDPs. This deviation leads to a curvature mismatch and acts as a “disassembly” signal for the BDPs as dictated by the BDP sensitivity factor (the exponential term χ in Equation 5). Accordingly, upon narrowing of the tubule, the higher the sensitivity factor χ, the faster the turnover of the BDPs, and hence the more that vesicle scission is facilitated. As Figure 6D shows, when the interfacial force is very large (>60 pN), it is capable of squeezing the interfacial boundary even if the BDPs are not disassembled; endocytosis would proceed normally even with prolonged BDP accumulation at the tubule (χ = 0). On the other hand, when the interfacial force is in an intermediate range (e.g., 30–60 pN), its action could be insufficient to overcome the bending resistance of the preferred membrane shape set by the BDPs. Given that the interfacial force will also dissipate in a short period of time (∼5 s, Figure 5A), a minimal level of curvature-dependent sensitivity in BDP accumulation is required to induce fast BDP turnover upon squeezing of the membrane tubule, relieving the bending resistance, and hence facilitating vesicle scission. This sets the lower threshold of the curvature-dependent sensitivity of BDP dynamics for successful vesicle scission. Note that the curvature sensitivity, χ, is central to the positive feedback between BDP recruitment and the local membrane deformation (Figure 4A). The above results imply that successful endocytosis requires that BDP binding feeds back positively with the underlying membrane shape.
During endocytosis, recruitment of the endocytic proteins is sequential and self-reinforcing, or autocatalytic [3],[4],[8],[9],[10],[11],[12],[35]. We propose that these features are properties of positive mechanochemical feedback loops between membrane curvature and the various reactions leading to vesicle formation and scission (Figures 2–6). To our knowledge, our model is the first of its kind that can coherently capture all of the key endocytic events in budding yeast. The dynamics predicted by the model fit well with time-lapse experimental measurements (Figure 1A). Moreover, the parameter diagrams in Figure 6 show that successful endocytosis can be realized over a broad range of parameter space. Thus the endocytic process is largely buffered against variations in the activities of specific molecular players.
Endocytosis in budding yeast evolves in a sequence of events that are explained by the model (as schematized in Figure 7A). As PIP2 accumulates at the endocytic site, it recruits coat proteins to the bud region that nucleate actin polymerization. Using anchorage to the coat proteins (e.g., Sla2), F-actin polymerization and myosin motor activity generate a pulling force that deforms the membrane into a tubule. The high curvature of the tubule in turn recruits BDPs that coat the tubule by binding to PIP2. The BDPs protect the PIP2 along the tubule from hydrolysis by the phosphatase. The coat proteins on the vesicle bud do not protect the PIP2 from hydrolysis as effectively, so a boundary region is created that develops a circumferential interfacial tension. This tension exerts a squeezing force on the phase boundary, which further increases the curvature at the bud neck, which in turn increases the hydrolysis there. Thus a positive feedback loop arises between membrane curvature and PIP2 hydrolysis rates at the interface, the result of which is the rapid growth of the interfacial force leading to vesicle scission (Figure 5). Furthermore, the positive feedback loop between the curvature-sensing and deforming activities of the BDPs ensures rapid turnover of the BDPs, facilitating timely vesicle scission. After scission, PIP2 is hydrolyzed all over the membrane surface, promoting disassembly of the entire endocytic apparatus. Therefore, it is the two intertwined positive feedback loops (Figure 4) that ensure rapid, robust, and timely endocytosis in budding yeast.
Our model depicts endocytosis at the level of functional modules, rather than at the level of particular proteins; the model enables us to discern the general features of the process and to dissect how the sub-processes fit together. As different proteins can play the same functional role in different organisms, our model can be extended to account for the endocytosis in other organisms. We have applied this framework to endocytosis in mammalian cells; the predictions from our model are largely consistent with experiments and provide further mechanistic insight, suggesting that similar principles may dictate the dynamics and robustness of protein recruitment, and the vesicle scission mechanism.
Our model predicts that the main profile of the endocytic membrane in mammalian cells is a constricted coated pit instead of the tubular structure in yeast. The interfacial force generated by lipid phase segregation is sufficient to pinch off the vesicle, and actin is largely dispensable while the membrane-deforming dynamin GTPase and clathrin are essential. We schematize our main findings of mammalian endocytosis in Figure 7B and relegate the detailed discussions to Protocol S1.
The model reproduces the behavior of observed endocytic mutant phenotypes and predicts several phenotypes that have not yet been studied in experiments. We predict that yeast endocytosis will be hindered if BDP protection of PIP2 on the tubule is either too weak or too strong, which is testable by BDP mutant analysis. Weak protection of PIP2 would reduce the PIP2 difference and, hence, the interfacial squeezing force. On the other hand, the more persistently the BDPs coat the tubule, the more resistant the tubule will be to the further squeezing from the interfacial force (phenotype 2 in Figure 6). This is because BDPs prefer a well-defined membrane shape (tubules of 30 nm diameter). In addition to rapid BDP assembly, therefore, BDP disassembly concurrent with vesicle scission is also essential for endocytosis.
The role of BDPs in vesicle scission suggests an explanation for dynamin mechanism that contrasts with the conventional view of dynamin as a pinchase (see Section F in Protocol S1 for a detailed discussion of dynamin). Dynamin disassembly precedes membrane fission [51],[52], which suggests that dynamin may act to disrupt local membrane structure, perhaps through generation of a phase boundary. Disassembly would be required to release the underlying membrane, allowing a line tension to constrict the vesicle neck to drive scission.
Successful endocytosis also entails three constraints on PIP2 hydrolysis rates, all of which lie at the heart of the mechanochemical feedback loop and can be tested by in vivo and in vitro experiments. First, the PIP2 hydrolysis rate must be curvature-dependent (see Figure S2). Second, it must be faster than the response time scale of the membrane deformation (Figure 6A). Third, it must be slower than the time scale for assembling the endocytic apparatus (Figure 6B). We predict that when the PIP2 hydrolysis rate drops below a threshold, endocytosis will cease, but the endocytic membrane invagination will persist (phenotype 1 in Figure 6). Thus, the phosphatase not only uncoats proteins from the endocytic vesicle, but it also is essential for vesicle scission. This dual function makes sense because endocytosis is a sequential process: each step paves the way for the next one. The coat proteins on the bud must disassemble upon—or shortly after—vesicle scission. Uncoating is essential for the fusion of endocytic vesicle with early endosomes and coat protein recycling. This prediction provides a fresh perspective on the functions of phosphatase/lipase in endocytosis in yeast as well as in mammalian cells, e.g., synaptojanin in neurons [36].
Given the small number of proteins present at each endocytic site at different times in the process (∼10–100) [10],[53], it would appear that the process should be very stochastic. Typically, stochastic protein recruitment arises from variations in the assembly “source signal” and in the number of proteins being recruited. The rapid sequential recruitment of endocytic proteins, such as the BDPs and phosphatase, implies a highly cooperative process: the Hill coefficient for BDP recruitment by actin is >6 as inferred from [9],[10]. Thus, without compensating mechanisms, small variations in the source signal would be amplified to large uncertainties in recruitment. And yet the timing of endocytic protein recruitment is very robust, and endocytosis proceeds smoothly. The effects of small variations in protein levels and activity could be overcome if extremely specific protein-protein interactions acted as a template for recruitment, which requires the free energy decrease for protein binding to be well above the level of thermal fluctuations, i.e., >10 kBT.
Our model implies an alternative mechanism: using local membrane curvature as the source signal; i.e., to assemble and disassemble BDPs. If we add random noise to Equations 1–5 and Equation 6 to mimic the instantaneous fluctuations in protein numbers and membrane shape fluctuations, respectively, endocytosis remains stable up to 20%–30% variation in the maximum levels for each module (unpublished data). The reason for this stability is the small diameter of the endocytic invagination (∼50 nm). On this scale, the membrane is quite stiff, and so the membrane curvature will not fluctuate much because of the energy penalty associated with stochastic fluctuations in membrane shape (∼100 kBT) [54],[55]. Moreover, since a curvature mismatch increases the free energy associated with BDP binding, the membrane curvature modulates the BDP recruitment rate via a Boltzmann factor (Equation 5). Thus, the local membrane curvature is instantaneously stable throughout the process and dictates the timing and location of BDP assembly and disassembly accurately despite stochastic fluctuations. Hence, the mechanochemical feedback has a build-in robustness that ensures successful endocytosis.
In the future, much experimental and theoretical work will be required to test and refine our model. Here we discuss aspects of our model for budding yeast endocytosis that we have not yet addressed. A related discussion for mammalian cells is presented in Section F in Protocol S1.
For our model, the key to promoting rapid vesicle scission was to invoke positive feedback between growth of the interfacial force and curvature-dependent PIP2 hydrolysis at the interfacial boundary, resulting in a sharp dip in the local PIP2 concentration at the interface. For this mechanism, all that is needed is to induce a localized membrane deformation (i.e., higher mean curvature) at a specific site along the membrane tubule. This in turn will trigger a positive feedback effect on PIP2 hydrolysis. There are many ways in which a localized membrane deformation can be generated. In this paper, we only entertained one scenario, in which the initial squeezing of the membrane at the interfacial boundary is the result of an initial PIP2 level difference (lower in the bud region) due to BDP protection of PIP2 hydrolysis on the tubule. However, other scenarios are also feasible. For instance, as phenotype 4 shows, even without BDPs, the impact from normal actin/myosin force could deform the membrane neck so as to invoke positive feedback and hence a large interfacial force. Although in this case the interfacial force is too short-lived to drive vesicle scission, this scenario nonetheless suggests other avenues to generate a sufficiently strong and persistent force. Also, it could be that the coat proteins protect PIP2 on the bud more effectively than the BDPs protect PIP2 on the tubule. This will result in a higher PIP2 level at the bud relative to the tubule, which could equally well induce an interfacial force. Although this scenario seems less likely due to the apparent concentration of the phosphatase at the bud tip, clearly experimental work is needed to determine how yeast pinch off endocytic vesicles in the absence of dynamin.
Also, studies on the mechanisms that recruit PIP2 phosphatases to endocytic sites are needed. In fact, actin has been shown to recruit the phosphatase via the actin-binding protein Abp1 [11],[35], although this effect alone cannot account for the full phosphatase recruitment to the endocytic site in yeast [11]. What is not clear is whether actin or the actin-dependent membrane curvature, or the combined effects, are responsible for PIP2 phosphatase recruitment. In our model, we treated PIP2 phosphatase recruitment as curvature dependent without delving into the specific contributions of direct actin-mediated recruitment versus indirect membrane curvature-dependent recruitment. We can show that the curvature-dependence of PIP2 phosphatase recruitment is not essential for efficient endocytosis as long as the effective phosphatase recruitment rate is neither too fast nor too slow as compared to PIP2 synthesis (Figure S4) and the hydrolysis rate is curvature-dependent. Future experimental studies must mechanistically address the contributions of BDPs, actin, and coat proteins in the vesicle formation process.
In summary, our model is based on the notion that the local curvature of the endocytic membrane is both slave to, and master over, the accompanying biochemical reaction pathways. The coupling between curvature and biochemical reactions orchestrates a robust sequence of events leading to vesicle scission. Formulating the model in terms of functional modules allowed us to look beyond the molecular details and explore the larger features of how membrane dynamics and biochemical reactions fit together during endocytosis. This scheme can quantitatively describe clathrin-mediated endocytosis in budding yeast and the analogous process in mammalian cells. Thus, our model can serve as a unified framework for dissecting endocytosis in general.
We incorporate the qualitative ingredients of the model into a set of quantitative equations. The detailed assumptions and the choices of the parameters are given in Protocol S1. Equations 1–5 describe the dynamics of the chemical reactions of the functional modules on the surface of the endocytic membrane. Levels of functional modules are expressed as the coverage fraction (mole fraction). We assume that the endocytic membrane has cylindrical symmetry. The local spatial coordinates along the membrane surface represent the arc length s with unit length 1 nm. The local membrane shape is uniquely defined by the tangent angle φ(s) and the radius r(s) (see Figure 2). The bud region is defined by the arc length s = 0–100; the tubule region is defined by s = 101–500.
PIP2 dynamics in the bud region (Notation: P):(1a)
PIP2 dynamics in the tubule region:(1b)
Enzyme (lipid phosphatase or lipase) dynamics (Notation: E):(2)
Coat protein dynamics in the bud region (Notation: C):(3)
Actin dynamics in the bud region (Notation: A):(4)
BDP dynamics in the tubule region (Notation: B):(5)
In Equations 1–5, Ω(s) and Ω(R)(s) are the mean curvature and the curvature in radial direction of the local membrane invagination, respectively, which are defined by local membrane orientation φ(s) and radius r(s) (see Figure 2 and Protocol S1 for their formula). and are the preferred curvatures by coat proteins at the bud and by the BDPs at the tubule, respectively. ΩC and ΩB are the preferred curvatures for the bud region and the tubule region, respectively, when they are fully covered by their corresponding proteins (C = 1, B = 1). The key mechanochemical couplings are: the PIP2 hydrolysis rate linearly depends on the local membrane curvature in Equation 1; BDP recruitment rate depends exponentially on its fit to the local membrane curvature in Equation 5. Furthermore, term in Equation 5 represents the actin-aided BDP recruitment, where is the average actin level at the endocytic site (see Protocol S1 for details).
The feedback between the chemical reactions and the membrane shape is specified by how the local chemical reactions directly control the membrane dynamics. The membrane dynamics is governed by Equation 6:(6)Here, F[φ (s)] is the Helfrich-like free energy for the endocytic membrane, which is characterized by the membrane bending energy and surface tension that specify the energy penalty associated with membrane deformations. Γ is the relative timescale of the membrane dynamics compared to the local chemical reactions.
Equation 6 describes the membrane dynamics affected by the interfacial force λ, the spontaneous curvatures , and the pulling force f in the bud region, which are all controlled by the local chemical reactions. The interfacial force λ is a function of the PIP2 level difference across the interface between the bud region and the tubule region, , where λ0 is the interfacial force constant and s = 100 is the interfacial boundary position (see Figure 2). Note that the pulling force on the bud region must anchor to the coat protein to be effective. We neglect protein diffusion in Equations 1–5 and the in-plane hydrodynamics of membrane flow in Equation 6. The justifications for these assumptions are given in Protocol S1.
|
10.1371/journal.pgen.1000883 | Multiple Signals Converge on a Differentiation MAPK Pathway | An important emerging question in the area of signal transduction is how information from different pathways becomes integrated into a highly coordinated response. In budding yeast, multiple pathways regulate filamentous growth, a complex differentiation response that occurs under specific environmental conditions. To identify new aspects of filamentous growth regulation, we used a novel screening approach (called secretion profiling) that measures release of the extracellular domain of Msb2p, the signaling mucin which functions at the head of the filamentous growth (FG) MAPK pathway. Secretion profiling of complementary genomic collections showed that many of the pathways that regulate filamentous growth (RAS, RIM101, OPI1, and RTG) were also required for FG pathway activation. This regulation sensitized the FG pathway to multiple stimuli and synchronized it to the global signaling network. Several of the regulators were required for MSB2 expression, which identifies the MSB2 promoter as a target “hub” where multiple signals converge. Accessibility to the MSB2 promoter was further regulated by the histone deacetylase (HDAC) Rpd3p(L), which positively regulated FG pathway activity and filamentous growth. Our findings provide the first glimpse of a global regulatory hierarchy among the pathways that control filamentous growth. Systems-level integration of signaling circuitry is likely to coordinate other regulatory networks that control complex behaviors.
| Signal integration is an essential feature of information flow through signal transduction pathways. The mechanisms by which signals from multiple pathways become integrated into a coordinated response remain unclear. We show that multiple pathways that regulate filamentous growth converge on a differentiation-dependent MAPK pathway. Our findings indicate that more extensive communication occurs between signaling pathways that control the filamentation response than has previously been appreciated. We suggest that global communication hierarchies regulate information flow in other systems, particularly higher eukaryotes where multiple pathways typically function simultaneously to modulate a complex response.
| Signal transduction pathways regulate the response to extracellular stimuli. Complex behaviors frequently require the action of multiple pathways that act in concert to reprogram cell fate. In metazoan development for example, a highly regulated network of interactions between evolutionarily conserved pathways like Notch and EGFR coordinates every facet of cell growth and differentiation [1]. An important question therefore is to understand how different pathway activities are coordinated during complex behaviors. Addressing this question is increasingly problematic because signaling pathways operate in vast interconnected web-like information networks [2]. Miscommunication between pathways is an underlying cause of diseases such as cancer [3], and therefore it is both critically important and extremely challenging to precisely define the regulatory connections among signaling pathways.
The budding yeast Saccharomyces cerevisiae undergoes a variety of different responses to extracellular stimuli as a result of the function of evolutionarily conserved signal transduction pathways. In response to nutrient limitation, yeast undergoes filamentous growth [4],[5],[6], a cellular differentiation response in which changes in polarity, cell-cycle progression, and gene expression induce the formation of branched chains of interconnected and elongated filaments. The filamentous cell type is widely regarded as a model for differentiation [7],[8],[9],[10], and in pathogens like Candida albicans, filamentous growth is a critical aspect of virulence [11],[12],[13].
A number of different pathways are required for filamentous growth (Figure 1A, left panel). These include a MAPK pathway commonly referred to as the FG pathway (Figure 1A, right panel [14],[15],[16]), the RAS pathway [4],[17], the target of rapamycin or TOR pathway [18], the RIM101 pathway [19],[20],[21], the retrograde pathway (RTG [7]), the inositol regulatory transcription factor Opi1p [22], and a global glucose control protein kinase Snf1p [23],[24],[25]. It is not clear whether these different pathways function together or independently to regulate filamentous growth. This question is compounded by the fact that several hundred other proteins have been implicated in the filamentation response [7],[26],[27].
To identify new aspects of filamentous growth regulation, we developed a screening approach to identify regulators of the MAPK pathway that controls filamentous growth. The FG pathway is regulated by the signaling mucin Msb2p [28], a cell-surface glycoprotein [29] that mediates signaling through the RHO guanine nucleotide triphosphatase (GTPase) Cdc42p [30]. Msb2p is processed in its extracellular domain by the aspartyl protease Yps1p, and release of the extracellular domain is required for FG pathway activation [31]. By measuring release of the extracellular domain of Msb2p in complementary genomic collections, we identified new regulators of the FG pathway. Unexpectedly, many of the major filamentation regulatory pathways (RAS, RIM101, OPI1, and RTG) were found to be required for MAPK activation. Our study indicates that the pathways that control filamentous growth are connected in co-regulatory circuits, which brings to light a systems-level coordination of this differentiation response.
To identify new regulators of the FG pathway, secretion of the extracellular domain of Msb2p [31] was examined using a high-throughput screening (HTS) approach in complementary genomic collections. Similar approaches have identified regulators of protein trafficking by mis-sorting and secretion of carboxypeptidase Y [32],[33],[34]. An ordered collection of 4,845 mutants deleted for nonessential open reading frames (ORFs, [35]) was transformed with a plasmid carrying a functional epitope-tagged MSB2-HA fusion gene, and transformants were screened by colony immunoblot to identify mutants with altered Msb2p-HA secretion (Figure 1B). Computational methods were used to quantitate, normalize, and compare secretion between mutants, which allowed ranking by the level of secreted Msb2p-HA (Table S3). As a result, 67 mutants were identified that showed reduced secretion of Msb2p-HA (Figure 1C, yellow), and 58 mutants were identified that showed elevated secretion (Figure 1C, green).
The secretion of Msb2p was also examined using an overexpression collection of 5,411 ORFs under the control of the inducible GAL1 promoter [36]. This collection allows examination of essential genes and can be assessed in the Σ1278b background, in which filamentous growth occurs in an Msb2p- and MAPK pathway-dependent manner [37]. Approximately 390 genes were identified that influenced the secretion of Msb2p-HA when overexpressed (Table S4). The two screens identified few common genes (Figure 1D, 1.5% overlap), which is not entirely surprising given that gene overexpression does not necessarily induce the same (or opposite) phenotype as gene deletion [38] and because the two backgrounds exhibit different degrees of filamentous growth [37]. Significant overlap was observed at the level of gene process/function (Figure 1E, 89% overlap), which resulted in classification of genes into different functional categories (Figure S1; Tables S3 and S4). Introduction of an MSB2-lacZ reporter showed a high correlation between genes that affect Msb2p secretion and MSB2 expression (Figure 1F; compare 1d to MZ). Because MSB2 is itself a target of the FG pathway [28], many of the genes identified likely influence the activity of the FG pathway.
In total, 505 genes were identified that influenced Msb2p secretion, which might represent an underestimate due to the stringent statistical cutoff employed. This unexpectedly large collection suggests that Msb2p is subject to extensive regulation, although presumably many of these genes exert their effects indirectly. To enrich for genes that specifically regulate the FG pathway, secondary tests were performed. In one test, the secretion profile of Msb2p was compared to the secretion profile of two other mucins, the signaling mucin Hkr1p [39],[40] and transmembrane mucin Flo11p [41],[42]. Almost half the genes were common to multiple mucins (44%, Figure 1G) and may function in the general maturation of large secreted glycoproteins. In a second test, the secretion profile of Msb2p was compared to a genomic screen for genes that when overexpressed influence the expression of a FG pathway-dependent reporter (Figure 1H). These tests eliminated general regulators of mucin maturation/trafficking and enriched for potential MAPK regulatory proteins (∼72 candidate genes).
Several mutants were identified that were expected to influence Msb2p secretion. Mutants lacking FG pathway components (see Figure 1A), which are required for MSB2 expression in a positive feedback loop [28], showed a defect in Msb2p-HA secretion (ste20Δ, ste50Δ, ste11Δ, and ste7Δ; Table S3). The mutant lacking the aspartyl protease Yps1p, which processes Msb2p and is required for release of the extracellular domain [31], was also identified (yps1Δ; Table S3). A subset of the genes that influence Msb2p secretion but not its expression might function through regulating expression of the YPS1 gene, which is highly regulated [43]. The cell-cycle regulatory transcription factors Swi4p and Swi6p [44],[45],[46],[47],[48], were also found to regulate Msb2p secretion (Table S3B and S7). The swi4 and swi6 mutants had different phenotypes in Msb2p-HA secretion (Table S3), which suggests that the Swi4p and Swi6p proteins may play different roles in regulating cell-cycle dependent expression of the MSB2 gene [49]. Some mating pathway-specific genes (STE5) were also identified, which may have an as yet unappreciated role in communication between the mating and FG pathways, which share a number of components [50].
As a proof-of-principle test, we disrupted fourteen genes that came out of the deletion screen in the Σ1278b background and tested for defects in Msb2p-HA secretion and FG pathway signaling. The test showed a >70% recovery rate based on phenotype and identified a novel connection between the tRNA modification complex Elongator and MSB2 expression, establishing this protein complex as a novel regulator of the MAPK pathway [51]. Therefore, secretion profiling is a valid approach to identify established and potentially novel regulators of the FG pathway.
To identify new genes that regulate the FG pathway, ∼50 candidate genes were disrupted in wild-type strains of the Σ1278b background, and the resulting mutants were tested for effects on FG pathway activity. To distinguish between mutants that influence filamentous growth from those that have a specific effect on FG pathway activity, a transcriptional reporter (FUS1) was used that in Σ1278b strains lacking an intact mating pathway (ste4) is dependent upon Msb2p and other FG pathway components including the transcription factor Ste12p (Figure 2A and 2B [28]). A number of potential MAPK regulatory proteins were identified by this approach, many of which have been implicated in regulating filamentous growth through their functions in other pathways.
Mitochondrial retrograde signaling (the RTG network), which is responsible for mitochondrial communication with the nucleus, was required for FG pathway signaling. The RTG network responds to the integrity of the mitochondria and nutrient state, specifically the availability of certain amino acids. The proteins Rtg1p, Rtg2p, and Rtg3p are the main targets of the pathway, and these factors initially induce the expression of genes involved in the TCA cycle [52]. Rtg1p, Rtg2p, and Rtg3p were required for FG pathway signaling (Figure 2A). These factors are under the control of Mks1p, which inhibits Rtg3p translocation to the nucleus, preventing expression of target genes [53]. Mks1p is subsequently under the control of Rtg2p, which will bind and sequester Mks1p and prevent its interaction with Rtg3p [54]. Consistent with its inhibitory role in the RTG pathway, Mks1p had an inhibitory role on FG pathway activity (Figure 2A and 2B). Rtg2p, in turn, is negatively regulated by the Tor1p complex via Lst8p [55]. Several mitochondrial components, including ribosomal subunits and enzymes of the TCA cycle, showed varied expression and signaling defects in the screen (Tables S3 and S4; Figure S7A). Deletion of Tor1p did not have the same effect as Mks1p (Figure 2A and 2B), but it has been shown that RTG can function independently of Tor1p inputs [56]. Consistent with a connection to RTG, the activity of the FG pathway was sensitive to certain amino acids such as glutamate (Figure S2).
Components of the Rim101p pathway [19], including the transcriptional repressor Rim101p and Dfg16p, which is required for processing and activation of Rim101p [20],[21], were required for FG pathway activity (Figure 2A). The Rim101p pathway is required for pH-dependent invasive growth, and the activity of the FG pathway was slightly sensitive to pH levels (Figure S2). Other regulators of the FG pathway included components of the RAS pathway (see below) and the inositol regulatory transcription factor Opi1p (Figure 2A [57],[58]), which has recently been tied to filamentous growth regulation [22].
The discovery that many filamentous growth regulatory pathways impinge on the FG pathway suggests a systems-level coordination between the pathways that regulate filamentous growth. We directly tested other filamentation regulatory proteins. Filamentous growth is regulated by the global glucose-regulatory protein Snfl1p [23],[24],[25]. Snf1p was not required for FG pathway activity (Figure 2A). The negative regulators Fkh1p/Fkh2p [59] and Nrg1p/Nrg2p [23],[60], which function to inhibit invasive growth, did not influence FG pathway activity (Figure 2A). Therefore, many but not all inputs into filamentous growth regulation also regulate the FG pathway. This finding may begin to account for the large number of genes identified by secretion profiling that impinge on the FG pathway.
The RAS pathway has previously been implicated in FG pathway regulation [61]. However, systematic genomic analyses have failed to substantiate a connection between the two pathways [8],[9],[10],[62],[63],[64],[65], and a prevailing consensus is that the pathways function independently and converge on common targets [66],[67]. We found that the GTPase Ras2p [67],[68] was required for the expression of FG pathway-dependent reporters (Figure 3A), which indicates that the RAS pathway does regulate the FG pathway.
To determine where in the FG pathway that Ras2p functions, genetic suppression analysis was performed using alleles of MSB2 and SHO1 that hyperactivate the FG pathway [31]. Previous genetic suppression analysis indicates that Msb2p functions above Sho1p in the FG pathway [28] (Figure 1A). A hyperactive allele of SHO1 (SHO1P120L) bypassed the signaling defect of ras2Δ, whereas the activated allele MSB2Δ100–818 did not (Figure S3A and S3B), indicating that Ras2p functions above Sho1p and at or above the level of Msb2p in the FG pathway. This result led us to test whether Ras2p regulates the expression of the MSB2 gene. Overexpression of MSB2 (GAL-MSB2) bypassed the agar-invasion defect (Figure 3B, left panels) and the cell elongation defect of the ras2Δ mutant (Figure 3B, right panels; Figure S3A), which indicates that Ras2p regulates MSB2 expression. We confirmed that Ras2p was required for the expression of an MSB2-lacZ reporter, including the starvation-dependent induction of MSB2 expression (Figure 3C). This effect was independent of the FG pathway, which also regulates MSB2 expression [28], because it was observed from a reporter lacking the Ste12p recognition sequence (Figure 3C, MSB2AG-lacZ) that makes MSB2 expression Ste12p-insensitive (Figure 3C, ste12Δ). Secretion profiling also uncovered the alternative guanine nucleotide exchange factor Sdc25p [69],[70]. We confirmed that Sdc25p was required for starvation-dependent MSB2 expression (Figure 3C).
Ras2p might regulate MSB2 expression by modulating cAMP levels through activation of adenylate cyclase [71]. As shown in Figure 3C, bypass of the MSB2 expression defect of the ras2Δ mutant was observed in cells lacking the phosphodiesterase PDE2 [72]. The ras2Δ pde2Δ double mutant showed wild-type expression of the MSB2 gene (Table S7), and overexpression of the PDE2 gene inhibited FG pathway activity (Table S4). Sensitizing MSB2 expression to cAMP levels represents a direct nutritional tie into MAPK regulation (Figure S2) and may explain why a large collection of genes that function in nutrient sensing were identified by secretion profiling (23.5%; Tables S3, S4, S5).
Ras2p also functions in a pathway that regulates the activity of the transcription factor Flo8p, which converges along with the FG pathway on the common target FLO11 [8],[66],[73],[74],[75]. The Flo8p pathway is co-regulated by Ras2p and the glucose receptor-heterotrimeric G-protein Gpr1p-Gpa2p [76],[77],[78],[79],[80]. Gpr1p, Gpa2p, and Flo8p were not required for MSB2 expression (Figure 3C, shown for flo8) or FG pathway signaling (Figure S2B). Ras2p also regulates a diverse collection of nutrient- and stress-related responses [81], and the ras2Δ mutant exhibits a number of phenotypes including glycogen accumulation, resistance to oxidative stress [82],[83], enhanced chronological lifespan [84],[85],[86], and resistance to oleate [87]. MAPK pathway mutants did not show these phenotypes (Figure S4A, S4B, S4C, S4D), which suggests that the FG pathway does not regulate Ras2p function. We therefore suggest that the two pathways function in a unidirectional regulatory circuit RAS -> MAPK.
Expression profiling using whole-genome DNA microarrays was used to validate the above results. The expression profiles of cells overexpressing MSB2 or containing a hyperactive allele (MSB2Δ100–818) were compared in wild-type cells, the ras2Δ mutant, and the ste12Δ mutant genome wide (Figure 3D). Most of the genes that showed MSB2-dependent induction (Figure 3D, red) were reduced in the ras2Δ and ste12Δ mutants (Figure 3D, green). Expression profiling identified new targets of the FG pathway, including genes that function in nutrient scavenging, respiration, and the response to stress (Figure 3D; yellow, N/S, Nutrient/Stress). Targets of the RIM101 pathway [19] were also identified (Figure 3D, purple). The induction of nutrient scavenging genes may represent a feed-forward loop where starvation induces FG pathway activation, which results in the induction of genes to sustain foraging. We note that the observed gene expression changes in the microarray profiling studies encompass both direct and indirect effects. To summarize, RAS regulates the FG pathway by regulating the starvation-dependent induction of MSB2 expression.
To further explore the regulation of MSB2 expression, transcriptional regulatory proteins identified by secretion profiling were examined. The Rpd3p(L) HDAC complex was identified as a strong positive regulator of MSB2 expression. Rpd3p(L) comprises a 12 subunit complex [88] that includes the bridging proteins Sin3p [89],[90] and Sds3p [91]. Rpd3p, Sin3p, and Sds3p were required for starvation-dependent MSB2 expression and to produce wild-type levels of the Msb2p protein (Figure 4A). Rpd3p, Sin3p, and Sds3p were required for the induction of FG pathway reporters (Table S7) and invasive growth (Figure S6A).
Rpd3p functions in distinct large and small complexes with different cellular functions. Rpd3p(S) recognizes methylated histone H3 subunits and functions to repress spurious transcription initiation from cryptic start sites in open reading frames [92]. Rpd3p(S) contains unique proteins Rco1p and Eaf3p. In contrast, Rpd3p(L) is enriched for different proteins including Rxt2p [92], which is required for Rpd3p(L) function [93]. Immunoblot analysis (Figure 4A), transcriptional reporters (Figure 4B), and invasive growth assays (Figure S6A), showed that Rxt2p was required for MSB2 expression and FG pathway activation. In contrast, Rpd3p(L) was not required for pheromone response pathway activation (Figure S5), which shares components with the FG pathway but induces different target genes [5],[94],[95] and which does not require Msb2p function [28].
Histone deacetylases typically repress gene expression by causing compaction of chromatin into structures inaccessible to transcription factors. With respect to MSB2 expression however, Rpd3p(L) had a positive role. We tested whether Rpd3p(L) negatively regulated an inhibitor of MSB2 expression, such as the transcription factor Dig1p [96],[97],[98] but were unable to find evidence to support this possibility (Figure S6B). Rpd3p(L) functions as a positive regulator of some genes, including targets of the high osmolarity glycerol response (HOG) pathway by direct association with the promoters of HOG pathway targets [99], and in the regulation of some mating-specific targets [100]. Therefore, Rpd3p(L) might positively regulate MSB2 expression by associating with the MSB2 promoter. Chromatin immunoprecipitation (ChIP) analysis identified Rpd3p(L) at the MSB2 promoter (Figure 4C, Sin3p-HA). In the sin3Δ mutant, less Ste12p was found at the MSB2 promoter (Figure 4C and 4D, compare wild type to sin3Δ). Msb2p and the FG pathway also control STE12 expression. Ste12p-HA protein levels were reduced in the sin3Δ mutant (Figure 4E), as a result of a decrease in STE12 expression determined by q-PCR (data not shown). Therefore, Rpd3p(L) positively regulates the FG pathway by promoting MSB2 expression. Rpd3p(L) may also indirectly promote FG pathway activity by the regulation of other transcription factors that influence MSB2 expression.
Most of the regulators of MSB2 expression [MAPK, RAS, RIM101, OPI1, and Rpd3p(L)] also regulate expression of FLO11 [22],[66],[101], a target “hub” gene required for cell adhesion during filamentous growth and biofilm formation [8],[18],[42],[66],[102],[103],[104],[105]. Because FLO11 is one of the most highly regulated yeast genes and a target of the FG pathway, we explored the regulation of the two genes in detail.
MSB2 and FLO11 expression were compared by quantitative PCR in a panel of signaling and transcription factor mutants. Phenotypic tests were used to measure FG pathway activity and Flo11p-dependent adhesion (Table S7). As expected, MSB2 and FLO11 expression showed similar dependencies on a number of pathways (Figure 5A, top panel). FLO11 expression was also dependent on genes that did not affect MSB2 expression (Figure 5A, middle panel), including primarily components of the Flo8p pathway. Unexpectedly, FLO11 expression was also dependent on genes that inhibited MSB2 expression, such as snf2Δ and msn1Δ (Figure 5A, bottom panel). Such inhibition might be explained by a negative-feedback loop that functions to dampen the FG pathway.
Like Msb2p, Flo11p is a secreted protein [41], which allowed an extensive comparison between the regulation of the two proteins in different conditions and genetic contexts. The levels of secreted proteins were examined in microbial mats, which expand in an Msb2p (MAPK pathway)- and Flo11p-dependent manner [105]. Msb2p-HA and Flo11p-HA showed similar dependencies on Ras2p and the FG pathway transcription factor Ste12p (Figure 5B). A “ring” pattern of Msb2p-HA and Flo11p-HA secretion was observed in the ras2Δ mutant (Figure 5B), which may indicate that Ras2p is required to sustain MSB2 and FLO11 expression in cells exposed to modest nutritional stress, like the interior of expanding mats. In contrast, Flo11p-HA secretion was dependent on proteins on which Msb2p-HA was not (Figure 5C; snf2Δ and msn1Δ). Msb2p-HA and Flo11p-HA showed opposing secretion patterns under some conditions in specific regions of mat interiors (Figure 5D) in agreement with the idea that the two genes are differentially regulated in some contexts. Therefore, expression, phenotypic, and secretion profiling corroborate the existence of a complex, partially overlapping regulatory circuit that governs the regulation of MSB2 and a primary MAPK target hub gene FLO11 (Figure S8).
In this study, we address the fundamental question of how different signal transduction pathways function in concert to reprogram cellular behavior. This question is relevant to understanding the complex cellular decision-making underlying many cellular responses. Most studies on signal transduction regulation focus on a single pathway, or on highly related pathways that potentially engage in cross talk [50]. Here we report one of the first attempts to gain a systems-level perspective on the global regulation of a MAPK pathway that controls a differentiation response. This effort relied on primarily genomic approaches including a new technique called secretion profiling, which measures release of a cell-surface MAPK regulatory protein. The “top-view” perspective we obtained has shown that many of the major regulatory proteins that control filamentous growth also control MAPK signaling. This finding is challenging to accept given that many of the pathways are currently viewed as separate entities. Nonetheless, this view is consistent with an emerging systems-level appreciation of pathway regulation in complex situations – cell differentiation, stem cell research, and cancer - where pathway interconnectedness drives the rationale for drug development and new therapeutic endeavors.
We specifically show that four major regulators of filamentous growth are also required for FG pathway signaling (Figure 5E). The connections between the RIM101, OPI1, and RTG pathways to FG pathway regulation are entirely novel. We further show that the RAS pathway and several other pathways [MAPK, Swi4p, Rpd3p(L)] converge on the promoter of the key upstream regulator (MSB2) of the FG pathway. Like other signal integration mechanisms [8],[66], the MSB2 promoter is a target “hub” where multiple pathways converge. This regulatory point makes sense from the perspective that Msb2p levels dictate FG pathway activity [28]. Given that the FG pathway exhibits multimodality in regulation of pathway outputs [39], its precise modulation is critical to induce an appropriate response.
The regulation of the FG pathway differs from that of a related MAPK pathway, the pheromone response pathway, where secreted peptide pheromones provide the single major input into activation [5],[106],[107]. The pheromone response pathway is not regulated by the RAS, RIM, or RTG pathways and does not appear to be influenced by the Rpd3p(L) HDAC, although some target mating genes do require Rpd3p(L) for expression [100]. Both the FG and mating pathways are regulated by positive feedback [28],[65],[108],[109], presumably to provide signal amplification. Therefore, the FG pathway is distinguished in that it is highly sensitized to multiple external inputs. Because our screening approach was highly biased to identify regulators that function at the “top” of the pathway (at Msb2p), it is likely that the FG pathway may be subject to additional regulation not identified by this approach.
The networks that regulate filamentation signaling pathways appear in some cases to be redundant. For example, RAS regulates FLO11 expression through Flo8p and the FG pathway (Figure 5E). RIM101 similarly controls FLO11 expression through Nrg1p/Nrg2p and FG regulation (Figure 5E). However, any one of the major filametation regulatory pathways when absent appears fully defective for the response. Such parallel processing of signaling networks might allow for “fine-tuning” of the differentiation response, or alternatively to synchronize all of the pathways to the global regulatory circuit. Such synchronization may coordinate different aspects of filamentous growth (cell-cell adhesion, cell cycle regulation, reorganization of polarity) into a cohesive response.
Filamentous growth occurs within a narrow nutritional range [25], between high nutrient levels that support vegetative growth and limiting nutrients that force entry into stationary phase. The finding that Ras2p controls the overall levels of MSB2 expression extends the initial connection between RAS and the FG pathway [61] in several important ways. First, the results explain how Ras2p connects to the FG pathway above Cdc42p. Second, the data provide a link between nutrition and FG pathway signaling at the level of cellular cAMP levels. Third, Ras2p does not appear to function as a component of the FG pathway. This conclusion is based on the conditional requirement for Ras2p in MSB2 regulation, the suppression of the ras2Δ signaling defect by loss of PDE2, and the placement of Ras2p above MSB2 in the FG pathway. Our results fit with the general idea that RAS controls a broad response to cellular stress that encompasses FG pathway regulation. Ras2p has recently been shown to regulate the MAPK pathway that controls sporulation [110], a diploid-specific starvation response. Ras2p may therefore function as a general regulator of MAPK signaling in response to nutritional stress.
Examples by which MAPK pathways control target gene expression by recruitment of chromatin remodeling proteins is relatively common and include Rpd3p(L) regulation of HOG [99],[111],[112] and mating pathway [100] target genes. However, the regulation of MAPK activity through chromatin remodeling proteins provides a hierarchical mechanism for global cellular reprogramming. Regulation of FG pathway activity by Rpd3p(L) may contribute to the establishment of a differentiated state. Once activated, FG pathway activity may be sustained by Rpd3p(L) to reinforce accessibility to the MSB2 promoter and formation of the filamentous cell type. Although the connection between Rpd3p(L) and nutrition is not entirely clear, Rpd3p(L) preferentially localizes to highly expressed genes, such as those required for anabolic processes [113]. Therefore, Rpd3p(L) may coordinate overall growth rate with the persistence of MAPK activity. Precedent for Rpd3p(L) HDACs in regulating developmental transitions comes from Drosophila DRpd3, which together with the Chameau HAT function as opposing cofactors of JNK/AP-1-dependent transcription during metamorphosis [114]. HDAC regulation of MAPK activity may therefore represent a general feature of MAPK regulation.
In conclusion, we have identified an unprecedented degree of regulation of a differentiation-dependent MAPK pathway by multiple regulatory proteins and pathways. Our findings open up new avenues for exploring the relationships between pathways and the extent of their cross-regulation with the ultimate goal of understanding all functionally relevant pathway interactions in a comprehensive manner.
Yeast strains are described in Table S1. Plasmids are described in Table S2. Yeast and bacterial strains were manipulated by standard methods [115],[116]. PCR-based methods were used to generate gene disruptions, GAL1 promoter fusions [117],[118], and insertion of epitope fusions [119], using auxotrophic and antibiotic resistant markers [120]. Integrations were confirmed by PCR Southern analysis and DNA sequencing. Plasmids pMSB2-GFP and pMSB2-HA have been described [31], as have plasmids pMSB2-lacZ and pMSB2AG-lacZ [39]. Plasmids containing FG pathway targets KSS1, SVS1, PGU1, and YLR042C fused to the lacZ gene were provided by C. Boone [65]. pFLO8 was provided by G. Fink [37]. Plasmid pIL30-URA3 containing FgTy-lacZ was provided by B. Errede [121], and pFRE-lacZ was provided by H. Madhani [16]. The positions of the hemagglutinin (HA) epitope fusions were 500 amino acid residues for the Msb2p protein [31], at 1015 residues for the Flo11p protein [41], and at 1000 residues for the Hkr1p protein [39].
The single cell invasive growth assay [25] and the plate-washing assay [14] were performed to assess filamentous growth. Budding pattern was based on established methodology [122], and confirmed for some experiments by visual inspection of connected cells [25]. Halo assays were performed as described [123]. Microbial mat assays were performed as described [105] by growing cells on low-agar (0.3%) YEPD medium. Oleate medium was derived from standard synthetic medium lacking amino acids that was supplemented with 0.1% yeast extract, 0.5% potassium phosphate pH 6.0, and oleic acid (Toyko Kasei Kogyo Co. TCI) at a final concentration of 0.125% (w/v) solubilized in 0.5% Tween-20. Antimycin A, from Streptomyces sp. (Sigma-Aldrich, St. Louis, MO) was used at 3 µg/ml. Oligomycin (Sigma-Aldrich, St. Louis, MO) was used at 3 µg/ml, and rapamycin (Sigma-Aldrich) was added at 20 ng/ml. β-galactosidase assays were performed as previously described [124]. For some experiments, β-galactosidase assays were performed in 96-well format by growing cells containing the MSB2-lacZ reporter to saturation in synthetic medium lacking uracil (SD–URA) to maintain selection for the plasmid in 96-well plates at 30°C. Inductions were performed in duplicate, and the average of at least two independent experiments is reported. All experiments were carried out at 30°C unless otherwise indicated.
The MATa haploid deletion collection [35] was transformed with a plasmid carrying a functional hemagglutinin (HA)-tagged MSB2 gene (pMSB2-HA; [31]) using a high-throughput microtiter plate transformation protocol [125]. The deletion collection was manipulated with the BioMek 2000 automated workstation (Beckman-Coulter, Fullerton CA). For some experiments, a 96-fixed pinning tool (V & P Scientific, 23 VP 408) and plate replication tool (V & P Scientific, VP 381) were used. Sterilization was performed by sequential washes in 5% bleach, distilled water, 70% ethanol, and 95% ethanol. Ethanol (5 µl of 95%) was added to each transformation mix (200 µl) to increase transformation efficiency [126]. Transformants were harvested by centrifugation in 96-well plates, resuspended in 30 µl of water, and transferred to synthetic medium containing 2% glucose and lacking uracil (SD-URA) medium in Omnitrays (VWR International Inc. Bridgeport NJ). Transformants were pinned to SD-URA for 48 h. Colonies were transferred to 96-well plates containing 100 µl of water and pinned to SD-URA medium overlaid with a nitrocellulose filter (0.4 µm; HAHY08550 Millipore) and incubated for 48 h at 30°C. Filters were rinsed in distilled water to remove cells and probed by immunoblot analysis. Cross-contamination was estimated at 0.8% based on growth in blank positions; ∼93% of the collection (4554 mutants) was examined.
For the overexpression screen, a collection of ∼5,500 overexpression plasmids [36] was examined in a wild-type Σ1278b strain containing a functional MSB2-HA gene integrated at the MSB2 locus under the control of its endogenous promoter (PC999). Plasmids were purified from Escherichia coli stocks by alkaline lysis DNA preparation in 96-well format. Plasmid DNA was transformed into PC999 using the high-throughput transformation protocol described above. Transformants were selected on SD-URA and screened by pinning to nitrocellulose filters on synthetic medium containing 2% galactose and lacking uracil (S+GAL-URA) to induce gene overexpression. Colonies were incubated for 2 days at 30°C. Filters were washed in a stream of water and probed by immunoblot analysis as described above. Candidate genes/deletions that were initially identified were confirmed by retesting. Approximately ∼35% of the deletion strains and 37% of the overexpression plasmids failed retesting and were considered false positives (Tables S3B and S4B). Comparative secretion profiling of Msb2p-HA, Hkr1p-HA, and Flo11p-HA was performed by transforming overexpression plasmids identified in the Msb2p-HA screen into strains that contain Hkr1p-HA (PC2740) and Flo11p-HA (PC2043). Transformants were pinned onto S-GAL-URA on nitrocellulose filters. Colonies were incubated for 48h, at which time filters were rinsed and evaluated by colony immunoblot analysis. The complete deletion collection transformed with pMSB2-HA (Table S3) and overexpression collection in strain PC999 (Table S4) were frozen in aliquots at −80°C and are available upon request.
To compare levels of secreted protein between samples, spot intensity was measured and normalized to colony size. Small colonies initially scored as undersecretors and large colonies (particularly at plate corners) scored as hypersecretors were eliminated or flagged for retesting. As an additional test, standard immunoblots were performed from cells grown in liquid culture. Supernatants (S) and cell pellets (P) were separated by centrifugation. In most cases (>80%), differential secretion by colony immunoblot was reflected by an altered S/P ratio by standard immunoblot analysis. The ImageJ MicroArray Profile.jar algorithm (http://www.optinav.com/download/MicroArray_Profile.jar) created by Dr. Bob Dougherty and Dr. Wayne Rasband, commonly used for Microarray analysis, was used as a plugin for the ImageJ program. Image intensity was determined in 96-panel format by inverting the image after background subtraction. Plate-to-plate variation was normalized by dividing the average intensity of all spots with the average intensity of each plate, and this factor was applied to the intensity of each spot. Heat maps for expression and secretion profiling were generated as described [127]. Classification of genes based on process/function was determined using GO ontology terms in publicly available databases including the Saccharomyces genome database (http://www.yeastgenome.org/). Analysis of human mucin genes was facilitated by the Human Protein Reference Database (http://www.hprd.org/).
DNA microarray analysis was performed as described [128],[129]. Wild type (PC538), GAL-MSB2 (PC1083), GAL-MSB2 ste12Δ (PC1079), GAL-MSB2 ras2Δ (PC2949), MSB2Δ100–818 (PC1516), MSB2Δ100–818 ste12Δ (PC1811), and MSB2Δ100–818 ras2Δ (PC2364) strains were grown in YEP-GAL for 6 h, at which point cells were examined by microscopy for the characteristic filamentation response. RNA was prepared by hot acid phenol and passage over an RNeasy column (Qiagen). Microarray construction, target labeling, and hybridization protocols were as described [130]. Sample comparisons were independently replicated at least 3 times from separate inductions. Fluoro-reverse experiments were used to identify sequence-specific dye biases. Arrays were scanned using a GenePix 4000 scanner (Axon Instruments). Image analysis was performed using GenePix Pro 3.0. Array features (i.e., spots) having low signal intensities or signals compromised by artifacts were removed from further analysis. Background subtracted Cy5/Cy3 ratios were log2 transformed and a Loess normalization strategy (f = 0.67) was applied for each array using S-Plus (MathSoft, Cambridge, MA). Each feature where the |log2 (ratio)|≥0.8, the corresponding gene was considered differentially expressed.
Differential-interference-contrast (DIC) and fluorescence microscopy using the FITC filter set were performed using an Axioplan 2 fluorescent microscope (Zeiss) with a PLAN-APOCHROMAT 100X/1.4 (oil) objective (N.A. 0.17). Digital images were obtained with the Axiocam MRm camera (Zeiss). Axiovision 4.4 software (Zeiss) was used for image acquisition and analysis. For Msb2p-GFP localization, cells were grown to saturation in selective medium to maintain plasmids harboring MSB2-GFP fusions. Cells were harvested by centrifugation and resuspended in YEPD medium for 4.5 h. Cells were harvested, washed three times in water, and visualized at 100X.
Immunoblots were performed as described [31]. To compare protein levels between strains, cells were grown to saturation in YEPD medium and subcultured into YEPD or YEP-GAL medium for 8 h. Culture volumes were adjusted to account for differences in cell number and harvested by centrifugation. Supernatant volumes were similarly adjusted. Cells were disrupted by addition of 200 µl lysis buffer (8 M Urea, 5 % SDS, 40 mM Tris-HCl pH 6.8, 0.1 M EDTA, 0.4 mg/ml Bromophenol blue and 1 % β-mercaptoethanol) and glass beads followed by vortexing for 5 minutes at the highest setting and boiling 5 min. Supernatants were examined by boiling in 1.5 volumes of lysis buffer for 5 min. For some experiments, cells were lysed in spheroplast buffer (1.2M sorbitol, 50 mM potassium phosphate pH 7.4, 1 mM MgCl, 250 ug/ml of zymolyase), and protein concentration was determined by Bradford assays (Bio-Rad, Hercules CA). Equal concentrations of protein were loaded into each lane. For some experiments, blots were stripped and re-probed using anti-actin monoclonal antibodies (Chemicon; Billerica, MA). Monoclonal antibodies against the HA epitope were used (12CA5). Proteins were separated by SDS-PAGE on 10% or 10%–20% gradient gels (Bio-Rad, Hercules CA) and transferred to nitrocellulose membranes (protran BA85, VWR International Inc. Bridgeport NJ). Membranes were incubated in blocking buffer (5% nonfat dry milk, 10 mM Tris-HCl pH 8, 150 mM NaCl and 0.05% Tween 20) for 1 hr at 25°C. Nitrocellulose membranes were incubated for 18 hr at 4°C in blocking buffer containing primary antibodies. ECL Plus immunoblotting reagents were used to detect secondary antibodies (Amersham Biosciences, Piscataway NJ). Immunoblots of proteins secreted from mats were performed by growing cells on nitrocellulose filters on low-agar (0.3%) YEPD medium.
ChIP assays were performed as described [131]. Strains PC3021 (STE12-HA), PC3353 (sin3Δ, STE12-HA) and PC3579 (SIN3-HA) were grown in YEPD medium for 8 h. Cross-linking was performed with 1% formaldehyde for 15 min at 25°C. Cells were collected by centrifugation and washed twice in PBS buffer. Cells were resuspended in ChIP lysis buffer (Upstate, Billerica, CA), and lysed by Fast Prep 24 (MP) for one cycle at 6.5 for 45 sec. After puncturing the bottom of Fast Prep tubes with a 22 gauge needle, lysates were collected by centrifugation. DNA was sheared by sonication on a Branson Digital Sonifier at setting 20% amplitude, 15 pulses for 20 sec, 55 sec. rest. Pull downs were performed using a ChIP assay kit (Upstate) with anti-HA antibodies with 10% input sample set aside for a control. qPCR was used to determine relative amount of immunoprecipitated specific DNA loci in IP, Input, and Mock (no antibody) samples. The housekeeping ACT1 gene was used to normalize quantification in qPCR reactions. Data are expressed as IP/Input where ΔΔCt = (Ct IP_MSB2-Ct IP_ACT1)-(Ct Input_MSB2-Ct Input_ACT1). Primers used were MSB2 promoter forward 5′-CGATAGCTGATAGACTGTGGAGTCG-3′ and reverse 5′- CTGGCAACGCCCGACGTGTCTAGCC-3′, MSB2 intergenic region forward 5′- TGACCAAACTTCGACTGCTGG-3′ and reverse 5′- AGCTGCTGATGCAGTGGTAA-3′; and ACT1 forward 5′- GGCTTCTTTGACTACCTTCCAACA-3′ and reverse 5′- GATGGACCACTTTCGTCGTATTC-3′. ChIP pull downs were also visualized by gel electrophoresis.
Total RNA was isolated from 25 ml cultures grown in YEP GAL for 8 h using hot acid phenol extraction. cDNA synthesis was carried out using 1 µg RNA and the iScript cDNA Synthesis Kit (Bio-Rad; Hercules CA) according to the manufacturer's instructions. One tenth of the synthesized cDNA was used as the template for real-time PCR. 25 ul real time PCR reactions were performed on the BioRad MyiQ Cycler with iQ SYBR Green Supermix (Bio-Rad). RT qPCR was performed using the following amplification cycles: initial denaturation for 8 min at 95°C, followed by 35×cycle 2 (denaturation for 15 sec at 95°C and annealing for 1 min at 60°C). Melt curve data collection was enabled by decreasing the set point temperature after cycle 2 by 0.5°C. The specificity of amplicons was confirmed by generating the melt curve profile of all amplified products. Gene expression was quantified as described [132]. Primers were based on a previous report [133] and were FLO11 forward 5′- GTTCAACCAGTCCAAGCGAAA-3′ and reverse 5′- GTAGTTACAGGTGTGGTAGGTGAAGTG-3′and those described for ChIP assays above. All reactions were performed in duplicate and average values are reported.
|
10.1371/journal.pgen.1005274 | The Human Blood Metabolome-Transcriptome Interface | Biological systems consist of multiple organizational levels all densely interacting with each other to ensure function and flexibility of the system. Simultaneous analysis of cross-sectional multi-omics data from large population studies is a powerful tool to comprehensively characterize the underlying molecular mechanisms on a physiological scale. In this study, we systematically analyzed the relationship between fasting serum metabolomics and whole blood transcriptomics data from 712 individuals of the German KORA F4 cohort. Correlation-based analysis identified 1,109 significant associations between 522 transcripts and 114 metabolites summarized in an integrated network, the ‘human blood metabolome-transcriptome interface’ (BMTI). Bidirectional causality analysis using Mendelian randomization did not yield any statistically significant causal associations between transcripts and metabolites. A knowledge-based interpretation and integration with a genome-scale human metabolic reconstruction revealed systematic signatures of signaling, transport and metabolic processes, i.e. metabolic reactions mainly belonging to lipid, energy and amino acid metabolism. Moreover, the construction of a network based on functional categories illustrated the cross-talk between the biological layers at a pathway level. Using a transcription factor binding site enrichment analysis, this pathway cross-talk was further confirmed at a regulatory level. Finally, we demonstrated how the constructed networks can be used to gain novel insights into molecular mechanisms associated to intermediate clinical traits. Overall, our results demonstrate the utility of a multi-omics integrative approach to understand the molecular mechanisms underlying both normal physiology and disease.
| Biological systems operate on multiple, intertwined organizational layers that can nowadays be accesses by high-throughput measurement methods, the so-called ‘omics’ technologies. A major aim in the field of systems biology is to understand the flow of biological information between the different layers at a systems level in both health and disease. To unravel the complex mechanisms underlying those molecular processes and to understand how the different functional levels interact with each other, an integrated analysis of multiple layers, i.e. a ‘multi-omics‘ approach is required. In our present study, we investigate the relationship between circulating metabolites in serum and whole-blood gene expression measured in the blood of individuals from a population-based cohort. To this end, we constructed a correlation network that displays which transcript and metabolite show the same trend of up- and down-regulation. We derived a functional characterization of the network by developing a novel computational analysis. The analysis revealed systematic signatures of signaling, transport and metabolic processes on both a regulatory and a pathway level. Moreover, integrating the network with associations to clinical markers such as HDL-cholesterol, LDL-cholesterol and TG identified coordinately activated pathways or modules which might help to assess the molecular machinery behind such an intermediate phenotype.
| Blood is a connective tissue, which not only ensures nutrient and oxygen supply of all organs of the human body, but also the communication between them. Among the variety of key tasks performed by blood are immunological functions through white blood cells. Due to its diverse functionality, blood is heterogeneous and complex in its composition. Besides cellular constituents, which can be roughly divided into red and white blood cells, blood mainly consists of plasma. Plasma represents the aqueous phase containing proteins, peptides, signaling molecules and steroid hormones, but also other metabolites (e.g. carbohydrates, amino acids and lipids) which are consumed and released by the organs. This unique composition of blood, agglomerating both metabolic and transcriptional variation carrying molecular signatures of system-wide processes, together with its minimally invasive accessibility, makes blood a widely used system for integrative biomedical research [1,2].
With the development of high-throughput omics technologies for different levels of molecular organization, a systematic analysis of biological mechanisms underlying the functionality (or dysfunctionality) of a system became possible. In the case of transcriptomics data, an established framework to systematically investigate the constituents of involved biological processes and their interactions are network-based approaches, where pairwise associations between molecular entities (nodes) are modeled as network edges. Such studies commonly identify context-specific functional modules [3], but also global co-expression networks [4] from different organisms [5] and cell types [6]. When particularly focusing on the blood system, several studies investigated the co-regulation of transcripts either from single white blood cell types or whole blood samples. For example, regulatory networks [7,8] or global gene co-expression networks [9–11] were constructed from B- and T-cells to investigate pathways and mechanisms involved in the immune response. Further examples using whole blood data include the identification of disease related gene networks [12,13] or molecular signatures of distinct human vaccines captured in blood transcriptional modules [14].
Similarly, for metabolomics data a variety of studies extensively analyzed interactions between metabolites in various tissues, conditions and species [15–17]. Regarding blood measurements, we and others recently systematically characterized molecular interactions in the blood metabolome [18–21]. Utilizing Gaussian graphical models and serum metabolomics data from more than 1000 participants of a population cohort we were able to show that correlations between circulating blood metabolites resemble known metabolic pathways [22]. Furthermore, we have shown that these data-derived metabolic networks can be useful in a variety of applications, e.g. for the functional annotation of unknown metabolites [23] or to identify sex-specific serum metabolome differences [24].
The integration of multiple omics measurements (e.g. gene expression levels and metabolite concentrations) is an area of active research with many successful applications investigating the interplay between multiple organizational layers of a biological system [25–28]. However, only few studies with large sample sizes focused on a combined analysis in human blood. One recent example is the work of Inouye et al, who analyzed whole blood transcriptomics data in combination with blood lipid measurements and metabolites from a Finnish population cohort [29]. In their study, the authors associated a module of highly co-expressed genes with 134 blood metabolic markers in the context of heart disorders and identified a link between the immune system and circulating metabolites. The study by Inouye et al was among the first to provide clear evidence for this immune system link in blood, suggesting that gene expression in white blood cells is responsive to changing blood metabolite levels. Thus, it can be concluded that even if not cell-specific, the signals derived from whole blood data still reflect organism-wide processes. This is also in line with previous studies conducted on whole blood transcriptomics or metabolomics data separately [1,30,31].
The aim of the present study was to make use of the joint power of metabolomic and transcriptomic profiling to comprehensively characterize the complex interplay between serum metabolomics and whole-blood transcriptomics data. While serum metabolomics represent a footprint of whole-body processes, blood transcriptomics data will mainly reflect immune system processes through white blood cells. To this end, we analyzed metabolomics and transcriptomics measurements of 712 individuals from the German population study KORA (“Kooperative Gesundheitsforschung in der Region Augsburg”), comprising 440 metabolites and 16,780 genes after filtering. We constructed a global correlation network to elucidate the complex interplay and regulation between these omics layers (Fig 1A). The correlation analysis takes advantage of the naturally occurring variation from individual to individual, which we assume to carry a systematic footprint of the coregulation of metabolites and mRNAs. Such an integrative approach was recently termed “systems genetics”, providing a global view on the information flow between the various biological scales [32]. We deliberately left out an analysis of metabolite-metabolite and transcript-transcript correlations, which were rigorously investigated in the above-mentioned earlier studies. Instead, we specifically sought to assess the interconnection and information flow between the two omics layers.
The manuscript is organized as follows: In the first part, we systematically characterize the blood metabolome-transcriptome interface (BMTI) using different strategies. First, we manually investigated the strongest associations and provide evidence from literature wherever possible. Moreover, using a Mendelian randomization (MR) approach, we examined potential causal relationships between metabolites and transcripts. Second, using the most recent genome-wide human metabolic network Recon 2 [33], we systematically analyzed correlations between metabolites and transcripts at a pathway level (Fig 1B). Third, we developed a novel network clustering approach based on functional annotations, leading to a pathway interaction network (PIN) that allows for fast functional interpretation of the BMTI and furthermore provides insights into the cross-talk among distinct molecular pathways (Fig 1C). In the second part of this manuscript, we demonstrate how the identified networks can be used as a resource to further investigate the link between metabolism and gene regulation by two different applications. First, we investigated whether a common regulatory signature is observable from transcripts connected to the same metabolite or to metabolites that are part of the same metabolic pathways. For this purpose, we analyzed promoter regions of the genes for overrepresented transcription factor binding sites (Fig 1D). Second, we integrated the metabolome-transcriptome and the pathway interaction network with associations to high density lipoprotein cholesterol (HDL-C), low density lipoprotein cholesterol (LDL-C) and triglycerides (TG), which are well-known risk factors for cardiovascular disease [34]. To this end, we mapped the results of linear regressions between these clinical lipid parameters with metabolites and mRNAs onto the networks (Fig 1E). Finally, we demonstrate the potential of our systems genetics approach to generate novel hypothesis by combining results from all separate analysis steps and establish an association between the branched-chain amino acid pathway and the levels of plasma TG and HDL-C.
All network results are available to the scientific community as interactive versions in graphml and Cytoscape format (S1 Dataset).
For this study, we focused on a subset from the KORA F4 cohort with simultaneously available metabolomics and transcriptomics data. After quality control and filtering, the data set comprised of 712 human blood samples (354 males, 358 females) with gene expression data of 16,780 uniquely mapping gene probes and metabolite concentrations of 440 metabolites (Fig 1A, see Materials and Methods for details). 186 of these 440 metabolites were not chemically identified, which is marked by a metabolite name starting with “X-”throughout this manuscript. Both gene expressions and metabolite concentrations were log transformed and adjusted for age and sex effects. Pairwise Spearman’s rank correlations between the measured mRNAs and metabolites were then calculated. We used this correlation method to account for possible non-linear associations and to ensure robustness against outliers. Note that for this particular dataset, Spearman and Pearson correlations produced almost identical results (S1 Fig). S1 Table provides a full list of identified significant associations between blood metabolites and transcripts.
Metabolite-mRNA Spearman correlation coefficients were symmetrically distributed around zero (mean:−4.5 × 10−4 ± 0.0433, Fig 2A) with a maximum absolute correlation value of ρ = 0.56. Moreover, the distribution of inter-omics correlations showed a rather narrow shape, indicating generally lower correlations when compared to the intra-omics correlations (mRNA-mRNA, metabolite-metabolite). The metabolite-metabolite distribution was strongly skewed for positive correlation values, which is in accordance with our previous findings on a different metabolomics panel [22]. In contrast, the mRNA-mRNA distribution displayed a broad and symmetric distribution of correlation values (Fig 2A).
We then generated a weighted bipartite network of metabolites and transcripts by constructing an edge between a pair of metabolite and transcript if the respective correlation was significant with a false discovery rate (FDR) of 0.01. This corresponded to an absolute correlation cutoff of ~0.181 and a p-value threshold at 1.07 × 10−6. Obviously, the number of edges in a correlation network heavily depends on the chosen threshold. It has been shown in previous studies that a biologically reasonable threshold can be found by investigating network density as a function of the correlation cutoff value [35]. According to that study, a cutoff value slightly above the minimal density combined with a decreasing number of nodes and edges leads to biologically meaningful results. As indicated in Fig 2B, a clear decline in the number of included nodes and edges can be observed for increasing correlation threshold levels beginning between correlation values of 0.15 and 0.25. Minimal network density was reached for a correlation threshold value between 0.13 and 0.18 (S2 Fig). Notably, applying the above-mentioned conventional statistical significance threshold to our data set precisely coincides with the network density-based threshold described by [35].
The resulting network, subsequently called the blood metabolome-transcriptome interface (BMTI), consisted of 636 nodes (114 metabolites, 522 transcripts) and 1109 edges, corresponding to a total network connectivity of ~0.015% (Fig 2B and 2D). Out of the total number of edges, 63% (699) were positive correlations and 37% (410) were negative correlations. The metabolite showing the highest degree was mannose, with significant correlations to 98 different transcripts. In contrast, the mRNA with the highest connectivity was SLC25A20 with 37 metabolites attached (S1 Table).
We used data from the DILGOM study, which included NMR metabolomics as well as transcriptomics data for 518 individuals, for independent replication of our correlations. In total, 17 metabolites (11 amino acids, 3 lipids, 2 carbohydrates and 1 belonging to the energy metabolism) overlapped between the KORA F4 dataset and the DILGOM study, which allowed us to investigate the replication of 211 edges (~19% of the BMTI). 61 out of the 211 edges (~29%) reached a nominal significance (p-value < 0.05) in the DILGOM study of which 38 (~18%) remained significant after multiple testing correction (FDR < 0.05, see S2 Table).
To investigate the possible origins of the metabolite-transcript correlations, we compared all genes represented in the BMTI with 1) two a priori defined blood cell type-specific marker gene lists, and 2) a database of more general tissue gene expression signatures (see Materials and Methods). For the first part, we used a list of genes derived from Palmer et al. [36] comprising 907 specifically expressed genes for 5 different blood cell types (leukocytes only) and a second list derived from the HaemAtlas as generated by Watkins et al. [37] comprising 1,716 genes characterizing 9 different blood cell types. For the second part, we used the HECS database from Shoemaker et al. [38] containing information for more than 6,000 genes and 84 tissues. Both comparisons in 1) showed that most of the BMTI genes (85% and 67%, respectively) were not specifically attributable to any blood cell type (see Fig 2C and S3 Table). The remaining genes could be assigned to the respective measured cell types, with granulocytes making up the largest blood cell faction in both cases (8% and 20%, respectively) and only minor signals for the other blood cell types. A similar result was observed when comparing the BMTI genes to the HECS database. 52% of the BMTI genes showed no tissue specificity, while 12 out of the 15 strongest tissue signatures where either blood cells or blood related tissues (S3 Table).
As a first step to characterizing the BMTI, we performed a manual literature lookup for the strongest absolute correlations in the network (Fig 1B). In the following, we provide a detailed discussion of the 25 strongest edges (Table 1). Notably, most of the top 25 identified associations reflect biochemically reasonable interactions like transport processes of lipids, but also regulatory signatures between signaling metabolites and transcription factors.
The strongest association in the dataset was observed between cortisol and DNA-Damage-Inducible Transcript 4 (DDIT4, ρ = 0.55, p-value = 7.70 × 10−59), which are known to play a role in stress response [39]. Cortisol is a glucocorticoid whose release is mainly induced by exogenous stress. Via binding to the glucocorticoid nuclear receptor (GR, official gene symbol NR3C1), it regulates various cellular processes like carbohydrate metabolism and the immune system by direct activation of target genes [40]. Remarkably, DDIT4 was identified as a GR target gene in mouse hepatocytes [41], rat hippocampus [42] and also in human peripheral blood lymphocytes [43] delivering a potential explanation of an indirect association for the observed correlation. Another GR target gene associated to cortisol is Suppressor Of Cytokine Signaling 1 (SOCS1, ρ = 0.36, p-value = 2.19 × 10−23), a major constituent of the cytokine signaling pathway and inflammatory response [44]. We observed further top 25 correlations involving cortisol for Kruppel-Like Factor 9 (KLF9) and Dual Specificity Phosphatase 1 (DUSP1) (ρ = 0.34, p-value = 5.20 × 10−21; ρ = 0.34, p-value = 7.58 × 10−21, respectively). KLF9 is a ubiquitously expressed transcription factor involved in the regulation of diverse biological processes like cell development and differentiation in adipogenesis [45]. DUSP1 is an enzyme involved in the response to environmental stress [46]. Interestingly, for both transcripts, a cortisol-dependent regulation was already observed in epidermal cells [47] and peripheral blood mononuclear cells [48].
Another metabolite showing several strong associations to blood transcripts was 1-monoolein, which belongs to the class of monoacylgylcerols. This particular class of metabolites are bioactive compounds recently identified to be involved in various signaling processes of the immune system [49,50]. The source of 1-monolein in humans is not fully understood. Experiments in rodents suggest that dietary 1,3-diacylglycerols are preferentially digested to 1-monoacylglycerols and free fatty acid in the small intestine, making dietary 1,3-diacylglycerols containing an oleoyl moiety at position sn-1 or sn-3 a plausible source of 1-monoolein [51]. In our analysis, 1-monoolein showed a strong negative correlation to four transcripts – GATA Binding Protein 2 (GATA2), Histidine Decarboxylase (HDC), Solute Carrier Family 45, Member 3 (SLC45A3) and Chromosome 1 Open Reading Frame 186 (C1ORF186) (ρ between -0.41 and -0.32, p-values between 2.01 × 10−29 and 4.72 × 10−18). HDC is a cytosolic enzyme that catalyzes the conversion of histidine to histamine and thus represents an important immune system trigger molecule [52]. In addition, GATA2, a key regulator of gene expression in hematopoietic cells [53], C1ORF186 and SLC45A3, two membrane-bound proteins, were all identified to play a role in the immune response [13].
Carnitine-acylcarnitine translocase (SLC25A20) occurred in 15 of the 25 top ranked correlations. This gene encodes an enzyme which transports acylcarnitines, i.e. the transport variant of fatty acids, into the mitochondria for subsequent ß-oxidation. Interestingly, the majority of SLC25A20-associated metabolites among our top 25 correlations belonged to the class of long chain fatty acids (11 long chain fatty acids, 2 essential fatty acids, 1 medium chain fatty acid, 1 ketone body), which is in accordance with its function as a lipid transporter. Of note, among the metabolites associated with SLC25A20 beyond the top 25 correlations were also 5 acylcarnitines, although at lower correlation values.
We observed a significant, negative correlation between isoleucine and ATP-Binding Cassette Sub-Family G Member 1 (ABCG1, ρ = −0.32, p-value = 3.39 × 10−18). It has been shown previously that circulating levels of branched-chain amino acids (BCAAs) affect a variety of metabolic processes such as glucose and lipid metabolism [54]. ABCG1 is a major player of lipid metabolism, controlling the transfer of cholesterol from peripheral macrophages to exogenous HDL [55]. Interestingly, an association between circulating BCAA levels and plasma HDL-C levels was also observed in a recent population study [56] and in a previous paper on the same population cohort used in the present study [57].
To assess whether metabolite-transcript links in BMTI contain causal effects, we performed a Mendelian randomization analysis [58]. For each metabolite-mRNA edge, we tested both the causal directions metabolite→mRNA and mRNA→metabolite given that adequate instrumental variables were available. As instruments we used SNP lists from previously published GWAS studies. After filtering for strong instrumental variables, we were left with 15 SNPs identified by a metabolomics GWAS study [23] associated to 16 metabolites in the BMTI. Moreover, for 157 mRNAs in the network, we selected 192 SNPs from [59]. In total, we tested the causal relationship of 440 BMTI edges (~40%) of which 60 could be tested bi-directional. In the BMTI, 42 metabolite-mRNA pairs (19 mRNA→metabolite; 23 metabolite→mRNA) showed a nominally significant (p-value < 0.05) causal effect. At an FDR of 0.05, none of the tested pairs remained significant (S4 Table).
In order to further reveal the underlying mechanisms determining the observed associations, we systematically analyzed whether correlating pairs of metabolites and transcripts (i.e. enzymes) correspond to the structure of the underlying metabolic network. Specifically, we investigated if strong metabolite-transcript edges of the BMTI tend to be in close proximity within biochemical pathways. All pairwise associations between metabolites and transcripts were mapped to their corresponding network nodes in the Human Recon 2 metabolic network reconstruction [33]. As a measure for metabolic network proximity, the length of the shortest path connecting each metabolite-enzyme pair was determined (Fig 3A). This measure is based on the common assumption that the shortest connection between two network entities corresponds to the biologically reasonable one [22,60]. To avoid potential biologically meaningless shortcuts, we removed co-factors and currency metabolites prior to the analysis (see Methods section for details and S5 Table for a list of removed metabolites).
We could map 121 metabolites and 1,467 enzymes out of the 254 metabolites with known identity and 16,780 transcripts onto the metabolic network, respectively. While most pairwise correlation coefficients were closely distributed around zero for all investigated network distances, a distinct pattern was observable for statistically significant correlations. The majority of significantly correlating pairs accumulated at short distances and was dominated by positive correlations (Fig 3B). To determine the significance of this observation, one-tailed Fisher’s exact tests were performed by either considering each distance individually or by aggregating all pairs up to a particular distance. The latter aggregation analysis combines all transcript-metabolite pairs which are reachable up to a certain number of steps (biochemical reactions) in the metabolic network. For both cases, we observed a substantial overrepresentation of significantly correlating pairs at short distances (Fig 3C). The strongest signals are observed for pairs that take part either directly in the same reaction (d = 0) or for those which are just one reaction apart (d = 1). For the cumulative distances we also observed significant enrichment up to a distance of d = 2 reactions. Proportions of significant and non-significant pairs per distance are given in S3 Fig and a detailed view on an exemplary path of length 2 is depicted in S4 Fig
To further characterize the underlying biochemical pathways, we calculated frequencies of functional annotations from Recon 2 among the significant associations for pathway distances 0 to 2 (Fig 3D and S5 Fig). At a distance of 0, we identified mainly transport reactions (67%) accompanied by reactions belonging to lipid metabolism (bile acid synthesis 11%, fatty acid oxidation 11%) and carbohydrate metabolism (pyruvate metabolism 11%). The transport reactions can be further subdivided into extracellular transport (45%), or mitochondrial transport (11%) and peroxisomal transport (11%). Similar signals can be found at distances of 1 and 2, where we additionally identified reactions belonging to energy metabolism (oxidative phosphorylation 27%) and amino acid metabolism (histidine metabolism 11%, glutamate metabolism 4%).
Taken together, the BMTI captured a systematic signal of metabolite-enzyme associations to be in close proximity when mapped onto a global metabolic network. Moreover, the strongest signals found for pathway distances of 0, 1 or 2 reflect distinct metabolic reactions mainly belonging to lipid, energy and amino acid metabolism, and transport mechanisms.
Up to this point, our analysis was a reaction-centered approach limited to single edges only, thereby neglecting the global network structure and cross-talk between pathways captured in the BMTI. To derive a comprehensive functional description of the biological modules included in the BMTI, we developed a novel approach based on functional annotations which provides an integrated view on cellular processes. Briefly, the approach consists of three steps: First, we used pathway annotations to define groups of functionally related metabolites and transcripts. For metabolites, we used metabolic pathway annotations provided with the metabolomics dataset, and for transcripts we downloaded the Gene Ontology (GO) slim annotations. Second, an aggregated z-score (aggZ-score) was calculated for each functional category. Third, we calculated correlations between aggZ-scores of all functional categories. A schematic overview of this multi-step approach is provided in S6 Fig and described in more detail in the Material and Methods. A full list of the resulting categorical correlations can be found in S6 Table.
We again constructed a network (the pathway interaction network, PIN) by drawing edges between significantly correlated categories. Interestingly, even when applying a stringent Bonferroni-corrected threshold (α = 0.01, p-value ≤ 2.2 × 10−6) this resulted in an overly dense connected network of 166 nodes and 1220 edges. To generate a visually interpretable version of this network, an ad-hoc stringent threshold of p-value ≤ 1.0 × 10−11 was applied to draw the network. This resulted in a PIN consisting of 113 nodes (93 GO terms, 20 metabolic pathways) connected by 244 edges (196 positive correlations, 48 negative; Fig 4A). Remarkably, we observed a high conformity between linked metabolic pathways and gene annotations. For example, the metabolic pathway “carnitine metabolism” was connected to the biological processes “lipid metabolic process” and “transmembrane transport”. Moreover, it was linked to the cellular component “mitochondrion”, indicating transport processes of fatty acids into the mitochondrion for subsequent ß-oxidation. Further biologically reasonable pairs were “Valine, Leucine and Isoleucine metabolism” and “Glutamate metabolism” attached to “cellular nitrogen compound metabolic process”. As a last example, “Steroid/Sterol” was connected to “response to stress” and “signal transducer activity”, pointing to an interaction between hormones and regulation of gene expression. In the following, we examine two selected category-category relationships in detail, including the individual metabolites and gene transcripts that gave rise to the association.
The BMTI contains a prominent “flower-like” network topology, i.e. many transcripts associated to a single metabolite. We therefore asked whether these coordinated changes around a metabolite and also the network topology can be explained by common transcriptional regulatory processes through transcription factors (TFs, Fig 1D). For the following analysis, we only considered metabolites linked to at least 3 transcripts. We analyzed the promoter regions of all connected genes for an enrichment of known transcription factor binding sites (TFBS) derived from the Jaspar database [68]. This resulted in significantly enriched transcription factor binding motifs for 46 single metabolites, 24 subpathways and 7 superpathways. The Methods section provides a detailed explanation of the process. A summary of all enriched TFBS can be found in S7 Table.
In total, out of the 205 binding motif matrices used in the analysis, 189 reached a significant enrichment in at least one of the metabolite-derived gene sets, indicating a generally prevalent common regulation. Across all lists of enriched TFBS identified from our network, the motifs that occurred most frequently were Sterol Regulatory Element Binding Transcription Factor 2 (SREBF2), Peroxisome Proliferator-Activated Receptor Gamma (PPARG; Jaspar motifs PPARG and PPARG::RXRA) and Nuclear Factor, Interleukin 3 Regulated (NFIL3). SREBF2 is a major regulator of cholesterol metabolism [69] while PPARG is known to be activated by fatty acid ligands, thereby regulating fatty acid ß-oxidation and other processes [70]. NFIL3 is a regulator specifically found in activated T cells, natural killer (NK) cells, and mast cells, involved in the regulation of the immune response and the circadian rhythm [71].
Branched-chain amino acids were among the metabolites most strongly connected to SREBF2 targets. Specifically, the transcripts correlating with isoleucine and valine show high enrichment of SREBF2 binding sites (p-value = 5.83 × 10−8 and p-value = 2.36 × 10−10, respectively; S7 Table). Moreover, considering all 172 genes associated to at least one metabolite from the entire branched-chain amino acid pathway (“Valine, leucine and isoleucine metabolism”) yielded significantly enriched binding sites for SREBF1 and SREBF2 (p-values 6.78 × 10−10 and 9.11 × 10−10, respectively; S7 Table). Both SREBs are important regulators in lipid homeostasis, including cholesterol and fatty acid biosynthesis, further indicating a regulatory cross-link between HDL-C, TG and BCAA metabolism.
The highly interlinked network topologies of both the blood metabolome-transcriptome interface and the pathway interaction network suggest a strong coregulation between the different metabolites, processes, and pathways. As a second step of coregulation analysis, we inferred the number of pairwise shared significant TFBS to determine the extent of coregulation between single metabolites and metabolic pathways (S8 Table). At the single metabolite level, we found a maximum number of 27 shared TFs between histidine and X-03094 (S7 Fig). Moreover, this highly connected unknown metabolite shared 14 TFs with another unknown metabolite (X-12442) and with a peptide (HWESASXX). For the metabolic subpathways, we observed an overlap between “histidine metabolism” and the group of “long chain fatty acids” and between “glycolysis, gluconeogenesis, and pyruvate metabolism” and the group of “fibrinogen cleavage peptides” (11 shared TFs each; Fig 5A). On the level of superpathways, the highest number of shared TFBS was 4, identified between “carbohydrate” and “peptide metabolism” (S8 Fig). Overall, we found that TF binding sites are shared to a large extent, indicating a complex coregulation not only within but also between different processes and pathways.
To gain further insight into this coregulation, we determined transcription factors which also occur as transcripts in the BMTI. 165 out of the 189 transcription factors with available binding motif were contained in the filtered data set. Only 12 of these transcription factors displayed a significant correlation to any metabolite and are thus included in the BMTI. This observation is not completely unexpected given that TFs are regulated to a large extend at a post-transcriptional level [72]. Interestingly, for two out of these 12 TFs, we also observe enriched binding sites in the promoter region of the other genes connected to the same metabolite, i.e. a “triad” network motif consisting of a metabolite, a transcription factor and its target genes (Fig 1D, S7 Table).
The first transcription factor is B-cell CLL/Lymphoma 6 (BCL6), a transcriptional repressor involved in the STAT-dependent interleukin 4 response of B-cells [73]. BCL6 is negatively correlated with methionine and tyrosine in our network (Fig 5B). The TFBS enrichment analysis using all 15 genes connected to methionine within the BMTI resulted in a significant overrepresentation of the BCL6 binding motif (p-value = 5.71 × 10−09, 82% of the 15 promoter sequences contained at least one occurrence of the motif), while no significant enrichment was observable for the genes connected to tyrosine. The second motif was identified around Nuclear Receptor Subfamily 4, Group A, Member 2 (NR4A2), which was associated to 7 metabolites in our network. The 22 neighboring genes of one of those metabolites, kynurenine, showed significantly enriched binding sites for this transcription factor (p-value = 3.79 × 10−09, 73% of the 22 promoter sequences contained at least one occurrence of the motif; see Fig 5C and S7 Table).
As a final analysis step, we sought to use the BMTI and the PIN to infer novel insights into the molecular mechanisms and pathways underlying complex traits. To this end, we associated the nodes of both networks with intermediate clinical phenotypes (Fig 1E, Table 2). As already stated earlier, we chose the levels of HDL-C and LDL-C, as well as concentrations of blood triglycerides (TG), known risk factors for a variety of diseases. We performed multiple linear regression analyses with HDL-C, LDL-C and TG blood parameters as response variables and all 440 metabolites and 16,780 transcripts as explanatory variables. All models were corrected for sex and age. Statistical significance was defined by a Bonferroni adjusted p-value cutoff at 2.9 × 10−6 (α = 0.05). We then projected the −log10 transformed p-values from this regression as colors onto the corresponding nodes in the two networks. Similarly, the analysis was performed using aggZ-scores of pathways / GO terms as explanatory variables and mapped to the PIN (Fig 6 and S1 Dataset). Note that we presented similar approaches in the past for metabolomics-only networks [24,74].
In total, regression analyses yielded 233 (54 metabolites, 179 mRNAs), 28 (28 metabolites, 0 mRNAs) and 1,124 (49 metabolites, 1,075 mRNAs) statistically significant associations for HDL-C, LDL-C and TG, respectively. Of those associations, 64%, 28% and 25%, were contained in the BMTI, respectively (see S9 Table for a complete list of associations). We only observed few LDL-C metabolite associations, which can be mainly summarized in the “Glycerolipid metabolism” and “Carnitine metabolism”, while none were observable for the transcripts (Fig 6E and 6F, S9 Table). We will therefore focus on network associations for HDL-C and TG in the following.
Examination of the networks for HDL-C and TG revealed localized regions of similar associations, which reflect potentially co-regulated modules (Fig 6A and 6C). Interestingly, when compared to each other, there appeared to be an antagonistic pattern of associations for HDL-C and TG, which is in accordance with an overall negative correlation of the two traits (ρ = −0.53). This opposing pattern also holds for the categorical networks (Fig 6A–6D and S10 Table). To confirm this observation statistically, we utilized an approach to compare the different networks suggested by Floegel et al. [74]. Basically, we calculated the Spearman correlation of the association measures between the different clinical traits. This resulted in a strong negative correlation between the BMTI-HDL-C and the BMTI-TG network (ρ = −0.84) which was even more pronounced between the PIN-HDL-C and PIN-TG networks (ρ = −0.94, S9 and S10 Figs). A similar pattern of opposing associations between HDL-C and TG phenotypic traits was already described in previous studies, which suggested a pleiotropic, heritable relation between the two lipid and lipoprotein measures, potentially regulated by a common, complementary mechanism [13,75].
In the following, we will discuss exemplary pathway mechanisms identified in the phenotype networks. ABCG1 and ABCA1, known constituents of the reverse cholesterol transport necessary for the proper formation of plasma HDL-C [55], were positively correlated with HDL-C (p-value = 4.37 × 10−12 and p-value = 2.92 × 10−8, respectively). At the pathway level, processes like “generation of precursor metabolites and energy” or “catabolic process” are negatively associated with HDL-C, while “nucleic acid binding transcription factor activity” and “signal transducer activity” are positively associated (Fig 6D). An inverse pattern can be seen for TG, where positive associations predominate and processes like “generation of precursor metabolites and energy” or “catabolic process” are strongly positively associated (Fig 6A and 6B).
Given the known association between HDL/TGs and branched-chain amino acids [57,76], we investigated the phenotypic networks to further examine this metabolic class. First, we examined the edge between isoleucine and ABCG1 within the BMTI-HDL-C network. As already mentioned, ABCG1 was strongly positively associated to HDL-C levels, while we found that isoleucine was significantly negatively associated to the concentration of HDL-C (β = −4.30, p-value = 5.80 × 10−19). Moreover, gamma-glutamyl variants of BCAAs belonging to “gamma-glutamyl metabolism” (β = −4.84, p-value = 3.15 × 10−14) and “Valine, leucine and isoleucine metabolism” (β = −4.66, p-value = 9.17 × 10−11) displayed profound negative associations to HDL-C (Fig 6D and S10 Table), further validating a connection between HDL-C and BCAA metabolism. For triglycerides, we observed an inverse relationship with BCAAs and BCAA-related pathways (Fig 6B, S9 and S10 Tables).
We constructed a global network model across two levels of biological information by integrating cross-sectional omics data from a large-scale population cohort. The dataset was based on circulating metabolites from plasma and transcriptional variation derived from whole blood. This analysis exploited the naturally-occurring variation caused by genetic variation, environmental and behavioral influences from a natural population over multiple layers of organization. Such an approach was recently referred to as ‘systems genetics’, enabling the systematic exploration of information flow between the different biological scales [32].
As mentioned in the introduction, blood is a heterogeneous tissue containing a series of distinct cell-types. In this study, we utilized whole blood transcriptomics data from unsorted cells, leading to a complex mixture of transcriptional signals in the transcriptome dataset [36]. Similarly, the levels of circulating metabolites are strongly influenced by metabolically active organs [31], but also by metabolites from blood cells and those taken up from the environment. The comparison to known cell-type specific markers further suggested that a substantial amount of the signals are derived from specific blood cells. However, the analysis also showed that the majority of the BMTI contained transcripts are not assigned to any cell-type. Thus, we assume that the metabolite-mRNA associations captured in the BMTI mainly reflect cell-type unspecific processes involved in the fundamental maintenance of cellular function, besides some processes specifically related to immune functions.
Independent replication of the BMTI edges was investigated using data from the DILGOM study. Out of 211 possible associations, we were able to replicate 29% at a nominal significance and 18% after multiple testing correction (FDR<0.05). This relatively low number of replicated associations might have various reasons. For example, 1) The DILGOM study used an NMR-based metabolomics platform in contrast to the mass spectrometry-based methodology used in KORA. 2) The smaller sample size of the DILGOM study might limit the power to detect existing associations between metabolites and transcripts. 3) Differences in laboratory procedures and protocols or the population structure can affect replication across cohorts. Future high-powered studies with more similar measurement platforms can further address the stability of metabolite-transcript correlations across studies.
A comprehensive analysis of the strongest associations between transcripts and metabolites clearly revealed biologically reasonable relationships, such as signaling and transport mechanisms. Many identified associations, e.g. between cortisol and DDIT4 or between SLC25A20 and multiple long chain fatty acids, were in consent with known signaling or metabolic pathways. Others support and extend results from previous studies. As one example, nearly all transcripts included in the lipid-leukocyte (LL) module identified by Inouye et al [29] were among the top scoring association pairs. For instance, we were able to confirm associations between CP3A, FCER1A, GATA2, HDC, MS4A2, and SLC45A3, core genes of the LL module, and leucine, isoleucine, and several lipids (see S1 Table). In addition, we found associations which, to the best of our knowledge, have not been described before. These include associations between 1-monoolein and GATA2, a key regulator of hematopoiesis, or SLC45A3, a known diagnostic marker for prostate cancer [77]. The identified associations extend the current knowledge about the connection between system-wide metabolism and immunity-related pathways.
Causal inference of the metabolite-mRNA associations using Mendelian randomization yielded no statistically significant results. There are various possible reasons for this negative outcome. First, there might be no causal effect in either direction between the investigated transcripts and metabolites. Besides that, the lack of significant findings could also be caused by the limitations of Mendelian randomization. For instance, MR is known to require large numbers of samples to detect true causal relationships, and the power in our study (n = 712) might have been too low [58]. We therefore decided to leave a more detailed discussion and analysis of causal effects to future, high-powered studies.
Comparison of the blood metabolome-transcriptome interface with the most recent human genome-scale metabolic reconstruction [33] allowed to assess whether transcript-metabolite correlations also recapitulate known biochemical reactions at a systematic level. We were able to show that strong associations between enzymes (represented by their respective transcripts) and metabolites are significantly accumulated at shorter pathway distances (Fig 3B and 3C), which is consistent with previous studies [60,78,79]. Further functional characterization identified transport, energy, lipid and amino acid subsystems to be predominately present at short pathway distances (Fig 3D and S5 Fig). This observation may reflect metabolic proximity through the uptake of metabolic nutrients by metabolically active blood cells. For instance, in our analysis we found signatures for all three major sources for energy production: lipids, proteins (in terms of amino acids) and carbohydrates indicating an active use of fuel molecules for energy generation by the blood cells.
Our model-based analysis has several limitations. Obviously, any such analysis is heavily dependent on the quality of the underlying metabolic reconstructions, which are still far from being complete [80]. This incompleteness, together with a prevalent inconsistent nomenclature of metabolites allowed us to map only 121 out of 254 measured metabolites onto the metabolic network model. Another limitation is the incomplete coverage of the metabolome, which is owed to the capabilities of currently available technologies. In this study we used measurements of 440 metabolites, which corresponds to just ~10% of the estimated human serum metabolome [81]. Nevertheless, we believe that despite incomplete pathway mappings, our observations further indicate that combined metabolomics and transcriptomics data from human blood reflect reaction signatures of system-wide biological processes.
To further functionally characterize the blood metabolome-transcriptome interface at a global level, we developed a network approach based on functional annotations. To this end, we aggregated z-score transformed measurements of metabolites and transcripts into their corresponding metabolic pathways and gene ontology categories, respectively. This approach allowed us to calculate correlation values between different functional categories, rather than between single metabolites and transcripts only. From these associations, we generated a pathway interaction network (PIN) of associated metabolic pathways and Gene Ontology terms, substantially reducing the complexity of the original network and thus facilitating functional interpretations. Detailed inspection of the PIN revealed that correlating nodes resembled not only signatures of well-known biological processes, like the carnitine shuttle, but also suggested novel interactions such as a crosstalk between monoacylglycerols and immune system processes. Taken together, the pathway interaction network enabled us to verify and elevate observations from the single reaction level (see model-based analysis) onto a pathway level. Moreover, we are now able to explore associations between biological processes/pathways across different biological scales including those that are not necessarily covered by metabolism, such as signaling or transcriptional processes.
Given the high interconnectivity of the BMTI and the PIN, we asked whether these associations contain information about regulatory interactions across the different metabolite classes and pathways. Enrichment analysis of transcription factor binding sites in the promoter regions of the genes contained in our network identified regulatory signatures for transcripts associated to the same metabolite, which are additionally largely shared between metabolites belonging to different metabolic pathways (Fig 5, S7 and S8 Figs). There is a series of possible explanations for this observation. On the one hand, our findings could indicate that single metabolites/transcripts are fulfilling multiple roles, thus sharing several biochemical pathways. On the other hand, it might reflect regulatory interactions operationally linking different metabolic pathways. In depth investigation of 12 transcription factors included in the BMTI additionally revealed two “triad” network motifs between transcription factors BCL6 and NR4A2, their target genes and the metabolites methionine and kynurenine, respectively. Remarkably, in a study conducted on mice fed a methionine and choline deficient diet, a significant reduction in the expression of BCL6 was observed [82]. It is widely known that metabolites can act as intermediates in cellular signaling, thereby also regulating gene expression, and together with our findings we suggest that characteristics of metabolic regulation are captured in the BMTI. However, from a correlation network, the detection of an association between a metabolite and a transcript does not necessarily imply a regulatory relationship nor can a conclusion be drawn about the directionality of the relationship. Yet, a combined analysis might offer the opportunity to identify novel molecular mechanisms behind cellular regulation that need to be validated further by experimental evidence.
Besides transcriptional regulation mediated by TFs, a substantial fraction of transcripts are expected to be regulated by epigenetic processes [83]. Comparing 1,350 reported methylation-metabolite associations from a recent epigenome-wide association study [31] with our results surprisingly revealed only a single overlapping hit: X-03094 and the MAN2A2 transcript correlated in our study and also displayed a comparable methylation-metabolite association in the EWAS study. This sparse overlap could be explained by a phenomenon termed “phenotypic buffering” [32], where effects in one organizational layer (e.g. epigenetics) are not detectable anymore on the next layer (e.g. transcriptomics). A detailed explanation of this observation is beyond the scope of the present paper and needs further investigation.
Further following the scheme of a systems genetics approach, we integrated the two identified networks with intermediate clinical trait data. To this end, we investigated the relationships between changing levels of HDL-C, LDL-C and TG and all measured metabolites and transcripts, metabolic pathways and GO terms (Fig 6). A similar study already described an association between a gene-module derived from whole blood transcriptomics data and circulating lipid parameters including apolipoprotein B (APOB), HDL-C and triglycerides (TG) from a Finnish population cohort [29]. Our systematic analysis identified a large number of metabolites, transcripts, metabolic pathways, and functional GO categories that are all associated with the levels of circulating lipids. These findings further strengthen the assumption of a close link between system-wide metabolism, reflected by circulating metabolites and clinical lipid markers, and intracellular gene regulatory processes of blood cells. In addition, an opposite pattern between HDL-C and TG associations (Fig 6A–6F) was observed from the phenotype networks which supports a previously suggested antagonistic regulation of both clinical traits [75,84]. However, the precise molecular mechanism behind this regulation is not entirely known, and our results might provide a basis for future studies to gain novel insights into the regulatory mechanisms of intermediate physiological phenotypes.
Combining results from all analysis steps allows for novel hypothesis generation. For example, for the well-known interactions between HDL-C, TG and BCAAs [57,76], we discovered a potential regulatory pattern on different biological scales. In our first analysis step, we identified a strong negative association between the branched-chain amino acid isoleucine and ABCG1, a major constituent of lipid homeostasis and cholesterol metabolism [55,85]. Second, at a more global level, the phenotype networks revealed an inverse association between HDL-C and TG, and also between HDL-C, TG and BCAAs (BCAAs are positively associated to TG, negatively to HDL-C, see S9 Table). Third, in the TFBS enrichment analysis we were able to identify a clear regulatory signature of SREBPs in the vicinity of BCAAs, which are known to regulate cholesterol metabolism, indicating a potential coregulation between BCAAs and cholesterol metabolism at the transcriptional level. Interestingly, a combined study using cultured hepatocytes in a branched-chain amino acid-rich medium and obese mice showed that BCAAs directly induce the expression of SREBP1C which leads to hypertriglyceridemia, further supporting the suggested regulatory cross-link between HDL-C, TG and BCAAs [76]. This link is of particular interest since all three molecular traits have been associated to various diseases such as coronary artery disease, obesity and diabetes type II [86–88] and our observations might contribute to further decipher their underlying mechanisms.
In summary, our study highlights the potential of a systems genetics approach for understanding interactions across multiple biological scales – in this case circulating metabolites and blood cellular gene expression—and how those insights can be used to generate novel hypothesis on mechanisms underlying common diseases.
The Cooperative Health Research in the Region of Augsburg (KORA) study is a series of independent population-based epidemiological surveys and follow-up studies of participants living in the region of Augsburg, southern Germany [89,90]. In this paper, cross-sectional data from 712 participants of the KORA F4 population cohort was used for whom metabolite concentration, gene expression data and genotyping information were available. This subpopulation contains combined fasting serum metabolomics and whole blood transcriptomics measurements of 354 males and 358 females aged 62–77 years (mean 68.82 ± 4.31). All participants are residents of German nationality identified through the registration office and written informed consent was obtained from each participant. The study was approved by the local ethics committee (Bayerische Landesärztekammer). Detailed descriptions of blood sample acquisition and experimental procedures for the metabolomics and transcriptomics data, and clinical trait measurements can be found in [59,91–93]. Briefly, metabolic profiling was performed by Metabolon, Inc. using ultrahigh-performance liquid-phase chromatography and gas-chromatography separation, coupled with tandem mass spectrometry. In total, 517 serum metabolites were measured, thereof 293 with known chemical identity and 224 unidentified metabolites (“unknowns”). All identified metabolites were assigned to one out of eight superpathways and one out of 61 subpathways by Metabolon, Inc., representing two different levels in the metabolic pathway classification hierarchy (see S5 Table for a full list of annotations). Gene expression profiling was performed using total RNA extracts from whole blood samples on Illumina Human HT-12 v3 Expression BeadChips. Genotyping was carried out using the Affymetrix GeneChip array 6.0. A detailed description of the experimental procedures and preprocessing of the genetic data can be found in [92].
Replication of the significant metabolite- mRNA associations identified in the KORA dataset was carried out with the Finish DILGOM cohort dataset which included whole blood NMR metabolomics data as well as transcriptomics data for 518 individuals. A detailed description of the sample acquisition as well as data preparation can be found in [13,29].
To ensure data quality, metabolites with more than 50% missing values were excluded, leaving 440 metabolites (254 knowns and 186 unknowns) for further analysis. The remaining metabolite concentrations were log-transformed, since testing for normality indicated that for most cases the log-transformed concentrations were closer to a normal distribution than the untransformed values [23]. For gene expression arrays, quality control and imputation of missing values of the raw intensities was performed as described in [94]. Briefly, the initial preprocessing of the raw intensity data was done with GenomeStudio V2010.1. Raw probe level data was then imported to R and further preprocessed by log transformation and quantile normalization using the ‘lumi’ package [95] from the Bioconductor open source software (http://www.bioconductor.org). To account for technical variation, gene expression intensities data were adjusted for RNA amplification batch, RNA integrity number and sample storage time. Only probes with the annotation flag QC_COMMENT “good” as provided in the updated Illumina Human HT-12 v3 BeadChip annotation file were considered for analysis [94]. In addition, probes mapping to gonosomal chromosomes were removed. Out of 48,803 probes on the Illumina Human HT-12 v3 array, 24,818 passed these filtering criteria.
The metabolite-transcript interface was constructed based on Spearman’s correlation coefficients between the concentrations of all possible metabolite-transcript pairs (24,818x440) across the individuals of the study cohort. Correlation calculation was performed separately for each variable pair, only considering samples without missing values for the metabolites. Statistical significance of correlations was determined at an FDR of 0.01 [96], corresponding to an absolute correlation value of 0.1816 and an adjusted significance level of 1.07 × 10−06. To get a unique network node per gene, redundant probes matching the same gene were removed. One representative probe per gene was chosen based on the maximum correlation strength to any metabolite, leaving 16,781 unique probes for subsequent analysis. It has to be noted that the applied significance level was still calculated on the whole dataset (including multiple matching probes per gene) to properly account for multiple testing. Network density was calculated as described in [35]. More precisely, for a stepwise increasing correlation threshold, the ratio between the total number of observed edges and all possible edges was calculated. Significant correlations between metabolites and transcripts were visualized as a bipartite graph using yEd graph editor (yWorks GmbH, Tuebingen; http://www.yworks.com).
BMTI genes were mapped to three published lists of tissue- and cell-specificity based on gene expression microarrays from purified cells or tissues. The first two marker gene lists were taken from Palmer et al. [36], who defined markers for B-cells, CD4+ T-cells and CD8+ T-cells, lymphocytes and granulocytes, and from the HaemAtlas as generated by Watkins et al. [37], who reported markers for CD19+ B-cells, CD4+ T-cells and CD8+ T-cells, CD14+ monocytes, CD56+ NK cells, CD66b+ granulocytes, erythroblasts and megakaryocytes. The third marker list was downloaded from the CTen website: http://www.influenza-x.org/~jshoemaker/cten/db_info.php and comprised markers for 84 different human tissues/cell types [38]. The three lists together with the analysis results are provided in S3 Table.
Estimation of causal effects within the BMTI was performed using a Mendelian randomization (MR) approach [58]. A total of 224 candidate SNPs reported as lead association signals at genome-wide significance in two recent GWAS studies for 16 metabolites and 186 mRNAs (BMTI contained) were preselected as instrumental variables[23,59]. To ensure the validity of the instrumental variables, only candidate SNPs that showed a significant association with a trait (metabolite or gene expression level) at an FDR of 0.05 in our data were considered for further analysis (32 SNPs were removed). Associations between SNPs and traits were assessed using linear regressions with age and sex as covariates. To further avoid potential confounding, all candidate SNPs were checked for pairwise linkage disequilibrium using the SNiPA tool [97]. None of the remaining 192 SNPs were in LD. Based on the metabolite-mRNA edges in the BMTI, 550 SNP-metabolite(Met)-mRNA and SNP-mRNA-Met sets were defined, covering 44% of all edges contained in the BMTI. Causal relationships SNP→Met→mRNA and SNP→mRNA→Met were estimated, i.e. whether changes in the metabolite level cause changes in the transcript level and vice versa. Causal effects of both models were calculated using the Wald ratio method [98]:
β^Met→mRNA=β^SNP→mRNAβ^SNP→Metandβ^mRNA→Met=β^SNP→Metβ^SNP→mRNA,
where
β^Met→mRNA
and
β^mRNA→Met
are the causal effects, and
β^SNP→mRNA
and
β^SNP→Met
are regression coefficients of the respective mRNA or metabolite levels on SNPs, under a simple linear model with age and sex as adjustment variables. 95% Confidence intervals and p-values of the causal effects were calculated by sample bootstrapping with 10,000 repetitions. Q-values were calculated to control the false discovery rate (FDR). Summary information for the utilized SNPs together with detailed results of the MR approach can be found in S4 Table.
Metabolic reactions were extracted from the consensus metabolic reconstruction “Recon 2”, v. 02 available at http://humanmetabolism.org as of October 2013 [33]. Compartmental information was removed by merging shared nodes and reactions between different compartments. To avoid potential biologically meaningless shortcuts between network nodes, co-factors and currency metabolites were excluded from the metabolic network prior to the distance calculation (see S5 Table for a full list of removed metabolites). Measured metabolites and transcripts were mapped onto the corresponding network nodes based on KEGG IDs or HMDB identifier for metabolites, and Entrez Gene IDs for transcripts. Distances between all mapped pairs of metabolites and transcripts were defined as the shortest path in the network, i.e. the minimal number of reaction steps between them. For instance, a distance of zero between a transcript and metabolite indicates that the metabolite is a direct reactant of the reaction catalyzed by the particular enzyme encoded by the transcript. A distance of one indicates that the enzyme-encoding transcript catalyzes a directly connected reaction, which takes a product of the particular metabolite as input, and so on. A distance of infinity (Inf) was assigned if the respective metabolite and transcript were disconnected in the pathway network. Moreover, a “not mapped” (NM) distance was assigned if either the metabolite or the transcript could not be mapped to Recon 2. Note that the network was treated as undirected, i.e. all reaction directions were ignored.
Functional annotations were retrieved from two different sources. For transcripts, the generic GO Slim catalogue was downloaded from Gene Ontology (GO, http://www.geneontology.org/GO.slims.shtml). Generic GO Slim is a broad and non-redundant subset of all Gene Ontology terms consisting of 148 unique terms covering all three GO domains (cellular component, molecular function and biological process; [99]). The three root terms cellular component, molecular function and biological process and terms with no annotations for any of the 16,781 transcripts were removed, resulting in 140 terms for further analysis. For metabolites, the subpathway annotations were used (see above). Metabolic pathways (MP) with less than two metabolites were excluded from the analysis, leaving 48 metabolic pathways.
To aggregate the components belonging to a specific annotation term and to derive a score for each of these functional categories, the average of the associated z-score normalized gene expression profiles or metabolite concentrations was calculated according to
aggZCj=1|C|∑i∈CZi,j
where C corresponds to a metabolic pathway or GO term, i enumerates all members in this set, and Zi,j is the z-score of the gene/metabolite with index i in sample j. Spearman’s rank cross-correlation between the aggZ-Scores of all possible GO-MP combinations was then calculated (note that Pearson correlation yielded similar results, see S11 Fig). Since it is known that many biological processes include distinct branches often fulfilling complementary tasks controlled by mutual regulation, a consideration of all pathway members simultaneously could obscure the calculation of the aggZ-Score. A similar problem might occur due to the generic property of the GO-terms or metabolite classes used here, often including functionally rather distinct molecules. To account for this, only those members of the two categories were considered for z-score calculations which share at least one mutual edge within the reconstructed network for the respective GO-MP combination (see S6 Fig for more details). Finally, significant associations between the functional annotation pairs were visualized as a bipartite pathway interaction network (PIN).
Linear regression analysis was performed with age and sex as covariates:
y=β0+β1*x+β2*age+β3*sex+ϵ
where y is the concentration of HDL, LDL or TG over all individuals, β0 is the intercept, β1–3 are regression coefficients, x is a vector of expression/concentration values of a particular gene/metabolite and ϵ is a normally distributed error term. In the same way, the association of annotations (GO and MP) was tested with all three phenotypic traits using the aggZ-Score for the particular annotation of x. Note that for this analysis, aggZ-Scores were calculated only on those members of a particular annotation that are also contained in the BMTI. Each network node was then color-coded with the −log10(p-value) × sign(β1), where the p-value and β1 were derived from the linear regression with the respective metabolite, gene or annotation category. To assess statistical significance of the determined associations, a Bonferroni-corrected threshold of 0.05/(16,780 × 440) ≈ 2.9 × 10−6 was applied.
To investigate regulatory signatures in the BMTI, an enrichment analysis of transcription factor binding sites was performed. Sets of input sequences were created from the neighbors of each metabolite with a degree ≥ 3 (at least 3 connected genes). Analogously, the pathway interaction network was used to construct sequence sets based on the neighborhood of a metabolic pathway node. For each set of input sequences, a separate search for overrepresented TFBS was performed with the sequences of all remaining genes as background model. Promoter regions (-2,000 bp to +200 bp relative to the TSS) and TSS positions of all genes were extracted from the UCSC database using the R package BSgenome.Hsapiens.UCSC.hg19 version 1.3.1. Position-specific weight matrices of the transcription factor binding motifs were taken from the vertebrate collection of the Jaspar database version 5.0 alpha [68]. Enrichment analysis was performed with the TFM-Explorer command line tool [100]. The p-value threshold to determine significance of the motifs in all input sets was set to 1.0 × 10−7 which lies in the recommended optimal range given the numbers of input sequences we used in this study (mean number of input sequences: 62) [101]. The authors showed that for a fixed false positive rate of 10%, the optimal p-value threshold was 1.0 × 10−7 for a dataset of 100 input sequences.
|
10.1371/journal.pcbi.1006236 | Optimal multi-source forecasting of seasonal influenza | Forecasting the emergence and spread of influenza viruses is an important public health challenge. Timely and accurate estimates of influenza prevalence, particularly of severe cases requiring hospitalization, can improve control measures to reduce transmission and mortality. Here, we extend a previously published machine learning method for influenza forecasting to integrate multiple diverse data sources, including traditional surveillance data, electronic health records, internet search traffic, and social media activity. Our hierarchical framework uses multi-linear regression to combine forecasts from multiple data sources and greedy optimization with forward selection to sequentially choose the most predictive combinations of data sources. We show that the systematic integration of complementary data sources can substantially improve forecast accuracy over single data sources. When forecasting the Center for Disease Control and Prevention (CDC) influenza-like-illness reports (ILINet) from week 48 through week 20, the optimal combination of predictors includes public health surveillance data and commercially available electronic medical records, but neither search engine nor social media data.
| In the United States, seasonal influenza causes thousands of deaths and hundreds of thousands of hospitalizations. The annual timing and burden of the flu season vary considerably with the severity of the circulating viruses. Epidemic forecasting can inform early and effective countermeasures to limit the human toll of severe seasonal and pandemic influenza. With a growing toolkit of sophisticated statistical methods and the recent explosion of influenza-related data, we can now systematically match models to data to achieve timely and accurate warning as flu epidemics emerge, peak and subside. Here, we introduce a framework for identifying optimal combinations of data sources, and show that public health surveillance data and electronic health records collectively forecast seasonal influenza better than any single data source alone and better than influenza-related search engine and social media data.
| Seasonal influenza epidemics annually result in significant global morbidity and mortality [1], and influenza pandemics can cause catastrophic levels of death, social disruption, and economic loss [2]. Early detection and forecasting of both emergence and peak epidemic activity can inform an effective allocation of resources, surge planning, and public health messaging [1, 3–5]. Thus, public health and scientific communities have prioritized the development of influenza forecasting technologies [6–11].
There are a growing number and variety of readily available disease-related data sources that may ultimately be integrated into or even replace traditional systems. The Center for Disease Control and Prevention (CDC) relies on data from two primary national influenza surveillance systems: (1) the U.S. World Health Organization (WHO) and National Respiratory and Enteric Virus Surveillance System (NREVSS) collaborating laboratories (henceforth, WHO US) and (2) the US Outpatient Influenza-like Illness Surveillance Network (ILINet). Recently, Meaningful Use [12], a CDC led effort, is advancing the expansion of syndromic surveillance systems such as ESSENCE to address a broader set of infectious disease surveillance objectives [13–15].
Novel data sources for outbreak surveillance are also arising outside of public health. Notably, researchers at Google launched the Google Flu Trends service (GFT) in 2008 to provide real-time estimates of influenza prevalence based on disease-related search activity [16]. They showed that time series tracking the volumes of influenza-related Google searches closely mirrored influenza data from ILINet. However, it failed to capture the emergence of the 2009 H1N1 pandemic and fell short in subsequent influenza seasons [17–21], resulting in the termination of the program in August 2015 by the company. Epidemic-related data have also been extracted from not only search engines [22] but also interactive web-based applications (e.g., Flu Near You, InfluenzaNet) [23] and online social platforms such as Twitter (e.g., MappyHealth) [24, 25], Facebook [24–29], and Wikipedia [30]. While most of these data sources contain broad information, epidemic related data is passively mined and filtered. There are, however, a few participatory systems that directly solicit health data from voluntary participants [23]. For example, InfluenzaNet, has over 50 000 volunteers from ten European countries [23]. While many of these sources have been shown, individually, to estimate and predict influenza activity, we have yet to build forecasting models based on systematic comparisons and integration of complementary data.
Given the real-time availability of GFT at multiple geographic scales (from city to continental), many of the early forecasting methods used GFT as a test bed. Notably, Shaman et al. [8] pioneers the use of Kalman filters to predict seasonal GFT dynamics from historical GFT and humidity data and Nsoesie et al. [31] couples a simulation optimization method with a network-based epidemiological model to forecast regional influenza peaks. Another study forecasts GFT from a combination of GFT, temperature, and humidity data in a specific metropolitan area (Baltimore), and demonstrates that the integration of multiple data sources can improve forecast accuracy [7].
More recent forecasting efforts have directly targeted CDC ILINet, rather than GFT, using a variety of predictor data sources. Brooks et al. [6] apply a novel simulation-based Bayesian forecasting framework to forecast one season of ILINet from prior ILINet data. Their method first constructs prior distributions of seasonal flu curves by stochastically combining and transforming features of past flu seasons. As a season emerges, it updates the posterior distribution based on real-time observations and uses importance sampling to generate forecasts. Two other studies forecast ILINet from alternative data sources—one evaluates the predictive performance of Google, Twitter, and Wikipedia, individually [32], and the other considers a multi-linear combination of internet source, digital surveillance, and electronic medical records data [33].
Such data sources vary considerably in both availability and reliability. Some are available in near-real time, whereas others are lagged by days or weeks; some deeply sample geographic or socioeconomic slices of a population, whereas others provide representative but sparse samples of an entire population. In particular, internet and social media data can be misleading, particularly during newsworthy epidemiological events [34–36], but potentially provide a valuable real-time window into emerging events when combined with validated public health or medical data sources. Optimization allows us to systematically balance such trade-offs and quantify the informational content and complementarity of different categories of data. We argue that, for a given forecasting task, candidate data sources should be evaluated and integrated based on clear performance metrics, which may include, for example, measures of forecast accuracy or precision at one or across multiple time points.
Here, we introduce an optimization method for designing robust multi-source epidemic forecasting systems and apply it forecasting seasonal flu in the US. Our framework is intended to be plug-and-play, allowing researchers to evaluate large combinations of data sources with respect to their own forecasting model and performance metrics. In our case study, the candidate data sources include thousands of time series data sources from public health surveillance systems, electronic health records systems (EHR), search engines, and other website and social media applications. Our forecasting model is an extension of the flexible Bayesian machine learning method introduced in [6], modified to combine multiple predictors. Finally, our objective function considers overall similarity between historical data and out-of-sample forecasts, averaging across 16 recent flu seasons. Unlike recent multi-source forecasting studies (such as [33]), we present a framework to rigorously evaluate much larger sets of candidate data sources both at the national and regional level and select complementary combinations that maximize forecast performance metrics. This approach not only yields more accurate forecasts, but provides quantitative insight into the relative utility of data sources.
We use greedy optimization with forward selection to iteratively identify combinations of predictor data sources that collectively result in the most accurate forecast for a target data source. Our approach consists of three steps, as shown in Fig 1. First, we individually forecast candidate data sources using an empirical Bayesian framework. Second, we use linear models to combine such individual forecasts into grand forecasts of a target time series. Finally, we build an optimal forecasting system (i.e., collection of predictor data sources) by sequentially adding candidate data sources that most improve the accuracy of historical out-of-sample forecasts of the target. Next sections describe these steps in detail.
We use RMSE to evaluate forecasts and thereby select informative combinations of data sources. It measures the difference between predicted and actual time series, as given by
RMSE s = 1 n ∑ w = 1 n ( x w - y w ) 2 (5)
where xw and yw denote the observed and predicted values of the target data source, respectively, at week w of the season, for w = {1, 2, …, n}. Post selection, we evaluate the quality of the forecasts using two additional metrics that address the timing and magnitude of the epidemic peak. Specifically, the peak week error (PWE) of a given season is the absolute difference between predicted and actual peak week, as given by
PWE s = | p - p ˜ | (6)
where p and p ˜ denote the weeks during which the observed and predicted time series, respectively, hit their maximum values. The peak magnitude error (PME) of a given season is the ratio of the absolute difference between the maximum observed and predicted values of the time series and the maximum observed value, as given by
PME s= | h - h ˜ | h (7)
where h and h ˜ denote the maximum values reached by the observed and predicted target time series, respectively.
We analyzed several different sets of candidate data sources, with the goal of identifying subsets of data sources that provide accurate and timely forecasts of ILINet. For each round of data evaluation, we separately predicted each season between 1997 and 2014, excluding the 2009-2010 H1N1 pandemic. For simplicity, we assumed that all 16 seasons span from the 40th calendar week of a given year to the 20th calendar week of the subsequent year. For each season in each data source, we assume that we observe values during the first nine weeks of the season (i.e., the 40th through 48th calendar week) and then forecast ILINet levels for the remainder of the flu season.
Each experiment resulted in an optimized surveillance system, that is, a list of data sources prioritized by the order in which they were selected during optimization. We compare the optimized surveillance systems using three metrics that evaluate the accuracy of the overall (RMSE) and peak (PWE and PME) forecasts.
First, we consider an optimized system consisting of five data sources selected from among all 453 local, regional and national data sources, and compare it to two baseline systems–one using only ILINet to forecast itself and another using a combination of ILINet and WHO laboratory data to forecast ILINet (Table 1). ILINet is selected as the single most informative predictor when evaluated in conjunction with only WHO laboratory data or with all 453 available sources. The fully optimized system combines ILINet with WHO and three Athena state- and regional-level data sources (no internet-based data sources is chosen), suggesting that proprietary electronic medical record data may provide a more reliable source of real-time epidemiological data than freely available internet source data. In comparing the ILINet plus WHO system to the fully optimized system (All), we find that Athena data improves performance only marginally relative to the addition of all four data sources, which together reduce the historical RMSE by roughly 15%.
The optimization selected Athena data from HHS region 8, Illinois, and Georgia, from among all 435 Athena candidate time series. To assess the value of such local, state and regional data, we conducted an additional experiment, restricting the selection to only US-level candidate data sources. The resulting system includes two national Athena data sources (i.e., absolute and percent ILI visits across all facilities) and WordPress flu activity (Table 1). It yields better forecasts than the public health baselines, but is inferior to the optimized system that includes state and regional data.
While ILINet and WHO data are consistently selected as the most informative data sources, they tend to have greater time lags than some of the other real-time candidate data sources. To evaluate the viability of a real-time system using alternative national-level data, we optimized two additional systems, one excluding ILINet and the other excluding both ILINet and WHO data. Without ILINet, WHO is selected as the single most informative source and combined with four different national-level Athena data sources tracking flu-related visits and prescriptions (Table 1). The forecasts decline only slightly relative to systems that include ILINet. However, when both ILINet and WHO data are excluded, the expected performance drops considerably. For comparison, we optimized systems for forecasting state-level ILINet (California, New York, and Texas), and found that national-level surveillance data (ILINet and WHO US) are always selected among the top three most informative data sources, with forecasts enhanced by a variety of state and regional athenahealth variables. (See Table in S1 Table)
The best five-source system (optimized from all available data sources) consistently produces accurate historical out-of-sample forecasts, as shown in Fig 2. After observing only the first nine weeks of the flu season, the system is able to predict the remaining 24 weeks of the season with an average RMSE under 1%. The forecasted 95% credible interval contained the historical ILINet value in 87% of all weeks across all 16 forecasts. However, the 2002-2003 and 2003-2004 forecasts capture the peaks but considerably overestimate prevalence towards the ends of the seasons (12 weeks out of 24 lie outside the 95% credible interval). Excluding these two seasons, 92.9% of all historical weeks fall within the forecasted 95% interval. In the system optimized from all national-level data sources except ILINet, accuracy drops to 66% of all historical weeks contained in the credible intervals. (See S1 Fig for detailed results).
Although these systems were optimized solely to minimize RMSE, the resulting forecasts perform quite well with respect to predicting the timing and magnitude of the epidemic peak. In over 85% of the seasons, the forecasts predict the peak to occur within two weeks of the actual peak; in over 85%, the predicted height of the peak is within 20% of its actual height. Since the Athena predictors are only available between 2011 and 2014, they provide no information for the first 13 of the 16 seasons. Consequently, we see a reduction in RMSE for the three most recent forecasts.
Performance curves for this optimized system indicate that additional data sources, beyond the five included, are not expected to improve performance considerably, according to our empirical results show in Fig 3. On their own, ILINet and WHO are the strongest predictors of future ILINet activity. Although the Athena data sources exhibit poor individual performance, they substantially improve forecast accuracy when combined with ILINet and WHO. The hierarchical selection method was thus able to integrate complementary data sources into a multi-source system that is expected to provide more reliable forecasts than single-source systems. This is also true for systems which exclude ILINet and WHO as candidate predictors. (See S2 Fig for detailed results).
We also build out-of-sample forecasts of ILINet using ILINet and WHO as predictors, using only (1) three years (2011-2014) and (2) five years of training data (2008-2014) to build the Bayesian prior distributions. In the original out-of-sample forecasts, we used 15 of the 16 available seasons to build priors for forecasting the remaining season. (See S3 and S4 Figs for more details). Performance increased with the duration of the training data, with average RMSE decreasing from 0.69 to 0.64 to 0.56 as we increase the training period from three to five to fifteen years. However, even the poorest set of forecasts (based on three years of training) are decent. In addition, we note that the original experiments selected Athena Health data as highly informative predictors, despite only being available for three years (2011-2014).
There are a growing number of powerful methods for forecasting seasonal and pandemic flu (e.g. [6, 45]). To achieve earlier and more accurate predictions of epidemic emergence, growth, peaks and burden, researchers are developing sophisticated statistical methods–some adapted from mature forecasting sciences like meteorology [8]–and creatively leveraging diverse sources of predictor data. The increasing public availability of disease-related data sources is promising yet daunting, with annually, hundreds of thousands of influenza-related tweets [42], several millions of page hits on Wikipedia to influenza-related pages [30], thousands of influenza-related blog posts on Wordpress [40] and hundreds of thousands of hospital and clinic visits. While many studies have demonstrated the promise of surveillance [46] and forecasting from novel data sources [33], we do not yet have rigorous methods for evaluating the utility of such data or identifying effective combinations of data for particular models and forecasting goals.
Over several years, we have developed a general framework for addressing exactly this challenge [20, 46, 47]. For any public health surveillance goal, the approach is designed to systematically evaluate up to thousands of candidate data sources and identify complementary combinations of predictors that achieve the stated goal. For example, we have identified optimal zip codes for seasonal flu surveillance and early detection of pandemic flu in Texas [48], selected informative clinics for dengue surveillance in Puerto Rico [47], and developed software for optimal selection and integration of surveillance data sources for the Defense Threat Reduction Agency’s (DTRA’s) Biosurveillance Ecosystem (BSVE) [49].
In this study, we have used this framework to design multi-source surveillance systems for accurate forecasting of seasonal influenza, and, in the process, rigorously assess the performance and complementarity of diverse data sources. To do so, we combined two previously published methods. The first is an empirical Bayes strategy for forecasting seasonal flu from a single data source [6]. Rather than imposing strong assumptions about transmission dynamics, it assumes that the forecasting target (typically, the currently emerging flu season) will roughly resemble past seasons in terms of the shape, peak week, peak magnitude, and pace of the epidemic curve. By combining and perturbing these features from prior seasonal data, we simulate distributions of plausible (hybrid) flu curves. Then, as a season unfolds, we predict future weeks by extrapolating from variates that most resemble recent activity. To forecast flu (target) from multiple data sources (predictors), we make empirical Bayes forecasts of each predictor separately and combine them into a target forecast using a linear model previously fit to historical predictor and target data. The second method is a greedy optimization that sequentially selects a maximally informative set of data sources to achieve a specified goal [47, 50]. In our case, the candidate providers are a diverse set of public health, commercial health-care, internet query and social media data sources. Our public health goal is accurate forecasting of seasonal flu starting in calendar week 48.
The field has primarily focused on the development of statistical models that predict seasonal dynamics on multiple geopolitical scales, and only secondarily considered the quality of predictor data. Test bed data are often selected based on convenience. Until recently, Google Flu Trends data was free and abundant at multiple scales, and thus a popular choice [7, 10, 20, 31]. A few studies have integrated multiple different types of data and shown that, for short-term forecasting (one to three weeks ahead), the combination of all independent flu predictors performs better than using single source [33]. However, they have not systematically optimized the combination of data sources or quantified their relative contributions to forecast accuracy, as we have done here. Our study confirms that multi-source forecasting can outperform single-source forecasting, but only when complementary sources are identified and systematically integrated.
We optimized forecasting models from three classes of data–traditional public health surveillance data, electronic health records (EHR) from a data services company, and data aggregated from the influenza-related internet search and social network activity. A priori, each has pros and cons. Official surveillance systems are designed for the purpose of monitoring and predicting flu activity, and thus may provide more accurate and robust signals than the alternatives. However, surveillance data tends to be sparse and time-lagged. Internet source data can be abundant and immediately available, but provides only correlated activity that can be highly susceptible to extrinsic perturbations such as media events and modifications to source websites [34, 35]. EHR data has the combined advantages of real-time availability and access to multi-dimensional flu data at various geographic scales. However, it is not freely available and may require statistical corrections for sampling biases.
Our analyses provide quantitative insights into harnessing these trade-offs for forecasting. First, when data sources are evaluated individually, we find that public health surveillance data yields the most accurate forecasts, followed by EHR data, and internet-source data trailing far behind. Second, optimized combinations of data sources (with or without ILINet) provide far better forecasts than any individual data source alone. Third, EHR data are always selected before internet-source data to augment public health data, suggesting that EHR’s provide a more valuable source of complementary information. Forth, when CDC and WHO data are excluded, the optimal EHR and internet-source systems are unable to achieve comparable forecasting performance. Fifth, state-level EHR data improves forecasts significantly more than national-level EHR data.
While we believe that these insights are robust, they may reflect specific assumptions of our model, and not apply to other diseases, forecasting methods, or objective functions. First, the superior performance of the public health data source is likely biased by our choice of ILINet as the gold standard forecasting target. If we had instead sought to forecast athenahealth or GFT time series, these data sources may have been selected as their own top predictors. However, we believe that this choice of target is justified, as it is the only data source specifically designed to estimate flu prevalence in the US. Along with WHO it always selected as a top predictor for selected level forecasts. Second, we follow Brooks et al. [6] in assuming uniform distributions for peak height and peak week, constrained by historical observations. This might limit forecasting accuracy for seasons with atypically high, low, early or late peaks. To address this, one could assume distributions that include low probability extreme departures from past seasons.
We emphasize that this framework is designed to select optimal combinations of data sources for any combination of predictor data sources, multi-linear forecasting method and objective function. As a case study, we built optimal combinations of data sources for forecasting seasonal flu using a published univariate Bayesian empirical framework ([6]) that we extended to forecast with multiple data sources. The optimized systems provide reliable forecasts of the overall seasonal trends and epidemic peak, in most of the 16 historical out-of-sample evaluations. The data-driven selection of informative predictors revealed that public health surveillance data is invaluable for flu forecasting, and that, when rigorously integrated into forecasting models, proprietary electronic health record data can significantly increase accuracy, to a greater degree than freely available internet data. The same optimization framework, forecasting method and RMSE objective function could be readily applied to designing high performing multi-linear forecasting systems for other diseases, for which we have amble historic data, such as Dengue [51–54] and Chikungunya [55]. By modifying the objective function, we can alternatively build systems for forecasting early transmission dynamics or clinical severity of emerging outbreaks.
|
10.1371/journal.pcbi.1002294 | Probabilistic Inference in General Graphical Models through Sampling in Stochastic Networks of Spiking Neurons | An important open problem of computational neuroscience is the generic organization of computations in networks of neurons in the brain. We show here through rigorous theoretical analysis that inherent stochastic features of spiking neurons, in combination with simple nonlinear computational operations in specific network motifs and dendritic arbors, enable networks of spiking neurons to carry out probabilistic inference through sampling in general graphical models. In particular, it enables them to carry out probabilistic inference in Bayesian networks with converging arrows (“explaining away”) and with undirected loops, that occur in many real-world tasks. Ubiquitous stochastic features of networks of spiking neurons, such as trial-to-trial variability and spontaneous activity, are necessary ingredients of the underlying computational organization. We demonstrate through computer simulations that this approach can be scaled up to neural emulations of probabilistic inference in fairly large graphical models, yielding some of the most complex computations that have been carried out so far in networks of spiking neurons.
| Experimental data from neuroscience have provided substantial knowledge about the intricate structure of cortical microcircuits, but their functional role, i.e. the computational calculus that they employ in order to interpret ambiguous stimuli, produce predictions, and derive movement plans has remained largely unknown. Earlier assumptions that these circuits implement a logic-like calculus have run into problems, because logical inference has turned out to be inadequate to solve inference problems in the real world which often exhibits substantial degrees of uncertainty. In this article we propose an alternative theoretical framework for examining the functional role of precisely structured motifs of cortical microcircuits and dendritic computations in complex neurons, based on probabilistic inference through sampling. We show that these structural details endow cortical columns and areas with the capability to represent complex knowledge about their environment in the form of higher order dependencies among salient variables. We show that it also enables them to use this knowledge for probabilistic inference that is capable to deal with uncertainty in stored knowledge and current observations. We demonstrate in computer simulations that the precisely structured neuronal microcircuits enable networks of spiking neurons to solve through their inherent stochastic dynamics a variety of complex probabilistic inference tasks.
| We show in this article that noisy networks of spiking neurons are in principle able to carry out a quite demanding class of computations: probabilistic inference in general graphical models. More precisely, they are able to carry out probabilistic inference for arbitrary probability distributions over discrete random variables (RVs) through sampling. Spikes are viewed here as signals which inform other neurons that a certain RV has been assigned a particular value for a certain time period during the sampling process. This approach had been introduced under the name “neural sampling” in [1]. This article extends the results of [1], where the validity of this neural sampling process had been established for the special case of distributions with at most order dependencies between RVs, to distributions with dependencies of arbitrary order. Such higher order dependencies, which may cause for example the explaining away effect [2], have been shown to arise in various computational tasks related to perception and reasoning. Our approach provides an alternative to other proposed neural emulations of probabilistic inference in graphical models, that rely on arithmetical methods such as belief propagation. The two approaches make completely different demands on the underlying neural circuits: the belief propagation approach emulates a deterministic arithmetical computation of probabilities, and is therefore optimally supported by noise-free deterministic networks of neurons. In contrast, our sampling based approach shows how an internal model of an arbitrary target distribution can be implemented by a network of stochastically firing neurons (such internal model for a distribution , that reflects the statistics of natural stimuli, has been found to emerge in primary visual cortex [3]). This approach requires the presence of stochasticity (noise), and is inherently compatible with experimentally found phenomena such as the ubiquitous trial-to-trial variability of responses of biological networks of neurons.
Given a network of spiking neurons that implements an internal model for a distribution , probabilistic inference for , for example the computation of marginal probabilities for specific RVs, can be reduced to counting the number of spikes of specific neurons for a behaviorally relevant time span of a few hundred ms, similarly as in previously proposed mechanisms for evidence accumulation in neural systems [4]. Nevertheless, in this neural emulation of probabilistic inference through sampling, every single spike conveys information, as well as the relative timing among spikes of different neurons. The reason is that for many of the neurons in the model (the so-called principal neurons) each spike represents a tentative value for a specific RV, whose consistency with tentative values of other RVs, and with the available evidence (e.g., an external stimulus), is explored during the sampling process. In contrast, currently known neural emulations of belief propagation in general graphical models are based on firing rate coding.
The underlying mathematical theory of our proposed new method provides a rigorous proof that the spiking activity in a network of neurons can in principle provide an internal model for an arbitrary distribution . It builds on the general theory of Markov chains and their stationary distribution (see e.g. [5]), the general theory of MCMC (Markov chain Monte Carlo) sampling (see e.g. [6], [7]), and the theory of sampling in stochastic networks of spiking neurons - modelled by a non-reversible Markov chain [1]. It requires further theoretical analysis for elucidating under what conditions higher order factors of p can be emulated in networks of spiking neurons, which is provided in the Methods section of this article. Whereas the underlying mathematical theory only guarantees convergence of the spiking activity to the target distribution , it does not provide tight bounds for the convergence speed to (the so-called burn–in time in MCMC sampling). Hence we complement our theoretical analysis by computer simulations for three Bayesian networks of increasing size and complexity. We also address in these simulations the question to what extent the speed or precision of the probabilistic inference degrades when one moves from a spiking neuron model that is optimal from the perspective of the underlying theory to a biologically more realistic neuron model. The results show, that in all cases quite good probabilistic inference results can be achieved within a time span of a few hundreds ms. In the remainder of this section we sketch the conceptual and scientific background for our approach. An additional discussion of related work can be found in the discussion section.
Probabilistic inference in Bayesian networks [2] and other graphical models [8], [9] is an abstract description of a large class of computational tasks, that subsumes in particular many types of computational tasks that the brain has to solve: The formation of coherent interpretations of incomplete and ambiguous sensory stimuli, integration of previously acquired knowledge with new information, movement planning, reasoning and decision making in the presence of uncertainty [10]–[13]. The computational tasks become special cases of probabilistic inference if one assumes that the previously acquired knowledge (facts, rules, constraints, successful responses) is encoded in a joint distribution over numerous RVs , that represent features of sensory stimuli, aspects of internal models for the environment, environmental and behavioral context, values of carrying out particular actions in particular situations [14], goals, etc. If the values of some of these RVs assume concrete values (e.g. because of observations, or because a particular goal has been set), the distribution of the remaining variables changes in general (to the conditional distribution given the values ). A typical computation that needs to be carried out for probabilistic inference for some joint distribution involves in addition marginalization, and requires for example the evaluation of an expression of the form(1)where concrete values (the “evidence”or “observations” have been inserted for the RVs , . These variables are then often called observable variables, and the others latent variables. Note that the term “evidence” is somewhat misleading, since the assignment represents some arbitrary input to a probabilistic inference computation, without any connotation that it represents correct observations or memories. The computation of the resulting marginal distribution requires a summation over all possible values for the RVs that are currently not of interest for this probabilistic inference. This computation is in general quite complex (in fact, it is NP-complete [9]) because in the worst case exponentially in many terms need to be evaluated and summed up.
There exist two completely different approaches for solving probabilistic inference tasks of type (1), to which we will refer in the following as the arithmetical and the sampling approach. In the arithmetical approach one exploits particular features of a graphical model, that captures conditional independence properties of the distribution , for organizing the order of summation steps and multiplication steps for the arithmetical calculation of the r.h.s. of (1) in an efficient manner. Belief propagation and message passing algorithms are special cases of this arithmetical approach. All previously proposed neural emulations of probabilistic inference in general graphical models have pursued this arithmetical approach. In the sampling approach, which we pursue in this article, one constructs a method for drawing samples from the distribution (with fixed values for some of the RVs, see (1)). One can then approximate the l.h.s. of (1), i.e., the desired value of the probability , by counting how often each possible value for the RV occurs among the samples. More precisely, we identify conditions under which each current firing state (which records which neuron has fired within some time window) of a network of stochastically firing neurons can be viewed as a sample from a probability distribution that converges to the target distribution . For this purpose the temporal dynamics of the network is interpreted as a (non-reversible) Markov chain. We show that a suitable network architecture and parameter choice of the network of spiking neurons can make sure that this Markov chain has the target distribution as its stationary distribution, and therefore produces after some “burn–in time”samples (i.e., firing states) from a distribution that converges to . This general strategy for sampling is commonly referred to as Markov chain Monte Carlo (MCMC) sampling [6], [7], [9].
Before the first use of this strategy in networks of spiking neurons in [1], MCMC sampling had already been studied in the context of artificial neural networks, so-called Boltzmann machines [15]. A Boltzmann machine consists of stochastic binary neurons in discrete time, where the output of each neuron has the value or at each discrete time step. The probability of each value depends on the output values of neurons at the preceding discrete time step. For a Boltzmann machine a standard way of sampling is Gibbs sampling. The Markov chain that describes Gibbs sampling is reversible, i.e., stochastic transitions between states do not have a preferred direction in time. This sampling method works well in artificial neural networks, where the effect of each neural activity lasts for exactly one discrete time step. But it is in conflict with basic features of networks of spiking neurons, where each action potential (spike) of a neuron triggers inherent temporal processes in the neuron itself (e.g. refractory processes), and postsynaptic potentials of specific durations in other neurons to which it is synaptically connected. These inherent temporal processes of specific durations are non-reversible, and are therefore inconsistent with the mathematical model (Gibbs sampling) that underlies probabilistic inference in Boltzmann machines. [1] proposed a somewhat different mathematical model (sampling in non-reversible Markov chains) as an alternative framework for sampling, that is compatible with these basic features of the dynamics of networks of spiking neurons.
We consider in this article two types of models for spiking neurons (see Methods for details):
A key step for interpreting the firing activity of networks of neurons as sampling from a probability distribution (as proposed in [3]) in a rigorous manner is to define a formal relationship between spikes and samples. As in [1] we relate the firing activity in a network of spiking neurons to sampling from a distribution over binary variables by setting(2)(we restrict our attention here to binary RVs; multinomial RVs could in principle be represented by WTA circuits –see Discussion). The constant models the average length of the effect of a spike on the firing probability of other neurons or of the same neuron, and can be set for example to .
However with this definition of its internal state () the dynamics of the neural network can not be modelled by a Markov chain, since knowledge of this current state does not suffice for determining the distribution of states at future time points, say at time . This distribution requires knowledge about when exactly a neuron with had fired. Therefore auxiliary RVs with multinomial or analog values were introduced in [1], that keep track of when exactly in the preceding time interval of length a neuron had fired, and thereby restore the Markov property for a Markov chain that is defined over an enlarged state set consisting of all possible values of and . However the introduction of these hidden variables , that keep track of inherent temporal processes in the network of spiking neurons, comes at the price that the resulting Markov chain is no longer reversible (because these temporal processes are not reversible). But it was shown in [1] that one can prove nevertheless for any distribution for which the so-called neural computability condition (NCC), see below, can be satisfied by a network of spiking neurons, that defines a non-reversible Markov chain whose stationary distribution is an expanded distribution , whose marginal distribution over (which results when one ignores the values of the hidden variables ) is the desired distribution . Hence a network of spiking neurons can sample from any distribution for which the NCC can be satisfied. This implies that any neural system that contains such network can carry out the probabilistic inference task (1): The evidence could be implemented through external inputs that force neuron to fire at a high rate if in , and not to fire if in . In order to estimate , it suffices that some readout neuron estimates (after some initial transient phase) the resulting firing rate of the neuron that represents RV .
In contrast to most of the other neural implementations of probabilistic inference (with some exceptions, see for example [17] and [18]) where information is encoded in the firing rate of the neurons, in this approach the spike times, rather than the firing rate, of the neuron carry relevant information as they define the value of the RV at a particular moment in time according to (2). In this spike-time based coding scheme, the relative timing of spikes (which neuron fires simultaneously with whom) receives a direct functional interpretation since it determines the correlation between the corresponding RVs.
The NCC requires that for each RV the firing probability density of its corresponding neuron at time satisfies, if the neuron is not in a refractory period,(3)where denotes the current value of all other RVs, i.e., all with . We use in this article the same model for a stochastic neuron as in [1] (continuous time case), which can be matched quite well to biological data according to [19]. In the simpler version of this neuron model one assumes that it has an absolute refractory period of length , and that the instantaneous firing probability satisfies outside of its refractory period , where is its membrane potential (see Methods for an account of the more complex neuron model with a relative refractory period from [1], that we have also tested in our simulations). The NCC from (3) can then be reformulated as a condition on the membrane potential of the neuron(4)
Let us consider a Boltzmann distribution of the form(5)with symmetric weights (i.e., ) that vanish on the diagonal (i.e., ). In this case the NCC can be satisfied by a that is linear in the postsynaptic potentials that neuron receives from the neurons that represent other RVs :(6)where is the bias of neuron (which regulates its excitability), is the strength of the synaptic connection from neuron to , and approximates the time course of the postsynaptic potential caused by a firing of neuron at some time ( assumes value 1 during the time interval , otherwise it has value ).
However, it is well known that probabilistic inference for distributions of the form (5) is too weak to model various important computational tasks that the brain is obviously able to solve, at least without auxiliary variables. While (5) only allows pairwise interactions between RVs, numerous real world probabilistic inference tasks require inference for distributions with higher order terms. For example, it has been shown that human visual perception involves “explaining away”, a well known effect in probabilistic inference, where a change in the probability of one competing hypothesis for explaining some observation affects the probability of another competing hypothesis [20]. Such effects can usually only be captured with terms of order at least 3, since 3 RVs (for 2 hypotheses and 1 observation) may interact in complex ways. A well known example from visual perception is shown in Fig. 1, for a probability distribution over 4 RVs , where is defined by the perceived relative reflectance of two abutting 2D areas, by the perceived 3D shape of the observed object, by the observed shading of the object, and by the contour of the 2D image. The difference in shading of the two abutting surfaces in Fig. 1A could be explained either by a difference in reflectance of the two surfaces, or by an underlying curved 3D shape. The two different contours (RV ) in the upper and lower part of Fig. 1A influence the likelihood of a curved 3D shape (RV ). In particular, a perceived curved 3D shape “explains away” the difference in shading, thereby making a uniform reflectance more likely. The results of [21] and numerous related results suggest that the brain is able to carry out probabilistic inference for more complex distributions than the order Boltzmann distribution (5).
We show in this article that the neural sampling method of [1] can be extended to any probability distribution over binary RVs, in particular to distributions with higher order dependencies among RVs, by using auxiliary spiking neurons in that do not directly represent RVs , or by using nonlinear computational processes in multi-compartment neuron models. As one can expect, the number of required auxiliary neurons or dendritic branches increases with the complexity of the probability distribution for which the resulting network of spiking neurons has to carry out probabilistic inference. Various types of graphical models [9] have emerged as convenient frameworks for characterizing the complexity of distributions from the perspective of probabilistic inference for .
We will focus in this article on Bayesian networks, a common type of graphical model for probability distributions. But our results can also be applied for other types of graphical models. A Bayesian network is a directed graph (without directed cycles), whose nodes represent RVs . Its graph structure indicates that admits a factorization of the form(7)where is the set of all (direct) parents of the node indexed by . For example, the Bayesian network in Fig. 1B implies that the factorization is possible.
We show that the complexity of the resulting network of spiking neurons for carrying out probabilistic inference for can be bounded in terms of the graph complexity of the Bayesian network that gives rise to the factorization (7). More precisely, we present three different approaches for constructing such networks of spiking neurons:
We will show that there exist two different neural implementation options for each of the last two approaches, using either specific network motifs or dendritic processing for nonlinear computation steps. This yields altogether 5 different options for emulating probabilistic inference in Bayesian networks through sampling via the inherent stochastic dynamics of networks of spiking neurons. We will exhibit characteristic differences in the complexity and performance of the resulting networks, and relate these to the complexity of the underlying Bayesian network. All 5 of these neural implementation options can readily be applied to Bayesian networks where several arcs converge to a node (giving rise to the “explaining away” effect), and to Bayesian networks with undirected cycles (“loops”). All methods for probabilistic inference from general graphical models that we propose in this article are from the mathematical perspective special cases of MCMC sampling. However in view of the fact that they expand the neural sampling approach of [1], we will refer to them more specifically as neural sampling.
We show through computer simulations for three different Bayesian networks of different sizes and complexities that neural sampling can be carried quite fast with the help of the second and third approach, providing good inference results within a behaviorally relevant time span of a few hundred ms. One of these Bayesian networks addresses the previously described classical “explaining away” effect in visual perception from Fig. 1. The other two Bayesian networks not only contain numerous “explaining away” effects, but also undirected cycles. Altogether, our computer simulations and our theoretical analyses demonstrate that networks of spiking neurons can emulate probabilistic inference for general Bayesian networks. Hence we propose to view probabilistic inference in graphical models as a generic computational paradigm, that can help us to understand the computational organization of networks of neurons in the brain, and in particular the computational role of precisely structured cortical microcircuit motifs.
We present several ways how probabilistic inference for a given joint distribution , that is not required to have the form of a order Boltzmann distribution (5), can be carried out through sampling from the inherent dynamics of a recurrent network of stochastically spiking neurons. All these approaches are based on the idea that such network of spiking neurons can be viewed –for a suitable choice of its architecture and parameters –as an internal or “physical model” for the distribution , in the sense that its distribution of network states converges to , from any initial state. Then probabilistic inference for can be easily carried out by any readout neuron that observes the resulting network states, or the spikes from one or several neurons in the network. This holds not only for sampling from the prior distribution , but also for sampling from the posterior after some evidence has become available (see (1)). The link between network states of and the RVs is provided by assuming that there exists for each RV a neuron such that each time when fires, it sets the associated binary RV to 1 for a time period of some length (see Fig. 1C). We refer to neurons that represent in this way a RV as principal neurons. All other neurons are referred to as auxiliary neurons.
The mathematical basis for analyzing the distribution of network states, and relating it to a given distribution , is provided by the theory of Markov chains. More precisely, it was shown in [1] that by introducing for each principal neuron an additional hidden analog RV , that keeps track of time within the time interval of length after a spike of , one can model the dynamics of the network by a non-reversible Markov chain. This Markov chain is non-reversible, in contrast to Gibbs sampling or other Markov chains that are usually considered in Machine Learning and in the theory of Boltzmann machines, because this facilitates the modelling of the temporal dynamics of spiking neurons, in particular refractory processes within a spiking neuron after a spike and temporally extended effects of its spike on the membrane potential of other neurons to which it is synaptically connected (postsynaptic potentials). The underlying mathematical theory guarantees that nevertheless the distribution of network states of this Markov chain converges (for the “original” RVs ) to the given distribution , provided that the NCC (4) is met. This theoretical result reduces our goal, to demonstrate ways how a network of spiking neurons can carry out probabilistic inference in general graphical models, to the analysis of possibilities for satisfying the NCC (4) in networks of spiking neurons. The networks of spiking neurons that we construct and analyze build primarily on the model for neural sampling in continuous time from [1], since this continuous time version is the more satisfactory model from the biological perspective. But all our results also hold for the mathematically simpler version with discrete time.
We exhibit both methods for satisfying the NCC with the help of auxiliary neurons in networks of point neurons, and in networks of multi-compartment neuron models (where no auxiliary neurons are required). All neuron models that we consider are stochastic, where the probability density function for the firing of a neuron at time (provided it is currently not in a refractory state) is proportional to , where is its current membrane potential at the soma. We assume (as in [1]) that in a point neuron model the membrane potential can be written as a linear combination of postsynaptic potentials. Thus if the principal neuron is modelled as a point neuron, we have(8)where is the bias of neuron (which regulates its excitability), is the strength of the synaptic connection from neuron to , and approximates the time course of the postsynaptic potential in neuron caused by a firing of neuron . The ideal neuron model from the perspective of the theory of [1] has an absolute refractory period of length , which is also the assumed length of a postsynaptic potential (EPSP or IPSP). But it was shown there through computer simulations that neural sampling can be carried out also with stochastically firing neurons that have a relative refractory period, i.e. the neuron can fire with some probability with an interspike interval of less than . In particular, it was shown there in simulations that the resulting neural network samples from a slight variation of the target distribution , that is in most cases practically indistinguishable.
Before we describe two different theoretical approaches for satisfying the NCC, we first consider an even simpler method for extending the neural sampling approach from [1] to arbitrary distributions : through a reduction to order Boltzmann distributions (5) with auxiliary RVs.
It is well known [15] that any probability distribution , with arbitrarily large factors in a factorization such as (7), can be represented as marginal distribution(9)of an extended distribution with auxiliary RVs , that can be factorized into factors of degrees at most . This can be seen as follows. Let be an arbitrary probability distribution over binary variables with higher order factors ). Thus(10)where is a vector composed of the RVs that the factor depends on and is a normalization constant. We additionally assume that is non-zero for each value of . The simple idea is to introduce for each possible assignment to the RVs in a higher order factor a new RV , that has value 1 only if is the current assignment of values to the RVs in . We will illustrate this idea through the concrete example of Fig. 1. Since there is only one factor that contains more than 2 RVs in the probability distribution of this example (see caption of Fig. 1), the conditional probability , there will be 8 auxiliary RVs , , …, for this factor, one for each of the 8 possible assignments to the 3 RVs in . Let us consider a particular auxiliary RV, e.g. . It assumes value 1 only if , , and . This constraint for can be enforced through second order factors between and each of the RVs and . For example, the second order factor that relates and has a value of 0 if and (i.e., if is not compatible with the assignment ), and value 1 otherwise. The individual values of the factor for different assignments to , and are introduced in the extended distribution through first order factors, one for each auxiliary RV . Specifically, the first order factor that depends on has value (where is a constant that rescales the values of the factors such that for all assignments to , and ) if , and value 1 otherwise. Further details of the construction method for are given in the Methods section, together with a proof of (9).
The resulting extended probability distribution has the property that, in spite of deterministic dependencies between the RVs and , the state set of the resulting Markov chain realized through a network of spiking neurons according to [1] (that consists of all non-forbidden value assignments to and ) is connected. In the previous example a non-forbidden value assignment is and . But is also a non-forbidden value assignment. Such non-forbidden value assignments to the auxiliary RVs corresponding to one higher order factor, where all of them assume value of 0 regardless of the values of the RVs provide transition points for paths of probability that connect any two non-forbidden value assignments (without requiring that 2 or more RVs switch their values simultaneously). The resulting connectivity of all non-forbidden states (see Methods for a proof) implies that this Markov chain has as its unique stationary distribution. The given distribution arises as marginal distribution of this stationary distribution of , hence one can use to sample from (just ignore the firing activity of neurons that correspond to auxiliary RVs ).
Since the number of RVs in the extended probability distribution can be much larger than the number of RVs in , the corresponding spiking neural network samples from a much larger probability space. This, as well as the presence of deterministic relations between the auxiliary and the main RVs in the expanded probability distribution, slow down the convergence of the resulting Markov chain to its stationary distribution. We show however in the following, that there are several alternatives for sampling from an arbitrary distribution through a network of spiking neurons. These alternative methods do not introduce auxiliary RVs , but rather aim at directly satisfying the NCC (4) in a network of spiking neurons. Note that the principal neurons in the neural network that implements neural sampling through introduction of auxiliary RVs also satisfy the NCC, but in the extended probability distribution with second order relations , whereas in the neural implementations introduced in the following the principal neurons satisfy the NCC in the original distribution . In Computer Simulation I we have compared the convergence speed of the methods that satisfy the NCC with that of the previously described method via auxiliary RVs. It turns out that the alternative strategy provides an about fold speed-up for the Bayesian network of Fig. 1B.
Assume that the distribution for which we want to carry out probabilistic inference is given by some arbitrary Bayesian network . There are two different options for satisfying the NCC for , which differ in the way by which the term on the r.h.s. of the NCC (4) is expanded. The option that we will analyze first uses from the structure of the Bayesian network only the information about which RVs are in the Markov blanket of each RV . The Markov blanket of the corresponding node in (which consists of the parents, children and co-parents of this node) has the property that is independent from all other RVs once any assignment of values to the RVs in the Markov blanket has been fixed. Hence = , and the term on the r.h.s. of the NCC (4) can be expanded as follows:(11)where(12)The sum indexed by runs over the set of all possible assignments of values to , and denotes a predicate which has value 1 if the condition in the brackets is true, and to 0 otherwise. Hence, for satisfying the NCC it suffices if there are auxiliary neurons, or dendritic branches, for each of these , that become active if and only if the variables currently assume the value . The current values of the variables are encoded in the firing activity of their corresponding principal neurons. The corresponding term can be implemented with the help of the bias (see (8)) of the auxiliary neuron that corresponds to the assignment , resulting in a value of its membrane potential equal to the r.h.s. of the NCC (4). We will discuss this implementation option below as Implementation 2. In the subsequently discussed implementation option (Implementation 3) all principal neurons will be multi-compartment neurons, and no auxiliary neurons are needed. In this case scales the amplitude of the signal from a specific dendritic branch to the soma of the multi-compartment principal neuron .
The second strategy to expand the log-odd ratio on the r.h.s. of the NCC (4) uses the factorized form (10) of the probability distribution . This form allows us to rewrite the log-odd ratio in (4) as a sum of log terms, one for each factor , , that contains the RV (we write for this set of factors). One can write each of these terms as a sum over all possible assignments of values of the variables the factor depends on (except ). This yields(13)where is a vector composed of the RVs that the factor depends on –without , and is the current value of this vector at time . denotes the set of all possible assignments to the RVs . The parameters are set to(14)The factorized expansion in (13) is similar to (11), but with the difference that we have another sum running over all factors that depend on . Consequently, in the resulting Implementation 4 with auxiliary neurons and dendritic branches there will be several groups of auxiliary neurons that connect to , where each group implements the expansion of one factor in (13). The alternative model that only uses dendritic computation (Implementation 5) will have groups of dendritic branches corresponding to the different factors. The number of auxiliary neurons that connect to in Implementation 4 (and the corresponding number of dendritic branches in Implementation 5) is equal to the sum of the exponents of the sizes of factors that depend on : , where denotes the number of RVs in the vector . This number is never larger than (where is the size of the Markov blanket of ), which gives the corresponding number of auxiliary neurons or dendritic branches that are required in the Implementation 2 and 3. These two numbers can considerably differ in graphical models where the RVs participate in many factors, but the size of the factors is small. Therefore one advantage of this approach is that it requires in general fewer resources. On the other hand, it introduces a more complex connectivity between the auxiliary neurons and the principal neuron (compare Fig. 5 with Fig. 2).
We have tested the viability of the previously described approach for neural sampling by satisfying the NCC also on two larger and more complex Bayesian networks: the well-known ASIA-network [24], and an even larger randomly generated Bayesian network. The primary question is in both cases, whether the convergence speed of neural sampling is in a range where a reasonable approximation to probabilistic inference can be provided within the typical range of biological reaction times of a few 100 ms. In addition, we examine for the ASIA-network the question to what extent more complex and biologically more realistic shapes of EPSPs affect the performance. For the larger random Bayesian network we examine what difference in performance is caused by neuron models with absolute versus relative refractory periods.
We have shown through rigorous theoretical arguments and computer simulations that networks of spiking neurons are in principle able to emulate probabilistic inference in general graphical models. The latter has emerged as a quite suitable mathematical framework for describing those computational tasks that artificial and biological intelligent agents need to solve. Hence the results of this article provide a link between this abstract description level of computational theory and models for networks of neurons in the brain. In particular, they provide a principled framework for investigating how nonlinear computational operations in network motifs of cortical microcircuits and in the dendritic trees of neurons contribute to brain computations on a larger scale. Altogether we view our approach as a contribution to the solution of a fundamental open problem that has been raised in Cognitive Science:
“What approximate algorithms does the mind use, how do they relate to engineering approximations in probabilistic AI, and how are they implemented in neural circuits? Much recent work points to Monte Carlo or stochastic sampling–based approximations as a unifying framework for understanding how Bayesian inference may work practically across all these levels, in minds, brains, and machines ” [13].
We have presented three different theoretical approaches for extending the results of [1], such that they yield explanations how probabilistic inference in general graphical models could be carried out through the inherent dynamics of recurrent networks of stochastically firing neurons (neural sampling). The first and simplest one was based on the fact that any distribution can be represented as marginal distribution of a order Boltzmann distribution (5) with auxiliary RVs. However, as we have demonstrated in Fig. 3, this approach yields rather slow convergence of the distribution of network states to the target distribution. This is a natural consequence of the deterministic definition of new RVs in terms of the original RVs, which reduces the conductance [9], [30] (i.e., the probability to get from one set of network states to another set of network states) of the Markov chain that is defined by the network dynamics. Further research is needed to clarify whether this deficiency can be overcome through other methods for introducing auxiliary RVs.
We have furthermore presented two approaches for satisfying the NCC (3) of [1], which is a sufficient condition for sampling from a given distribution. These two closely related approaches rely on different ways of expanding the term on the r.h.s. of the NCC (4). The first approach can be used if the underlying graphical model implies that the Markov blankets of all RVs are relatively small. The second approach yields efficient neural emulations under a milder constraint: if each factor in a factorization of the target distribution is rather small (and if there are not too many factors). Each of these two approaches provides the theoretical basis for two different methods for satisfying the NCC in a network of spiking neurons: either through nonlinear computation in network motifs with auxiliary spiking neurons (that do not directly represent a RV of the target distribution), or through dendritic computation in multi-compartment neuron models. This yields altogether four different options for satisfying the NCC in a network of spiking neurons. These four options are demonstrated in Fig. 2, 4–6 for a characteristic explaining away motif in the simple Bayesian network of Fig. 1B, that had previously been introduced to model inference in biological visual processing [21]. The second approach for satisfying the NCC never requires more auxiliary neurons or dendritic branches than the first approach.
Each of these four options for satisfying the NCC would be optimally supported by somewhat different features of the interaction of excitation and inhibition in canonical cortical microcircuit motifs, and by somewhat different features of dendritic computation. Sufficiently precise and general experimental data are not yet available for many of these features, and we hope that the computational consequences of these features that we have exhibited in this article will promote further experimental work on these open questions. In particular, the neural circuit of Fig. 5 uses an implementation strategy that requires for many graphical models (those where Markov blankets are substantially larger than individual factors) fewer auxiliary neurons. But it requires temporally precise local inhibition in dendritic branches that has negligible effects on the membrane potential at the soma, or in other dendritic branches that are used for this computation. Some experimental results in this direction are reported in [31], where it was shown (see e.g. their Fig. 1) that IPSPs from apical dendrites of layer 5 pyramidal neurons are drastically attenuated at the soma. The options that rely on dendritic computation (Fig. 4 and 6) would be optimally supported if EPSPs from dendritic branches that are not amplified by dendritic spikes have hardly any effect on the membrane potential at the soma. Some experimental results which support this assumption for distal dendritic branches of layer 5 pyramidal neurons had been reported in [26], see e.g. their Fig. 1. With regard to details of dendritic spikes, these would optimally support the ideal theoretical models with dendritic computation if they would have a rather short duration at the soma, in order to avoid that they still affect the firing probability of the neuron when the state (i.e., firing or non-firing within the preceding time interval of length ) of presynaptic neurons has changed. In addition, the ideal impact of a dendritic spike on the membrane potential at the soma would approximate a step function (rather than a function with a pronounced peak at the beginning).
Another desired property of the dendritic spikes in context of our neural implementations is that their propagation from the dendritic branch to the soma should be very fast, i.e. with short delays that are much smaller than the duration of the EPSPs. This is in accordance with the results reported in [32] where they found (see their Fig. 1) that the fast active propagation of the dendritic spike towards the soma reduces the rise time of the voltage at the soma to less than a millisecond, in comparison to the 3 ms rise time during the propagation of the individual EPSPs when there is no dendritic spike. Further, in [22] it is shown that the latency of an action potential evoked by a strong dendritic spike, calculated with respect to the time of the activation of the synaptic input at the dendritic branch, is slightly below 2 ms, supporting the assumption of fast propagation of the dendritic spike to the soma.
We have focused in this article on the description of ideal neural emulations of probabilistic inference in general graphical models. These ideal neural implementations use a complete representation of the conditional odd-ratios, i.e. have a separate auxiliary neuron or dendritic branch for each possible assignment of values to the RVs in the Markov blanket in implementations 2 and 3, or in the factor in implementations 4 and 5. Hence, the required number of neurons (or dendritic branches) scales exponentially with the sizes of the Markov blankets and the factors in the probability distribution, and it would quickly become unfeasible to represent probability distributions with larger Markov blankets or factors. One possible way to overcome this limitation is to consider an approximate implementation of the NCC with fewer auxiliary neurons or dendritic branches. In fact, such an approximate implementation of the NCC could be learned. Our results provide the basis for investigating in subsequent work how approximations to these ideal neural emulations could emerge through synaptic plasticity and other adaptive processes in neurons. First explorations of these questions suggest that in particular approximations to Implementations 1,2 and 4 could emerge through STDP in a ubiquitous network motif of cortical microcircuits [33]: Winner-Take-All circuits formed by populations of pyramidal neurons with lateral inhibition. This learning-based approach relies on the observation that STDP enables pyramidal neurons in the presence of lateral inhibition to specialize each on a particular pattern of presynaptic firing activity, and to fire after learning only when this presynaptic firing pattern appears [34]. These neurons would then assume the role of the auxiliary neurons, both in the first option with auxiliary RVs, and in the options shown in Fig. 2 and 5. Furthermore, the results of [23] suggest that STDP in combination with branch strength potentiation enables individual dendritic branches to specialize on particular patterns of presynaptic inputs, similarly as in the theoretically optimal constructions of Fig. 4 and 6. One difference between the theoretically optimal neural emulations and learning based approximations is that auxiliary neurons or dendritic branches learn to represent only the most frequently occurring patterns of presynaptic firing activity, rather than creating a complete catalogue of all theoretically possible presynaptic firing patterns. This has the advantage that fewer auxiliary neurons and dendritic branches are needed in these biologically more realistic learning-based approximations.
Other ongoing research explores neural emulations of probabilistic inference for non-binary RVs. In this case a stochastic principal neuron that represents a binary RV is replaced by a Winner-Take-All circuit, that encodes the value of a multinomial or analog RV through population coding, see [34].
There are a number of studies proposing neural network architectures that implement probabilistic inference [15], [17], [18], [35]–[48]. Most of these models propose neural emulations of the belief propagation algorithm, where the activity of neurons or populations of neurons encodes intermediate values (called messages or beliefs) needed in the arithmetical calculation of the posterior probability distribution. With some exceptions [17], most of the approaches assume rate-based coding of information and use rate-based neuron models or mean-field approximations.
In particular, in [37] a spiking neural network model was developed that performs the max-product message passing algorithm, a variant of belief propagation, where the necessary maximization and product operations were implemented by specialized neural circuits. Another spiking neural implementation of the sum-product belief propagation algorithm was proposed in [36], where the calculation and passing of the messages was achieved in a recurrent network of interconnected liquid state machines [49]. In these studies, that implemented probabilistic inference with spiking neurons through emulation of the belief propagation algorithm on tree factor graphs, the beliefs or the messages during the calculation of the posterior distributions were encoded in an average firing rate of a population of neurons. Regarding the complexity of these neural models, as the number of required computational operations in belief propagation is exponential in the size of the largest factor in the probability distribution, in the neural implementations this translates to a number of neurons in the network that scales exponentially with the size of the largest factor. This complexity corresponds to the required number of neurons (or dendritic branches) in implementations 1, 3 and 5 in our approach, whereas implementations 2 and 4 require a larger number of neurons that scales exponentially with the size of the largest Markov blanket in the distribution. Additionally, note that the time of convergence to the correct posterior differs in both approaches: in the belief propagation based models it scales in the worst case linearly with the number of RVs in the probability distribution, whereas in our approach it can vary depending on the probability distribution.
Although the belief propagation algorithm can be applied to graphical models with undirected loops (a variant called loopy belief propagation), it is not always guaranteed to work, which limits the applicability of the neural implementations based on this algorithm. The computation and the passing of messages in belief propagation uses, however, equivalent computations as the junction tree algorithm [24], [50], a message passing algorithm that operates on a junction tree, a tree structure derived from the graphical model. The junction tree algorithm performs exact probabilistic inference in general graphical models, including those that have loops. Hence, the neural implementations of belief propagation could in principle be adapted to work on junction trees as well. This however comes at a computational cost manifested in a larger required size of the neural network, since the number of required operations for the junction tree algorithm scales exponentially with the width of the junction tree, and the width of the junction tree can be larger than the size of the largest factor for graphical models that have loops (see [9], chap. 10 for a discussion). The analysis of the complexity and performance of resulting emulations in networks of spiking neurons is an interesting topic for future research.
Another interesting approach, that adopts an alternative spike-time based coding scheme, was described in [17]. In this study a spiking neuron model estimates the log-odd ratio of a hidden binary state in a hidden Markov model, and it outputs a spike only when it receives new evidence from the inputs that causes a shift in the estimated log-odd ratio that exceeds a certain threshold, that is, only when new information about a change in the log-odd ratio is presented that cannot be predicted by the preceding spikes of the neuron. However, this study considers only a very restricted class of graphical models: Bayesian networks that are trees (where for example no explaining away can occur). The ideas in [17] have been extended in [18], where the neural model is capable of integration of evidence from multiple simultaneous cues (the underlying graphical model is a hidden Markov model with multiple observations). It uses a population code for encoding the log-posterior estimation of the time varying hidden stimulus, which is modeled as a continuous RV instead of the binary hidden state used in [17]. In these studies, as in ours, spikes times carry relevant information, although there the spikes are generated deterministically and signal a prediction error used to update and correct the estimated log-posterior, whereas in our approach the spikes are generated by a stochastic neuron model and define the current values of the RVs during the sampling.
The idea that nonlinear dendritic mechanisms could account for the nonlinear processing that is required in neural models that perform probabilistic inference has been proposed previously in [39] and [41], albeit for the belief propagation algorithm. In [39] the authors introduce a neural model that implements probabilistic inference in hidden Markov models via the belief propagation algorithm, and suggest that the nonlinear functions that arise in the model can be mapped to the nonlinear dendritic filtering. In [41] another rate-based neural model that implements the loopy belief propagation algorithm in general graphical models was described, where the required multiplication operations in the algorithm were proposed to be implemented by the nonlinear processing in individual dendritic trees.
While there exist several different spiking neural network models in the literature that perform probabilistic inference based on the belief propagation algorithm, there is a lack of spiking neural network models that implement probabilistic inference through Markov chain Monte Carlo (MCMC sampling). To the best of our knowledge, the neural implementations proposed in this article are the only spiking neural networks for probabilistic inference via MCMC in general graphical models. In [35] a non-spiking neural network composed of stochastic binary neurons was introduced called Boltzmann machine, that performs probabilistic inference via Gibbs sampling. The neural network in [35] performs inference via sampling in probability distributions that have only pairwise couplings between the RVs. An extension was proposed in [51], that can perform Gibbs sampling in probability distributions with higher order dependencies between the variables, which corresponds to the class of probability distributions that we consider in this article. A spiking neural network model based on the results in [35] had been proposed in [52], for a restricted class of probability distributions that only have second order factors, and which satisfy some additional constraints on the conditional independencies between the variables. To the best of our knowledge, this approach had not been extended to more general probability distributions.
A recent study [53] showed that as the noise in the neurons increases and their reliability drops, the optimal couplings between the neurons that maximize the information that the network conveys about the inputs become larger in magnitude, creating a redundant code that reduces the impact of noise. Effectively, the network learns the input distribution in its couplings, and uses this knowledge to compensate for errors due to the unreliable neurons. These findings are consistent with our models, and although we did not consider learning in this article, we expect that the introduction of learning mechanisms that optimize a mutual information measure in our neural implementations would yield optimal couplings that obey the same principles as the ones reported in [53]. While stochasticity in the neurons represents a crucial property that neural implementations of probabilistic inference through sampling rely on, this study elucidates an important additional effect it has in learning paradigms that use optimality principles like information maximization: it induces redundant representation of information in a population of neurons.
The existing gap between abstract computational models of information processing in the brain that use MCMC algorithms for probabilistic inference on one hand, and neuroscientific data about neural structures and neural processes on the other hand, has been pointed out and emphasized by several studies [12], [13], [54], [55]. The results in [1] and in this article propose neural circuit models that aim to bridge this gap, and thereby suggest new means for analyzing data from spike recordings in experimental neuroscience, and for evaluating the more abstract computational models in light of these data. For instance, perceptual multistability in ambiguous visual stimuli and several of its related phenomena were explained through abstract computational models that employ sequential sampling with the Metropolis MCMC algorithm [55]. In our simulations (see Fig. 10) we showed that a spiking neural network can exhibit multistability, where the state changes from one mode of the posterior distribution to another, even though the Markov chain defined by the neural network does not satisfy the detailed balance property (i.e. it is not a reversible Markov chain) like the Metropolis algorithm.
Our models postulate that knowledge is encoded in the brain in the form of probability distributions , that are not required to be of the restricted form of order Boltzmann distributions (5). Furthermore they postulate that these distributions are encoded through synaptic weights and neuronal excitabilities, and possibly also through the strength of dendritic branches. Finally, our approach postulates that these learnt and stored probability distributions are activated through the inherent stochastic dynamics of networks of spiking neurons, using nonlinear features of network motifs and neurons to represent higher order dependencies between RVs. It also predicts that (in contrast to the model of [1]) synaptic connections between neurons are in general not symmetric, because this enables the network to encode higher order factors of .
The postulate that knowledge is stored in the brain in the form of probability distributions, sampled from by the stochastic dynamics of neural circuits, is consistent with the ubiquitous trial-to-trial variability found in experimental data [56], [57]. It has been partially confirmed through more detailed analyses, which show that spontaneous brain activity shows many characteristic features of brain responses to natural external stimuli ([3], [58], [59]). Further analysis of spontaneous activity is needed in order to verify this prediction. Beyond this prediction regarding spontaneous activity, our approach proposes that fluctuating neuronal responses to external stimuli (or internal goals) represent samples from a conditional marginal distribution, that results from entering evidence for a subset of RVs of the stored distribution (see (1)). A verification of this prediction requires an analysis of the distributions of network responses –rather than just averaging –for repeated presentations of the same sensory stimulus or task. Similar analyses of human responses to repeated questions have already been carried out in cognitive science [60]–[62], and have been interpreted as evidence that humans respond to queries by sampling from internally stored probability distributions.
Our resulting model for neural emulations of probabilistic inference predicts, that even strong firing of a single neuron (provided it represents a RV whose value has a strong impact on many other RVs) may drastically change the activity pattern of many other neurons (see the change of network activity after 3 s in Fig. 8, which results from a change in value of the RV that represents “x-ray”). One experimental result of this type had been reported in [63]. Fig. 8 also suggests that different neurons may have drastically different firing rates, where a few neurons fire a lot, and many others fire rarely. This is a consequence both of different marginal probabilities for different RVs, but also of the quite different computational role and dynamics of neurons that represent RVs (“principal neurons”), and auxiliary neurons that support the realization of the NCC, and which are only activated by a very specific activation patterns of other presynaptic neurons. Such strong differences in the firing activity of neurons has already been found in some experimental studies, see [64], [65]. In addition, Fig. 10 predicts that recordings from multiple neurons can typically be partitioned into time intervals, where a different firing pattern dominates during each time interval, see [28], [29] for some related experimental data.
Apart from these more detailed predictions, a central prediction of our model is, that a subset of cortical neurons (the “principal neurons”) represent through their firing activity the current value of different salient RVs. This could be tested, for example, through simultaneous recordings from large numbers of neurons during experiments, where the values of several RVs that are relevant for the subject, and that could potentially be stored in the cortical area from which one records, are changed in a systematic manner.
It might potentially be more difficult to test, which of the concrete implementations of computational preprocessing for satisfying the NCC that we have proposed, are implemented in some neural tissue. Both the underlying theoretical framework and our computer simulations (see Fig. 8) predict that the auxiliary neurons involved in these local computations are rarely active. More specifically, the model predicts that they only become active when some specific set of presynaptic neurons (whose firing state represents the current value of the RVs in ) assumes a specific pattern of firing and non-firing. Implementation 3 and 5 make corresponding predictions for the activity of different dendritic branches of pyramidal neurons, that could potentially be tested through -imaging.
We have proposed a new modelling framework for brain computations, based on probabilistic inference through sampling. We have shown through computer simulations, that stochastic networks of spiking neurons can carry out demanding computational tasks within this modelling framework. This framework predicts specific functional roles for nonlinear computations in network motifs and dendritic computation: they support representation of higher order dependencies between salient random variables. On the micro level this framework proposes that local computational operations of neurons superficially resemble logical operations like AND and OR, but that these atomic computational operations are embedded into a stochastic network dynamics. Our framework proposes that the functional role of this stochastic network dynamics can be understood from the perspective of probabilistic inference through sampling from complex learnt probability distributions, that represent the knowledge base of the brain.
A Markov chain in discrete time is defined by a set of states (we consider for discrete time only the case where has a finite size, denoted by ) together with a transition operator . is a conditional probability distribution for the next state of , given its preceding state . The Markov chain is started in some initial state , and moves through a trajectory of states via iterated application of the stochastic transition operator (more precisely, if is the state at time , then the next state is drawn from the conditional probability distribution . A powerful theorem from probability theory (see e.g. p. 232 in [5]) states that if is irreducible (i.e., any state in can be reached from any other state in in finitely many steps with probability ) and aperiodic (i.e., its state transitions cannot be trapped in deterministic cycles), then the probability was the initial state) converges for to a probability that does not depend on . This state distribution is called the stationary distribution of . The irreducibility of implies that is the only distribution over the states that is invariant under the transition operator , i.e.(15)Thus, in order to generate samples from a given distribution , it suffices to construct an irreducible and aperiodic Markov chain that leaves invariant, i.e., satisfies (15). This Markov chain can then be used to carry out probabilistic inference of posterior distributions of given an evidence for some of the variables in the state . Analogous results hold for Markov chains in continuous time [5], on which we will focus in this article.
We use two types of neurons, a stochastic point neuron model as in [1], and a multi-compartment neuron model.
Let be a probability distribution(23)that contains higher order factors, where is a vector of binary RVs. are the factors that depend on one or two RVs, and are the higher order factors that depend on more than 2 RVs. is the vector of the RVs in the factor , is the vector of RVs that the factor depends on, and is the normalization constant. is the number of first and second order factors, and is the total number of factors of order 3 or higher. To simplify the notation, in the following we set , since this set of factors in will not be changed in the extended probability distribution.
Auxiliary RVs are introduced for each of the higher order factors. Specifically, the higher order relation of factor is represented by a set of auxiliary binary RVs , where we have a RV for each possible assignment to the RVs in ( is the domain of values of the vector ). With the additional sets of RVs we define a probability distribution as(24)We denote the ordered set of indices of the RVs that compose the vector as , i.e.(25)where denotes the number of indices in .
The second order factors are defined as(26)where denotes the component of the assignment to that corresponds to the variable , and is the Kronecker-delta function. The factors represent a constraint that if the auxiliary RV has value 1, then the values of the RVs in the corresponding factor must be equal to the assignment that corresponds to. If all components of are zero, then there is not any constraint on the variables. This implies another property: at most one of the RVs in the vector , the one that corresponds to the state of , can have value 1. Hence, the vector can have two different states. Either all its RVs are zero, or exactly one component is equal to 1, in which case one has . The probability for values of and that do not satisfy these constraints is .
The values of the factors in for various assignments of are represented in by first order factors that depend on a single one of the RVs . For each we have a new factor with value if , and otherwise. We assume that the original factors are first rescaled, such that for all values of and . We had to modify the values of the new factors by subtracting 1 from the original value , because we introduced an additional zero state for that is consistent with any of the possible assignments of .
The resulting probability distribution consists of first and second order factors.
In this neural implementation each principal neuron has a dedicated preprocessing layer of auxiliary neurons with lateral inhibition. All neurons in the network are stochastic point neuron models.
The auxiliary neurons for the principal neuron receive as inputs the outputs of the principal neurons corresponding to all RVs in the Markov blanket of . The number of auxiliary excitatory neurons that connect to the principal neuron is ( is the number of elements of ), and we index these neurons with all possible assignments of values to the RVs in the vector . Thus, for each state of values at the inputs we have a corresponding auxiliary neuron . The realization of the NCC is achieved by a specific connectivity between the inputs and the auxiliary neurons and appropriate values for the intrinsic excitabilities of the auxiliary neurons, such that at each moment in time only the auxiliary neuron corresponding to the current state of the inputs , if it is not inhibited by the lateral inhibition due to a recent spike from another auxiliary neuron, fires with a probability density as demanded by the NCC (3):(31)During the time when the state of the inputs is active, the other auxiliary neurons are either strongly inhibited, or do not receive enough excitatory input to reach a significant firing probability.
The inputs connect to the auxiliary neuron either with a direct strong excitatory connection, or through an inhibitory interneuron that connects to the auxiliary neuron. The inhibitory interneuron fires whenever any of the principal neurons of the RVs that connect to it fires. The auxiliary neuron receives synaptic connections according to the following rule: if the assignment assigns a value of 1 to the RV in the Markov blanket , then the principal neuron connects to the neuron with a strong excitatory synaptic efficacy , whereas if assigns a value of 0 to then the principal neuron connects to the inhibitory interneuron . Thus, whenever fires, the inhibitory interneuron fires and prevents the auxiliary neuron to fire for a time period . We will assume that the synaptic efficacy is much larger than the log-odd ratio value of the RV given according to the r.h.s. of (3). We set the bias of the auxiliary neuron equal to(32)where gives the number of components of the vector that are 1.
If the value of the inputs at time is , and none of the neurons fired in the time interval , then for an auxiliary neuron such that there are two possibilities. Either there exists a component of that is and its corresponding input , in which case the principal neuron of the RV connects to the inhibitory interneuron and inhibits . Or one has in which case the number of active inputs that connect to neuron do not provide enough excitatory input to reach the high threshold for firing. In this case the firing probability of the neuron is(33)and because of the strong synaptic efficacies of the excitatory connections equal to , which are by definition much larger than the log-odd ratio of the RV , it is approximately equal to 0. Hence, only the neuron with has a non-vanishing firing probability equal to (31).
The lateral inhibition between the auxiliary neurons is implemented through a common inhibitory circuit to which they all connect. The role of the lateral inhibition is to enforce the necessary refractory period of after any of the auxiliary neurons fires. When an auxiliary neuron fires, the inhibitory circuit is active during the duration of the EPSP (equal to ), and strongly inhibits the other neurons, preventing them from firing. The auxiliary neurons connect to the principal neuron with an excitatory connection strong enough to drive it to fire a spike whenever any one of them fires. During the time when the state of the input variables satisfies , the firing probability of the auxiliary neuron satisfies the NCC (3). This implies that the principal neuron satisfies the NCC as well.
Introducing an evidence of a known value of a RV in this model is achieved by driving the principal neuron with an external excitatory input to fire a spike train with a high firing rate when the observed value of the RV is 1, or by inhibiting the principal neuron with an external inhibitory input so that it remains silent when the observed value of the RV is 0.
We assume that the principal neuron has a separate dendritic branch for each possible assignment of values to the RVs , and that the principal neurons corresponding to the RVs in the Markov blanket connect to these dendritic branches.
It is well known that synchronous activation of several synapses at one branch, if it exceeds a certain threshold, causes the membrane voltage at the branch to exhibit a sudden jump resulting from a dendritic spike. Furthermore the amplitude of such dendritic spike is subject to plasticity [22]. We use a neuron model according to [23], that is based on these experimental data. The details of this multi-compartment neuron model were presented in the preceding subsection of Methods on Neuron Models. We assume in this model that the contribution of each dendritic branch to the soma membrane voltage is predominantly due to dendritic spikes, and that the passive conductance to the soma can be neglected. Thus, according to (22), the membrane potential at the soma is equal to the sum of the nonlinear active components contributed from each of the branches :(34)where is the nonlinear contribution from branch , and is the strength of branch (see [22] for experimental data on branch strengths). is the target value of the membrane potential in the absence of any synaptic input. The nonlinear active component (dendritic spike) is assumed to be equal to(35)where denotes the Heaviside step function, is the local activation, and is the threshold of branch . The amplitude of the total contribution of branch to the membrane potential at the soma is then .
As can be seen in Fig. 4, the connectivity from the inputs to the dendritic branches is analogous as in Implementation 2 with auxiliary neurons: from each principal neuron such that is in the Markov blanket of there is a direct synaptic connection to the dendritic branch if the assignment assigns to the value , or a connection to the inhibitory interneuron in case assigns the value 0 to . The inhibitory interneuron connects to its corresponding branch , and fires whenever any of the principal neurons that connect to it fire. The synaptic efficacies of the direct synaptic connections are assumed to satisfy the condition(36)where is the set of indices of principal neurons that directly connect to the dendritic branch , is the efficacy of the synaptic connection to the branch from , and is the threshold at the dendritic branch for triggering a dendritic spike. Additionally, each synaptic weight should also satisfy the condition(37)The same condition applies also for the efficacy of the synaptic connection from inhibitory interneuron to the dendritic branch .
These conditions ensure that if the current state of the inputs is , then the dendritic branch will have an active dendritic spike, whereas all other dendritic branches will not receive enough total synaptic input to trigger a dendritic spike. The amplitude of the dendritic spike from branch at the soma is(38)where is a positive constant that is larger than all possible negative values of the log-odd ratio. If the steady value of the membrane potential is equal to , then we have at each moment a membrane potential that is equal to the sum of the amplitude of the nonlinear contribution of the single active dendritic branch and the steady value of the membrane potential, which yields the expression for the NCC (4).
In this implementation a principal neuron has a separate group of auxiliary neurons for each factor that depends on the variable . The group of auxiliary neurons for the factor receives inputs from the principal neurons that correspond to the set of the RVs that factor depends on, but without . For each possible assignment of values to the inputs , there is an auxiliary neuron in the group for the factor , which we will denote with . The neuron spikes immediately when the state of the inputs switches to from another state, i.e. the spike marks the moment of the state change. This can be achieved by setting the bias of the neuron similarly as in (32) to where is the number of components of the vector that are equal to 1, is the efficacy of the direct synaptic connections from the principal neurons to and is a constant that ensures high firing probability of this neuron when the current value of the inputs is .
The connectivity from the auxiliary neurons to the principal neuron keeps the soma membrane voltage of the principal neuron equal to the log-odd ratio of ( = r.h.s. of (4)). From each auxiliary neuron there is one excitatory connection to the principal neuron, terminating at a separate dendritic branch . The efficacy of this synaptic connection is , where is the parameter from (13), and is a constant that shifts all these synaptic efficacies into the positive range.
Additionally, there is an inhibitory interneuron connecting to the same dendritic branch . The inhibitory interneuron receives input from all other auxiliary neurons in the same sub-circuit as the auxiliary neuron , but not from . The purpose of this inhibitory neuron is to shunt the active EPSP when the inputs change their state from to another state . Namely, at the time moment when the inputs change to state , the corresponding auxiliary neuron will fire, and this will cause firing of the inhibitory interneuron . A spike of the inhibitory interneuron should have just a local effect: to shunt the active EPSP caused by the previous state at the dendritic branch . If there is not any active EPSP, this spike of the inhibitory interneuron should not affect the membrane potential at the soma of the principal neuron .
At any time , the group of auxiliary neurons for the factor contributes one EPSP to the principal neuron, through the synaptic input originating from the auxiliary neuron that corresponds to the current state of the inputs . The amplitude of the EPSP from the sub-circuit that corresponds to the factor is equal to . If we assume that the bias of the soma membrane potential is , then the total membrane potential at the soma of the principal neuron is equal to:(39)which is equal to the expression on the r.h.s. of (13) when one assumes that . Hence, the principal neuron satisfies the NCC.
In this implementation each principal neuron is a multi-compartment neuron of the same type as in Implementation 3, with a separate group of dendritic branches for each factor in the probability distribution that depends on . In the group (corresponding to factor ) there is a dendritic branch for each assignment to the variables that the factor depends on (without ). The dendritic branches in group receive synaptic inputs from the principal neurons that correspond to the RVs . Each dendritic branch can contribute a component to the soma membrane voltage (where is like in Implementation 3 the branch strength of this branch), but only if the local activation in the branch exceeds the threshold for triggering a dendritic spike. The connectivity from the principal neurons corresponding to the RVs to the dendritic branches of in the group is such so that at time only the dendritic branch corresponding to the current state of the inputs receives total synaptic input that crosses the local threshold for generating a dendritic spike and initiates a dendritic spike. This is realized with the same connectivity pattern from the inputs to the branches as in Implementation 3 depicted in Fig. 4. The amplitude of the dendritic spike of branch at the soma should be where is the parameter from (13) and is chosen as in Implementation 3.
The membrane voltage at the soma of the principal neuron is then equal to the sum of the dendritic spikes from the active dendritic branches. At time there is exactly one active branch in each group of dendritic branches, the one which corresponds to the current state of the inputs. If we additionally assume that the bias of neuron is , then the membrane voltage at the soma has the desired value (39).
|
10.1371/journal.pgen.1004751 | Notch3 Interactome Analysis Identified WWP2 as a Negative Regulator of Notch3 Signaling in Ovarian Cancer | The Notch3 signaling pathway is thought to play a critical role in cancer development, as evidenced by the Notch3 amplification and rearrangement observed in human cancers. However, the molecular mechanism by which Notch3 signaling contributes to tumorigenesis is largely unknown. In an effort to identify the molecular modulators of the Notch3 signaling pathway, we screened for Notch3-intracellular domain (N3-ICD) interacting proteins using a human proteome microarray. Pathway analysis of the Notch3 interactome demonstrated that ubiquitin C was the molecular hub of the top functional network, suggesting the involvement of ubiquitination in modulating Notch3 signaling. Thereby, we focused on functional characterization of an E3 ubiquitin-protein ligase, WWP2, a top candidate in the Notch3 interactome list. Co-immunoprecipitation experiments showed that WWP2 interacted with N3-ICD but not with intracellular domains from other Notch receptors. Wild-type WWP2 but not ligase-deficient mutant WWP2 increases mono-ubiquitination of the membrane-tethered Notch3 fragment, therefore attenuating Notch3 pathway activity in cancer cells and leading to cell cycle arrest. The mono-ubiquitination by WWP2 may target an endosomal/lysosomal degradation fate for Notch3 as suggested by the fact that the process could be suppressed by the endosomal/lysosomal inhibitor. Analysis of The Cancer Genome Atlas dataset showed that the majority of ovarian carcinomas harbored homozygous or heterozygous deletions in WWP2 locus, and there was an inverse correlation in the expression levels between WWP2 and Notch3 in ovarian carcinomas. Furthermore, ectopic expression of WWP2 decreased tumor development in a mouse xenograft model and suppressed the Notch3-induced phenotypes including increase in cancer stem cell-like cell population and platinum resistance. Taken together, our results provide evidence that WWP2 serves as a tumor suppressor by negatively regulating Notch3 signaling in ovarian cancer.
| Notch pathway is important for many cellular activities, and its dysregulation leads to several diseases in humans, including cancer. Although Notch hyperactivity has been observed in many types of cancers, the interactome of Notch receptor remains largely unknown, especially for Notch3, which is involved in ovarian cancer pathogenesis. This article is the first study, to our knowledge, that delineates the Notch3 interacting network, and demonstrates that one of the Notch3 interacting proteins, WWP2, an E3 ubiquitin-protein ligase, plays a major role in negative regulation of Notch3 signaling in cancer cells. WWP2 locus was found to be deleted, and its mRNA down-regulated in a significant fraction of ovarian carcinomas. Ectopic expression of WWP2 reduced tumorigenicity of ovarian cancer cells, and counteracted Notch3-mediated phenotypes, including promotion of cancer stem-like cell phenotype and platinum resistance, further supporting its tumor suppressor role. The results from this study provide new insights into how Notch3 signaling contributes to cancer development, and should have implications for the design of Notch3-based cancer therapy.
| Notch signaling is a highly conserved cell-cell communication system present in multicellular organisms, and has been shown to be involved in cell fate decision, cell lineage specification, cell proliferation, and survival. Dysregulation of Notch signaling has been known to play a significant role in many diseases including cancer. However, its role as either an oncogene or a tumor suppressor is context- or tissue-type dependent, as both activating mutations and inactivating mutations have been identified in human cancers of different tissue lineages.
In contrast to Drosophila, which has only one type of Notch receptor, mammals contain four types of Notch receptors (Notch1, Notch2, Notch3, and Notch4), perhaps having evolved to support the demands of intricate signaling networks in higher order organisms. The four Notch receptors in mammals share several conserved protein structures which include a large ligand-binding extracellular fragment containing EGF-like repeats, followed by a membrane spanning region and an intracellular fragment containing a RAM domain, tandem ankyrin repeats, and a PEST domain. Notch signaling is initiated by ligand-receptor engagement, which triggers sequential juxta-plasma membrane protein cleavages by ADAM and γ-secretase proteases, leading to the release of intracellular cytoplasmic domain (ICD) fragments. The ICD is then translocated into the nucleus where it acts as a coactivator of RBPJ (CSL)-dependent gene expression. Notch-dependent transcription can be regulated by the quantity of ligand and receptor present on the cell surface, binding affinity of the receptor for its ligand, and protein stability of the ICD.
It has been thought that different Notch receptors share canonical functions but also differentially regulate cellular signaling. Although members of Notch receptor may complement the function of another in some circumstances, each receptor exhibits unique, context-dependent functions. For example, Notch1 receptor knockout causes an embryonically lethality in mice, while Notch3 receptor knockout does not cause the same lethality although the Notch3-null mice harbor minor, yet appreciable, postnatal defects in smooth muscle structure and vascular function [1], [2]. Notch3, but not other Notch receptors, are highly expressed in arterial smooth muscle cells and may regulate differentiation of arterial smooth muscle cells and capillary growth through a paracrine interaction between smooth muscle cells and endothelial cells. Mutations of single amino acid residues at the EGF-like domain of Notch3 affect the functions of vascular smooth muscle cells, leading to apoptosis and degeneration of these cells, and resulting in brain arteriopathy, recurrent strokes, and other symptoms of cerebral autosomal dominant arteriopathy with sub-cortical infarcts and leukoencephalopathy, (CADASIL) [3]. Notch3 gene amplification occurs in ovarian carcinoma, and rearrangement occurs in a subset of malignant glomus tumors [4], [5]. Recent large-scale genomic and expression analysis in ovarian high-grade serous carcinomas (HGSC) by the TCGA consortium has further confirmed the frequent amplification of the Notch3 locus and other aberrations including gene amplification and transcriptional upregulation in members of Notch signaling pathways [6]. Notch3 has been shown to be essential for tumor growth and development of drug resistance [7]. Therefore, targeting the Notch3 pathway represents a promising therapeutic strategy for cancers with dysregulated Notch3 signaling.
Despite the important roles of Notch3 in human diseases, the molecules that mediate or modulate Notch3 signaling remain to be identified. To this end, we have previously performed a systems biology approach using ChIP-chip and transcriptome analyses to identify Notch3 direct target genes in cancer cells [8]. The current study was aimed at identifying Notch3-binding proteins with the presumption that those proteins may modulate Notch3 function. We used a human proteome microarray to screen for N3-ICD protein binding partners and identified an array of Notch3-interacting proteins. Network analysis of the newly identified Notch3 interactome indicated the involvement of ubiquitination in regulating Notch3 signaling; therefore we performed functional studies to characterize an E3-ubiquitin ligase, WWP2, and demonstrated that it is a negative regulator of Notch3 in ovarian cancer.
The purpose of this study was to identify and characterize proteins that interacted with the Notch3 intracellular cytoplasmic domain (N3-ICD). First, we expressed and purified human recombinant N3-ICD protein tagged with the V5 epitope (rhN3-ICD-V5). The rhN3-ICD-V5 protein was then used as a “bait” to screen for binding proteins on the human proteome microarray, comprised of 16,368 individually purified, full-length human proteins [9]. This screen identified a number of N3-ICD-interacting proteins. Among them, RBPJ (CSL), a well-known transcriptional co-factor of the Notch receptors, was at the top of the N3-ICD binding partner list and displayed a high fluorescent binding score (Table S1). This observation indicated that protein microarray is a valid approach to rapidly and reliably identify Notch3 interacting proteins. We next performed Ingenuity Pathway Analysis to search for enriched functional links among the newly identified N3-ICD interactome. The results demonstrated that the enriched functional networks include gene expression, cell cycle, cell signaling, and cellular development (Table S2). As illustrated in Fig. 1A, ubiquitin C (UBC) is a major hub in the top functional network, suggesting the involvement of ubiquitination in regulating and/or mediating Notch3 signaling.
Since NEDD4 family of E3 ubiquitin ligases were previously reported to interact with Drosophila Notch receptors and regulate their signaling activity [10], [11], our discovery of WWP2, an E3 ubiquitin ligase belonging to the NEDD4 family, as a strong Notch3-interacting protein, is particularly interesting (Fig. 1A, Table S1). In the following study, we focused on determining the role of WWP2 in controlling Notch3 ubiquitination and receptor signaling.
To determine if WWP2 was a bona fide N3-ICD interacting protein, we first tested the interaction between rhN3-ICD-V5 and WWP2-GST proteins using an in vitro GST pull-down assay. The results demonstrated that rhN3-ICD-V5 was specifically pulled down by WWP2-GST in a dose-dependent fashion, but was not pulled down by GST-control peptides (Fig. 1B). Since WWP2 is a HECT-domain E3 ligase that regulates ubiquitin-dependent degradation of its substrates, we determined whether WWP2 promoted ubiquitination of Notch3 protein using in vitro ubiquitination assays. Recombinant WWP2-FLAG was incubated with rhN3-ICD-V5, ubiquitin (Ub), and the E1 and E2 enzymes. Ubiquitination of Notch3 protein was determined by immunoprecipitation using an anti-V5 antibody followed by Western blotting with an anti-ubiquitin antibody. We found that WWP2 induced heavy ubiquitination of N3-ICD (Fig. 1C). These data demonstrated that N3-ICD directly bound to WWP2, and that Notch3 was an ubiquitination substrate of WWP2.
To test whether WWP2 specifically bound to N3-ICD but not to other Notch receptors, we performed co-immunoprecipitation assays in HEK293T cells transiently transfected with the WWP2 expression construct along with construct carrying intracellular cytoplasmic domains (NICD) of each of the four human Notch receptors (Notch1–4). We observed that among four NICDs, WWP2 only co-immunoprecipitated with N3-ICD (Fig. 2A). WWP2 belongs to the “WW” domain-containing protein family, which is known to interact with their substrates through the “PPxY” motif [11]. Sequence analysis of the human Notch receptors has shown that the “PPxY” motif is present in the PEST-domain of N3-ICD (presented as PPPY) but is not present in other human Notch receptors [12]. To determine whether the PPPY motif of the N3-ICD was indeed required for the binding between WWP2 and N3-ICD, we generated an N3-ICD PPPA mutant (N3-ICD Y-A) by substituting tyrosine in this motif with alanine. In addition, we created a Notch1-Notch3 chimera by replacing Notch1 TAD and PEST domains with the PEST domain of Notch3 (Fig. 2B). Co-immunoprecipitation experiments were performed in HEK293T cells transiently co-transfected with each mutant construct and WWP2. The results demonstrated that a single point mutation in the PPPY motif of N3-ICD abolished its interaction with WWP2 (Fig. 2C). Furthermore, although N1-ICD cannot interact with WWP2, fusion of the N3-ICD PEST domain with the N1-ICD endowed the newly generated chimeric protein to interact with WWP2 (Fig. 2C). These results indicated that the C-terminal PEST-domain of Notch3 conferred the ability to interact with WWP2.
Reciprocally, we determined if the WW domain of WWP2, which was known to mediate protein-protein interaction, was responsible for interacting with N3-ICD. We created a truncation construct, WWD, which contains the four WW domains of WWP2 (Fig. 2D), and demonstrated that WWD was co-immunoprecipitated with the wild-type N3-ICD but not with the N3-ICD Y-A mutant (Fig. 2E). These experiments confirmed that the PPPY motif in the N3-ICD PEST domain mediated direct interaction with WWP2 via its WW domains.
Notch3 receptor is activated through two juxta-membrane protease cleavages: upon ligand stimulation, the alteration in configuration of membrane tethered Notch3 fragment (N3-TM) triggers the first cleavage by α-secretase of the ADMA family to generate N3-NEXT, which contains N3-ICD plus the transmembrane domain (Fig. 3A). Membrane-tethered N3-NEXT is then cleaved by γ-secretase within the transmembrane region to release the soluble intracellular fragment, N3-ICD. N3-ICD quickly translocates into the nucleus and regulates transcription of target genes through its interaction with co-factors including RBPJ. Although our protein microarray screen which used recombinant protein in a cell-free system suggests that WWP2 directly interacts with the final secretase cleavage product of Notch3, N3-ICD, at the cellular level, WWP2 is likely to interact with and ubiquitinate intermediate Notch3 fragments including N3-TM and N3-NEXT. To assess this possibility, we generated N3-TM-V5 and N3-NEXT-V5 expressing constructs and inserted a signal peptide sequence to the N-terminus of these constructs to ensure correct protein topology of these Notch3 fragments (Fig. 3A). In vivo ubiquitination status of these fragments was examined by performing co-transfection with flag-tagged WWP2 along with HA-tagged ubiquitin plasmids in 293T cells. The expression of each construct was confirmed by western blot, and the results demonstrated that similar amounts of Notch3 fragments were present in each experimental group (Fig. 3B). WWP2 and ubiquitin were also expressed at comparable levels (Fig. 3B). The level of ubiquitination of Notch3 fragments was measured by reciprocal immunoprecipitation with HA and V5 antibodies. The results demonstrated that all tested Notch3 fragments were ubiquitinated in cells co-transfected with WWP2; however, the ubiquitination level of N3-NEXT was most apparent (Top, Fig. 3C and Fig. S1A). There was a prominent band representing mono-ubiquitinated Notch3 protein and a weak high molecular weight smear corresponding to poly-ubiquitinated products. To test the possibility that the differential ubiquitination levels among the Notch3 fragments was due to their ability to encounter and interact with WWP2 in vivo, we performed co-immunoprecipitation experiments in 293T cells following co-transfection of the V5-tagged Notch3 fragment and flag-tagged WWP2. The results demonstrated that although protein expression levels were comparable among the three Notch3 fragments, N3-NEXT protein was more abundantly bound to WWP2 than N3-TM and N3-ICD, reflected by an intense protein pull-down band (Bottom, Fig. 3C). Reciprocal co-immunoprecipitation assays also confirmed the above finding (Fig. S1B).
To determine specificity of Notch3 ubiquitination by WWP2, we generated catalytically inactive WWP2 mutant constructs including C838A and WWP2ΔHECT (Fig. 3D). Flag-tagged wildtype or mutant WWP2 construct was co-transfected with N3-NEXT-V5 and HA-tagged ubiquitin into 293 cells. The expression of each construct was verified by western blot which demonstrated comparable levels of expression in different constructs (Fig. 3E). Co-immunoprecipitation showed a similar binding between N3-NEXT fragment and WWP2 proteins: wildtype WWP2, WWP2ΔHECT mutant and WWP2CA mutant (bottom panel Fig. 3E). In vivo ubiquitination was determined by reciprocal co-immunoprecipitation assays using V5 and HA antibodies. The results showed that N3-NEXT was readily mono-ubiquitinated by wild-type WWP2 but not by the catalytically inactive WWP2 mutants (Fig. 3F).
A previous study demonstrated that Notch3 protein degradation occurs in the endosome/lysosome compartments [13]; therefore, we determined whether Notch3 cleavage fragments accumulated in OVCAR3 and MCF7 cancer cells in the presence of a lysosomal inhibitor. Representative figures were shown in Fig. 4A, demonstrating that we were able to confirm this prior finding. Furthermore, we demonstrated that after lysosomal blockage by NH4Cl, the faster migrating band corresponding to post-secretase cleavage products of Notch3 accumulated in the membrane/cytosol fraction but not in the nuclear fraction (Fig. 4A). As a control, we used EDTA to trigger Notch3 cleavages and generate soluble N3-ICD. In this case, the released N3-ICD was detected exclusively in the nuclear fraction (Fig. 4B). To determine if WWP2 interacted with endogenous Notch3 fragments, we ectopically expressed WWP2-FLAG in OVCAR3 cells because WWP2 expression is relatively low in most tested ovarian cancer cell lines. Then we performed co-immunoprecipitation experiments using an anti-FLAG antibody and an antibody recognizing endogenous Notch3. The cells were assayed in the presence or absence of the lysosomal blocker, NH4Cl. Inhibition of lysosomal degradation increased detectable association of WWP2 with the secretase cleaved Notch3 fragments in the cytoplasmic fraction (Fig. 4C). This is likely attributable to increased levels of secretase cleaved Notch3 fragments after lysosomal blockage. We note that only the faster migrating band corresponding to secretase cleaved products of Notch3 was co-immunoprecipitated with WWP2, whereas the pre-cleaved form of Notch3, TM, was not. We also performed EDTA treatment and subcellular fractionation to test if endogenous N3-ICD interacted with WWP2. Co-immunoprecipitation experiments demonstrated that although abundant N3-ICD fragments were detected in the nuclear fraction, and WWP2 was also present in the nucleus, there was no detectable interaction between them under this condition (Fig 4C).
We also applied immunofluorescence staining to visualize cellular localization of WWP2 and Notch3 after treatment of a lysosomal blocker, NH4Cl. As shown in Fig. S2A, upon NH4Cl treatment, both WWP2 and Notch3 protein fragments co-localized predominantly in the cytoplasm. In contrast, when cells were treated by EDTA (Fig. S2B), Notch3 predominantly localized to the nucleus, most likely because of the induced cleavage by secretases to release N3-ICD (Western blot shown in Fig. 4B). However, under this condition, WWP2 was primarily localized to the cytoplasm, so there is minimal co-localization between WWP2 and N3-ICD. Therefore, the immunofluorescence result was consistent with the above co-immunoprecipitation study.
We have previously demonstrated Notch3 gene amplification and over-expression in ovarian high-grade serous carcinoma (HGSC) [4]. If WWP2 is a negative regulator of Notch3, its expression is expected to be down-regulated in ovarian cancer as compared to normal tissues. To test this hypothesis, we examined the copy number alteration and expression pattern of WWP2 in ovarian HGSCs using a large summarized TCGA dataset [14]. In the 554 ovarian HGSCs that have available copy number data, the majority (77.3%) of HGSCs harbored deletions at the WWP2 locus, including 410 hemizygous and 18 homozygous deletions (Fig. 5A). In addition, based on a GISTIC copy number analysis in ovarian HGSC samples, the WWP2 locus is in a significantly deleted region (q-value = 10−13, Fig. S3) [15]. To define the minimal region of homozygous deletion- a strategy commonly used to identify potential deleted tumor suppressor gene, we aligned the homozygously deleted regions of the 18 HGSC samples and found that the region spans from 69,942,965 bp to 69,976,479 bp on chromosome 16 (hg19), which only encompasses WWP2 gene. To determine if loss of WWP2 DNA copy number affects WWP2 transcription, we determined the relationship between the gDNA and mRNA copy numbers of WWP2 using all HGSCs available in the TCGA. The result demonstrated a significant positive correlation between WWP2 gDNA copy number and WWP2 mRNA expression levels in HGSCs (Fig. S4, Pearson's r = 0.6863, p<0.0001).
As protein ubiquitination plays a major role in regulating the steady state level of protein expression, we determined the relationship between Notch3 protein expression and WWP2 expression using semi-quantitative Western blot analysis. Immunoblotting was performed on ovarian cancer cell lines, ovarian HGSC tissues, primary cultures of ovarian carcinomas, and non-transformed epithelial cells derived from female reproductive organs (endometrium, fallopian tube and ovarian surface), and WWP2 and Notch3 protein expression levels were quantified using a densitometer. As shown in Fig. 5B, an increase in Notch3 expression level and a decrease in WWP2 level were observed in ovarian cancer cell lines as compared to normal non-transformed epithelial cells. The ratio of Notch3 expression level to WWP2 expression level in all samples was shown in Fig. 5C. There was a significantly higher Notch3 to WWP2 protein expression ratio than the normal epithelial cells in ovarian cancer tissues, primary cultures of ovarian cancer cells, and ovarian cancer cell lines. When expression levels of Notch3 was plotted against WWP2 expression levels in ovarian cancer tissues and primary cultures, an inverse correlation in protein expression levels between WWP2 and Notch3 was observed. The result indicated that ovarian HGSC tissues (Fig. 5D) and primary tumor cell cultures (Fig. 5E) with higher Notch3 expression levels tended to have lower expression levels of WWP2, and vice versa.
To determine whether reconstitution of WWP2 expression in cancer cells resulted in reducing Notch3 signaling activity and subsequently cellular proliferation, we performed proliferation assays in OVCAR3 and MCF7 cells transduced with pLPC control plasmid, N3-ICD, WWP2, or catalytically inactive mutant WWP2-C838A (WWP2-CA). Proliferation significantly decreased in both cell lines when WWP2 was ectopically expressed as compared to cells transfected with the control plasmid or cells expressing the WWP2-CA (Fig. 6A). We also determined if N3-ICD could reverse the anti-proliferative effect imposed by WWP2 by co-transfecting cells with N3-ICD and WWP2 expression plasmids. Our data demonstrated that N3-ICD counteracted the growth-inhibitory effect of WWP2, so the growth curve of cells co-transfected with N3-ICD and WWP2 was comparable to the curve of cells transfected with control plasmid (Fig. 6A). Using a Notch signaling reporter assay, we also observed reduced endogenous Notch signaling activity when WWP2 was expressed in MCF7 and OVCAR3 cells (Fig. 6B). Expression of the enzymatically inactive mutant WWP2-C838A in MCF7 and OVCAR3 cells also reduced Notch signaling activity, however to a lesser extent. To confirm the phenotypes observed in cancer cell lines, we performed experiments in primary ovarian tumor cultures. WWP2 cDNA or control vector was co-transfected with a Notch signaling reporter, pJH23A, into primary cell cultures and Notch signaling activity of WWP2 cDNA -transfected group was measured and the data was normalized to the data obtained from the control vector-transfected group. The normalized data were plotted against Notch3 expression levels measured in the same tumor samples (Fig. S5A). The result demonstrated that ectopic expression of WWP2 potently suppressed Notch signaling in cells with high levels of Notch3 (reflecting by greater reduction of Notch signaling activity in the Notch3-high cells) (Fig. S5B).
Next, we employed a loss-off-function approach to assess the contribution of WWP2 on Notch signaling. Expression of WWP2 in untransformed cell lines including OSE4 (an ovarian surface epithelial cell line) and FT2821 (a fallopian tube epithelial cell line) was down-regulated by using siRNAs. The knockdown efficiency of two different WWP2 targeting siRNAs was validated by Western blot (Fig. S6). To detect the effect on Notch signaling, cells were co-transfected with a Notch signaling reporter, pJH23A, or with a control vector pGL3. The data demonstrated that downregulation of WWP2 by siRNAs significantly increased Notch signaling activity in both untransformed cell lines (Fig. 6C).
To determine whether WWP2 expression affects tumor development, we ectopically expressed WWP2 in SKOV3-CR cells which were pre-established to develop carboplatin resistance (CR) phenotype and expressed abundant Notch3. SKOV3-CR cells transfected with the empty vector, pLPC, served as a control. The presence or absence of ectopic WWP2 expression in these transfected cells was validated by Western blot (Fig. 7A). A total of 3×106 cells were injected subcutaneously into the athymic nu/nu mice and tumor growth was monitored every three days. We observed that WWP2 expression led to a significant reduction in tumor formation and in fact as compared to the control group, tumors in the WWP2 expressing group were hardly detectable (Fig. 7B). Tumor weight measured at the endpoint of the study was also significantly lower in the WWP2-expressing group than in the control group (Fig. 7C).
Inhibition of Notch3 by siRNA was previously shown to lead to cell cycle arrest in the G2/M phase in OVCAR3 cells [8] and to cause G0/G1 arrest in MCF7 cells [16]. To determine if WWP2 expression leads to similar cell cycle arrests, WWP2 expression plasmid or pLPC control plasmid was transfected into OVCAR3 and MCF7 cells. Results of cell cycle analysis on these cells were consistent with results from Notch3 knockdown. Overexpression of WWP2 in OVCAR3 cells led to a significantly increased population of cells within the G2/M phase and a simultaneous loss of cells within the G0/G1 population (Fig. S7). Overexpression of WWP2 in MCF7 cells led to increased cell numbers within the G0/G1 phase and reduced numbers within G2/M, demonstrating the functionally diverse outcomes in different cellular backgrounds.
Previous studies by our group and others have shown that Notch3 upregulation is related to the recurrence of ovarian cancer and is associated with a poor prognosis [7]. In addition, Notch 3 overexpression increases the proportion of cancer stem cell-like cells (CSCs) and resistance to platinum-based therapy [7], [17]. To determine whether WWP2 can counteract these Notch3-mediated phenotypes, we performed flow cytometry and Hoechst 33342 dye efflux assay to measured side population which is enriched with CSCs. Ovarian cancer cells were transfected with empty vector, N3-NEXT, WWP2, or N3-NEXT+WWP2. The flow cytometry results demonstrated that overexpression of Notch3 increased the side population fraction but the percentage of cells in this population was reduced when WWP2 was also expressed. Thus, the presence of WWP2 counteracts Notch effect on promoting the CSC phenotype (Fig. 8). Similarly, expression of WWP2 decreased the Notch3-induced platinum resistance as evidenced by analyzing cell viability in OVCAR3 cells cultured with carboplatin (Fig. S8).
It has been established that Notch3 plays a fundamental role in a variety of cellular functions including cellular differentiation, organ development, and cancer pathogenesis. However, the mechanisms that regulate and mediate Notch3 function remain largely unknown. To address the issue, this study aimed to identify Notch3 interacting proteins using an unbiased, comprehensive protein chip-based approach. In this screen, the canonical Notch receptor co-factor, RBPJ, exhibited high binding affinity to Notch3 receptor, indicating the efficiency of the approach. In addition to RBPJ, a list of potential Notch3 receptor-interacting proteins was identified. These proteins are involved in functional networks including protein ubiquitination, cell signaling, and transcriptional regulation. Their identification provides new insight into the Notch3 signaling network. We chose one of the Notch3 interacting proteins, WWP2, in the ubiquitination pathway, for further investigation.
Protein ubiquitination is a posttranslational modification of a specific protein that leads to a unique fate through protein trafficking. For example, mono-ubiquitination of membrane proteins triggers their endocytosis and targeting to endosomes/lysosomes [18]; whereas K48-linked polyubiquitination is a signal for targeting cytosol proteins to the proteasome for degradation [18]. Genetic studies in Drosophila have demonstrated that the Nedd4 family of ubiquitin ligases, including Suppressor of Deltex and Nedd4, negatively regulate Notch receptor signaling [10]. In mammals, a homolog of the Nedd4 family, Itch/AIP4, was found to ubiquitinate membrane-bound Notch1 receptor (N1-TM) prior to its activation, and to target its endocytosis and endosomal trafficking [12], [19]. More recently, it has been demonstrated that in contrast to Drosophila Nedd4, mammalian Itch does not interact directly with Notch1 receptor. Its ability to negatively regulate Notch signaling is thought to be through interaction with the adaptor protein, Numb [20]. This result is not surprising because unlike Drosophila Notch or mammalian Notch3, mammalian Notch1 lacks the PPxY motif in the C-terminal region which is critical for direct interaction with E3 ubiquitin ligases in the Nedd4 family [12]. Compared to the membrane-tethered form, the soluble active form of Notch1, N1-ICD, has been shown to be directly polyubiquitinated by Sel-10 (FBXW7), a constituent of the SCF ubiquitin protein ligase complex, and was subsequently targeted to proteasome for degradation [21].
These prior studies demonstrated that Notch1 activity can be regulated by multiple ubiquitination-mediated mechanisms. Comparatively little is known about regulation of other Notch receptors such as Notch3. In this study, using comprehensive proteome microarray, we have identified WWP2, a NEDD4 family member of E3 ubiquitin ligase, as a negative regulator of Notch3 signaling in ovarian cancer. We have shown that WWP2 directly interacts with and mono-ubiquitinates post-secretase cleaved Notch3 fragments, promoting their sorting to and degradation in lysosomes, thereby suppressing Notch3 signaling activity in cancer cells. Among the four Notch receptors present in mammals, only Notch3 contains the PPxY motif within its N3-ICD PEST domain. The PPxY motif is evolutionary conserved from Drosophila to mammals and is responsible for direct interaction with the WW domain of WWP2. The PPxY motif in Drosophila Notch was shown to be responsible for direct interaction with Drosophila Nedd4 and targeting the receptor for endocytosis and endosomal sorting [11]. In this study, we found that WWP2 promotes a strong mono-ubiquitination pattern of post-α-secretase cleaved, membrane tethered fragments of Notch3, N3-NEXT. This ubiquitination pattern has been suggested to serve as the “seed” for endocytic protein interaction networks. Ubiquitination occurred to a lesser extent in the “resting form” N3-TM or the cytosolic soluble form, N3-ICD. We speculate that the favorable ubiquitination towards N3-NEXT is triggered by juxta-membrane α-secretase cleavage which leads to a conformational alteration in the Notch3 fragments and increases their accessibility to WWP2.
A recent study has described endocytosis of the Notch3 receptor and subsequent lysosomal degradation of both extracellular and intracellular fragments of Notch3 [13]. Accumulation of N3-ECD (extracellular fragments) as well as N3-ICD (intracellular fragments) was observed in the presence of the lysosomal inhibitors but not in the presence of the proteasome inhibitor. In contrast to its effect on Notch3, proteasome inhibitor leads to accumulation of intracellular fragments of Notch1 (N1-ICD). In this study, we have confirmed the above findings in an ovarian cancer cell line, OVCAR3, and in a breast cancer cell line, MCF7, both of which express abundant Notch3. In addition to Notch receptors, OVCAR3 and MCF7 also express high levels of Notch ligands including Jagged1 and Delta4, respectively, which may provide a constitutive cue for Notch3 signal activation. Treatment with NH4Cl, which suppresses lysosomal degradation of Notch3 fragments, significantly increased the interaction between WWP2 and post-secretase cleaved fragments of Notch3, suggesting that WWP2 regulates targeting of Notch3 to the endosome/lysosome compartment where the receptor is degraded. Since WWP2 contains a C2 domain primarily found in proteins regulating membrane trafficking and mediating binding to phospholipids, raising a possibility that the primary subcellular interaction site between WWP2 and Notch3 is located at juxta-plasma membrane region.
In addition to demonstrating WWP2 as a regulator for Notch3 trafficking and signaling activity, this study has established a tumor suppressor role of WWP2 in ovarian cancer. We reported that WWP2 is genetically deleted and down-regulated in a significant number of ovarian high-grade serous carcinomas and overexpression of WWP2 suppresses tumor development and causes cell cycle arrest. Furthermore, we demonstrated that WWP2 expression counteracts the Notch3 promoted phenotypes including CSCs and platinum resistance. The negative regulation of WWP2 on Notch signaling activity is more prominent in Notch3-overexpressing cancer cells. This could be because those cancer cells are more dependent on Notch signaling for maintaining cellular survival, promoting CSC population and developing platinum resistance. Since WWP2 targets Notch3 and promotes its degradation, enforced expression of WWP2 is expected to result in a more prominent anti-Notch effect in cells expressing abundant Notch3 (or Notch3-dependent cells) than in cells with low or minimal Notch3 expression (or Notch3-indepdenent cells). The reported negative regulatory function of WWP2 on Notch3 signaling could be exploited for designing new strategies to target Notch3 signaling pathway. For example, reagents upregulating the expression of WWP2 and gene therapy approach by re-introducing WWP2 into cancer cells are worth pursuing for the treatment of cancers that have developed dependence on Notch3 signaling.
The human proteome microarray was fabricated by spotting 16,368 unique, full-length human recombinant proteins in duplicate along with control proteins including IgG, GST, and histones as previously described [9]. Recombinant human N3-ICD-V5 (Met1663-Ala2321) purified from E. coli was previously reported [8]. Protein-binding assays on the human proteome microarrays were performed using protocols reported previously [9]. In brief, the microarray was first blocked with 2% BSA in 1× PBS for 2 hr at room temperature, incubated with purified N3-ICD proteins for 1 hr at room temperature, and washed in TBST, followed by incubation with mouse anti-V5 antibody (Invitrogen) at 1∶5000, and then incubated with 1∶1000 Alexafluor 647 conjugated goat anti-mouse antibody (Invitrogen) for detection. After drying, the protein microarray was scanned using a GenePix 4000B scanner (Molecular Devices, Sunnyvale, CA) and signal intensity was calculated as the ratio of median foreground and median background signals in the Cy5 channel.
To quantify the affinity of N3-ICD to each interacting protein on the microarray, we first calculated the mean and standard deviation of the signal intensity across all spots on the microarray. We obtained a normalized signal intensity score for any given protein spot using ratio of median foreground and median background fluorescence as described [22]. Positive interaction spots found in one unique row pool and one column pool that showed intensity score greater than 2.8 for both duplicate spots of any given protein were flagged for individual analysis. Table S1 lists the N3-ICD-interacting proteins with intensity scores greater than 7.0 that were positive in duplicate spots.
The cell lines used in this study included ovarian cancer cell lines OVCAR3, MPSC1, and SKOV3-CR, an immortalized ovarian surface epithelial cell line OSE4, an immortalized fallopian tube epithelial cell line FT2821, a breast cancer cell line MCF7, and an embryonic kidney epithelial cell line HEK293. SKOV3-CR was developed by incubating parental SKOV3 with low dosage (10 µM) of carboplatin for 3 months. As a result SKOV3-CR became resistant to carboplatin. OVCAR3, MPSC1, SKOV3-CR, OSE4 and FT2821 were maintained in RPMI1640 supplemented with 5% fetal bovine serum and antibiotics. MCF7 and HEK293 cells were maintained in DMEM supplemented with 5% fetal bovine serum and antibiotics. There was no evidence of mycoplasma contamination based on a PCR assay.
Expression plasmids were transfected into HEK293, OVCAR3, or MCF7 cells using Lipofectamine 2000 (Invitrogen). Cells were harvested 24 hr post-transfection for co-immunoprecipitation experiments. Cells were lysed in lysis buffer (50 mM Tris, pH 8.0, 150 mM NaCl, 1% NP40 supplemented with protease inhibitor cocktail (Thermo Scientific)). Lysates were then incubated with either anti-V5 agarose (Sigma, St. Louis, MO), or anti-FLAG M2 affinity gel (Sigma, St. Louis, MO). After five washes (3 times with lysis buffer and 2 times with TBS) precipitates were resuspended in Laemmli sample buffer containing 5% β-mercaptoethanol.
Samples were separated by 6% or 4–15% SDS-PAGE (Bio-Rad), and were transferred onto a PVDF membrane (Amersham) using a semi-dry transfer apparatus (Bio-Rad). The membrane was blocked with 5% non-fat dry milk (Bio-Rad) or 3% BSA (Sigma) in TBST (20 mM Tris-HCl, 0.5 M NaCl, 0.1% Tween 20) and incubated with a primary antibody, followed by washes with TBST. Subsequently, the membrane was incubated with horseradish peroxidase-conjugated secondary antibody (Jackson Laboratories, West Grove, PA) and detected with ECL developing solution (Thermo Scientific). Western blot was performed using antibodies as indicated.
To correlate protein expression of WWP2 and Notch3, we analyzed 51 ovarian HGSC tissues using Western blot analysis with specific antibodies. Normal counterparts of ovarian cancers including primary cultures of human endometrial epithelial cells (EMs) and fallopian tube epithelial cells (FTs)as well as ovarian surface epithelial cell lines (OSE4, OSE7, and OSE10) were included as controls. Cancer tissues and cells were lysed with lysis buffer (50 mM Tris, pH 8.0, 150 mM NaCl, 1% NP40), supplemented with protease inhibitor cocktail (Thermo Scientific). Cell lysates were subsequently subjected to Western blot analysis. The intensities of WWP2, Notch3, and GAPDH were measured using the ChemiDoc XRS and Image Lab software (Bio-Rad). The intensity of WWP2 or Notch3 was then normalized to the intensity of GAPDH and relative expression value was calculated using the following formula: Δ(IntA or B/IntGAPDH)/highest(IntA or B/IntGAPDH)×100.
The human WWP2 full length clone (ID:5588092) was purchased from Open Biosystems in pCMV-sport6, PCR amplified with Pfu Ultra II polymerase (Agilent) according to the manufacturer's protocol and cloned with EcoRI and XhoI into the pLPC-N-FLAG plasmid (purchased from Addgene). A construct coding for the four WW domains (WWD) was PCR amplified from full length WWP2 (Ala 274- Gly 478 (820 bp–1434 bp)) and cloned with EcoRI and XhoI into the pLPC-N-FLAG plasmid. WWP2 ΔHECT, coding for the CA2 domain and the first WW-repeat was purchased from Invitrogen (ultimate ORF clone IOH 4735) and cloned into pLPC-N-FLAG plasmid with the Gateway LR clonase II enzyme mix (Invitrogen). WWP2 C838A was generated by site directed mutagenesis (TG 2512–2513>GC (Cys>Ala)) using a kit from Agilent according to the manufacturer's protocol.
The N3-ICD construct has been described previously [7]. Notch1-ICD (5260–7665 (Val 1754-Lys 2555)) and Notch2-ICD (5260–7665 (Val 1754-Lys 2555)) were PCR amplified with Pfu Ultra II polymerase from cDNA transcribed from total RNA derived from SKOV3 and MCF7 cells, respectively, and cloned with NotI and BamHI into pCDNA6-A V5 HIS (Invitrogen). The Notch4 full length ORF clone 9021650 was purchased from Open Biosystems. Notch4-ICD (4399–6009 (Val 1467-Lys 2003)) was PCR amplified and cloned into pCDNA6-A V5 HIS. A Notch1/Notch3 chimera was developed by amplifying the RAM domain and ANK repeats of Notch1 (5260 bp–6342 bp) and the C-terminal PEST domain of Notch3 (6076 bp–6963 bp) with overlapping primers followed by a fusion PCR using the Notch1 forward primer and the Notch3 reverse primer. The fusion fragment was then cloned into the pCDAN-6-V5 expression plasmid with NotI and BamHI.
The human Notch3 full length clone was a kind gift from Dr. Michael Wang (University of Michigan). PCR amplification of N3-NEXT and N3-TM were performed with Pfu Ultra II polymerase (Agilent) and cloned into the pcDNA6-V5 plasmid (Invitrogen) with the addition of a V5 tag in the constructs. To add signal peptide to these constructs, PCR amplified N3-NEXT and N3-TM products were inserted in frame into the plasmid, pSF-CMV-NH2-InsulinSP1 (purchased from Oxford Genetics, UK). The final fusion products included the V5 tag from pcDNA6 and signal peptide from insulin. All constructs were confirmed by sequencing (Macrogen USA). The sequence of cloning primers is available upon request.
HEK293T cells were transfected with various combinations of Notch3 and WWP2 plasmids together with HA-tagged ubiquitin. Cells were harvested 24 hr after transfection and lysed in lysis buffer (50 mM Tris, 150 mM NaCl, 1% NP-40 with complete inhibitor (Roche)). Lysates were subjected to immunoprecipitation with anti-HA, anti-FLAG, or anti-V5 beads. Ubiquitination was analyzed by immunoblotting of the anti-HA precipitates with anti-V5 antibody (Bethyl).
The reactions were carried out at 37°C for 45 min in 25 µl of ubiquitination reaction buffer (40 mM Tris-HCl at pH 7.6, 2 mM DTT, 5 mM MgCl2, 0.1 M NaCl, 2 mM ATP) containing the following components: 250 µM ubiquitin, 100 nM E1 (UBE1), 200 nM UbcH5b (all from Boston Biochem). Purified FLAG-WWP2 was added to the reaction. rhN3 (0.5 µg) was used as a substrate in the reactions as indicated. After the ubiquitination reaction, lysis buffer was added, and samples were incubated with V5-agarose beads (Sigma) for 2 hr. The beads were washed five times with lysis buffer and TBS and boiled in SDS–PAGE loading buffer. Ubiquitination of rhN3 was monitored by Western blotting with anti-ubiquitin antibody (Cell Signaling). An aliquot of each reaction was used directly for Western blotting to demonstrate equal amounts of substrate and/or E3 ligase in the samples as indicated.
Bacterial-expressed GST–WWP2 or control GST bound to glutathione–Sepharose beads (Amersham) was incubated with rhN3-ICD for 1 hr at 4°C. The washed complexes were eluted by boiling in SDS sample buffer and separated by SDS–PAGE, and the interactions were analyzed by Western blotting. An aliquot of flow through was used to confirm depletion of rhN3 in presence of WWP2.
Separation of cell nuclei and membrane/cytosol was performed as previously described (Lee and Green, 1990) with the following modifications: Cells were scraped off and washed twice with cold PBS and once with 20 ml of buffer A (10 mM Tris-HCl, pH 7.4, 8.3 mM KCl, 1.5 mM MgSO4, 1.3 mM NaCl). Cells were then swollen on ice for 30 min in buffer A. Nuclei/membranes and cytosol were separated by passing the suspension eight times through a 23-gauge needle followed by 20 rounds through a glass-glass homogenizer. Nuclei and membranes were pelleted by centrifugation at 3000 g for 10 min. Supernatant (cytosolic fraction) was cleared by centrifugation at 10,000 g.
Nuclei and membranes were re-suspended in 1 ml of buffer B (buffer A supplemented with 0.5% NP-40 and 1 mM PMSF) and separated by passing the suspension again eight times through a 23-gauge needle followed by 20 rounds through a glass-glass homogenizer. The homogenate was centrifuged for 10 min at 900 g to pellet the nuclei. The supernatant (membrane fraction) was transferred to a fresh microfuge tube and again cleared by centrifugation at 10000 g for 15 min. The nuclear pellet was re-suspended in 1 ml of buffer C (buffer A containing 1 mM PMSF), and purity of nuclei was verified under a microscope. Resuspended nuclei were centrifuged at 1000 g for 10 min. The pellet was re-suspended in lysis buffer, sonicated briefly, and used for IP or prepared for Western blotting by addition of SDS-PAGE sample buffer. Alternatively, the nuclear enrichment kit from PIRCE was used to separate the nuclear fraction from membrane/cytosolic fraction according to manufacturer's protocol.
For cellular growth assay, cells stably transduced (with retroviral constructs) or transiently transfected (with Lipofectamine 2000; Invitrogen) with constructs as indicated were plated in 96-well plates at 5×103 cells/well. The cell number was determined 48 hr after plating based on the fluorescence intensity of SYBR green I nucleic acid staining (Molecular Probes, Eugene, OR) measured in a fluorescence microplate reader (Fluostar from BMG, Durham, NC). The data were expressed as mean ± SD from triplicate assays.
Cell cycle assay was performed using Guava cell cycle kit according to the manufacturer's protocol 48 hr after transient transfection or stable transduction of cell lines. Samples were analyzed in a Muse cell analyzer (EMD Millipore).
Cells were plated in 24-well plates and were transiently transfected (using Lipofectamine 2000) with 50 ng of pRL-Renilla, 0.5 µg of the Notch reporter pJH23A (4×wtCBF1Luc) luciferase construct (provided by D. Hayward, Johns Hopkins University), and 0.25 µg of various plasmid combinations as indicated in individual experiments. Assay analysis was carried out 24 h after transfection using the Dual-Glo Luciferase Reporter Assay (Promega) according to the manufacturer's protocol. Expression was normalized to pRL-Renilla Luciferase. The data were expressed as mean ± SD of triplicate assays.
Level 3 TCGA data on ovarian HGSC were retrieved from Broad Institute's Genome Data Analysis Center (http://gdac.broadinstitute.org/runs/stddata__2014_01_15/data/OV/20140115/). Data used include transcriptome on Affymetrix U133a platform (http://gdac.broadinstitute.org/runs/stddata__2014_01_15/data/OV/20140115/gdac.broadinstitute.org_OV.Merge_transcriptome__ht_hg_u133a__broad_mit_edu__Level_3__gene_rma__data.Level_3.2014011500.0.0.tar.gz) and somatic copy number alterations on Affymetrix SNP 6 platform (http://gdac.broadinstitute.org/runs/stddata__2014_01_15/data/OV/20140115/gdac.broadinstitute.org_OV.Merge_snp__genome_wide_snp_6__broad_mit_edu__Level_3__segmented_scna_minus_germline_cnv_hg19__seg.Level_3.2014011500.0.0.tar.gz).
Ovarian cancer cells were transfected with pLPC, pLPC-WWP2, SP-N3-NEXT or pLPC-WWP2+SP-N3-NEXT for 3 days, and cells were detached by trypsin, washed with PBS, then re-suspended in RPMI culture medium (supplemented with 2% FBS). The concentration of cells were adjusted to 1×106 cells/ml and Hoechst 33342 (Sigma) was added at a final concentration of 5.0 µg/mL either alone or in the presence of 50 µg/mL verapamil. After incubation at 37°C for 90 min, cells were washed with ice cold PBS and then subjected to flow cytometry analysis. Hoechst 33342 was excited with UV laser at 350 nm and fluorescence emission was measured with 405/BP30 and 570/BP20 optical filters for Hoechst blue and Hoechst red emission wavelengths, respectively. A 550-nm long-pass dichroic mirror (Omega Optical Inc., Brattleboro, VT) was used to separate the emission wavelengths.
To visualize the subcellular localization of WWP2 and Notch3, MCF7 cells were seeded on gelatin-coated coverslips in a 6-well plate and treated with 25 mM NH4Cl for 2 hr and 4 hr or with 2.5 mM EDTA for 30 min. Cells were fixed with 3.7% paraformaldehyde/PBS for 10 min, permeabilized with 0.5% triton X-100/PBS for 5 min, and blocked with 1% normal goat serum for 30 min. Cells were incubated with an anti-Notch3 antibody (1∶100 dilution; Santa Cruz Biotechnology) followed by incubation with an anti-rabbit Cy3-conjugated antibody (1∶500 dilution; Sigma). After washing with PBS, cells were incubated with anti-WWP2 antibody (1∶100 dilution; Santa Cruz Biotechnology) and followed by anti-mouse Alex488-conjugated antibody (1∶500; Cell Signaling) for 1 hr. DNA was counterstained with DAPI (Invitrogen) for 1 min. Fluorescence images were acquired with a fluorescence microscope (TE200, Nikon).
WWP2-specific small interfering RNAs (siRNAs) were purchased from Invitrogen. The target sequences of WWP2 siRNAs are: UAGACACGUCCGUUGGGCAGCUCUC and ACACGGGCUUCACCCUCCCUUUCUA. OSE4 and FT2821 cells were transfected with siRNAs at a final concentration of 100 nM using Lipofectamine RNAiMAX (Invitrogen) according to the manufacturer's protocol. Twenty-four hours after, cells were further transfected with 1 µg of Notch reporter plasmid, pJH23A (4×wtCBF1Luc), and 50 ng of pRL-Renilla for measuring the Notch signaling activity.
To determine whether WWP2 may act as a tumor suppressor, tumorigenic carboplatin-resistant SKOV3 cells (SKOV3-CR) were transiently transfected with a plasmid encoding WWP2 cDNA or a control plasmid, pLPC. Six-week-old female athymic nu/nu mice were subcutaneously inoculated with 3×106 tumor cells. Ten mice were included in each experimental group. Tumor size was measured every three days using a caliper and at the end of the study, tumors were carefully excised and weighted. Tumor volume based on caliper measurements were calculated by the modified ellipsoidal formula: Tumor volume = 1/2 (length×width×height).
|
10.1371/journal.pcbi.1004926 | Novel Computational Protocols for Functionally Classifying and Characterising Serine Beta-Lactamases | Beta-lactamases represent the main bacterial mechanism of resistance to beta-lactam antibiotics and are a significant challenge to modern medicine. We have developed an automated classification and analysis protocol that exploits structure- and sequence-based approaches and which allows us to propose a grouping of serine beta-lactamases that more consistently captures and rationalizes the existing three classification schemes: Classes, (A, C and D, which vary in their implementation of the mechanism of action); Types (that largely reflect evolutionary distance measured by sequence similarity); and Variant groups (which largely correspond with the Bush-Jacoby clinical groups). Our analysis platform exploits a suite of in-house and public tools to identify Functional Determinants (FDs), i.e. residue sites, responsible for conferring different phenotypes between different classes, different types and different variants. We focused on Class A beta-lactamases, the most highly populated and clinically relevant class, to identify FDs implicated in the distinct phenotypes associated with different Class A Types and Variants. We show that our FunFHMMer method can separate the known beta-lactamase classes and identify those positions likely to be responsible for the different implementations of the mechanism of action in these enzymes. Two novel algorithms, ASSP and SSPA, allow detection of FD sites likely to contribute to the broadening of the substrate profiles. Using our approaches, we recognise 151 Class A types in UniProt. Finally, we used our beta-lactamase FunFams and ASSP profiles to detect 4 novel Class A types in microbiome samples. Our platforms have been validated by literature studies, in silico analysis and some targeted experimental verification. Although developed for the serine beta-lactamases they could be used to classify and analyse any diverse protein superfamily where sub-families have diverged over both long and short evolutionary timescales.
| Beta-lactamases are bacterial proteins largely responsible for resistance to beta-lactam antibiotics and so pose a significant challenge to modern medicine. Whilst there are many studies cataloguing beta-lactamases, antibiotic screening has not always been consistent or comprehensive, causing confusion in the classification of these proteins and difficulty in recognising bacteria with different resistance profiles. We therefore developed strategies for automatically and consistently classifying distinct classes and types of beta-lactamases, having particular antibiotic resistance profiles. Our methods focus mainly on the sequences of the beta-lactamases, as for most new bacterial strains we will only know the sequence. We have classified all sequenced beta-lactamases stored in major public repositories into classes. We then mainly focus on the Class A beta-lactamases as these are responsible for most of the resistance to clinically relevant antibiotics. We applied methods to pinpoint key sequence sites where changes result in new antibiotic resistance properties. Understanding which sites confer resistance is important for recognizing whether new evolving strains can evade current antibiotic regimes. Our classification methods allowed us to classify 151 Class A serine beta-lactamase types and to recognize a new type of Class A beta-lactamase in a bacteria found in a drain sample.
| In this article we demonstrate the value of different clustering and analysis platforms for classifying an important superfamily of bacterial proteins, the beta-lactamases. Our approaches are based largely on the sequence properties of the relatives although structural information is considered for some analyses. The purpose of the classification was to aid the identification of functional determinants (FDs), i.e. residue sites influencing the functional properties of the relatives, where these properties relate to implementation of the catalytic mechanism or substrate profiles. In particular, we aimed to show that identification of these sites could aid in the prediction of phenotype for newly determined relatives not yet experimentally characterised.
Beta-lactamases represent the main bacterial mechanism of resistance to beta-lactam antibiotics and are a significant challenge to modern medicine. Beta-lactam antibiotics are characterised by the possession of a four-atom beta-lactam ring, as shown in red in the main categories of antibiotics (penicillins, cephalosporins, carbapenems and monobactams) in Fig 1. Beta-lactamases catalyse the hydrolysis of the bond between the nitrogen atom and the carbonyl group of the beta-lactam ring, breaking the ring open and thus inactivating the antibiotic. There is a large pool of naturally occurring beta-lactamases in environments such as the human gut that are selected for, mutated and transmitted horizontally into pathogenic bacteria following the introduction of new antibiotics [1].
All beta-lactamases are assigned the Enzyme Commission (EC) number 3.5.2.6 which is shorthand for “a member of the hydrolases, acting on carbon-nitrogen bonds, other than peptide bonds, in cyclic amides”. The EC functional classification scheme does not extend to more specific distinctions than this. The Gene Ontology (GO) [2] molecular function ontology term GO:0008800 represents “beta-lactamase activity” which is further subdivided into GO:0033250 “penicillinase activity” and GO:0033251 “cephalosporinase activity”. Both terms refer to activity against a broad range of chemically distinct antibiotics (i.e. having different “R-groups”) based on the penicillin and cephalosporin core structures shown in Fig 1a and 1c, which also includes ampicillin to illustrate an example penicillin “R-group” (Fig 1b). There are also other beta-lactam antibiotic core structures, such as that possessed by carbapenems which are commonly reserved as antibiotics of last resort to combat multi-resistant bacteria (see Fig 1d). The recent spread of carbapenemases, such as the New Delhi metallo-beta-lactamase NDM-1 is a cause for some alarm [3]. A frequently used term in the scientific literature, “broad spectrum” indicates that penicillins and cephalosporins are inactivated at the same rate, while the term “extended-spectrum” indicates the ability to inactivate third-generation cephalosporins with an oxyimino side chain as well as monobactams (see Fig 1e). Inhibitors such as clavulanic acid inhibit the activity of some beta-lactamases and are often used in treatments in conjunction with beta-lactam antibiotics.
An early classification of beta-lactamases by Ambler [4], based on sequence comparison and preliminary structural data grouped beta-lactamases into classes A and B. A class A structure (PDB 1BTL) was experimentally determined in 1987, providing structural evidence for the involvement of a key catalytic serine residue in the hydrolysis reaction [5]. In 1995, the first class B structure was experimentally determined (PDB 1BMC), which represented a new type of active site zinc-binding protein fold. Based on differences in sequence motifs, classes C and D have subsequently been added and revealed to possess the same protein fold and the same catalytic serine as the class A beta-lactamases.
The single domain serine beta-lactamases (Classes A, C and D) are revealed by structural and catalytic residue similarity to be closely related to the beta-lactam antibiotic targets, the DD-peptidases (also known as DD-transpeptidases). The serine beta-lactamases are thought to have evolved from the DD-peptidases about 2 billion years ago after fungi evolved the ability to synthesize beta-lactam antibiotics [6]. The DD-peptidases are involved in cross-linking bacterial cell walls, which is essential to their survival. The metallo-beta-lactamases (Class B) are a group of enzymes that are structurally unrelated to serine beta-lactamases and appear to have evolved independently of DD-peptidases [7].
Singh et al. [8] report a graph-based clustering of best bi-directional hits (generated using BLASTP) of beta-lactamase sequences that reproduces the four classes proposed by Ambler (A, B, C and D). They also suggest the possibility of two additional small groups that they classify as E and F, which seem to be more closely related to class B metallo-beta-lactamases than to the serine beta-lactamases. An online database “Dlact” is also reported but this does not seem to be available at the time of writing. Two other online databases do provide some limited information about beta-lactamase antibiotic resistance specificity: the ARDB Antibiotic Resistance Genes Database (http://ardb.cbcb.umd.edu/) [9] and the Beta-LActamase Database, BLAD (http://www.blad.co.in) [10].
Developing a simple tool or database for relating a sequence cluster or motif to antibiotic specificity is likely to be challenging. This is well illustrated by the Bush-Jacoby classification of beta-lactamase sub-types, where a different group can be assigned following the mutation of a single residue and by the study of Verma et al. [11]. In an extensive investigation of the physiochemical properties of class A beta-lactamases, Verma et al. [11] revealed that new antibiotic resistance activities, including those found in “extended-spectrum” beta-lactamases, are evolutionarily easy to achieve because they come about through small changes that do not globally affect structure nor the concomitant electrostatic properties (e.g. electrostatic network, pairwise energies, electrostatic network composition, residue charge, and per residue pKa shifts). They do, however, report a statistically significant correlation between global protein charge and antibiotic resistance specificity. Guthrie et al. [12] also report success with a network model used to identify co-evolving residues within the class A type TEM beta-lactamases. Triple mutant combinations are found that increase cefotaxime resistance. Mandage et al. [13] analyse residue conservation on the surface of beta-lactamases using the ConSurf [14] server but this property does not appear to relate clearly to antibiotic resistance specificity. The Livesay group have developed a Distance Constraint Model (DCM) to examine changes in protein stability and flexibility and this been applied to proteins from Class C serine beta-lactamases [15] and metallo-beta-lactamases [16].
The goal of the work reported here is to analyse sequence features of serine beta-lactamases at different levels of classification: 1) ‘Classes’–distinguishing different implementations of the mechanism of action; 2) ‘Types’ or sequence clusters; and 3) ‘Variants’, that provide a context within which to understand the subtle evolution of antibiotic resistance specificity.
Our FunFHMMer algorithm [17] identifies functional families (FunFams) that distinguish well the Class A, C, D serine beta-lactamases. Subsequent clustering of the Class sequences, using CD-HIT [18] based on an optimal sequence identity cut-off, largely reproduces well-characterised types within the Class A serine beta-lactamases. To identify key functional positions (e.g. catalytic residues) and FDs that vary significantly between different types, we developed the novel Active Site Structural Profile (ASSP) algorithm, which exploits both structure and sequence and uses parsimony to characterise residues in the enzyme active site, which are likely to have a functional role.
Over the last few decades, the introduction and overuse of Man-made antibiotics have driven the evolution of beta-lactamase variants with broader substrate profiles. In particular, novel variants in the Class A TEM-type are responsible for a significant proportion of clinically reported inhibitor resistance. We use another parsimony-based approach, Secondary Shell Parsimony Analysis (SSPA), to identify driver mutations in serine beta-lactamase Class A variants that confer resistance to Man-made beta-lactam antibiotics and beta-lactamase inhibitors. We examine the locations of these variant mutations relative to the conserved core of the active site and the FDs that distinguish the different classes and types.
In summary, we propose that the precise antibiotic resistance specificity and inhibitor resistance of serine beta-lactamases can be seen as a synthesis of various levels of classification: 1) implementation of the mechanism of action (distinguishing A, C, D classes); 2) a sequence cluster correlating with specificity (beta-lactamase type(s)); and 3) variant (beta-lactamase sub-type). We focus mainly on the Class A beta-lactamases, the class which currently has most clinical relevance, and apply our classification approach to identify Class A beta-lactamase types in all complete bacterial genome sequences in our comprehensive CATH-Gene3D resource [19,20]. Our classification approaches are then applied to find and examine novel types in microbiome samples from human gut and drain.
It is already known that beta-lactamases fall into two distinct structural superfamilies and this is supported by the results of our structure comparisons using SSAP [21,22]. Classes A, C and D (i.e. serine beta-lactamases) are assigned to CATH DD-peptidase/Serine beta-lactamase superfamily, (3.40.710.10), on the basis of both structural similarity and conservation of key catalytic residues in the active site. Class B metallo-beta-lactamases adopt a different structural fold and are assigned to CATH superfamily 3.60.15.10 (see S1 Fig).
The DD-peptidase/Serine Beta-Lactamase superfamily contains a large number of DD-peptidases. Although Class A, C and D beta-lactamases tend to have lower structural similarity with the DD-peptidases than with each other (see S1 Table), there is conservation of the structural core across this superfamily. In particular, the active site and catalytic serine, which is found in both DD-peptidases and the Class A, C and D beta-lactamases, superpose well (see S2 Fig).
In this study we focus on the classification and analysis of serine beta-lactamases. Whilst S1 Table and S3 Fig show that structural similarity can be used to distinguish Class A, C and D beta-lactamases, most beta-lactamases in public repositories and discovered by metagenome studies have not been structurally characterised yet. Therefore, we developed sequence-based approaches to distinguish these classes.
Serine beta-lactamases are thought to have evolved independently from the DD-peptidases three times (i.e. Class A, C, D beta-lactamases) more than 2 billion years ago [23]. We predicted 105,810 sequences from UniProt [24] and Ensembl [25] belonging to the CATH DD-peptidase/Beta-Lactamase superfamily (3.40.710.10) using our in-house Gene3D classification protocol [19,20]. This superfamily is moderately functionally diverse as summarised in S2 Table. All member domains are hydrolases and belong to three main “branches”: peptidase activity; hydrolase activity acting on carbon-nitrogen (but not peptide) bonds; and hydrolase activity acting on ester bonds. The region of the GO Molecular Function Ontology (MFO) Directed Acyclic Graph (DAG) that is encompassed by experimentally determined UniProt [24] annotations within this superfamily has eleven most specific terms, of which seven are leaf terms [2].
The DD-peptidases use the same mechanism of action as the beta-lactamases but the chemistry is a little different since an N-C peptide bond is being broken as opposed to a N-C bond in a cyclic amide. It is possible that the mechanism of action is as ancient as the fold itself and we expect similar mechanisms of actions are also used by the other main esterase “branch” in the GO molecular function ontology, as illustrated in S2 Table. All scissile bonds are characterised by delocalisation of electrons, which may be an essential feature of the mechanism of action. Thus, in this superfamily it appears that mechanism is the most conserved and evolutionarily ancient aspect, perhaps as old as the fold itself, and that its specific implementation, are secondary.
Simple pairwise sequence approaches (e.g. BLAST) can be used to recognise homologues with very closely related sequences (i.e. greater than 60% identity) in each class of beta-lactamases. However, since distant relatives in each class can share less than 30% sequence identity (see S3 Table) more sensitive techniques are needed to distinguish classes. Our FunFHMMer protocol [17] sub-classified the superfamily into distinct functional families (FunFams). Manual inspection of the UniProt [26] descriptions of the serine beta-lactamases confirmed that three FunFams captured well the three classes A, C and D respectively. Small manual adjustments result in complete agreement between FunFam classification and beta-lactamase classes. For each FunFam (i.e. Class A, C, D) we inspected the experimental annotations given in UniProt and removed those few sequences having non beta-lactamase annotations, e.g. having a DD-peptidase annotation. These comprised fewer than 2% of sequences within each FunFam. Two large and sequence diverse, functionally pure DD-peptidase FunFams are also automatically identified by FunFHMMer.
Almost every domain sequence that can be assigned to the DD-peptidase/Serine beta-lactamase superfamily has an SXXK motif that maps to equivalent structural locations when the domain structures are superposed (S2a Fig). There are 3 catalytic residues (Ambler residues serine 70, lysine 73 and lysine 234) that are common to all known DD-peptidases and beta-lactamases (see S2b Fig). There have been a number of studies examining how residue differences in these proteins account for their diverse substrates (linear versus cyclic peptides) but the mechanistic roles of the residues remain unclear apart from a few relatives [23,27].
Within each serine beta-lactamase class relatives have diverged considerably in sequence identity and in their phenotypes, e.g. the ability to degrade different ranges of beta-lactam substrates. Several classification approaches have been used to distinguish relatives. In particular, ‘types’ are commonly referred to in the literature and these groups tend to be associated with particular substrate profiles and efficacies. Another approach, based more on clinical phenotypes, e.g. resistance to specific beta-lactamase inhibitors, is the Bush-Jacoby classification. However, it is not always clear from the literature that the identified types and Bush-Jacoby (BJ) classes have been identified using the same standardised experimental screening against an explicit repertoire of compounds. For that reason, we derived a classification protocol, the results of which matched the ‘types’ and ‘BJ classes’ reported in the literature as far as possible, but which exploits standard sequence-based approaches that would be easy to replicate by other biomedical researchers.
Table 1 shows the sequence population of each serine beta-lactamase Class (i.e. the number of Gene3D sequence counts) and lists types that have been identified in the literature and that have at least ten annotated members, together with their UniProt annotations and a representative structural domain.
We first considered the Class A (3.40.710.10.blA) and Class C (3.40.710.10.blC) FunFams as these are sufficiently sequence diverse to benefit from a sequence-based classification that could ultimately be used to characterise changes in functional residues likely to be modifying the phenotypes. Furthermore, the sequence diversity was sufficient for HMMs derived for these classes to be powerful enough to recognise both close and remote homologues in metagenome sequences. Because the Class C FunFam (3.40.710.10.blC) only contains one major clinically significant type (and three sub-types) we focused on the class A FunFam (3.40.710.10.blA) that contains fifteen clinically significant types and which, as we demonstrate here, contains sufficient sequence information to accurately characterise changes in functional residues in the active site.
Because of their clinical significance, the type names: CTX-M, TEM, SHV, Z, L2, KPC, OXY, PER, OKP, GES, LEN, CfxA, RAHN, CARB and PSE, or variations thereon, are frequently used in the UniProt descriptions of the protein sequences and thus provided a guide for automatically subdividing the Class A FunFam into types [26]. We have only considered well-populated types having at least ten annotated sequences in CATH-Gene3D. 1,321 out of 2,154 (~60%) full-length Gene3D domain sequences assigned to the Class A FunFam are annotated with clinical type information in UniProt.
CATH-Gene3D domain sequence intra- and inter-type pairwise sequence identities are derived from the full FunFam alignment and their distributions are shown in S4 Fig. Edit distance from the UniProt annotation of types (number of split and merge operations) is calculated for a range of sequence identity cut-offs used in CD-HIT clustering and a minimum is found at 60% sequence identity (see S5 Table). Clustering with a 60% sequence identity cut-off is performed for all 2,154 Gene3D domain sequences and the resulting “cluster60” distributions of inter- and intra-cluster sequence identities are shown in Fig 4.
Using this cut-off, the 15 types highlighted in the literature fall into 9 cluster60s (i.e. 9 predicted types, see Table 2). The 60% cut-off for separation of function specificity are supported by other studies relating functional similarity to sequence identity [35,36] and may avoid over-fitting to the currently available annotation data and therefore over estimation of the number of types that can be found in nature. Where different types defined in the literature are merged into the same 60% sequence identity cluster, the Bush-Jacoby groups (i.e. resistance phenotype) associated with them tend to be very similar (see Table 2).
Using the 60% threshold to cluster CATH-Gene3D sequences in the Class A FunFam into types, we identified 151 types of which 142 are new types not reported in the scientific literature (ftp://ftp.biochem.ucl.ac.uk/pub/cath/v4_0_0/supplementary_files/151_types_uniprot_cath-gene3d.dat).
In order to explore the differences between the Class A types and understand changes in their substrate specificities and efficacies, we developed a new approach (the ASSP protocol, see Methods) to landscape the active site characteristics of these different groupings. Although FunFHMMer can identify conserved sites differing between pairs of types, because there are 151 types an optimisation strategy is needed to identify the specific residues differing between all types. Furthermore, some types have few relatives to date, most of which are recently diverged. FunFHMMer’s entropy-based approach works best in distinguishing residue sites conserved differently between groups over significant evolutionary time-scales. Comparing types that have recently emerged is challenging, since many residues appear to be conserved sites over these much shorter time scales and need to be considered as possible FDs. To narrow down the number of residue sites to consider, ASSP exploits structural information and uses a parsimony based approach to explore different combinations of residues in the active site that could be influencing the substrate and resistance profiles.
An initial Active Site Structural Profile (ASSP) was derived (see Methods) based on all 151 types identified in the CATH-Gene3D Class A FunFam. It comprised all those Ambler residues that lie within 8Å of the catalytic serine. This gave an ASSP with 31 positions. S6 Table shows the residues found at each ASSP position for each of the 9 predicted Class A types having clinical annotations in UniProt. For many positions in the ASSP, many types share the same residues.
The next steps of the ASSP method find the smallest combination of residue positions in this original ASSP for which all Class A types have different residues. In other words, the smallest combination of residues best able to capture the active site diversity of Class A types. To do this we analysed first two-residue, then three-residue, up to N-residue permutations of the 31 residue positions in the first stage ASSP to identify unique configurations of residues between all the types (see Methods for a schematic representation of the approach). For each N-residue configuration examined, the number of unique residue combinations across all types was counted. Subsequently, the distribution of these counts was plotted for each N. Z-scores (minimum and maximum) were calculated for each distribution (i.e. from the maximum or minimum number of types observed). Fig 5 shows the distribution of the number of observed configurations in the 151 types, for all 4,495 three-position (triplet) permutations of the 31 positions in the first stage ASSP.
Minimum and maximum Z-scores for configurations of up to eight positions (N = 8) are shown in Table 3 and it can be seen in Fig 6 that 7 residues in the configuration are necessary to fully distinguish all of the Class A types. The highest maximum Z-score occurs for a triplet configuration (N = 3). Although not all types have a unique configuration until N = 7 (see S1 Text for ASSP N = 7 residue configuration and S5 Fig for the N = 7 functionally important positions highlighted in the Class A serine beta-lactamase domain), the maximum Z-score for this number of residues in the configuration is not very significant, i.e. finding a unique configuration of these number of residues is not very unlikely. The line in Fig 6a rises steeply up to N = 3 but then takes a long time to level off and a triplet configuration distinguishes 114/151 (75%) of the predicted types with a highly statistically significant Z-score of 3.86. The lowest minimum Z-score for the configuration which captures positions common to all types is difficult to identify as the algorithm has not converged by 8 positions and is too computationally expensive to proceed to higher numbers of positions.
Based on the highest maximum Z-score in Table 3, FDs distinguishing between the types are given by a triplet consisting of Ambler positions 74, 129, and 244. We assume that this configuration of positions has been under strong selective pressure for long evolutionary periods to efficiently inactivate the wide variety of beta-lactam antibiotics that have been produced by fungi. The 8 positions giving the lowest minimum Z-score achieved in our analysis (i.e. residues conserved between all types which should include the known catalytic residues) together with the 3 positions likely to be FDs and differing in their composition between most of the types, are shown in Table 4 below.
The Type 1 Class beta-lactamases, which include the TEMs, are a highly populated type, capturing a significant proportion of clinically characterised beta-lactamases. Recent divergence of these enzymes has given rise to relatives with extended-spectrum beta-lactam resistance (i.e. ability to inactivate third-generation cephalosporins with an oxyimino side chain as well as monobactams) and inhibitor resistance (e.g. resistant to the inhibitors Clavulanic acid and Sulbactam). We were interested in exploring the mutations responsible for these clinically significant phenotypes. In this case, we are dealing with very recent divergence and many residue positions will appear conserved across the TEMs. Here, we wished to determine which mutations occurring in a variant TEM sequence, were contributing to the phenotype. However, reports in the literature of multiple driver mutations, some occurring remote from the active site (see S8 Table), meant that we could not restrict our analysis to active sites residues. We therefore developed another parsimony-based approach to identify driver mutations likely to be conferring these phenotypes. We validated our approach by examining how well our predictions agreed with experimentally confirmed genotype-phenotype data in the literature.
As well as using our approaches to analyse sites implicated in beta-lactam resistance, we also applied FunFHMMer and ASSP to search for novel Class A types in metagenomes sampled from human gut and a bathroom drain environment. Although BLAST can be used to detect known types (i.e. sequences having greater than 60% sequence identity to one of the Class A types identified using the CD-HIT clustering above), novel Class A types (i.e. having < 60% sequence identity) are difficult to distinguish from Class D beta-lactamases. Furthermore, microbiome sequences are sometimes incomplete and a preliminary analysis of BLAST matches revealed incomplete sequences with > 60% identity to a Class A beta-lactamase but lacking fragments of sequence containing the catalytic or FD residues, making it impossible to identify the type. Therefore, we used FunFHMMer to identify very safe matches within these microbiomes, which could then be subjected to experimental validation.
Sequences taken from thirteen human gut microbiomes (see Methods for details) were scanned against the HMM for the Class A FunFam using FunFHMMer. This identified 136 full length matches to Class A. These human gut microbiome beta-lactamase sequences clustered into 8 types, of which 7 were previously identified by our classification of Class A types above, and 3 of those 7 had clinical annotations. Therefore, 1 out of the 8 types found in gut microbiome sequences is novel, suggesting a reasonable level of novelty in the human gut metagenome. This new cluster, which is a singleton, has a unique FD triplet, FEV. However, the sequence lacked a signal peptide, suggesting that it may have evolved a different function and therefore it was not tested for activity.
Scans of sequences from our in-house drain metagenome data against the Class A FunFam HMM identified one match. This had 37% sequence identity to the closest Class A beta-lactamase in our CATH-Gene3D dataset, marking it out as a novel type. This was confirmed by the detection of a unique FD triplet, IQA (combination 1 in Table 8). This sequence was cloned and expressed in E. coli, and its activity was tested against a range of beta-lactam compounds known to be acted on by Class A beta-lactamases (see Methods). For this purpose, a qualitative agar-diffusion test was performed with the following antibiotics: amoxicillin, ampicillin, oxacillin, cloxacillin and carbenicillin at concentrations of 2, 5, 10 and 20 μg/ml. The size of zone of inhibition around 10 and 20 μg/ml of amoxicillin suggested that both with the native signal and the pelB signal, candidate beta-lactamase could give resistance to this antibiotic and that the one with native signal has higher activity. 5 different concentrations of amoxicillin were then tested: 10 (the lowest concentration that inhibited growth), 15, 20, 25, 30 μg/ml, all of which gave positive results. The agar-diffusion test was also performed with higher concentrations of ampicillin, oxacillin, cloxacillin and carbenicillin: 10, 20, 25, 50 μg/ml. The size of zones of inhibition suggests that the candidate beta-lactamase could also give resistance to ampicillin, again the protein with native signal has higher activity. The lowest concentration that inhibited growth was 25 μg/ml of ampicillin.
We were surprised that so few Class A matches were found in the drain microbiome sample. However, this could reflect the fact that the sequence samples lack important regions of the sequence and therefore fail to meet the strict Class A FunFam HMM inclusion threshold. We therefore examined 14 matches which failed to meet the inclusion threshold but which gave high scores against the Class A FunFam and significantly higher matches to Class A FunFams than to DD peptidases, Class C or Class D beta-lactamases.
These putative matches were examined for the following criteria: 1) contained all three motif regions identified by FunFHMMer for Class A beta-lactamases (see Methods for details), 2) contained a new combination of FD residues, and 3) had a bit score very close to the Class A inclusion threshold and very far from the DD-peptidase, and Class C and D inclusion thresholds. Three unique combinations of the FDs were found (see combinations 2, 3 and 4 in Table 8) suggesting that there are potentially three further novel types within this microbiome.
In conclusion, we have constructed a classification and analysis platform for beta-lactamases that applies a number of structure and sequence-based algorithms to distinguish beta-lactamases from DD-peptidases and to sub-classify classes and types of serine beta-lactamases. Importantly, our protocols search for residue sites likely to be exerting an influence on the function. This could relate to implementation of the catalytic mechanism or to the substrate profile. Our protocols provide a strategy for recognising previously unreported ‘types’, which could have novel resistance profiles and reveal emerging resistance to new drug regimes.
Although sequences sharing high sequence similarity (> 60%) to known serine beta-lactamases can easily be recognised by BLAST, in the twilight zone of sequence identity (< 30%) it is difficult to distinguish different classes of serine beta-lactamases from each other and from the DD-peptidases. Structural analyses can provide important clues, as we and others have reported, but few of the sequences emerging from high throughput studies e.g. metagenome studies, have structural data.
Therefore, our classification pipeline focused mainly on sequence data. Our FunFHMMer derived FunFams for the Class A, C and D beta-lactamases allowed us to recognise even very distant relatives of these beta-lactamase classes (< 20% sequence identity) as they capture distinct residue patterns associated with each class. Our results show that FunFHMMer was not only able to distinguish sequences with the beta-lactamase Gene Ontology (GO) term from sequences coming from other conflicting GO Molecular Function “branches” in the DD-peptidase superfamily, but also to separate FunFams corresponding to different implementations of the mechanism of beta-lactamase action i.e. separate the Class A, C and D beta-lactamases. Detailed analysis of the Functional Determinant (FD) residues differing between these classes revealed residue positions likely to be contributing to differences in the implementation of the catalytic mechanism. Many of these positions are validated by reports in the literature. Other FDs revealed by our method suggest sites that could be targeted to gain better understanding of the determinants separating the classes from each other and from the DD-peptidases.
The Class A beta-lactamases are the largest and most diverse class, responsible for most of the resistance to clinically relevant beta-lactams. We therefore decided to perform more detailed analyses of this Class. Fifteen clinically relevant types are reported in the literature, having largely different substrate profiles. However, it is not clear whether these assignments are based on standardised compound screening protocols. We found that using a sequence identity threshold of 60%, a value that corresponds to other studies identifying functionally related proteins [35,36], we obtained a good separation of the clinically reported types that also largely corresponded to similarity of Bush-Jacoby groups within each predicted type. Applying this threshold identified 151 types amongst the UniProt and Ensembl sequences assigned to the Class A FunFam in CATH-Gene3D, 142 more than reported in the literature.
Again, by revealing specific residue sites differing between the types and likely to be influencing the phenotypes (i.e. substrate profiles) we can provide a more refined analysis tool for classifying these types. FunFHMMer was not so suited to this task since some types are very recently diverged and because it is not designed to identify residues differing across multiple groups. We found that a simpler parsimony based approach (ASSP), that focused on residues close to the active site, could be used to find these FDs. Our ASSP predictions of catalytic sites showed significant agreement with catalytic positions reported in the literature, and the putative FDs were shown to be located very close to the catalytic residues or in the secondary shell. Further studies using docked substrates and using a substrate bound to an inactive mutant supported proximity of the FDs to the beta-lactam substrate. One of the positions makes a hydrogen bond with the beta-lactam and there are reports in the literature of its involvement with the catalytic activity. The other positions are more remote from the catalytic residues but located within the secondary shell of the active site where they may influence conformational rearrangements necessary to support changes along the reaction pathway.
Finally, we analysed variants in the TEM-Type Class A beta-lactamases, the type responsible for much of the clinically relevant resistance to beta-lactams. Again, the fact that some of these variants or ‘subtypes’ emerged very recently and that some driver mutations have been found quite far from the active site meant that a new strategy was needed. SSPA is not restricted to sites close to catalytic residues but examines all mutations. Validation against positions reported in the literature, showed that SSPA successfully identified 5 sites known to be associated with inhibitor resistance and 5 known to be associated with extended-spectrum resistance phenotype. Inspection of the SSPA predictions in 3D showed that many SSPA sites not yet experimentally verified lie close to ‘hot regions’ which are lying in or near the active site, or close to the omega loop which is thought to have a functional role.
We tested the validity of our SSPA approach by applying it to an important subtype in the beta-lactamase TEMs, i.e. mutants having a 2be phenotype in the Bush-Jacoby classification. However, the success of SSPA in identifying previously experimentally characterised sites suggests that it would be useful to apply SSPA to other subtypes which have sufficient genotype-phenotype data necessary for this approach.
We tested the ability of our Class A FunFam to recognise Class A serine beta-lactamases in two microbiome samples. A putative novel type was identified in the drain microbiome, which met the Class A FunFam inclusion threshold but which was likely to be a novel type as it shared less than 40% sequence identity to any Class A beta-lactamase in our Gene3D dataset and contained a unique FD triplet. Experimental validation confirmed its resistance to a range of compounds associated with Class A beta-lactamase activity. Much more extensive screening work can now be done to comprehensively explore its substrate range and how that differs from other known types.
Because of the stringency of the FunFam inclusion threshold, and the general poor quality of the metagenome sequences the matches reported in this study actually only represent about 2% of all the significant matches (E-value ≤ 0.0001) that were found. Manual analysis of a sample of these missed significant matches showed that fragments with key catalytic or FD residues were missing from the sequence. If the metagenomic data were of better quality, then we might reasonably expect to see at least an order of magnitude more novel beta-lactamase clusters.
In summary, we have developed a classification and analysis platform that allows us to separate relatives within the serine beta-lactamase superfamily according to their implementation of the mechanism of action and their substrate profiles. Our FunFHMMer method can separate the known beta-lactamase classes and identify those positions likely to be responsible for the different implementations of the mechanism of action in these enzymes, which emerged independently from DD-peptidases, three times during evolution. The ASSP algorithm detects FD sites which can help to classify the different Class A Types, whilst the SSPA algorithm detected sites conferring inhibitor resistance or extended-spectrum resistance phenotypes. Each algorithm has specific features designed to suit the nature of the dataset being analysed.
The FDs that we recognise can be used as fingerprints to classify new relatives and predict their likely resistance profiles. We tested the predictive value of our classification by uncovering and experimentally verifying a new Class A Type within a drain microbiome ie having a unique fingerprint of FD residues.
Finally, our parsimony based approaches for identifying FDs and for distinguishing driver from passenger mutations could obviously be applied to other protein superfamilies and one can imagine other medical applications where resistance to chemical challenges has emerged recently in evolution. For example, kinases implicated in certain cancers, which evolve resistance to drugs, and where residue configurations close to catalytic residues or other functional sites e.g. activation loops, could be analysed to detect driver mutations associated with different phenotypes, such as responses to drug treatments. Our functional family classification and analysis pipeline provides a strategy for detecting residue sites playing a functional role in the emergence of new phenotypes.
Domain structure representatives for each of the Class A, B, C and D beta-lactamases, and DD-peptidases were selected from our in-house CATH classification of protein domain superfamilies [20]. Each structural domain pair was compared using the in-house SSAP structure comparison algorithm [21,22]. The SSAP algorithm uses a well-established double dynamic programming algorithm to identify a reliable residue alignment between each pair of structures. A SSAP score is returned in the range of 0 to 100, where 100 indicates identical structures. The SSAP alignment was used as input to the ProFit algorithm (Martin, A.C.R., http://www.bioinf.org.uk/software/profit/), which superimposes the structures and calculates their RMSD.
For our analysis of beta-lactamase proteins we used the dataset of protein domains classified in our in-house Gene3D resource [19]. Gene3D is a sister resource of CATH [20] and version 12 comprises nearly 50 million domain sequences from UniProt version 2013_02 and Ensembl version 70, predicted to belong to CATH superfamilies. Domain sequences are assigned to a particular CATH superfamily following hmmscan scans against superfamily HMMs built from representative sequences (17).
An in-house automatic function classification method FunFHMMer [17] was used to sub-classify the CATH-Gene3D DD-peptidase/serine beta-lactamase superfamily into distinct functional families (FunFams). The superfamily sequences are initially clustered using the GeMMA agglomerative clustering algorithm [50] that creates a hierarchical tree of sequence relationships within the superfamily. GeMMA clusters close sequence relatives into starting clusters using CD-HIT [18]. Multiple sequence alignments for each starting cluster are built using MAFFT [51]. GeMMA then performs an iterative all-against-all profile-profile comparison of a set of clusters using COMPASS [52] followed by merging of the most similar clusters and realignment of the merged clusters by MAFFT. This iterative process continues until one cluster remains. The merging order is then used to build a hierarchical tree from the leaf nodes to the root rode. Once the tree has been generated, functional families (FunFams) are identified by FunFHMMer, which partitions the tree based on the identification of positions which are differentially conserved in different FunFams. Thresholds for partitioning superfamily trees have been optimised by validation against experimentally determined functions and functional sites [17].
Once FunFams have been identified, HMM profiles are built for each FunFam using HMMER version 3 [53]. Putative serine beta-lactamases can be identified by scanning query sequences against the Class A, C, D FunFam HMMs. Sequences are assigned to a particular FunFam provided they return a bit score that is greater than or equal to the inclusion threshold for that FunFam (14). FunFHMMer has been validated in silico [17] and independently validated for its performance in function prediction, ranking in the top 5 (out of 126 methods) in the international Critical Assessment of Protein Function Annotation [54] (CAFA) 2 experiment (Radivojac, P., personal communication).
FunFHMMer exploits the GroupSim [55] method to detect residue sites that are differentially conserved between FunFams. It was used to report sites differentially conserved between Class A, C, D FunFams and thus likely to play a functional role [17,56]. GroupSim takes an alignment containing pre-defined functional groups as input and provides a prediction score for each column in the alignment. The score ranges from 0 to 1, where any position in the alignment having a score greater than 0.65 may be a functional determinant (FD) [17].
To identify key FD residues between the three serine beta-lactamase classes (A, C and D) we built a three-way structural alignment of the corresponding FunFams. This was done by selecting representative sequences (at 60% sequence identity), with known structure, from each class and constructing a multiple alignment by performing successive pairwise structure alignments against the representative that best matches all other representatives. After this, hmmbuild from the HMMER package [53] was used to create an HMM for the structure-based alignment. Sequence relatives from the Class A, C, D FunFams were then aligned to the HMM using the hmmalign command from the HMMER package [53]. The resulting structure-based sequence alignment was then used for site analysis by applying GroupSim [55].
To sub-classify relatives in the serine beta-lactamase Class A FunFam into clusters corresponding to ‘types’ identified in the literature, the CD-HIT [18] algorithm was used. CD-HIT can very rapidly cluster protein sequences according to sequence identity at levels of similarity above about 40%. It is widely used in computational biology due to its speed and the reliability of its results.
In order to help understand the evolution of beta-lactamases, we characterised the extent and nature of the active site by the construction of Active Site Structural Profiles (ASSPs). These structure-based profiles were applied to the Class A serine beta-lactamases and first capture all residues within a threshold distance of well-characterised catalytic residues reported in the scientific literature. Subsequently, a parsimony-based approach identifies those residues (FDs) likely to have a role in modifying functional features between types. This approach helped to distinguish differences in key residue sites between Class A serine beta-lactamase types.
We decided to apply structural criteria in ASSP as a number of other methods have successfully explored residues lying close to catalytic residues to detect additional functionally important sites. For example, JESS [57], uses an initial active site template (constituting 2–5 amino acid residues) from the Catalytic Site Atlas (CSA) [58] to search for similar conformations of residues in other protein structures. For putative matches, residues within a 10 Å sphere are compared to calculate a local similarity score (SiteSeer score) that is used to rank the template match [59]. Similarly, the Evolutionary Trace method [60] identifies functionally important residues by partitioning a phylogenetic tree to identify subfamilies and focusing on highly conserved residues that lie within 4 Å of each other. Whilst the subfamily classification method, DASP (Deacon Active Site Profiler) [61,62], selects all residues within a 10 Å sphere of known catalytic residues which are then concatenated to build a structure based profile. Structural relatives having similar profiles are clustered into subfamilies and the subfamily profiles subsequently transformed into PSSMs and used to identify sequence relatives.
The first stage in the construction of the ASSPs is the analysis of the PDB data of a representative structure for the FunFam. The PDB 1SHV was chosen as the representative structure for ASSP analysis. This structure satisfied a number of criteria: it had a high score when compared to the HMM representing the FunFam; it is a wild-type sequence; it was expressed in reasonably physiological-like experimental conditions; and it has no bound ligand. 1SHV not only satisfied all of these criteria but its use of the standard Ambler residue numbering scheme helped with reference to the literature and analysis of mutation and phenotype data [63].
Details of the construction of the initial ASSP and its processing to produce the final ASSP is described in Figs 11 and 12.
The first plasmid borne beta-lactamase was identified in E. coli in Greece in 1963 and was named “TEM” after the patient from whom it was isolated [65]. Today it is the most commonly encountered beta-lactamase in Gram-negative bacteria and the TEM-1 subtype accounts for up to 90% of ampicillin resistance in E. coli. Mutation and phenotype data for variant TEM beta-lactamases are made available in Supporting Information by Guthrie et al. [12]. A parsimony-based approach was applied to this Guthrie dataset to distinguish driver from passenger mutations associated with the inhibitor-resistant (e.g. Clavulanic acid and Sulbactam) and extended-spectrum phenotypes (i.e. resistant to penicillins, cephalosporins and third-generation cephalosporins).
SSPA matrices were created for each of the two phenotypes where each column in the matrix represents a residue position where a mutation is found relative to the consensus sequence of the multiple alignment of all the variant TEM beta-lactamase sequences. Each variant possessing a distinct phenotype (i.e. inhibitor-resistant [12,30,32,42] or extended-spectrum phenotype [12,42–49]) occupies a row in the matrix. We then determine the minimum number of columns (i.e. putative driver mutations) for which one or more of these positions is mutated in every variant with a phenotype.
To identify novel Class A types we analysed two different microbiomes–gut and drain. Metagenome sequences were scanned against the Class A FunFam HMM. Sequences assigned to the Class A (i.e. meeting the inclusion threshold for the FunFam) were then compared against the sequences classified into the 151 types identified in this class to identify novel types having less than 60% sequence identity to sequences in any of these types.
Pre-processed gut metagenome sequences were obtained from the MG-RAST [66] and EBI Metagenomics [67] resources (S12 Table). Some of the MG-RAST and EBI microbiomes were already partially assembled into contigs but where this was not the case, MetaVelvet [68] was used for assembly to increase the chance of finding complete beta-lactamase domain sequences. Additional metagenome data derived from a bathroom drain and sequenced using Illumina MiSeq technology was processed by the Ward group at UCL (deposited in the EBI Metagenomics resource under project ID ERP011520). The paired-end reads were quality assessed and filtered using the Paired-End ToolKit (PETKit version 1.1b, http://microbiology.se/software/petkit/). Contiguous read assembly was performed on the clean reads using IDBA-UD [69]. Contig sequences were translated into protein sequences using a 6-frame translation with the tool Transeq from EMBOSS v6.6.0.0 [70]. Open-reading frames were predicted using Prodigal v2.6.2 [71].
Gene sequences from the drain environment and contig sequences from the human gut environments were scanned by FunFHMMer [17] against HMMs from the DD-peptidase/Serine beta-lactamase superfamily. If the resulting bit-score was greater than or equal to the inclusion threshold, the sequence was assigned to that FunFam [17]. Any sequence that was less than 80% of the average length of all sequences assigned to the FunFam was deemed a fragment and filtered out. Sequences sharing less than 60% sequence identity to any of the CATH-Gene3D Class A serine beta-lactamases were selected as potential novel types. To further refine matches likely to be novel types, metagenome-derived sequences giving a significant match to the Class A FunFam were aligned to the existing Class A alignment using the MAFFT algorithm [72]. Sequences long enough to contain the three main functional motifs [27,30] in Class A beta-lactamases, and capturing all the serine beta-lactamase catalytic residues (Motif 1: Ambler nos. 70–73 (SXXK); Motif 2: Ambler nos. 130–132 (SDN loop); Motif 3: Ambler nos. 234–236 (K[T/S]G)) and the FDs identified by the ASSP method (Ambler residue nos. 74, 129 and 244) were examined closely to analyse changes in residues. Those having a novel combination of the three FDs distinguishing the types, and not observed in any of the types classified in CATH-Gene3D [19] were considered for experimental validation.
A predicted gene encoding beta-lactamase, bla-29843, was amplified directly from the drain metagenomic DNA by a two–step PCR using a Phusion High-Fidelity DNA Polymerase (NEB) and conditions suggested by the manufacturer. The following PCR primers were used: forward, 5’- CATATGCGACGCGCCTCTCTCGTG– 3’ and reverse, 5’–GCGGCCGCGTTGACGGTAAGGAAATGGTCGTAAGCG– 3’. The blunt-ended PCR product was ligated into pCR-Blunt vector with a Zero Blunt PCR Cloning Kit (Invitrogen) followed by the transformation into chemically competent E. coli DH5α. pCR-Blunt vector containing bla-29843 gene was confirmed by DNA sequencing. This vector was further used as a template for PCR amplification with primers designed to incorporate 5′ NdeI restriction site followed by a pelB leader sequence and a 3′ NotI restriction site. The N-terminal pelB leader sequence was added to enable the periplasmic secretion of beta-lactamase via the Sec translocation machinery. Two PCR products were generated for bla-29843, one with its native N-terminal signal sequence and the other with the pelB leader sequence instead. The following PCR primers were used: (i) forward and reverse primers for bla-29843 with the native signal sequence were 5’- CATATGCGACGCGCCTCTCTCGTG—3’ and 5’—GCGGCCGCGTTGACGGTAAGGAAATGGTCGTAAGCG—3’ (ii) forward and reverse primers for bla-29843 with pelB sequence were 5’- TATACATATGAAATACCTGCTGCCGACCGCTGCTGCTGGTCTGCTGCTCCTCGCTGCCCAGCCGGCGATGGCCATGGCACCCGCAACAACGATCGCG– 3’ and 5’–GCGGCCGCGTTGACGGTAAGGAAATGGTCGTAAGCG– 3’. PCR products were purified and restriction cloned into NdeI and NotI sites of the bacterial expression vector pET-29a (+) (Novagen). The resulting vectors encode beta-lactamases containing an N-terminal leader sequence and a C-terminal poly-histidine tag preceded by 5 amino acids.
Expression of beta-lactamases was carried out in BL21 (DE3) pLysS E. coli cells (Invitrogen) harbouring pET29a- beta-lactamases vectors described above. To test susceptibility to antibiotics, diffusion in solid agar was used. All antibiotics (amoxicillin, ampicillin, oxacillin, cloxacillin, kanamycin) were purchased from Sigma except carbenicillin that was purchased from Invitrogen. Bacteria for lawn seeding were grown overnight at 37°C with shaking in Luria-Bertani (LB) medium supplemented with 50 μg/ml of kanamycin. Inoculum was spread on solid LB agar plates supplemented with 1mM IPTG. Holes were punched with a plastic tip and filled with the same amount of antibiotic solutions. Plates from three independent replicates were analyzed individually for the inhibition zone diameter. BL21 (DE3) pLysS E. coli cells carrying an empty pET29a vector were used as a negative control.
|
10.1371/journal.pgen.1008377 | BZU2/ZmMUTE controls symmetrical division of guard mother cell and specifies neighbor cell fate in maize | Intercellular communication in adjacent cell layers determines cell fate and polarity, thus orchestrating tissue specification and differentiation. Here we use the maize stomatal apparatus as a model to investigate cell fate determination. Mutations in ZmBZU2 (bizui2, bzu2) confer a complete absence of subsidiary cells (SCs) and normal guard cells (GCs), leading to failure of formation of mature stomatal complexes. Nuclear polarization and actin accumulation at the interface between subsidiary mother cells (SMCs) and guard mother cells (GMCs), an essential pre-requisite for asymmetric cell division, did not occur in Zmbzu2 mutants. ZmBZU2 encodes a basic helix-loop-helix (bHLH) transcription factor, which is an ortholog of AtMUTE in Arabidopsis (BZU2/ZmMUTE). We found that a number of genes implicated in stomatal development are transcriptionally regulated by BZU2/ZmMUTE. In particular, BZU2/ZmMUTE directly binds to the promoters of PAN1 and PAN2, two early regulators of protodermal cell fate and SMC polarization, consistent with the low levels of transcription of these genes observed in bzu2-1 mutants. BZU2/ZmMUTE has the cell-to-cell mobility characteristic similar to that of BdMUTE in Brachypodium distachyon. Unexpectedly, BZU2/ZmMUTE is expressed in GMC from the asymmetric division stage to the GMC division stage, and especially in the SMC establishment stage. Taken together, these data imply that BZU2/ZmMUTE is required for early events in SMC polarization and differentiation as well as for the last symmetrical division of GMCs to produce the two GCs, and is a master determinant of the cell fate of its neighbors through cell-to-cell communication.
| In the grasses, individual stomatal complexes comprise a pair of dumbbell-shaped guard cells associated with two subsidiary cells and the pore, which together play essential roles in the exchange of CO2 and O2, in xylem transport, and in transpiration. However, little is known about grass stomatal complex development. We have uncovered and characterized a key factor (BZU2/ZmMUTE) determining the formation of guard cells in Zea mays. Our data suggest that BZU2/ZmMUTE has a dual role, both as an important player in determining the formation of the guard mother cell, as well as being required for polarization and recruitment of the subsidiary mother cells.
| The development of the stomatal complex in maize (Zea mays), consisting of a pair of dumbbell-shaped guard cells (GCs) flanked by two subsidiary cells (SCs), provides an excellent model system to study the signals controlling fate determination of adjacent cells by intercellular communication. Unlike the two kidney-shaped GCs in eudicots, the four-celled stomatal complex in grasses may facilitate a faster response to environmental cues in order to optimize photosynthesis and water use [1, 2]. Highly differentiated GCs and SCs in grasses are generated by two asymmetric divisions [3]. The first asymmetric division generates a guard mother cell (GMC), which produces the extrinsic cues conveyed to the laterally adjacent subsidiary mother cells (SMCs). Upon perception of the signal, SMCs become spatially polarized with respect to the GMC, by localized accumulation of cortical F-actin at the SMC/GMC interface and migration of the pre-meiotic SMC nuclei to each interface [4]. The extrinsic cue eventually triggers an asymmetrical division of the SMCs, in an orientation that positions the smaller daughter cell, the SC, adjacent to the GMC [5, 6]. Finally, the GMC produces the two GCs by symmetrical division.
Several intrinsic factors that specify cell fate during the development of GCs have been identified. Moreover, an extrinsic signaling cascade has been dissected in the control of the number and plane of asymmetric divisions [7]. In Arabidopsis, three of basic helix-loop-helix (bHLH) transcription factors, SPEECHLESS (SPCH), MUTE and FAMA sequentially modulate the initiation, transition and differentiation events, respectively, regulating the cell lineage involved in the production of stomata [8–13]. The earliest acting bHLH protein, SPCH, functions in the transition from protoderm cells (PDCs) to meristemoid mother cells (MMCs) [10, 12]. Inhibition of SPCH function predominantly affects MMCs and meristemoids and their ability to divide asymmetrically [9, 10, 14], suggesting that SPCH is necessary for the asymmetric divisions that establish the stomatal lineage. The second bHLH protein, MUTE, which is required to terminate asymmetric division and stimulate transition from meristemoid to GMC identity, upregulates a series of cell-cycle genes to drive symmetric division of stomata [12, 15, 16]. Meristemoids of mute mutants undergo excessive amplifying divisions and fail to transition to a GMC [12]. Interruption of cell to cell signaling can instigate an extension or reduction of meristemoid division, consistent with the observation that epidermal cell number, including stomata, is a plastic trait that is monitored and adjusted based on internal and external cues [8, 12, 17, 18]. The third bHLH protein, FAMA, controls the final fate transition from GMC to GC, consisting of two distinguishable events: symmetric cell division of the GMC and GC transition [11].
Recently, stomatal initiation in the grass Brachypodium distachyon has been shown to involve the orthologs of Arabidopsis stomatal bHLH regulators. However, it appears that the function and behavior of the individual components and their interacting regulatory networks have diverged, acquiring specific functions for the unique stomatal development of grasses [19, 20]. BdMUTE, an ortholog of Arabidopsis MUTE in B. distachyon, is expressed in GMCs before moving to neighboring cells, and this mobility is a protein-intrinsic feature [21]. In maize, a leucine-rich receptor-like kinase (LRR-RLKs), PANGLOSS1 (PAN1), has been found to participate in transmission of the signals for establishing polarity in SMCs [22]. About 40% of SMCs in loss-of-function pan1 mutants had unpolarized nuclei and developed defective stomatal complexes, resulting from abnormal asymmetric cell division and SC patterning [22, 23]. Interestingly, PAN2 is required for the subsequent polarization of PAN1 and the Rho family GTPases (ROPs) [24, 25]. After PAN2 polarization, PAN1 and ROP2/9 are polarized, an actin patch forms at the SMC/GMC contact sites, and the SMC nuclei migrate to this site in an actin-dependent manner. pan2, pan1, rop2 and rop9 mutants have abnormal SCs, and their SMCs are defective in actin patch formation and nuclear migration [26, 27]. It has been proposed that the GMC may send a cue to the neighboring cells inducing the asymmetric cell division of the SMCs at adjacent sites to produce the SCs [21, 22]. However, the molecular mechanisms by which GMCs control polarization and cell fate of SMCs in maize remain unclear.
Here we identify BZU2/ZmMUTE (Bizui2, abbreviated to BZU2) gene, encoding a bHLH transcription factor, an ortholog of Arabidopsis MUTE protein. Loss-of-function BZU2/ZmMUTE mutants show no SCs and are defective in stomatal complexes. In addition, bzu2-1 displays defects in symmetric GMC division and lacks SMC polarization. We found that BZU2/ZmMUTE controls the expression of several essential genes involved in the formation of the stomatal complex, and binds to the promoters of PAN1 and PAN2 for amplifying the expression of the polarity program in SMCs. Our finding suggests that BZU2/ZmMUTE functions as an important player both acting to regulate the GMC to GC fate transition, and serving as an intercellular signal required for polarization and recruitment of the SMCs.
To isolate stomatal response deficient mutants, we produced an extensive collection of mutagenized maize lines. Pollen from the maize inbred line Mo17 was treated with ethyl methane sulfonate (EMS) and the M1 progeny were self-pollinated to produce M2 mutant seeds. We screened ~45,000 of these M2 at the seedling stage for drought-sensitivity, using a far infrared thermal imaging approach [28, 29]. We identified a number of mutants displaying an unusually high leaf-surface temperature and designated these as bizui which means closed mouth (abbreviated to bzu).
One of these recessive mutants, viz. bzu2-1, was studied in detail. It constitutively displays a hot leaf surface phenotype as compared to the wild-type (Mo17) (Fig 1A). In addition, the average leaf temperature of bzu2-1 mutants was ~1°C higher than wild-type, which reflects an inability of bzu2-1 leaves to appropriately regulate transpirational water loss from their surfaces (Fig 1B). The bzu2-1 mutant produced pale, highly hydrated, translucent leaves, and the seedlings died about 14 days after germination (Fig 1C). Upon examination of leaf structure, we observed that the normal mature four-celled stomatal complexes, comprising two dumb-bell shaped GCs adjacent to two SCs, were absent from bzu2-1 mutants (Fig 1D). As a consequence, the water loss from the leaf surfaces of bzu2-1 mutants was significantly slower than that of the wild-type (Fig 1E). We subsequently characterized chlorophyll fluorescence emission, comparing the maximum quantum yield (Fv/Fm) of wild-type and mutant plants. We found that leaves of the bzu2-1 mutant have a lower Fv/Fm value than that of wild-type (Fig 1F and 1G), which implies that the photosynthetic activity is dramatically decreased in 8-day-old seedlings of the bzu2-1 mutant. These observations indicate that the bzu2 mutant lacks the normal physiological functions of stomata, and is consistent with the lethal seedling phenotype.
As illustrated in Fig 2A, in the wild-type, protodermal cells (PDC) undergo an asymmetric division producing a smaller daughter cell, which is the precursor of an apical GMC, and a larger basal interstomatal cell. Following the differentiation and elongation of the GMCs, the two flanking cells acquire SMCs identities prior to the formation of the GCs. The SMCs then undergo another asymmetric division, producing the two SCs prior to a final division of the GMC, which divides symmetrically and longitudinally to produce the two GCs. The stomatal complexes with dumbbell-shaped GCs immediately adjacent to two triangular SCs are linearly separated by rectangular interstomatal cells.
In bzu2-1 mutants, the early phase of stomatal development seems to be normal, with PDCs being able to produce GMCs. However, the longitudinal symmetric division of GMCs is altered in the bzu2-1 mutants, with asymmetric or irregular directional divisions occurred (a transverse or diagonal longitudinal division), resulting in the production of short columns of elongated cells, which lack hallmarks of GC morphology (Fig 2A). Subsequently, none of GMCs develop into normal stomatal complexes (Fig 2B). Instead, in the bzu2-1 mutant, 48%, 27% and 25% of GMCs were further divided into 2-, 3- or 4-cell groups of undifferentiated cells, respectively (Fig 2C). Our data imply that one function of BZU2 may be to suppress ectopic or premature guard mother cell divisions.
Interestingly, in bzu2-1 mutants, GMCs also failed to induce flanking cells to form SMCs, resulting in an almost complete absence of SCs. Examination of 802 stomata in bzu2-1 mutants showed that ~95% of the abnormal GCs had no SCs, and ~5% remaining had one single abnormal SC (S1 Fig). The lack of SCs suggests that absence of BZU2 may abrogate the acquisition of cell fate of SMC precursors adjacent to GMCs (Fig 2A and S1 Fig).
We further examined the spatial interactions between the GMCs and SMCs during stomatal development via light and fluorescence microscopy. In the wild-type, GMCs induce the formation of SMCs, which exhibit polarity as defined by the anisotropic positioning of the SMC nucleus with respect to the SMC/GMC contact surface. In this process, an important step in SMCs polarization is the formation of a dense patch of F-actin, which presumably mediates nuclear migration or anchoring during cell division [22]. In the wild-type, when actin-patches appeared at the SMC/GMC contact sites, the SMCs nucleus will migrate to the contact sites, and then the SMCs undergo one asymmetric division to form SCs (Fig 3A). Our data demonstrated that 91.3% of the SMCs nuclei migrated to the SMC/GMC contact surface in the wild-type, the cells therefore being polarized and highly anisotropic. In contrast, 84.4% of the SMCs examined in bzu2-1 mutants were non-polarized, with the nucleus being centrally located (Fig 3A and 3B). These results show that BZU2 either acts as an intrinsic signal in SMC development, or controls the expression of intrinsic factors that determine SMC polarity and, ultimately, cell fate.
In order to characterize BZU2 at the molecular level, we crossed the bzu2-1 mutant (carried in the Mo17 background) to the inbred line B73, thereby generating reciprocal F1 hybrid progeny. The F2 population resulting from F1 self-crossing was screened based on the bzu2-1 mutant phenotype (S2 Fig). Initial mapping indicated that the BZU2 locus co-segregated with the simple sequence repeat (SSR) markers bnlg1863 (recombination rate 2.1%) and umc1858 (recombination rate 12.5%) in Bin 8.03 of chromosome VIII. Several rounds of fine-mapping narrowed this locus to a 0.69 Mb region between SSR markers 79M15 (79.01 M) and 79M45 (79.70 M) containing fifteen genes (Fig 4A). After sequencing, we identified the BZU2 gene as GRMZM2G417164 (Fig 4B), which encodes a bHLH transcription factor exhibiting sequence similarity to Arabidopsis, rice and Brachypodium MUTE (Fig 4D and S3 Fig). bzu2-1 contains a 4 nucleotide ‘AGCT’ insertion at the position 390 bp downstream of the start of transcription generating a premature STOP codon (Fig 4B and 4C). The bzu2-1 mutant phenotype was further confirmed by targeted gene knockouts of BZU2/ZmMUTE using the CRISPR/Cas9 system. Stomatal phenotypes of three independent CRISPR/Cas9-mutated lines (bzu2-2, bzu2-3 and bzu2-4) are identical to bzu2-1, confirming BZU2/ZmMUTE as GRMZM2G417164 (Fig 4E and 4F and S4 Fig). When a GFP- BZU2/ZmMUTE fusion chimera was transiently expressed in tobacco (Nicotiana tabacum L.), we were able to show accumulation of GFP-BZU2/ZmMUTE protein within the nucleus (S5 Fig), which is consistent with the putative function of BZU2/ZmMUTE as a transcription factor.
The interesting role of BZU2/ZmMUTE in controlling neighbor cell fate prompted us to further investigate the behavior of BZU2/ZmMUTE protein in the early developmental stages of SCs and GCs. In the transgenic reporter plants, the fluorescence of ZmMUTEp:YFP-ZmMUTE was first detected during the stomatal development of early GMCs, and then moved to SMCs in the SMC establishment stage. The YFP-ZmMUTE signal remains strong until the young GCs become mature GCs (Fig 5). Unexpectedly, as compared to the expression pattern of BdMUTE, BZU2/ZmMUTE is more specifically expressed in the SMC establishment stage (Fig 5).
To further test the capability of ZmMUTE to move, YFP-fused BZU2/ZmMUTE, as well as its homologues, BdMUTE, OsMUTE, and AtMUTE, were expressed in rice (Oryza sativa L.). OsMUTEp:nls-YFP was only expressed in the GMC at the developmental stages leading from GMC formation to SMC division (S6A Fig), indicating that OsMUTE promoter is active during development of GMC and SMC [13]. Interestingly, as for OsMUTEp:YFP-BdMUTE, OsMUTEp:YFP-OsMUTE and OsMUTEp:YFP-ZmMUTE, the fluorescence of the YFP-fused MUTEs were first detected in the early GMCs, then appeared in the SMCs (S6B–S6D Fig, white arrows). Similar to ZmMUTEp:YFP-ZmMUTE (Fig 5), after the division of the GMCs, the signals of OsMUTEp:YFP-ZmMUTE were also observed in young GCs and SCs, finally disappearing from the GCs and SCs at maturity (S6C Fig). In contrast, the fluorescence of OsMUTEp:YFP-AtMUTE was very weak in the early GMCs and not detected in SMCs (S6E Fig).
Meanwhile, we compared the expression patterns of these different YFP-fused MUTE coding sequences driven by the AtMUTE promoter in Arabidopsis. AtMUTEp:AtMUTE-YFP was observed exclusively in Arabidopsis GMCs (Fig 6A and 6F), but AtMUTEp:YFP-ZmMUTE and AtMUTEp:YFP-BdMUTE were observed both in GMCs and in neighboring cells (Fig 6B, 6C and 6F). Since divergent functions of MUTE have been reported in different species [10, 21], we performed sequence alignments, the results of which indicate that MUTE proteins from different species have a variable C-terminal region, whereas the N-terminal region is relatively conserved (Fig 4D). To further verify the characterization of the C-terminal regions of BZU2/ZmMUTE and BdMUTE, we generated AtMUTEp:YFP-ZmMUTE-ΔC and AtMUTEp:YFP-BdMUTE-ΔC transgenic plants in Arabidopsis which lack the 190–219 and 208–237 amino acids in ZmMUTE and BdMUTE, respectively. The fluorescence of AtMUTEp:YFP-ZmMUTE-ΔC is only detected in the GMCs and expressed in both nucleus and cytoplasm. In contrast, AtMUTEp:YFP- BdMUTE-ΔC is detected in the GMCs and restricted in nucleus, which is similar to the pattern seen with AtMUTEp:AtMUTE-YFP (Fig 6D–6F). At the same time, we checked the expression patterns of OsMUTEp:YFP-ZmMUTE-ΔC and OsMUTEp:YFP-BdMUTE-ΔC in rice. As shown in S6F and S6G Fig, YFP-ZmMUTE-ΔC and YFP-BdMUTE-ΔC are only located in the early GMCs, and not in the SMCs and young SCs.
However, it is obvious that the localization (for ZmMUTE-ΔC, Fig 6D and S6F Fig), or the intensity (for BdMUTE-ΔC, Fig 6E and S6G Fig), does not correspond to that of the wild-type, suggesting that the protein does not properly interact with its binding partner for normal localization and/or stability. To further confirm whether a truncated BZU2/ZmMUTE protein, from which the 30 amino acids of the C-terminus had been deleted, can still bind the E-box motifs within the promoters of PAN1 (-27) and PAN2 (-187, -200), the yeast one-hybrid assay and electrophoretic mobility shift assays (EMSA) were used to test the activity of ZmMUTE-ΔC binding to short nucleotide fragments (26–37 bp) containing the E-box motifs. The results of both assays show that BZU2/ZmMUTE lacking these C-terminal amino acids is similar in behavior to BZU2/ZmMUTE (S7 Fig and S8 Fig). These data imply that the C terminus of BZU2/ZmMUTE and BdMUTE might be necessary for their characteristic mobility. Together, these results support that BZU2/ZmMUTE is mobile and necessary for SMC formation and asymmetric division in normal development of the maize stomatal complex.
In order to further assess the role of BZU2/ZmMUTE in stomatal development and, in particular, define its interactions with other cellular components, we performed RT-qPCR to compare, for bzu2-1 mutants and wild-type, the transcript levels of genes previously reported to be involved in stomatal and leaf development (Fig 7A). PAN2 is polarized in premitotic SMCs [24]. After PAN2 polarization, PAN1 and ROP proteins are polarized, and an actin patch forms at the GMC/SMC interface [22, 23]. ROP2/9 functions downstream of PAN1 to promote the premitotic polarization of SMCs [25], and the premitotic SMC nucleus migrates to this site in an actin-dependent manner. Liguleless1 (LG1) accumulation at the site of ligule formation and in the axil of developing tassel branches, functions in the leaf shape and tassel architecture [30]. Low levels of the PAN1, PAN2 and ROP2/9 transcripts were observed in bzu2-1 mutant plants as compared to wild-type. Furthermore, these genes related to stomatal development are down-regulated at very early seedling stages (S9 Fig), since stomata production in grass initiates at the leaf base with a longitudinal gradient of development and differentiation toward the tip. This therefore suggests that the reduction of the transcript levels of these genes is not simply due to the leaves dying. For example, ROP2 transcript abundance in the wild-type was 6.7 times higher than that in bzu2-1. Transcripts of the Brick1 (BRK1) and BRK3 genes, required for the formation of epidermal cell lobes as well as for actin-dependent cell polarization events of subsidiary mother cells [26, 31, 32], were also found at lower levels in bzu2-1 plants. The expression of SCARECROW (SCR) gene in rice and maize was observed in leaf primordia and in young leaves, which is required for asymmetric cell divisions of GMCs [13, 26, 33]. The transcript level of SCR1 was significantly reduced in bzu2-1 as compared to wild-type. However, the transcript levels of LG1, a SQUAMOSA PROMOTER-BINDING proteins 1 and 2 like gene, in bzu2-1 mutants was comparable to that in the wild-type.
More importantly, it is also known that the function of FAMA and OsFAMA is conserved between dicots and monocots in the regulation of the final symmetric division of the GMCs [13, 34]. In rice, c-osfama (OsFAMA mutant in rice, generated by CRISPR/Cas9) mutants showed the stomatal complex consisted of four swollen cells, two GCs and two SCs, occasionally a stoma lacking one SC was observed, however in c-osfama the entry division and GMC differentiation stage even the stomatal density is the same compared with wild-type. OsFAMA controls the cell fate transition from GMCs to GCs and SMCs to SCs and affected SMC asymmetrical division [13]. We noticed that, in bzu2-1, the expression of ZmFAMA is reduced significantly (Fig 7A), similarly to the manner in which OsFAMA is dramatically downregulated in the c-osmute mutant [13]. Therefore, the absence of ZmFAMA or OsFAMA is simply a result of the lineage abortion. These data clearly demonstrate that the genes of asymmetric and symmetric division in bzu2-1 mutants are impaired in stomatal development, which is consistent with the proposed role of BZU2/ZmMUTE as a master regulator of stomatal differentiation. Therefore, it is speculated that the abnormal formation of SCs and the symmetric division of GMCs in bzu2-1 mutants might be specifically due to downregulation of PAN1, PAN2, and ZmFAMA (Fig 7A and S9 Fig).
Previous studies have shown that bHLH transcription factors can specifically bind to the E-box cis-element and regulate the expression of targets (Fig 7C) [35]. Therefore, we performed motif enrichment analysis by MEME [36], found that five and three E-box cis-elements are located in the regions within -750 - -1 in the promoter of PAN1 and PAN2, respectively (Fig 7C). This prompted us to assess the DNA binding specificity of BZU2/ZmMUTE. First, we obtained transgenic plants of BZU2/ZmMUTEp:YFP-BZU2/ZmMUTE, and used for ChIP-qPCR experiments using anti-GFP (ab290, Abcam) polyclonal antibody (the expression of YFP-BZU2/ZmMUTE was confirmed by Western blot using anti-GFP antibody) (S10 Fig). Our results showed a strong enrichment in promoter fragments for PAN1 and PAN2, as compared to negative controls (the wild-type in the presence of anti-GFP). However, no enrichment was observed in the negative control using the regions lacking E-boxes in the promoters of PAN1 and PAN2 (Fig 7B). These data are consistent with the ChIP-qPCR results using the native antibody of BZU2/ZmMUTE (S11 Fig). Second, we employed the yeast one-hybrid assay to directly examine interactions between BZU2/ZmMUTE and 3–4 repeats of short nucleotide fragments (26–37 bp) containing the E-box motifs from the promoters of PAN1 (-27) and PAN2 (-187, -200). Activation of BZU2/ZmMUTE the LacZ reporter gene was seen for PAN1 sequence motifs, but was not detected using sequences from a second putative E-box contained in the PAN2 promoter (-645) (Fig 7C–7E). Mutation of the E-box motif eliminated LacZ expression, providing confirmation of the binding specificity of BZU2/ZmMUTE. Finally, the experiment of EMSA further indicated that BZU2/ZmMUTE binds to the E-box motifs in vitro (Fig 7F), which is consistent with the yeast one-hybrid data. Taken together, our results suggest that PAN1 and PAN2 are two direct targets of BZU2/ZmMUTE.
In this study, we found that BZU2/ZmMUTE encoding a bHLH transcription factor is an ortholog of AtMUTE. AtMUTE plays an essential role in the transition of the meristemoid to GMC by repressing the stem cell activity of the meristemoid and inducing guard mother cell formation [10, 12]. Loss-of-function bzu2-1 mutants can form normal GMCs, but fail to undergo a symmetric division to generate two guard cells, as in wild-type. This indicates that the GC precursors in bzu2-1 are similar to wild-type, with the exception of stomatal development. Consequently, eight-day-old seedlings of bzu2-1 displayed etiolated phenotypes and, at the stage of three leaves, the seedlings subsequently died (Fig 1). Unlike AtMUTE mutants, the longitudinal symmetric division of the GMC in bzu2-1 mutants is replaced by an asymmetric or irregular directional division (transverse or diagonal longitudinal division), resulting in the production of 2–4 short columns of elongated cells (Fig 2B and 2C). In bzu2-1 mutants, the stomata are defective and thus the functions of gas exchange and water loss are impaired. Our data suggest that BZU2/ZmMUTE acts as a switch that controls stomatal development.
Normally, in maize, premitotic SMCs polarize toward the GMC in response to hypothetical cues coming from the adjacent GMCs. This process involves migration of the nucleus toward the GMCs and a distinct enrichment of cortical F-actin at the point of interaction of the GMCs and SMCs when the cell files are forming. Subsequently, SMCs divide asymmetrically to produce subsidiary cells flanking the GMC, which in turn divides to produce a guard cell pair to form stomatal complex [4]. In bzu2-1 mutants, a lack of nuclear polarization and/or actin accumulation at the GMC/SMC interaction area was observed. Thus, the GMCs failed to induce the flanking cells to form SMCs. bzu2-1 mutants share similar features with BdMUTE mutants of B. distachyon, but are different in several respects (Fig 2). Both mutations of MUTE in Zea mays and B. distachyon result in an abnormal SMC formation and polarization. BdMUTE mutants fail to recruit SCs and instead produce dicot-like two-celled stomata [21]. As compared to BdMUTE, where partial guard cells form without SCs, bzu2-1 is completely defective in guard cell and SMC formation. These data suggest that the function of BZU2/ZmMUTE transcription factor in maize is different from BdMUTE, since absence of the former appears to disrupt SC and GC formation in a more direct and severe manner than does absence of BdMUTE. In addition, our results suggest the C-terminal is necessary for the mobile nature of both BdMUTE and BZU2/ZmMUTE (Fig 6).
Even though previous work in B. distachyon has established the mobile nature of BdMUTE, the molecular mechanisms of how BdMUTE controls SCs formation are still unknown [21]. In fact, the mechanisms governing SMC polarization to allow establishment of asymmetric division are also largely unknown. We provide several lines of evidence suggesting that BZU2/ZmMUTE participates in the regulation of SMC development, and of PAN1 and PAN2, which are early regulators of SC precursor and of SMC polarization by cooperatively promoting polarization of the actin cytoskeleton and nuclei in these cells [22]. Firstly, our data show that BZU2/ZmMUTE can bind to the E-box of the PAN1 and PAN2 promoters (Fig 7 and S7 Fig); consistent with this, transcript levels for PAN1 and PAN2 are severely downregulated in bzu2-1 mutants as compared to the wild-type. Unexpectedly, yeast one-hybrid and EMSA data show that BZU2/ZmMUTE is not able to activate P3 of the promoter of PAN2 in vitro, but the ChIP-qPCR data indicate that BZU2/ZmMUTE can activate P3 of the PAN2 promoter in vivo. We speculate that BZU2/ZmMUTE combines with additional unknown factors to activate P3 of the PAN2 promoter in vivo (Fig 7, S7 Fig and S8 Fig), by analogy to the observation that the DNA binding specificity of AtMUTE depends on its dimerization partner ICE1 or SCRM2 [37]. Furthermore, a number of genes involved in the regulation of stomatal development are down-regulated in bzu2-1 mutants as compared to the wild-type (Fig 7A). Finally, mutations in BZU2/ZmMUTE disrupted the actin-based patch attachment of GMCs with SMCs, which can mis-orient the deposition of new cell walls (Fig 3A and 3B) [38–41]. Actin plays an important role in the spatial regulation of asymmetric cell division. For example, actin-dependent relocation of the nucleus during G1 to a defined cortical site of SMCs during stomatal complex formation is one of the early events of asymmetric cell division in Tradescantia [42] which is followed by formation of a dense actin patch at this site [43, 44].
Combining the existing knowledge and the results presented, here we propose a model in which BZU2/ZmMUTE plays a role partially similar to what has been previously described, but that includes some different roles in maize stomatal development (Fig 8). In the base of maize leaves, protodermal cells undergo one asymmetric division to form the GMCs. At a specific stage of GMC development (Fig 5), these cells could send BZU2/ZmMUTE as an extrinsic cue to neighboring cells connecting non-sister GMCs and SMCs, thereby acting as an important regulator of the early players in SC development (e.g. PAN1, PAN2). After BZU2/ZmMUTE-mediated induction, PAN1 and PAN2 (and possibly other additional factors) accumulate at the SMC/GMC interface, working together with F-actin to induce SMC polarity and nuclear migration towards the GMC proximal site [26]. Following polarization, SMCs undergo one asymmetric division to form a SC and an epidermal cell. In the final stage of stomatal development, BZU2/ZmMUTE performs an additional role controlling the symmetric division of the GMCs to produce the two GCs. It is important to note that the role of BZU2/ZmMUTE in the biogenesis of SCs is not limited to the control of PAN1/2 expression, since SC formation is not entirely disrupted in pan1 or pan2 mutants, whereas SCs are completely absent in bzu2-1 mutants [22]. Thus, these imply that BZU2/ZmMUTE, acting as a GMC-derived polarizing signal, moves to neighboring cells (Fig 5). This, in turn, initiates the expression of the polarity program by regulating the expression of the genes required for nuclear polarization and polarized actin accumulation at the GMC contact sites (Fig 7A). These data indicate that BZU2/ZmMUTE may play a role in the modulation of gene expression at earlier stages of SC precursor development. In summary, our data support a critical role for BZU2/ZmMUTE in the regulation of SC development and GC maturation. Further studies will be required to dissect the functions of BZU2/ZmMUTE in the determination of GMC fate and in the initiation of intercellular signaling required for the recruitment of the SMCs during stomatal development. More insights could emerge from characterization of the additional mutants obtained in our screen that show different defects in GC and SC development.
Maize plants were grown at the experimental station of Henan University in the Kaifeng experimental field, Henan Province, and the Sanya experimental field, Hainan Province.
To isolate stomatal development deficient mutants, we established an extensive collection of EMS-mutagenized maize plants, as follows: the pollen of maize inbred line Mo17 was treated with EMS, T1 population seeds were sowed in the soil, and T1 plants were self-crossed to obtain T2 population seeds. We screened the T2 mutagenized population for the phenotype of an altered leaf surface temperature, using a far-infrared imaging instrument. In 3,000 lines, represented by ~45,000 T2 population seeds, we found a single lethal mutant, and named bzu2-1 (called bizui, closed mouth, bzu2-1). Genetic analysis indicates BZU2 is a qualitative trait gene, with a segregation ratio of 3:1 in the F2 generation (Table 1).
The homozygous bzu2-1 (-/-) mutant is lethal. We use the heterozygous bzu2-1 (+/-) crossed with B73 for generation of the reciprocal F1 population. The F2 population resulting from the self-crossed F1, and a map-based cloning population, was screened for the bzu2-1 phenotype from the F2 population (S2 Fig). Preliminary mapping of BZU2 used 306 plants from the F2 population, derived from a cross between Mo17 and B73. The 384 SSR markers (SIGMA catalog number M4193) selected from the Maize Genetics and Genomic Database cover the entire genome with an average of 20 cM units of map distance between every two SSR markers. More SSR markers that were genetically mapped on IBM2 2008 Neighbors Frame were used, and BZU2 was mapped between SSR markers bnlg1863 (recombination rate 2.1%) and umc1858 (recombination rate 12.5%). Based on Maize B73_RefGen_v3 (http://www.maizesequence.org), the BZU2 was mapped to the chromosome VIII between bins 8.03 and 8.04. Therefore, all BAC contigs in bins 8.03–8.04 were exploited to develop new polymorphic markers. To develop more SSR markers, SSR Hunter 1.3 [45] was used to search for SSR sequences present in bins 8.03–8.04. SSRs and their flanking sequences about 150 bp were then aligned with NCBI nucleotide BLAST (http://www.ncbi.nlm.nih.gov) (high-throughput genomic sequences: HTGSs). Only single sequences were used as SSR markers and amplified by PCR. In fine mapping, ~7,000 plants from F2 population deprived from a cross between Mo17 and B73, 63 polymorphic markers were used, BZU2 was mapped in SSR markers 79M15 (79.01 M) and 79M45 (79.70 M). Further analysis and sequencing confirm that GRMZM2G417164 was located between these two markers, AGCT insert in the + 390 bp of GRMZM2G417164.
CRISPR constructs were designed using the vector system and following the design protocol [46], and was done by Genovo Biotechnology Co (Xi’an, Shanxi). The PAM sites were chosen at the + 289 bp downstream of the start codon of BZU2/ZmMUTE genome region, because there is no intron in BZU2/ZmMUTE. The sequence of the gRNA is CCTGTCATGATCAAGGAGCTCGC (S1 Table). To genotype CRISPR-induced mutations, we amplified a 668-bp fragment including the guide RNA site by PCR from the genome of transgenic seedlings [47, 48], and the PCR products were sequenced. We obtained three CRISPR/Cas9 mutant lines: bzu2-2, bzu2-3 and bzu2-4. A 5 bp and 25 bp deletion was detected behind PAM site in bzu2-2 line and bzu2-3 respectively, while a 1 bp insertion was detected behind PAM site bzu2-4 line. In T0 mutant plants, the phenotype of homozygous mutants lines was comparable to bzu2-1 (Fig 4C). The heterozygous T0 mutants lines were planted in the field to get seeds. In the T1 populations, homozygous mutants lines also showed a similar phenotype to that of bzu2-1.
Reporter constructs to be transformed into Oryza sativa Japonica were generated using the In-fusion cloning with the monocot binary expression vector pIPKb003 and to be transformed in Arabidopsis thaliana were generated using the Gateway Recombination Cloning Technology with plant expression vector pGWB504. AtMUTE promoter and genome sequences were amplified from the Arabidopsis (Col-0) genome, OsMUTE promoter and genome sequences were amplified from the Japonica genome, and BdMUTE sequence was amplified from the Brachypodium distachyon genome. All genomic DNA samples were produced using the Plant Genomic DNA Kit (TIANGEN). RNA samples were produced using the TRIzol extraction method and corresponding cDNA was obtained by M-MLV reverse transcriptase (M1705, Promega).
The Pri1 and Pri2 primers were used to clone the OsMUTE promoter from the rice genome. We then amplified tag-YFP using forward primers Pri3 and Pri4 (for AtMUTE), Pri3 and Pri7 (for OsMUTE), Pri3 and Pri10 (for BZU2/ZmMUTE), and Pri3 and Pri14 (for BdMUTE), respectively, which have sequences homologous with the OsMUTE promoter, and using four reverse primers (Pri4, Pri7, Pri10, Pri14) to get four tag-YFP PCR products which has homologous sequence with the AtMUTE, OsMUTE, BZU2/ZmMUTE and BdMUTE. Then we used primers Pri5 and Pri6 to clone the AtMUTE ORF carrying a STOP codon, primers Pri8 and Pri9 to clone the OsMUTE ORF with a STOP codon, and the primers Pri11 and Pri12 to clone the ZmMUTE ORF with a STOP codon. Primers Pri11 and Pri13 were used to clone the ZmMUTE ORF with a deletion of 30 AA in the C terminal replaced by a STOP codon, and primers Pri15 and Pri16 to clone the BdMUTE ORF with a STOP codon. Primers Pri15 and Pri17 were used to clone the BdMUTE ORF with a deletion of 30 AA in the C terminal replaced by a STOP codon. The three PCR products were employed for in-fusion cloning using the vector pIPKb003. OsMUTEp:YFP-AtMUTE, OsMUTEp:YFP-OsMUTE, OsMUTEp:YFP-BdMUTE and OsMUTEp:YFP-BZU2/ZmMUTE were produced according to the procedures described in the manual of the Clone Express® MultiS One Step Cloning Kit (Vazyme Biotech Co., Ltd, China).
Pri18 and Pri19 were used to clone the whole genomic fragment of AtMUTEp:AtMUTE from the Arabidopsis thaliana (Col-0) genome. The tag-YFP fragment was amplified using Pri20 and Pri21, Pri20 bearing homology to AtMUTE and Pri21 providing an attB2 site. The fusion gene AtMUTEp:AtMUTE-YFP was then generated by overlap PCR using primers Pri18 and Pri21, and then BP recombined into pDONR207, next LR recombined into destination vector pGWB504. For AtMUTEp:YFP-BZU2/ZmMUTE and AtMUTEp:YFP-BdMUTE, primers Pri18 and Pri22 were used to clone the AtMUTE promoter from the Arabidopsis (Col-0) genome. Pri22 provided sequence homologous to tag-YFP. With primer Pri23 and reverse primers (Pri24, Pri25), this amplified the tag-YFP PCR products all carrying sequences homologous to BZU2/ZmMUTE and BdMUTE. We then used primers Pri26 and Pri27 to clone the BZU2/ZmMUTE ORF carrying a STOP codon, and primers Pri26 and Pri28 to clone the BZU2/ZmMUTE ORF with 30 AA at the C terminal being replaced with a STOP codon. Primers Pri29 and Pri30 were similarly used to clone the BdMUTE ORF with a STOP codon, and primers Pri29 and Pri31 to clone of BdMUTE ORF with 30 AA at the C terminal being replaced with a STOP codon. All reverse primers amplified ORFs carrying Gateway attB2 sites. The inframe gene fusions AtMUTEp:YFP-BZU2/ZmMUTE, AtMUTEp:YFP-BZU2/ZmMUTE-ΔC, AtMUTEp:YFP-BdMUTE, and AtMUTEp:YFP-BdMUTE-ΔC were generated by overlap PCR using the forward primer Pri18 and four reverse primers (Pri27, Pri28, Pri30, Pri31), then BP recombined into pDONR207, followed by LR recombined into the destination vector pGWB504. AtMUTEp:AtMUTE-YFP, AtMUTEp:YFP-BZU2/ZmMUTE, AtMUTEp:YFP-BdMUTE, AtMUTEp:YFP-BZU2/ZmMUTE-ΔC and AtMUTEp:YFP-BdMUTE-ΔC were produced through Gateway Recombination Cloning (Invitrogen).
BZU2/ZmMUTEp:YFP-BZU2/ZmMUTE was inserted into the maize genome via Agrobacterium mediated transformation (see below). The promoter of BZU2/ZmMUTE being amplified from the wild-type (Mo17) genome using primers Pri32 and Pri33, the YFP ORF lacking the stop codon amplified using primers Pri34 and Pri35, and the ORF of BZU2/ZmMUTE including the stop codon amplified from the cDNA of Mo17 with primers Pri36 and Pri37. The three PCR products were employed for In-fusion cloning with the pCM3300 vector. The primers used in this work are listed in S1 Table.
Reporter constructs of OsMUTEp:YFP-AtMUTE, OsMUTEp:YFP-OsMUTE, OsMUTEp:YFP-BdMUTE, OsMUTEp:YFP-BZU2/ZmMUTE, OsMUTEp:YFP-BZU2/ZmMUTE-ΔC and OsMUTEp:YFP-BdMUTE-ΔC were transformed into Oryza sativa Japonica calli with EHA105 Agrobacterium, as previously described [49]. AtMUTEp:AtMUTE-YFP, AtMUTEp:YFP- BZU2/ZmMUTE, AtMUTEp:YFP-BdMUTE, AtMUTEp:YFP-BZU2/ZmMUTE-ΔC and AtMUTEp:YFP-BdMUTE-ΔC were transformed into Arabidopsis (Col-0) using Agrobacterium strain GV3101 [50].
Maize transgenic lines were produced via Agrobacterium-mediated transformation [51]. Maize seedlings were grown in the greenhouse or in the experimental field. Ears containing immature embryos, between 1.0 to 2.0 mm in length along the axis and optimal for transformation, were collected 8 to 13 days after pollination. The immature embryos were submerged in an Agrobacterium tumefaciens suspension contained in a 2.0 mL Eppendorf tube at room temperature for 1 h. The solution was then removed, and the embryos transferred onto fresh co-cultivation solid medium with the scutellum face up, and were incubated in darkness at 25°C for 2–3 days. After that, the calli were transferred onto fresh screen solid medium, screening three times for a period of 2 weeks each. The Type I calli that further proliferated were transferred to shoot regeneration medium, and incubated under continuous illumination (5,000 lux) at 25°C for 14–30 days. The emerging shoots were transferred onto root regeneration medium, and were incubated under continuous illumination (5,000 lux) at 25°C for 14–30 days. The rooted seedlings were transferred to pots containing appropriately supplemented soil for growth in the greenhouse for 3–4 months to collect progeny seeds.
We typically analyzed at least three independent lines in the T0 generation (depending on how many independent lines were recovered upon regeneration) and confirmed the observed expression pattern in at least three T1 individuals if the transgenics were fertile and produced seeds. T1 maize transgenic plants were used in this study. The images were acquired using a Leica SP8 confocal microscope, the cellular membranes being counterstained with propidium iodide (PI, red) in maize and Arabidopsis, and FM4-64 (red) in rice.
To determine the stomatal phenotype in wild-type and bzu2-1 mutants (Fig 1D), epidermal strips were peeled from mature leaves of wild-type and bzu2-1 plants, and were observed using a Zeiss Axioskop II microscope equipped with differential interference contrast optics. To obtain images characterizing the process of stomatal complex development (Fig 2A), around ~1.5 cm of segments from the leaf base were excised from 8-day-old seedlings. The tissues were cleared in Herr's solution (lactic acid:chloral hydrate:phenol:clove oil:xylene (2:2:2:2:1, by weight)) [52]. The stomatal development process was studied using Zeiss Axioskop II microscope equipped with differential interference contrast optics.
For F-actin observation, ~1.5 cm (the section of stomatal complex develops from stomatal lineage cell to mature stomatal complex is about 1.5 cm base in the leaf from maize seedling root node, indicated by the analysis of stomatal complex development using microscopy) of basal leaf segments, excised from 8-day-old seedlings, were cut into 0.2 cm wide x 0.5 cm long strips, and fixed for 30 min at room temperature in a solution comprising 4% paraformaldehyde, dissolved in 50 mM PEM (50 mM PIPES, 2.5 mM EGTA, 2.5 mM MgCl2). The strips were washed three times for 5 min in 50 mM PEM, and were permeabilized by submersion 20 mins in 50 mM PEM containing 5% DMSO and 1% Triton X-100. After three further washes in 50 mM PEM, the sections were incubated for 1.5 h in 50 mM PEM solution containing 90 nM AlexaFluor 488-phalloidin (dissolved in DMSO, Invitrogen/Molecular Probes) at room temperature. Images were acquired using a Zeiss LSM710 confocal microscope.
For observation of the polarized nuclei of the GMCs, ~1.5 cm of second or third basal leaf segments excised from 8-day-old seedlings was cut into 0.2 cm wide x 0.5 cm long strips directly stained with 4 μg/mL Hoechst 33258 (94403-1ML, SIGMA) dissolved in water for 15 min at room temperature. Images were acquired using a Leica SP8 Confocal Microscope.
A peptide corresponding to amino acids 192–206 of the BZU2/ZmMUTE protein (GQDTAEQKPQAEENH) was synthesized, conjugated to KLH, and used for polyclonal antibody production in rabbits by the Hanlin Biotechnology Co. (Shijiazhuang, Hebei).
ChIP-qPCR experiments were carried out using the Magna ChIP kit (MAGNA001, Millipore) with minor modification [53, 54]. Samples (0.4 g) of the stomatal development zones of 8-day-old seedlings were collected, and then immersed in buffer A (0.4 M sucrose, 10 mM Tris [pH 8], 1 mM EDTA, 1 mM PMSF) containing 1% formaldehyde, and were subjected to four eight-minute cycles under vacuum, until the materials became translucent. The materials were then transferred to fresh buffer A containing 0.1 M glycine, and incubation was continued for 16 min at 4°C. The materials were then washed, and frozen in liquid nitrogen. Samples (approximately 0.4 g) of the materials were ground for each immunoprecipitation, and were resuspended in 1 mL lysis buffer (50 mM HEPES [pH 7.5], 150 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.1% deoxycholate, 0.1% SDS, 1 mM PMSF, and 1 x Roche Protease Inhibitor Cocktail), followed by immunoprecipitation with an anti-GFP antibody (ab290, Abcam). Immunoprecipitated products were resuspended in 50 μL elution buffer. Approximately 3–5 μL was used for ChIP-qPCR. Each immunoprecipitation was performed three times independently, with the wild-type being used as the control. The primers for ChIP-qPCR are listed in S1 Table.
The plasmids pB42AD-BZU2/ZmMUTE and pB42AD-BZU2/ZmMUTE-ΔC, and wild-type and mutated PAN1-P1:LacZi, PAN2-P2:LacZi, PAN2-P3:LacZi were co-transformed into yeast strain EGY48 using standard transformation techniques. Transformants were grown on dropout plates containing X-gal (5-Bromo-4-chloro-3-indolyl-β-D-galactopyranoside) for blue color development. pB42AD-SPL9 (squamosa promoter binding protein-like 9) reacting with DFR (dihydroflavonol reductase) [55] was used as a positive control, and combinations with the empty pB42AD vector were used as negative controls. The PAN1 and PAN2 primers are listed in S1 Table.
The EMSA was conducted using the LightShift™ Chemiluminescent EMSA Kit (Thermo Fisher Scientific) according to the manufacturer’s protocol. The coding sequence of BZU2/ZmMUTE was cloned into pGEX-2TK. Recombinant BZU2/ZmMUTE protein was expressed and purified from BL21 E. coli. The probes of the PAN1/2 promoter were obtained by gene synthesis and biotin-labeled at their 5’ terminal. Biotin-unlabeled probes of the same sequences were used as competitors. The probe sequence is described in S1 Table.
The amino acid sequences of BZU2/ZmMUTE and other homologous proteins were retrieved from NCBI (https://www.ncbi.nlm.nih.gov/). Alignments of BZU2/ZmMUTE (GRMZM2G417164) (in wild-type and bzu2-1), BdMUTE (LOC100843821), OsMUTE (Os05g51820) and AtMUTE (At3g06120) were conducted using the MUSCLE and BOXSHADE programs. Construction of phylogenic trees using the neighbor-joining method and confirmation of tree topology by bootstrap analysis (5,000 replicates) was performed with MEGA6 software (using default settings except for the replicates of bootstrap value) [56].
Samples were taken from ~1.5 cm of the base of leaves of 8-day-old maize seedlings for RT-qPCR analysis. Total RNA was isolated using the TRIzol reagent (Life Technologies) according to the manufacturer’s protocol. Reverse transcription into cDNA was done with 2 μg total RNA in 25 μL reverse transcription mixture, using M-MLV Reverse Transcriptase (M1705, Promega). The cDNA was diluted to 100 μL, and 1 μL diluted cDNA was used as the template for quantitative RT-PCR analysis. The maize ZmUbiquitin 2 gene was used as an internal standard to normalize expression of the tested genes. Quantitation was performed using at least three independent biological replicates [57]. The primers used for RT-qPCR are listed in S1 Table.
The coding sequences of BZU2/ZmMUTE was cloned into pGreen0280 (35S:BZU2/ZmMUTE-GFP) for the analysis of subcellular localization. 35S:BZU2/ZmMUTE-GFP, empty vector, and 35S:H2B-mCherry were cotransfected into tobacco leaves. Green and red fluorescence was imaged using the Zeiss LSM710 confocal microscope 24 h after Agrobacterium-mediated infiltration of tobacco (Nicotiana benthamiana) leaves [58]. The primers used are listed in S1 Table.
Previous studies indicated that the E-box cis-element is the conserved CANNTG. The cis-element was produced by the Multiple Expectation Maximization for Motif Elicitation MEME Suite web server [36].
|
10.1371/journal.pcbi.1002400 | Comparison of Insertional RNA Editing in Myxomycetes | RNA editing describes the process in which individual or short stretches of nucleotides in a messenger or structural RNA are inserted, deleted, or substituted. A high level of RNA editing has been observed in the mitochondrial genome of Physarum polycephalum. The most frequent editing type in Physarum is the insertion of individual Cs. RNA editing is extremely accurate in Physarum; however, little is known about its mechanism. Here, we demonstrate how analyzing two organisms from the Myxomycetes, namely Physarum polycephalum and Didymium iridis, allows us to test hypotheses about the editing mechanism that can not be tested from a single organism alone. First, we show that using the recently determined full transcriptome information of Physarum dramatically improves the accuracy of computational editing site prediction in Didymium. We use this approach to predict genes in the mitochondrial genome of Didymium and identify six new edited genes as well as one new gene that appears unedited. Next we investigate sequence conservation in the vicinity of editing sites between the two organisms in order to identify sites that harbor the information for the location of editing sites based on increased conservation. Our results imply that the information contained within only nine or ten nucleotides on either side of the editing site (a distance previously suggested through experiments) is not enough to locate the editing sites. Finally, we show that the codon position bias in C insertional RNA editing of these two organisms is correlated with the selection pressure on the respective genes thereby directly testing an evolutionary theory on the origin of this codon bias. Beyond revealing interesting properties of insertional RNA editing in Myxomycetes, our work suggests possible approaches to be used when finding sequence motifs for any biological process fails.
| RNA is an important biomolecule that is deeply involved in all aspects of molecular biology, such as protein production, gene regulation, and viral replication. However, many significant aspects such as the mechanism of RNA editing are not well understood. RNA editing is the process in which an organism's RNA is modified through the insertion, deletion, or substitution of single or short stretches of nucleotides. The slime mold Physarum polycephalum is a model organism for the study of RNA editing; however, hardly anything is known about its editing machinery. We show that the combination of two organisms (Physarum polycephalum and Didymium iridis) can provide a better understanding of insertional RNA editing than one organism alone. We predict several new edited genes in Didymium. By comparing the sequences of the two organisms in the vicinity of the editing sites we establish minimal requirements for the location of the information by which these editing sites are recognized. Lastly, we directly verify a theory for one of the most striking features of the editing sites, namely their codon bias.
| RNA editing describes the process in which individual or short stretches of nucleotides in a messenger or structural RNA are inserted, deleted, or substituted. As a consequence, the final RNA product translated into a protein or functional by itself is different from its genomic template in organisms with RNA editing. RNA editing is widely spread across species, including plants, mammals, slime molds, viruses and many other organisms [1]–[7]. In some organisms, RNA editing is essential for their survival while for others it provides another layer of fine tuning the genetic program. Although some distinct editing mechanisms have been identified, in many instances the mechanisms of RNA editing are not understood at all [2], [5], [7].
A high level of RNA editing has been observed in the mitochondrion of the slime mold Physarum polycephalum [1], [4], [8]–[10]. In this organism, the mRNA of nearly every mitochondrial protein coding gene is edited at a rate of approximately one out of 25 nucleotides, while structural RNAs are edited at a rate of on average one out of every 40 nucleotides [1], [4], [8]. The by far most frequent editing type in Physarum is the insertion of individual Cs. However, the mitochondrion of Physarum performs a whole set of other editing types, including insertion of individual Us, insertion of certain dinucleotide pairs, deletion of nucleotides and substitutions of Cs by Us [10]–[12]. It has been shown in vivo that RNA editing in Physarum is extremely accurate [13], i.e., that nearly every transcript is completely edited at exactly the correct position.
While the machinery inside the mitochondrion of Physarum recognizes the editing sites with extreme precision, we know neither the mechanism by which these editing sites are recognized nor what machinery is actually performing the editing. It is challenging to decipher the code that determines the editing sites and identify the machinery that performs the actual editing. As far as the machinery is concerned, it has been determined that editing in Physarum is co-transcriptional, i.e., that the RNAs are edited as they are synthesized [14], [15]. Thus, the RNA editing machinery should be part of the RNA polymerase itself or very closely associated with it. As far as the location of the editing sites is concerned, it has been determined that the recognition of editing sites and the actual editing are two independent processes [16]. In order to understand the RNA editing machinery, it is necessary to identify how the RNA editing machinery knows which sites to edit. It is known that only the DNA in close proximity of the site is necessary in order to obtain editing [17]. Rhee et al. [18] demonstrated that DNA necessary for C insertion is within 9 or maybe 10 base pairs on either side of the editing site. But no sequence patterns have been identified that could explain how the sites are recognized. Although no patterns have yet been identified, computational methods for the prediction of insertional editing sites have been developed [19]–[21]. These prediction methods do not require any knowledge about the mRNA. They have been shown to predict the protein sequence with an accuracy of as much as 90% [19].
Here we present the comparison of insertional RNA editing in two related organism in Myxomycetes: Physarum polycephalum and Didymium iridis. Sequence information on mitochondrial genes of Didymium iridis has recently become available [22]–[24] and can be compared to the also recently determined complete edited transcriptome of Physarum polycephalum [25]. Being able to compare sequences from two related organisms promises better understanding of the editing machinery that is vital to the successful function of these organisms.
Our contributions include the prediction of RNA editing sites in Didymium genes based on the knowledge of the Physarum mitochondrial transcriptome, the investigation of sequence conservation in the vicinity of editing sites, and an analysis of codon bias. We use a computational approach to “predict” RNA editing sites in 15 Didymium genes for which the editing sites are known and find that using the Physarum protein sequence can greatly increase the prediction accuracy of Didymium editing sites. We also predict 7 new Didymium genes and their editing sites, and the prediction results suggest one of these genes may be unedited. We investigate the sequence conservation in the vicinity of editing sites between two organisms. Our data implies that a local RNA editing recognition mechanism that is based only on the information contained in any combination of the 18 nucleotides in immediate vicinity of an insertional editing site, even one that uses a different recognition agent (such as a guide RNA or a protein) for every single site, is unlikely. In addition, we show that if such a mechanism exists it has to use nearly all of the 18 positions to specify the site. Finally, we examine the codon position bias in C insertional RNA editing of these two organisms. A strong relationship between the strength of the codon bias and the overall sequence conservation is reported: more conserved genes tend to have more significant codon bias. This result verifies a previous mutation-selection theory for the codon bias.
The recognition mechanism of insertional RNA editing in Myxomycetes is an example where searches for common sequence motifs in a single organism have failed in spite of a very large number of training sites leading to the conclusion that site specific recognition mechanisms must be at work. This situation is not specific to the case of insertional RNA editing but can occur generically in any search for biological sequence motifs. Thus, the work presented here is not only interesting in terms of the specific results on insertional RNA editing in Myxomycetes but also much more broadly in terms of strategies to be employed if biological sequence motif searches in individual organisms fail.
The first issue we address is to what extent the recently achieved complete knowledge of the edited Physarum transcriptome [25] improves computational prediction of genes and editing sites in the Didymium mitochondrial genome. Computational prediction of insertional RNA editing in the absence of a reference transcriptome has been presented before [11], [19], [20]. The method uses a position specific scoring matrix (PSSM) built from protein sequences from other organisms. It finds the editing sites for a given genomic sequence that translate to the putative protein with maximum similarity to the protein family described by the PSSM. With this method, we prepared predictions of editing sites in Didymium using the PSSMs made before the Physarum transcriptome was known [11], [19], [21]. Since the edited Physarum transcriptome is now known [25], new predictions utilizing that information were created and compared to these baseline PSSM predictions. Given that Physarum and Didymium both exhibit RNA editing and are closely related, the purpose of this process was to see how much the new predictions, which include transcriptome information from Physarum, improve when compared to the baseline.
In order to be able to evaluate and compare the prediction quality we applied our prediction methods to mitochondrial genes in Didymium for which the editing sites had already been determined experimentally [22], [23]. In total we created five predictions of editing sites for every gene. The first was the baseline prediction using the PSSM developed before the Physarum transcriptome was known as described above. The second was a Physarum based PSSM. The Physarum based PSSMs were created from the NCBI website by starting a PSI-BLAST search [26] with the homologous Physarum protein sequence for each gene rather than with a homologous protein sequence from a more distant organism as in the creation of the original PSSMs. Three iterations of a protein PSI-BLAST search were run for each of the sixteen genes for which the Didymium editing sites are known (see Methods section). During this process we manually excluded the Didymium protein for which the prediction was being made from the model building. Since the first round of PSI-BLAST did not find any homologs when starting from the Physarum protein sequence for atp8, we could not create a Physarum based PSSM for atp8 and excluded it from further analysis. The remaining three predictions did not use a PSSM summarizing the properties of a whole family of homologs. Instead, the plausibility of a putative Didymium protein sequence (generated by inserting Cs into the Didymium genomic sequence and translating the result) was quantified by aligning the putative Didymium protein directly to the known Physarum homolog. Since alignment scores depend on the scoring matrix and different matrices are tuned toward different evolutionary distances, we prepared one prediction each using the BLOSUM62, BLOSUM75, and BLOSUM90 matrices.
We scored the accuracy of a prediction by counting the number of correct and incorrect predictions made by each prediction method (see Methods). We report the results as a percentage of editing sites in each category relative to the number of predictions the method included overall. The percentage of predicted editing sites is a better indicator than the absolute number since the computational model often does not make predictions for editing sites near the ends of genes. The results for each of the five prediction methods among all fifteen considered genes are shown in Figure 1(a).
Figure 1(a) indicates that the PSSMs created without using Physarum generate the worst predictions showing that the inclusion of the Physarum genome does increase the accuracy of the prediction; a finding that is expected due to the similarities between the organisms and their shared RNA editing. Interestingly, the results show that predictions using the Physarum protein alone outperform the predictions using either of the two PSSMs. Among the predictions that use only the Physarum protein the prediction accuracy increases as the BLOSUM matrices are tuned toward more closely related organisms. This result implies that the organisms are so similar that the inclusion of the genetic information of other organisms into the PSSMs actually decreases the accuracy of the editing site predictions.
Since the prediction method relies on sequence homology we wanted to determine the influence of sequence similarity on the prediction quality. To this end we separated the fifteen genes into a more conserved and a less conserved group (see Table S1) based on the nucleic acid conservation of the second codon position between the known Physarum and Didymium mRNA sequences. The prediction accuracy for all five methods in each of these two groups is shown in Figure 1(b) and (c).
The overall trend in prediction accuracy is the same for the more highly conserved and the more diverged set of genes. However, although one might have expected that using the Physarum protein works better for genes where the two organisms have diverged less from each other, the improvement in prediction accuracy by using the Physarum protein sequences is actually bigger for the more diverged set of genes than for the more conserved set of genes. We rationalize this by arguing that genes with less conservation between Physarum and Didymium are generally under less evolutionary pressure and thus will also have diverged more between the Myxomycetes in general and the other organisms used to build the baseline PSSMs. Thus, including proteins from other organisms in the prediction hurts the prediction accuracy more for less conserved genes.
Encouraged by the quality of the predictions on the already known Didymium genes, we proceeded to use the method to search for other genes in the Didymium mitochondrial genome and to predict the editing sites that are present in those genes. In order to do this, the genes first had to be located within the partial Didymium mitochondrial genome available to us. This was accomplished by scoring segments of the partial genome against the corresponding proteins from Physarum as described before [20]. We used only the Physarum sequence and the BLOSUM90 matrix since this approach performed best on the known genes as described above. Once the location of the genes were identified as the segments with the highest score, the editing sites were predicted. The results are shown in graphical form for each of the eight genes we identified in Figure 2. The predicted mRNA sequences with C insertions indicated as upper case C's are given in Table S2. We note that the predictions for nad2 only include part of the gene; a region at the 5′ end is not present as it is missing from the partial genomic DNA sequence available to us.
The two genes that stand out by their very low number of editing sites are nad3 and rpS11. Indeed, nad3 is already known to be unedited [24]. Our prediction resulted in a single editing site in nad3 toward the end of the gene. While this addition of a single predicted editing site was not expected, it is understandable since the prediction of editing sites becomes more challenging toward the ends of the gene. Six editing sites were found in rpS11; one was found near the 5′ end, three in close proximity of each other in the middle, and two were at the 3′ end. Because of the low number of editing sites and the striking pattern of the predicted editing sites we hypothesize that rpS11 is also unedited just like nad3. The additional predicted editing sites at the end are easily understood based on the overall low prediction accuracy at the end of genes. Since the three predicted editing sites in the middle of the gene are close to each other they can also be a prediction artifact; omitting them would only change the protein sequence over the range of 5 amino acids. We verified that omitting the three editing sites would not create an in frame stop codon in the middle of the protein. Thus, is is plausible that the edited rpS11 mRNA could have been reverse transcribed and inserted into the genome of Didymium as it has been hypothesized for nad3 [24]. We note, however, that while nad3 also has a very much reduced number of editing sites in Physarum (around one every nucleotides rather than the usual one every nucleotides), rpS11 shows the normal level of editing in Physarum.
While the graphs presented in the preceeding section convey the successes of the various computational methods for predicting the editing sites in the genes of Didymium with known editing sites, there can be no similar comparison of the successes of the predictions in the new genes shown here as their exact editing sites are not known. However, sequence alignment of the new genes does show that of these genes, rpS4 and rpS11 fall into the less conserved group while cox3, nad1, nad2, nad4, and nad5 (as well as nad3) all fall into the more conserved category. Using this information, an estimate about the success of the BLOSUM90 computational method can be made for the new genes based the known results of the BLOSUM90 prediction method for the sequenced genes. This estimate results in Figure S1 which show the expected number of correct sites, the expected number of sites one or two sites from the actual editing site, and the expected number of more wrongly predicted sites with the associated errors. A true assessment of the success of these predictions will of course have to wait until these RNAs are fully sequenced and their editing sites are known.
As indicated in the introduction, one of the major questions to be resolved is how the RNA editing machinery knows which sites to edit. Previous studies [10], [11], [25] have looked for sequence patterns in Physarum alone. One property within the mitochondrial genome of Physarum is that editing sites have a strong preference to occur after a combination of a purine and a pyrimidine [10], [11]. However, many editing sites do not follow this pattern and many purine-pyrimidines are not followed by an editing site. Thus, this pattern alone cannot explain the extremely reliable recognition of editing sites and the problem of editing site recognition remains unsolved.
The absence of discernable sequence patterns among the Physarum editing sites might suggest that every site (or small groups of them) are recognized individually. Such mechanisms exist in the kinetoplastids in the form of guide RNAs [27], [28] and in plant chloroplasts and mitochondria in the form of PPR proteins [6], [7], [29]. If every editing site is recognized individually, no sequence pattern will emerge when comparing the sequences surrounding all the editing sites in one organism consistent with previous studies in Physarum [10], [11], [25]. However, when comparing organisms at sequences surrounding their shared editing sites, sequence positions that play a role in site recognition are under increased evolutionary pressure and should thus show more conservation across species than sequence positions not involved in editing site recognition. Thus, instead of looking at one organism at a time, we here used the edited genes of two related organisms with insertional RNA editing, Physarum and Didymium, and examined the patterns of sequence conservation between the two organisms. We looked at the nucleotide identities at fixed positions relative to the editing site and investigated whether these nucleotides were conserved between the two organisms or not. In this analysis, we tried to identify positions relative to the editing site with statistically significantly increased degree of conservation from the background, which would indicate functional importance.
We studied the sixteen genes for which the editing sites are known in both organisms as described in the Methods section. Comparing the complete mRNA sequences of the two organisms, we determined the overall degree of conservation for the first, second and third codon position. This yielded the background frequencies or the “expected” frequencies at the first, second, and third codon position. Table 1 presents these background frequencies for conservation between Physarum and Didymium. The degree of conservation in these genes is relatively high and may not leave enough room to be significantly increased. Thus, we also studied as another group the subset of the 8 less conserved among the 16 genes (i.e., the genes the background frequency at the 2nd codon position of which is less than 85%, see the Methods section). It can be seen in table 1 that the two groups share similarities in their background levels of conservation that are to be expected: the second codon position has the highest conservation, while the third codon position is the least conserved. However, there is a clear difference in the amount of conservation between the two groups as expected by construction of the less conserved group.
In order to study the vicinity of the editing sites shared by the two organisms, we first identified those C insertional editing sites that are shared and unambiguous (i.e., at least in one of the two organisms the neighboring nucleotides are not Cs). Because of the variations of background frequencies among different codon positions, these editing sites were separated by codon position. Table 2 shows the number of these shared editing sites for each codon position.
A previous study demonstrated that the DNA necessary for C insertion is contained within 9 or maybe 10 base pairs on either side of the editing site [18]. Thus, we first determined the conservation information of the flanking sequences within a window of 9 positions upstream and downstream of each of the shared editing sites. Then we examined the difference between the observed sequence conservation in the vicinity of the shared insertional editing sites (at positions −9 to +9 relative to editing sites) and the background conservation.
Figure 3 shows the observed and the expected degree of conservation for positions −9 to +9 (relative to the shared editing site) for all genes separately for the shared editing sites at the first and third codon position (we do not show the data for the second codon position because of the small number of these editing sites which results in very low statistical significance). Results for the less conserved group are similar to the results for all genes (see Figure S2). From these figures, we can see that both the observed frequency and the background frequency are position dependent with codon position being the dominant factor. The observed frequency is higher than the background frequency at some positions, while at other positions the observed frequency is lower.
In order to see whether these variations are statistically significant, we calculated the probabilities for observing increased or decreased sequence conservation in the vicinity of the shared insertional editing sites based on the binomial distribution (see details in the Methods section). Figure 4 shows those probabilities for positions −9 to +9 for all genes. No significant -values are obtained (to take into account multiple testing, we use as the -value cut off), which implies that there are no statistically significant variations between the observed frequencies and the background frequencies. In spite of the larger room for increased conservation, the results for the less conserved group are similar to the results for all genes (see Figure S3).
We also extended our study to positions that are further away from the editing site than 9 nucleotides. For these positions, the analysis is complicated by the fact that additional editing sites can occur between the position to be studied and the editing site of interest thereby mixing different codon positions at the same position relative to the editing site of interest [10]. We circumvent this problem by eliminating all primary editing sites from the analysis that have an additional editing site between the primary site and the position we are interested in. The disadvantage of this approach is that as one studies positions that are further and further away from the primary editing site, there are less and less sequences that contribute to the analysis and thus the statistical power decreases. In practice, we reached a limit of contributing sequences at a distance of nucleotides for editing sites at the third codon position and at a distance of nucleotides for editing sites at the first codon position. However, even for these distances no statistically significant increase of sequence conservation was found (data not shown). This suggests that the information on editing site location is not contained within the sequence in the immediate vicinity of the editing site at least at the level of statistical significance set by our sample size.
This leaves us with the conundrum that on the one hand Rhee et al. [18] demonstrate experimentally that only 9 nucleotides of DNA on either side of the editing site are required for editing and on the other hand our results suggest that there is no statistically significant pattern of conservation within 9 (or even more) base pairs on either side of the editing site. Thus, we propose several possible explanations for the discrepancy between Rhee et al.'s findings and our results.
The first and second explanation lead us to consider how much increase in conservation for recognition sites of editing events we should see given the size of the mitochondrial genome of Physarum (62862 bp). Due to the extreme precision of RNA editing in Physarum, the recognition site of each editing event should be unique in the mitochondrial genome of Physarum. According to Rhee et al., the 9 nucleotides immediately upstream DNA and 9–10 nucleotides immediately downstream DNA of the editing sites are necessary and sufficient for editing site recognition. If the actual information on the editing site position is stored within these nucleotides this implies that the pattern recognized within the set of 18–19 nucleotides should occur at the rate of at most 1/60000 in a random DNA sequence.
Based on our calculations (see Methods section), the lowest conservation for a set of 19 nucleotides that still allows specification of a site within the genome is 80.7%, i.e., at least we should see 80.7% conservation in a 19 nucleotide region responsible for editing site recognition. To test whether a conservation of 80.7% or more would show up as a significant difference between the expected frequencies and the background frequencies at our sample size, we set 80.7% (the lowest expected conservation) as the “observed frequency” for positions −9 to +9 (i.e., we used 80.7% to replace the real observed frequencies). Then we calculated the -values for observing increased sequence conservation relative to the background frequencies in Table 1. For putative motif positions at the third codon position in the vicinity of editing sites at the third codon position we found a highly significant (compared to the cutoff of ) -value of . Thus, according to this analysis, we should have seen statistically significant variations between the actual observed frequencies and the background frequencies at our sample size even if the editing machinery does not recognize the same positions relative to the editing site at all of the sites. We thus conclude that the observed degree of conservation is significantly lower than what is to be expected when only the 9 nucleotides upstream and 10 nucleotides downstream of the editing sites contain the information for editing site recognition even if different sites use different combinations of the 19 nucleotides to specify the editing site location. These studies therefore suggest that the first and second of the hypotheses above can be ruled out.
As another test of which aspect of sequences around editing sites could determine the editing position, we tested the specificity of sequences around editing sites. In practice, we started by looking only at sequences immediately downstream of editing sites and examined the uniqueness of these sequences in Physarum, that is, we tested for every sequence of nucleotides (-mer) downstream of an unambiguous C insertion site in the known transcriptome of Physarum if this -mer only occurs downstream of C insertional editing sites, but does not occur following non-edited sites of the sequences. This analysis is especially powerful since the full transcriptome has recently been determined by a high throughput sequencing experiment [25] thereby giving complete access to all editing and non-editing sites for this analysis that compares all editing sites to all non-editing sites in all transcripts.
Since it is unknown if editing site recognition occurs at the DNA or RNA level, we tested the -mers in both the unedited sequences and the edited sequences. We asked which is the largest for which we can still find a -mer that occurs at least once immediately downstream of an unambiguous C insertion site and at least once in a position that is definitely not preceded by an editing site. Both on the RNA and on the DNA level the largest -mer we found was a -mer. Thus, we conclude that a mechanism that uses only the downstream sequence of an editing site to specify the editing event, even if it is a different mechanism for every editing site, must use at least nucleotides downstream of the editing site. Similarly, we found one -mer combination that is not unique for unambiguous C insertional editing sites when testing the unedited sequences and one -mer when testing the edited sequences when studying only sequences immediately upstream of the unambiguous editing sites.
Given that Rhee et al. found that the 9 or maybe 10 nucleotides of DNA both downstream and upstream of the editing sites are responsible for the editing event, we also investigated the specificity of all the possible nucleotide combinations upstream and downstream of the unambiguous C insertional editing sites within 9 nucleotides on either side (describing them as , where is the number of nucleotides upstream of the editing site and is the number of nucleotides downstream of the editing site with and ). Since in this case the unambiguous C is inserted inside the motif, testing the uniqueness of these motifs is different between the unedited sequences and the edited sequences. It is the same as before for the unedited sequences. For the edited sequences, we asked if the motif without the inserted C occurs anywhere in the transcriptome (in addition to the at least one occurence with the inserted C). We found possible combinations that are not unique for the unambiguous C insertional editing sites for both unedited sequences and edited sequences up to 9+5, 8+7, 7+8 and 6+9 nucleotide combinations. Therefore, the recognition region (if it exists) includes at least 9+6, 8+8, or 7+9 positions in agreement with Rhee et al.'s finding [18] that the whole 9+9 nucleotides are required for editing. It is important to note that since this particular analysis does not rely on the comparison of two organisms but uses only Physarum data it even in the case of hypothesis 3 (that the mRNA itself templates the editing sites) implies that the recognition has to involve at least the 9+6, 8+8, or 7+9 nucleotides surrounding an editing site. We would like to conclude by noting that while we did not find a non unique 9+9-mer which would rule out the “9+9” (or any larger) model (the identity of the 9 nucleotides both downstream and upstream of the editing sites carry the information for the editing event), we would not have expected to find one on statistical grounds alone since the probability for an 18-mer to occur in the mitochondrion of Physarum is extremely small even if there is no biological reason (uniqueness of the recognition sequence) that prevents it from occurring. We want to emphasize again, that the conclusions in this section do not rely on a common mechanism that simultaneously recognizes all editing sites in one organism but apply even to mechanisms that recognize every editing site individually such as guide RNAs or site recognition proteins. Of course, all these considerations only exclude the information for the editing site positions to be stored within the identities of the 9+5, 8+7, 7+8, or 6+9 nucleotides surrounding the editing site - it is still possible that the DNA in the immediate vicinity of the edited site carries a “marker” that is placed based on information elsewhere in the genome.
For editing sites within the coding regions, a significant codon bias is known. It has been found that in the mitochondrial genome of Physarum, the third codon position has the largest number of C insertional editing sites, while the second codon position has the lowest number [9], [10], [25], [30]. As shown in Table 2 the codon bias is also significant for the shared C insertional editing sites in both groups of genes.
A previous study proposed an evolutionary model which explains this codon position bias [31]. The general idea of this model is the following. During the proliferation of Physarum, nucleotide mutations (including substitutions, insertions, and deletions) occur at random positions in the mitochondrial DNA sequence. In the case of random deletions, the offspring can not survive because of the incorrect protein sequence since the mutated DNA sequence is out of frame. However, the editing machinery sometimes may insert back nucleotides to the positions of deletions and preserve the correct reading frame. In this case, the offspring can survive and proliferate. This idea of random creation of new editing sites is also consistent with phylogenetic data [32].
The net effect of a nucleotide deletion followed by the creation of a new insertional editing site is that the original nucleotide will be replaced by a C. The genetic code is organized such that the third codon position is the most irrelevant for the identity of the amino acid while the second codon position is the most relevant. Therefore, the third codon position is the least sensitive to nucleotide changes to C generated in editing events while the second codon position is the most. Thus, random deletions at the third codon position will have the highest survival rate and the lowest for the second codon position.
According to this model, the codon position bias in Physarum is mainly a consequence of random mutations with selection at the protein level [31]. This implies that genes that are under stronger selection should have a stronger codon position bias in their editing sites as well. Since for this study we have two organisms, we can directly determine the strength of selection on each gene from the sequence conservation. Thus, to test the theory proposed in [31], we examined the relationship between the strength of the codon bias and the overall sequence conservation in both Physarum and Didymium.
As described in the Methods section, the 16 genes were divided into several groups according to their overall sequence conservation at the second (most conserved) codon position between Physarum and Didymium. The detailed group information is shown in Table S1 (since the conservation at the second position of the 16 genes ranges from 60% to 100%, we separated them into four groups by splitting the range from 60% to 100% into four intervals of equal length). We used the ratio of the number of third codon position editing sites and the number of second codon position editing sites as a measure of codon bias, and used the overall sequence conservation at the second codon position as a measure of the conservation within different genes. We examined all unambiguous C insertional editing sites in Physarum and Didymium, i.e., shared editing sites as well as editing sites specific for either of the organisms. The solid black squares in Figure 5 illustrate the relationship between the codon bias and the conservation at the second codon position. In order to reduce statistical fluctuations, we also considered a grouping of the genes into only two groups by combining data of all genes with conservation between 60% and 80% at the second codon position into one group and the genes with conservation between 80% and 100% into the other group. Figure 5 shows that, whether the 16 known genes were separated into four groups or into two groups based on their conservation at the second codon position, genes with higher conservation at the second codon position have a higher ratio of . This difference is significant even when statistical errors within the ratios are taken into account. This demonstrates that genes under stronger selection (or with higher conservation) should have a stronger codon position bias in their editing sites; thus reinforcing the theory that codon bias is a consequence of evolutionary pressure on the protein sequence.
In order to increase the statistical significance it would be beneficial to include more genes in the study. Given our work presented above, the seven newly predicted Didymium genes and nad3 are likely candidates to add to the study. The problem with this idea is that the predicted mRNA sequences most likely deviate from the (unknown) true mRNA sequences, which might affect the accuracy for both the overall sequence conservation and the codon bias. Since we have the predicted Didymium mRNA sequences for all the 16 genes for which the actual mRNA sequences are known, we can test how much the estimates of overall sequence conservation and the codon bias differ between the predicted sequences and the true sequences.
To this end, we aligned the predicted Didymium mRNA sequences and the real Physarum mRNA sequences (see Table S3 for accession numbers) to obtain the conservation at the second codon position. Then we plotted the conservation data between the predicted Didymium mRNA and the real Physarum mRNA versus the conservation data between the real Didymium mRNA and the real Physarum mRNA. (see Figure 6(a)). We found, that the overall conservation at the second codon position for the predicted sequences (predicted Didymium mRNA and real Physarum mRNA) and the real sequences (real Didymium mRNA and real Physarum mRNA) are very close to each other except for possibly two genes – atp8 and atp9 – which are much shorter than the other genes. This implies that estimating the overall sequence conservation from the predicted Didymium mRNA sequences is a valid procedure since the difference in conservation by using the predicted sequences and the real sequences is small.
In the same way, we compared the codon bias between the predicted sequences and the real sequences. We treated the predicted Didymium sequences in the same way as the real Didymium sequences before (we examined all unambiguous C insertional editing sites and used four and two groups). As can be seen from Figure 6(b), the codon biases for the predicted sequences and the real sequences are equal within the error bars. This suggests that codon biases calculated from the predicted Didymium sequences can reasonably be used in lieu of exactly known codon biases. However, the agreement between predicted and true sequences is not as strong for the codon bias as it is for the conservation at the second codon position.
Since the overall sequence conservation and the codon bias show only small deviations between the predicted sequences and the true sequences, we can add the seven newly predicted Didymium genes and nad3 to the codon bias analysis. In the same way as described for the 16 known genes, we analyzed the codon bias of these eight genes using the predicted Didymium sequences and real Physarum sequences. Since the determination of conservation at the second codon position from predicted sequences is more robust with respect to prediction errors than the determination of codon bias (see Figure 6) we performed this analysis twice. First, we only used the (known) codon bias in Physarum for the eight predicted genes (indicated in Figure 5 as 16 known genes + 8 genes in Physarum), thus only using the predicted Didymium mRNA sequences to determine the overall conservation for group division of each gene, but not using the codon bias in the predicted Didymium mRNA sequences which is less robust with respect to prediction errors than that for overall conservation. Second, we also included the predicted editing sites in Didymium for the eight additional genes in the analysis (indicated in Figure 5 as 16 known genes + 8 genes in Physarum and Didymium). In this case, all unambiguous C insertional editing sites in Physarum and Didymium for all 24 genes are counted.
Figure 5 illustrates the relationship between codon bias and the overall conservation for the 16 known genes (already described above), 16 known genes + 8 genes with editing sites in Physarum and 16 known genes + 8 genes with editing sites in Physarum and Didymium with 2 and 4 groups, respectively. As can be seen from this figure, the strength of codon bias of the 24 genes (including the known genes and predicted genes) is not as strong as in the 16 known genes. However, given the reduced error bars the dependence of codon bias on selection pressure remains statistically significant. We have thus shown that more conserved genes have more significant codon bias in all unambiguous C insertional editing sites in Physarum and Didymium as suggested by the previous theory [31].
The mitochondrial genomes of two related organisms with insertional RNA editing, Physarum polycephalum and Didymium iridis were studied. Sixteen genes and their mRNA sequences from the two organisms were included in this study: atp1, apt6, atp8, atp9, cox1, cox2, cytb, nad4L, nad6, nad7, rpL2, rpL16, rpS3, rpS7, rpS12, and rpS19. All the sequences were downloaded from GenBank; see Table S4 for accession numbers. For several of our studies the sixteen genes were divided into groups according to their overall sequence conservation at the second codon position between Physarum and Didymium, which was obtained by aligning the mRNA sequences of each gene between the two organisms. Table S1 indicates for each gene which group it was assigned to.
The predicted editing sites were scored as either correct, one away, two away, or three or more sites away from the actual editing sites by comparison with the known mRNA sequences. We only scored C insertion sites, i.e., we ignored predicted insertion sites in close vicinity of thymine, adenine, guanine, or dinucleotide insertion sites in the known mRNA sequences. Also recorded was the number of editing sites included in each prediction due to occasionally missed editing sites at the beginning or at the end of a gene. Omissions of editing sites at either end of a gene was caused by not having a significant number of bases either before or after the input basis sequence. Therefore, the missed editing sites in these instances were due to the lack of information input into the computational method which results in poor conservation of the protein sequence of the gene. Thus, missed editing sites at the beginning and end of a gene sequence were not scored. While these types of predictions were not scored, occasionally an editing site would be missed or added by the prediction in the interior of a gene. Missed or added interior editing sites most often occurred in threes which preserves the reading frame and is most likely to conserve the protein sequence; interior missed or added editing sites were scored as three or more sites from the actual site.
For each gene, four sequences (Physarum-DNA, Physarum-mRNA, Didymium-DNA, and Didymium-mRNA) were aligned in Clustal X [33]. From these alignments, the C insertional editing sites that are unambiguous (i.e., at least in one of the two organisms the neighboring nucleotides are not Cs) and shared between Physarum and Didymium were identified. The flanking sequences within the window of 9 positions upstream and downstream of each of these editing sites in both organisms were investigated. Then these flanking sequences were turned into patterns of “0” and “1” where “1” means that the two organisms have the same base at the same relative positive and “0” means that they do not. The shared and unambiguous editing sites were separated by codon position.
Comparing the mRNA sequences of the two organisms, we obtained the overall conservation information for the first, second and third codon position by counting the “1”s in each codon position across the whole genes. This yielded the background frequencies or the expected frequencies of “1”s at the first, second, and third codon position.
In order to see whether variations from background were statistically significant, the probabilities for observing increased (or decreased) sequence conservation in the vicinity of the shared insertional editing sites were calculated. These probabilities were calculated based on the binomial distribution: The background frequency or the expected frequency of “1”s at the codon position is . The total number of shared editing sites at the 'th codon position is . For a specific position in the vicinity of the shared editing site, we can easily identify its codon position (see Table S5) and the actual number of “1”s in these samples. Thus, the observed frequency of “1” at this specific position is . Therefore, the probability of the observed increased sequence conservation is:If the observed frequency of “1” is less than the “expected” frequency, the -value was calculated analogously as the probability of observing the decreased sequence conservation.
In order to determine the codon bias in insertional RNA editing, the number of third codon editing sites and the number of second codon editing sites were counted. The codon bias was then quantified as their ratio . If we assume that the error for is just counting error given by the square root of (i.e., ), the statistical error of is
In order to know how much increase in conservation for recognition sites of editing events we should see in the mitochondrial genome of Physarum (62862 bp, NC_002508) if the region containing the editing site information is limited to the 18–19 nucleotides surrounding an editing site identified in Rhee et al. [18], we calculated the lowest conservation for a set of 19 nucleotides that allows specification of a site within the genome.
Since the effect of GC content in the Physarum mitochondrial genome is strong (the GC content is approximately 25% [34]), we considered the frequency that a set of 19 nucleotides occurs in a random DNA sequence with the same length (62862 bp) as well as the same GC content as the mitochondrial genome of Physarum. In such a sequence, the probability of two nucleotides being equal by chance is . Therefore, the probability for two sets of 19 nucleotides in the sequence described above being the same is (i.e., the occurring rate for such a combination is ), which is much lower than one in Physarum's mitochondrial genome. Thus, we relax the constraints on a set of 19 nucleotides that will still specify the editing site (decrease the number of nucleotides that are fixed) as long as the frequency of the relaxed constraints is not (much) higher than . We found that the occurring rate of a motif in which only 9 of the 19 nucleotides are fixed (and the other nucleotides could occur randomly) is , which is close to one per Physarum mitochondrial genome. We thus conclude that to uniquely specify a site by a 19 nucleotide motif, at least 9 of these nucleotides have to be fixed while the others can be variable.
We do not know which 9 (or more) of the 19 nucleotides are fixed for a given editing site, but we can calculate the average conservation generated by these fixed nucleotides. This average conservation for a set of 19 nucleotides is calculated as following: The conservation for each of the 9 fixed nucleotides is 100% while it is at least 63.3% (the lowest conservation between Physarum and Didymium we obtained, see Table 1) for the 10 random nucleotides. Thus, the average conservation is at least .
|
10.1371/journal.pcbi.1004699 | Josephin Domain Structural Conformations Explored by Metadynamics in Essential Coordinates | The Josephin Domain (JD), i.e. the N-terminal domain of Ataxin 3 (At3) protein, is an interesting example of competition between physiological function and aggregation risk. In fact, the fibrillogenesis of Ataxin 3, responsible for the spinocerebbellar ataxia 3, is strictly related to the JD thermodynamic stability. Whereas recent NMR studies have demonstrated that different JD conformations exist, the likelihood of JD achievable conformational states in solution is still an open issue. Marked differences in the available NMR models are located in the hairpin region, supporting the idea that JD has a flexible hairpin in dynamic equilibrium between open and closed states. In this work we have carried out an investigation on the JD conformational arrangement by means of both classical molecular dynamics (MD) and Metadynamics employing essential coordinates as collective variables. We provide a representation of the free energy landscape characterizing the transition pathway from a JD open-like structure to a closed-like conformation. Findings of our in silico study strongly point to the closed-like conformation as the most likely for a Josephin Domain in water.
| Proteins are fascinating molecular machines capable of organizing themselves into well-defined hierarchical structures through a huge number of conformational changes to accomplish a wide range of cellular functions. Protein conformational changes may be characterized by transitions from a low-energy conformation to another. Computer simulations and in particular molecular modelling may be able to predict protein transition dynamics and kinetics, thus playing a key role in describing protein tendencies towards specific conformational rearrangements. Approaching this problem from an energetic point of view is of great importance especially in case of amyloidogenic proteins, given the intimate interconnection between the functional energy landscape and aggregation risk. In this work we have employed molecular modelling techniques to shed light into conformational dynamics and kinetics of the Josephin Domain, part of the protein Ataxin 3, which is responsible for the spinocerebbellar ataxia 3, also called Machado Joseph Disease. In greater detail, we have employed enhanced sampling approaches to provide an estimation of the free energy landscape characterizing the transition pathway among several known molecular arrangements of the Josephin Domain.
| Proteins are fascinating molecular machines capable of organizing themselves into well-defined hierarchical structures through a huge number of conformational changes, in order to accomplish a wide range of cellular physiological functions. From an energy landscape point of view, protein conformational changes may be characterized by transitions from a low-energy conformation to another. In this connection, computational approaches have widely demonstrated their utility by providing important insights into the protein conformational features [1–5]. Molecular Dynamics simulations, and in particular enhanced sampling techniques, are able not only to predict protein transition pathways, but also to quantify the free-energy landscape along selected reaction coordinates, thus playing a key role in describing protein tendencies towards specific conformational rearrangements. Approaching this problem from an energetic point of view is of great importance especially in case of amyloidogenic proteins, given the intimate interconnection between the functional energy landscape and aggregation risk [6].
The Josephin Domain (JD), i.e. the N-terminal domain of Ataxin 3 (At3) protein, is an interesting example of competition between physiological function and aggregation risk [6,7]. In fact, the fibrillogenesis of Ataxin 3 is responsible for the spinocerebbellar ataxia 3, also called Machado Joseph Disease (MJD). Structurally, At3 is composed of a structured globular N-terminal region (i.e. the JD, residues Met1-Arg182 in the human protein), combined with a more flexible C-terminal tail that contains the polyQ tract and the Ubiquitin Interacting Motifs (UIM) [8,9].
The expansion of polyglutamine (polyQ) tract in Ataxin 3 (so-called expanded At3) is considered a cause for protein misfolding and aggregation, but the underlying mechanism remains to be elucidated.
Although it is known that the polyQ tract is necessary for kinetic instigation of an aggregation mechanism [10–14], several experimental studies support the hypothesis that JD structural stability could play a major role in determining the aggregation features and toxicity of polyQ proteins [7,15–21].
In this regard, experimental evidences have suggested a two-stage pathway for At3 fibrillogenesis: the first, JD-mediated and the second, polyQ-dependent [19,22,23]. Fibrillar aggregates of both not-expanded At3 and isolated JD have shown markedly similar morphological and mechanical properties, suggesting a leading role for the JD in the mechanism of fiber formation [17]. Moreover, inhibition of JD self-association by a small heat-shock protein significantly slows down expanded At3 aggregation [24]. For these reasons, the role of JD has been the subject of a robust debate in the past [6,7,18–21,25,26].
To date, several JD models solved by NMR are available in the literature (PDB entry 2JRI [27], 1YZB [28], 2AGA [29] and 2DOS [30]—UNIPROTID: P54252). Differences in the available models are located in the hairpin region (region α2-α3, residues Val31-Leu62). In particular, the 1YZB and 2JRI models are characterized by a “half-open” and “open” L-shape hairpin conformation, respectively. On the other hand, the “closed” 2AGA and “half-closed” 2DOS models exhibit the hairpin region packed against the rest of the globular structure [30]. Whereas all the above-mentioned NMR data have demonstrated the existence of several different available conformations for the JD, issues concerning i) the likelihood of JD achievable conformational states in solution, and ii) the role played by environmental conditions (such as the solution’s pH) and interacting physiological partners (such as ubiquitin) in JD conformational arrangement are still unresolved. Specifically, results from a recent characterization of the JD free energy landscape using MD simulations suggested the open-like model as the most representative of the JD structure in solution [31]. Nevertheless, other previous experimental and computational studies strongly support the idea that JD has a highly flexible extended hairpin in dynamic equilibrium between open and closed states [1,30]. In a very recent in silico study, the early stage of the JD-JD dimerization mechanism [1] has been investigated by MD and indicates that the JD-JD binding might play a role in determining the kinetics of hairpin opening/closure. However, the previous computational investigation is limited to a classical MD approach with a relatively short simulation timescale [1].
In principle, to prove the JD open-like or the JD closed-like configuration as favored, it would be necessary to show not only that i) there is more sampling in one state during a classical MD, but also that ii) several transitions between states are sampled during the simulation. Hence, an accurate evaluation of the JD conformational changes requires a longer simulation time-scale and robust sampling methods. In this regard, enhanced sampling methods represent a powerful tool to improve the sampling efficiency of classical MD [32–40], including those that artificially add an external driving force to guide the protein from one structure to another [38,41]. Moreover, reducing the dimensionality of the trajectory obtained from MD simulations can help identify the dominant modes in the motion of the molecule [38,41].
Motivated by the still open debate regarding the most representative JD structural arrangement [1,30,31,42,43], we have carried out an investigation on JD conformational changes using both unbiased MD and Metadynamics guided by essential coordinates. In this work, we provide an estimation of the free energy landscape characterizing the transition pathway from a JD open-like to a closed-like structure (which is henceforth called the folding pathway). The findings of our in silico study strongly suggest the closed conformation as the most likely for a Josephin Domain in water.
The 1YZB model [28,42] of JD was selected as the starting point for the present work. The rationale for this choice is in the experimental work of Nicastro et al. [42] indicating a satisfactory validation of the 1YZB model through the application of an arsenal of tools for checking the quality, accuracy and mutual consistency of the structures available.
The 1YZB model was solvated in a dodecahedron box where the minimum distance between the protein and the edge of the box was 1 nm, resulting in a molecular system of about 50,000 interacting particles. The net charge of the system was neutralized by the addition of Cl− and Na+ ions. AMBER99-ILDN force-field [44–46] and water TIP3P model [47] were chosen to describe the system’s topology. Particle-Mesh Ewald method with a short-range cut off of 1.2 nm was applied to treat electrostatics. A cut-off of 1.2 nm was also applied to Lennard-Jones interactions.
The system was minimized by the steepest descent energy minimization algorithm (1000 steps). Then, in order to increase the statistics of MD data, five replicas, differing in initial atom velocities, were created from the minimized system. In particular, for each replica, a random velocity taken from a Maxwell-Boltzmann distribution at 310 K was assigned to every atom of the system (i.e. JD, water and ions). A position-restrained and production MD simulations were carried out as described in the following. Two subsequent MD simulations (500 ps, and 100 ps, respectively) were run in the NPT ensemble, applying position restraints of 1000, and 100 kJ/mol/nm2, respectively, to the JD Cα atoms. System temperature was set at 310 K by using the v-rescale [48] thermostat with a coupling time step of 0.1 ps. Moreover, in NPT simulations a Berendsen barostat [49] was also employed with a reference pressure of 1 atm and a coupling time step of 1.0 ps. A third position restrained dynamics simulation (100 ps) was carried out by applying a force constant of 10 kJ/mol/nm2in the NVT ensemble at 310 K. Finally, an unrestrained production MD of 500 ns was run in the NVT ensemble at 310 K, as done in several previous Molecular Dynamics studies [50–52]. GROMACS 4.6 package was employed for all MD simulations and data analysis [53]. Ensemble data taken from all production MD trajectories of the above-mentioned five replicas (each simulated for 500 ns) were used for JD conformational analysis. The Visual Molecular Dynamics (VMD) package [54] provided the visual inspection of the simulated systems. Dedicated GROMACS tools were used for quantitative analyses in terms of Root-Mean-Square Deviation (RMSD) and Root-Mean-Square Fluctuation (RMSF). The secondary structure of the protein has been calculated by the STRIDE software [55] on several snapshots along the simulation time.
The identification of JD conformational transitions from open to closed JD conformations has been carried out by employing quantities which have already been demonstrated to be meaningful in describing the JD transition pathway: the radius of gyration (RG) and the hairpin angle [1,31].
Given that the NMR models (1YZB, 2JRI, 2DOS, and 2AGA) considered in this work present a different number of residues, the RG has been calculated by considering all the residues in common among the above mentioned PDB models. In detail, the residue range 1Met-171Asp (according to 1YZB numbering) has been chosen.
The hairpin angle was calculated from the centers of mass of the Cα atoms from three distinct JD regions: globular subdomain (residues 111–113, 122–125 and 162–165), hinge (residues 32–35) and loop (residues 45–48 and 58–61) [31] (Section 3 in S1 Text).
Principal Component Analysis (PCA) was applied to classical MD trajectories. PCA is an established method which allows to elucidate large-scale and low-frequency modes, respectively, yielding collective motions directly related to a specific molecular event [56]. In detail, after the alignment of the JD Cα Cartesian coordinates, the covariance matrix was calculated and diagonalized (Section 2 in S1 Text).
The free energy landscape representing the JD folding pathway was investigated by means of Metadynamics [58,59], a powerful technique for enhancing the sampling in MD simulations and reconstructing the free-energy surface as a function of few selected collective variables (CVs). The first eigenvector derived from the PCA was used as CV for a well-tempered Metadynamics simulation of 500 ns starting from the open-like 1YZB model [16]. The JD model was prepared for Metadynamics by applying system minimization and position restraint dynamics, as described above for the classical MD.
To perform Metadynamics simulations, a Gaussian width of 0.1 was used. Along the simulation, the initially prescribed Gaussian deposition rate value of 0.2 kJ/mol·ps was used and it gradually decreased on the basis of an adaptive scheme, with a bias factor of 20. The setting of Gaussian width and deposition rate was done on the basis of a well-established scheme [37,40]. In particular, the Gaussian width value was of the same order of magnitude as the standard deviation of the collective variable, calculated during unbiased simulations (production MD). The authors have also verified that the maximum force introduced by a single Gaussian distribution is smaller than the typical derivative of the free energy.
The estimation of the free energy profile was performed by employing the reweighted-histogram procedure [60,61], taking into account for the following collective variables: the projection along the first PCA eigenvector, the JD’s RG, the hairpin angle and the alphaRMSD variable. More specific information about the definition of the CVs, the convergence of the Metadynamics simulations and the free energy reconstruction is reported in Section 3 in S1 Text. GROMACS 4.6 package patched with PLUMED was employed for metadynamics simulations and data analysis [57].
As stated above, five independent replicas of a single JD in explicitly modeled water and ions were simulated for 500 ns. Structural conformational properties and stability were initially checked by monitoring the time evolution of the RMSD and secondary structure (Section 1 in S1 Text). The data generated indicated that a reasonable stability of the above-mentioned quantities has been reached in all cases in the last 100 ns of the production run of the MD simulations. Moreover, the JD secondary structure showed to be highly conserved throughout the whole simulation time (Section 1 in S1 Text).
The time evolution of the RG calculated over the classical MD trajectories (Fig 1A) reveals the JD transition for all replicas, from a half-open (starting configuration 1YZB) to a closed or half-closed conformation, characterized by RG lower than 1.6 nm (Fig 1A) and a hairpin angle lower than 80° (Fig 1B). Several intermediate half-open and half-closed conformations are explored during the MD simulation (Fig 1C). Moreover, no transition from the reached JD closed-like to the open-like structure has been detected during the simulated time.
By analyzing the same data in the form of a distribution plot (Fig 2), it is possible to observe that sampled structures in the stability region of the simulation (400–500 ns) are far from open like JD arrangement.
Secondly, it is worth mentioning the different distribution shape when considering the whole trajectory (0–500 ns, black dashed line in Fig 2) and the trajectory in the stability region (400–500 ns, red line in Fig 2). Specifically, a curve comparison of both the RG and hairpin angle distribution indicates how intermediate states have a tendency to converge toward closed like arrangements. The importance of using both the RG and hairpin angle to analyze the JD arrangement is demonstrated by looking at the NMR range values labeled in Fig 2. Namely, whereas the RG helps in discerning between half-open (1YZB) and open (2JRI) JD arrangement (Fig 2A) the hairpin angle perfectly distinguish between closed (2AGA) and half-closed (2DOS) JD (Fig 2B).
To reduce the high-dimensionality of the MD trajectory and to identify the dominant molecular phenomena related to the hairpin closure, PCA was applied. After the alignment of the JD Cα atoms, the MD trajectory was filtered to show only the motion along the first eigenvector, calculated by covariance matrix diagonalization. More information on PCA and the eigenvector values is reported in Section 2 in S1 Text. The JD Root Mean Square Fluctuation (RMSF) calculated over the filtered trajectory (Fig 3) shows that, as expected, the first PCA eigenvector effectively captures the hairpin motion (RMSFhairpin>0.5 nm).
The first eigenvector derived from the PCA was used as collective variable (CV) for a well-tempered Metadynamics simulation. Analyzing the free energy profiles reported in Fig 4A as a function of the RG, two energy wells of 36 kJ/mol and 4 kJ/mol, located at RG values of 1.55 nm and 1.78 nm, respectively, can be identified. This result is also confirmed by reweighting the free energy profile as function of the hairpin angle (Fig 4B). In this case, the deepest free energy minimum (36 kJ/mol) is found to be located at a value of the hairpin angle equal to 63°. A second minimum (5 kJ/mol) is found to be located at a value of the hairpin angle equal to 100°.
As expected, the RG and hairpin angle values corresponding to the free energy wells (Fig 4A) are in agreement with the distribution peaks obtained from the unbiased MD simulations in the stability region (400–500 ns) shown in Fig 2. This finding confirms the reliability of our Metadynamics results given that the free energy minima are expected to identify the most energetically favorable configuration.
An overall picture of the JD free energy landscape is provided in Fig 5B, showing the 2D color map of the free energy profile as a function of the RG and hairpin angle. Again, the free energy minima are expected to match the most energetically favorable JD configurations. Hence, it is interesting to compare the free energy map with the JD configurations sampled by classical MD (Fig 5A). Fig 5A also provides the snapshots derived from the JD models available in the literature. Interestingly, 2AGA, 2DOS and 2JRI models lie in regions regularly sampled by classical MD, and characterized by absolute or relative free energy minima. The most sampled configurations by classical MD, corresponding in term of RG and hairpin angle to the 2AGA model, is also the deepest energy minimum in the free energy landscape. Similar characteristics between metadynamics lowest energy state and the 2AGA model are also highlighted by contact maps reported in Section 2 in S1 Text. On the contrary, values of the RG and hairpin angle corresponding to the 1YZB, i.e., starting structure of our simulations, are merely sampled by classical MD and far from the absolute energy minimum in Metadynamics.
The characterization of the free energy landscape in a protein folding pathway represents a significant contribution to both experimental and theoretical approaches, given the intimate interconnection between the functional energy landscape and aggregation risk [6].
The JD folding pathway is an issue still debated in the literature [30,31,42] since a decisive proof of the most likely JD conformation has not been provided yet. Several JD models, solved by NMR, are available in the literature: open (2JRI) [27], half-open (1YZB [28], closed (2AGA) [29] and half-closed (2DOS) [30]. The above mentioned models have been questioned and debated in the recent literature on both computational and experimental studies. In agreement with previous research in this area [30] we have recently observed the metastable behavior of the JD, which is dynamically able to switch between an open-like and closed-like structure during the dimerization [1]. However, in order to predict the JD free energy landscape, the conformational ensemble provided by the classical MD simulation is not adequate. In particular, data from MD with a timescale of hundreds of ns show that the JD closed-like state is achievable starting from an open-like state whereas a transition from a completely closed-like to an open-like state has never be detected (Fig 1 and Section 1 in S1 Text). The classical MD simulation sampling is in this case insufficient, because when the system is trapped in the energy minimum characterized by the closed-like structure, thermal fluctuations are not enough to overcome the energy barrier (around 36 kJ/mol) needed to switch to an open-like configuration. A classical MD simulation might never be able to get out of such a deep energy minimum. To overcome this limitation, inherent to the classic MD, Metadynamics can be used to sample large-scale protein transitions as demonstrated by some relevant and pioneering papers in the field [62–64].
Our work brings together the efficient sampling of Metadynamics with a MD-PCA-based dimensionality reduction method. In particular, PCA was used to elucidate the transition pathway between the JD open-like and closed-like models, and Metadynamics was performed to estimate the corresponding free-energy landscape. Nevertheless, this computational approach was already successfully applied earlier [38,65–69], thus confirming its promise as a successful strategy for investigating conformational changes in complex biomacromolecules.
A limitation of the presented approach is, in fact, that preliminary information regarding the molecular transition is needed in order to calculate the essential coordinates: the transition from a JD open-like to a closed-like conformation allowed us to obtain the CV to guide the Metadynamics method.
Our findings confirm that the JD hairpin region, which protrudes out into the solvent, can be responsible for an extensive conformational change, switching between the open-like conformation and the closed-like one. As suggested in previous works [1,30], the hairpin mode of motion mainly consists of movement of region α3 (Asp57-Leu62) toward α6 (Asp145-Glu158) (Figs 1C and 3).
Interestingly, in our simulations, when the JD is alone in water environment the most stable configuration is characterized by the hairpin packed against the globular core, in agreement with models 2AGA [29] (Figs 4 and 5). However, an energy well of about -5 kJ/mol has been also detected corresponding to the open conformation, in agreement with the 2JRI [27] model. In general, such free energy minima should predict the most favorable conformational state. In fact, the RG values corresponding to the free energy minima are consistent with the peaks of the RG distribution of the unbiased simulation at equilibrium (Figs 1, 4 and 5), indicating the reliability of the presented approach in describing the JD free energy landscape. Surprisingly the JD half-open structure, with the RG and hairpin angle mainly corresponding to the 1YZB is the less sampled structure even during classical MD simulations.
In this connection, it is important to clarify that our work is not oriented to evaluate the “quality” of produced NMR which has already been checked with proper methodologies [42]. Instead, a first novel aspect of our work is that the employed approach has demonstrated to explore the state space by granting several transitions among closed and open JD conformations.
Data reported in recent literature [31] indicated the JD open structure (particularly referring to the 1YZB model) as the most likely JD conformation in water. Moreover, it has been emphasized how the 2AGA ensemble data were far away from the calculated free energy minima [31]. However, it is worth mentioning that, all the previous computational works [1,31,43] did not show transitions between open and closed JD and viceversa. In our opinion, such transitions are required to claim that a specific protein conformational state is characterized by lower free energy minima than another one. In fact, in the partial section of the free energy profile corresponding to the JD open-like structure, our results are in agreement with the above mentioned recent work [31]. However, the representation of the whole free energy profile describing both open and closed structures demonstrates the closed arrangement as the most likely for a Josephin Domain in water environment.
Nonetheless, there may be conditions under which the open-like state is stabilized (e.g. in the context of the full-length protein or in the presence of a physiological partner). For example, we have recently shown how the JD conformational state might be affected by the presence of another interacting JD [1] as well as by an inorganic surface [43]. In addition, as already suggested [30], the JD free energy landscape could be influenced by other environmental conditions, such as temperature and pH.
Further investigations are planned and will help in clarifying the influence of JD functional partners and environmental factors affecting the JD conformational arrangement. This information may be relevant not only to better understand the physiological function of the Josephin Domain, but also to provide insight into molecular phenomena characterizing the pathological nature of spinocerebellar ataxia 3.
|
10.1371/journal.ppat.1003647 | Transcriptional Regulation of Culex pipiens Mosquitoes by Wolbachia Influences Cytoplasmic Incompatibility | Cytoplasmic incompatibility (CI) induced by the endosymbiont Wolbachia pipientis causes complex patterns of crossing sterility between populations of the Culex pipiens group of mosquitoes. The molecular basis of the phenotype is yet to be defined. In order to investigate what host changes may underlie CI at the molecular level, we examined the transcription of a homolog of the Drosophila melanogaster gene grauzone that encodes a zinc finger protein and acts as a regulator of female meiosis, in which mutations can cause sterility. Upregulation was observed in Wolbachia-infected C. pipiens group individuals relative to Wolbachia-cured lines and the level of upregulation differed between lines that were reproductively incompatible. Knockdown analysis of this gene using RNAi showed an effect on hatch rates in a Wolbachia infected Culex molestus line. Furthermore, in later stages of development an effect on developmental progression in CI embryos occurs in bidirectionally incompatible crosses. The genome of a wPip Wolbachia strain variant from Culex molestus was sequenced and compared with the genome of a wPip variant with which it was incompatible. Three genes in inserted or deleted regions were newly identified in the C. molestus wPip genome, one of which is a transcriptional regulator labelled wtrM. When this gene was transfected into adult Culex mosquitoes, upregulation of the grauzone homolog was observed. These data suggest that Wolbachia-mediated regulation of host gene expression is a component of the mechanism of cytoplasmic incompatibility.
| Wolbachia are maternally inherited bacteria that manipulate invertebrate reproduction. Cytoplasmic incompatibility is embryo death that occurs when males carrying Wolbachia mate with females that do not, or that carry a different Wolbachia variant; its mechanism is poorly understood. In Culex mosquitoes, in the presence of Wolbachia a gene related to a Drosophila melanogaster gene, grauzone, which has been shown to act as a regulator of the meiotic cell cycle, showed an elevated level of expression. When lower levels of expression were achieved through RNA interference, embryo hatch rates were affected and the stage of development at which embryo death occurs was altered. To find Wolbachia genes that influence cytoplasmic incompatibility, we compared the genomes of two variants of Wolbachia from Culex that produce cytoplasmic incompatibility with one another. Although most segments of these genomes were very similar, one newly identified gene is predicted to be a regulator of gene transcription. We cloned this gene into a plasmid, expressed it in adult mosquitoes and found higher levels of expression of the Culex grauzone homolog. This suggests that the Wolbachia transcriptional regulator may play an important role in manipulating the host in order to induce cytoplasmic incompatibility.
| The intracellular maternally inherited bacterium Wolbachia pipientis, a widespread endosymbiont of invertebrates [1], can influence reproduction in arthropods. The most common manipulation is cytoplasmic incompatibility (CI). Sperm from Wolbachia-infected males are modified during maturation, prior to the loss of Wolbachia, such that aberrant events in the male pronucleus [2]–[7] lead to embryo developmental arrest when these sperm fertilize eggs from uninfected females. However, progeny are rescued when both parents carry compatible Wolbachia and therefore, infected females have a selective advantage under this unidirectional pattern of CI.
The Culex pipiens group of sibling species of mosquito, in which CI was first discovered, provides a model system with useful features for examining the genetic differences that underlie CI. Even though only one designated strain of Wolbachia (wPip) is present in Culex pipiens, complex crossing types including both unidirectional and bidirectional CI [8]–[14] occur between populations. Compatibility or partial CI is most often observed, but certain lines will be completely incompatible in one or both crossing directions with a majority of other C. pipiens lines. Understanding the basis of this complexity has been a long standing problem.
Several studies have shown that Wolbachia and its products can influence host gene transcription, notably a major upregulation of immune genes in transinfected naïve mosquitoes [15]–[21], which can contribute to the inhibition of arboviruses and Plasmodium parasites [18], [20], [22]. Differential regulation of several candidate genes has also been observed in Drosophila [23]–[25]; these genes have been hypothesized to form part of the CI phenotype, but so far it has not been possible to confirm a role in incompatibility generation.
Cytological studies in Drosophila and Nasonia have revealed that aberrant events in the male pronucleus in an incompatible fertilization include abnormal histone H3.3 and H4 deposition, prolonged or incomplete DNA replication, delayed chromosome condensation/segregation and nuclear envelope breakdown [2]–[7]. It has been suggested that cell cycle defects in the male pronucleus in CI embryos could be due to disruption of cell cycle regulators or the induction of checkpoints that control entry into mitosis, possibly at the metaphase to anaphase transition [5], [6]; these defects are rescued by Wolbachia in females. The D. melanogaster gene grauzone (grau) encodes a zinc finger transcription factor that plays a role in the regulation of the female meiotic cell cycle [26], [27]. Meiotic arrest at metaphase I is released by egg activation, resulting in completion of the two meiotic divisions; mothers mutant for grau are sterile and lay eggs presenting aberrant chromosomal segregations at meiosis I, which arrest their development in metaphase of meiosis II [26]. Among the many cell cycle regulating genes that have been described, grau is one of only two genes known to be involved in regulation of metaphase II. Given the parallels between abnormal chromosome segregations in both CI embryonic arrest and in grau mutant embryos, and the similarities between metaphase-anaphase transition in meiosis II and mitosis, we considered this gene a candidate for involvement in CI generation.
Not much is known about which Wolbachia genes are responsible for the manipulation of host early embryogenesis. The lack of a transformation system for this intracellular symbiont, which cannot be cloned or cultured outside of host cells, has meant that systematic testing of the phenotypic effects of candidate genes has not been possible. The Pel line of Culex quinquefasciatus from which a wPip genome sequence was generated [28] has been shown to be bidirectionally incompatible with a Culex molestus line [29]; comparative genome analyses between these incompatible Wolbachia strains could therefore be used to determine differences between them that may be involved in the generation of different crossing types. Comparisons with the JHB line wPip genome [30], [31] are also possible. Therefore, the aims of this study were to identify a homolog of grau in Culex and test its potential involvement in the mechanism of CI, and in parallel to purify, sequence and analyse the wPipMol genome in order to attempt to identify Wolbachia genes that may be involved in manipulation of the host, particularly in transcriptional modification.
We identified homologs of the D. melanogaster gene grau in Culex quinquefasciatus (a member of the C. pipiens complex) by interrogation of the genome database [31], which revealed the presence of two paralogs: CPIJ005623 and CPIJ015950. These two paralogs are identical with the exception of a 56 nucleotide (nt) indel between nt 994-1049 of CPIJ005623 which is absent in CPIJ015950, one nt change leading to one amino acid substitution and one nt deletion near the exon1-exon2 border of CPIJ015950 leading to two amino acid substitutions. PCR analysis revealed that only CPIJ005623 was present in the Culex lines used in this study (Figure S1), suggesting that they may just be allelic variants.
Relative transcription of CPIJ005623 was measured in members of the C. pipiens complex using quantitative reverse transcription PCR (qRT-PCR). We first looked at the transcription of the gene in whole adults over time to understand its expression dynamics in individual mosquitoes that carry Wolbachia and their antibiotic-treated Wolbachia-free genetic counterparts. In Wolbachia-infected C. quinquefasciatus Pel adult females, CPIJ005623 transcription showed a peak of over two-fold upregulation compared to the Wolbachia-cured line PelU at 4 days post pupal eclosion (dpe), which then decreased to similar levels as the PelU line by 8 dpe (Figure 1A). Differences in CPIJ005623 transcript levels were found, with an interaction between Wolbachia infection status (wis) and developmental time (Two-way Anova: wis:time- F = 7.66, Df = 1, p<0.01), where the difference between wis contributes most strongly to the interaction (Two-way Anova: wis- F = 13.53, Df = 1, p<0.001). In males (Figure 1B), there was an interaction between Wolbachia infection status (wis) and developmental time (Two-way Anova: wis:time- F = 5.79, Df = 1, p = 0.025) which was independent from any difference seen at 6 dpe between the Wolbachia-infected and uninfected males (Welch's t-test: t = 0.78, Df = 4.833, p = 0.4716).
We then focused our analysis at single time points where the expression was found to be highest, separately in ovaries (4 dpe) and testes (2 dpe) (the tissues where CI is expressed). A similar pattern of upregulation (Wilcoxon rank-sum: p<0.05) of CPIJ005623 in Wolbachia-infected ovaries (Figure 1C) and testes (Figure 1D) was seen for two different pairs of C. quinquefasciatus infected (Pel and Cxq) versus cured (PelU and CxqT) lines, and for an infected Culex molestus line. A higher level of upregulation in C. molestus Italy ovaries was observed than in Pel females.
Given the upregulation of CPIJ005623 detected in the Italy Wolbachia-infected line compared to uninfected females, we knocked down the gene in infected females using dsRNA to investigate whether this change may be involved in the rescue function of CI. Restoring transcription to uninfected levels in these females might result in developmental arrest in embryos from crosses with Wolbachia-infected males (Figure 2A). Knockdown (KD) levels were assessed in dissected ovaries and a reduction of approximately 50% was detected in the transcription of CPIJ005623 in iCPIJ005623-injected females compared to iLacZ-injected control-KD females at 4 days post injection (dpi) which were similar to PelU ovarian expression levels (Figure 2B). An increase in the number of unhatched eggs was detected in a compatible cross of Italy males with iCPIJ005623 females (48.7%) compared to control-KDs (37.9%) (Table 1- Welch's t-test: t = 8.975, Df = 39.861, p<0.0001).
In Culex and other Diptera, arrest can occur at an early (stage I) or late (stage II+III) stage in embryo development [3], [12], [32] and these stages are identified by distinctive embryo morphologies (Figure 2C). The proportions of stage I and stage II/III embryos that did not hatch in the compatible cross after iCPIJ005623 and control-KD differed (GLM: χ2 = 28.957, Df = 1, p<0.0001), with a higher proportion of embryos arresting at stage I in iCPIJ005623 females (Figure 2D).
A lower level of upregulation was observed for CPIJ005623 in Pel Wolbachia-infected versus uninfected females (Figure 1B). To understand if grau upregulation is part of a generalized Culex pipiens group rescue response, we performed KD analysis followed by a compatible cross in C. quinquefasciantus Pel females (Figure 3A). Knockdown levels were assessed in dissected ovaries and a reduction of up to 40%, similar to uninfected levels was detected in the transcription of CPIJ005623 compared to control-KDs at 4 days post injection (dpi) (Figure 3B). A higher proportion of embryos arresting at stage I for iCPIJ005623 females was detected than for control-KD females (Figure 3C, Table 1) (GLM: χ2 = 23.841, Df = 1, p<0.0001). Although in Pel females unhatched rates were not affected after grau homolog KD, the shift from late to early embryo mortality is conserved in the two Culex lines.
Since CPIJ005623 was also observed to be upregulated in infected versus uninfected males, we tested if this upregulation had a role in inducing CI in crosses between infected males and uninfected females. We failed to obtain an antibiotic-treated Wolbachia uninfected C. molestus line due to the low egg production of this blood meal independent autogenous line. As the upregulation detected was fairly similar and not statistically different between the Wolbachia infected lines tested (Figure 1C) we focused our attention to the C. quinquefasciatus Pel line. The effect of iCPIJ005623 in Pel sperm was assessed over time (Figure 4A), in order to determine if CPIJ005623 transcriptional upregulation produces phenotypic effects either early in spermatogenesis or in mature sperm, given that the average time for spermatogenesis in C. pipiens is 9–11 days [33]. No effect was observed on the survival of embryos produced by Wolbachia-uninfected mothers mated with iCPIJ005623 Wolbachia-infected males. In Culex, when uninfected-females mate with Wolbachia infected-males, most embryos arrest at an early stage and no developmental embryo progression occurs. Knocking down CPIJ005623 in Pel males, by as much as 80% and lower than CPIJ005623 expression levels in the uninfected line (Figure 4B), had no significant effect on developmental progression in unidirectional CI crosses.
The effect of knockdown of CPIJ005623 in an incompatible cross was then examined using C. quinquefasciatus Pel and the C. molestus Italy lines, for which zero egg hatch is observed in both crossing directions. When CPIJ005623 was knocked down in Pel males which were crossed to Italy females, no significant effect on hatch rate was observed, but the percentage of embryos that reached later developmental stages was higher at the two earlier time points tested (GLM: χ2 = 65.975, Df = 1, p<0.0001, T1-Figure 4C; χ2 = 33.079, Df = 1, p<0.0001, T2-Figure 4D) compared to the progeny of iLacZ injected controls (Table 2).
A double knockdown (double-KD) experiment in both Italy males and Pel females was also performed, since upregulation of CPIJ005623 was observed in Pel-infected vs Pel-uninfected ovaries (Figure 1B). A KD of 40% was detected in Italy testes using qRT-PCR analysis (Figure 4G). Crossing results show no significant difference in hatch rates between all four double-KD combinations; however, as before, StageI:StageII/III proportions are significantly different (GLM: χ2 = 23.222, Df = 3, p<0.0001) in all crosses where CPIJ005623 transcription was reduced irrespective of the gender of the individual (Figure 4H, Table 2).
We purified and sequenced the genome of the wPipMol strain present in C. molestus [29] and together with the previously sequenced wPipPel [28] and wPipJHB [30], [31] performed a three-way comparative genomic analysis. Crossing experiments revealed that the JHB line of C. quinquefasciatus is bidirectionally incompatible with Mol, while JHB and Pel are compatible with each other in both crossing directions (Table S1). Therefore the pattern of genome variation that is initially of most interest, with respect to different crossing types in the C. pipiens group, are any differences that may exist between wPipMol and wPipPel where wPipPel and wPipJHB are identical. The wPipPel and wPipMol genomes proved to be highly similar and the great majority of genes have identical sequences; only thirty-three non-synonymous SNPs (single nucleotide polymorphisms) were identified in total, in twenty-three genes (Table S2, S3). However, three genes in three inserted or deleted regions were present in wPipMol but absent in wPipPel and wPipJHB (Table 3). One of these newly identified genes, ankM1, encodes ankyrin (ANK) repeats, and ankM2 encodes ANK and Tetratricopeptide (TPR) repeats, both of which function in mediating protein-protein interactions [34]–[36]. The ankM1 gene is a homolog of WD0596 from Wolbachia strain wMel of Drosophila melanogaster [37]. The other newly identified gene, labeled wtrM, contains DNA binding domains associated with transcriptional regulators. Five inserted or deleted regions each comprising one to five genes were identified in wPipPel and wPipJHB but absent in wPipMol (table 3), all of which are within or adjacent to prophage regions.
As a preliminary guide to which of the inserted or deleted regions most warranted further experimental investigation, we examined patterns of presence or absence in some other C. pipiens group populations from different geographical locations using PCR (Table S4). The Wolbachia transcriptional regulator gene wtrM was the only gene that occurred solely in the two C. molestus lines examined, Mol and Italy, which generate bidirectional CI with a selection of other C. pipiens group line with which they have been crossed (Table S1). This gene is a member of a family of Wolbachia transcriptional regulator genes, with seven genes in the wPipPel genome and six in the D. melanogaster wMel Wolbachia genome (Figure 5A). One of the genes missing in wPipMol but present in wPipPel, WP0457, is also a member of this family of transcriptional regulators. WP0457 is identical to WP1341, located in a highly similar cluster of prophage-associated genes. Both of these loci are located close to patatin family phospholipase genes, which encode bacterial virulence factors that can disrupt cell membranes [38], [39].
We hypothesised that the protein products of the Wolbachia transcriptional regulator wtrM might be secreted, translocated to the host nucleus and modify host gene transcription as a means to manipulate the host. To investigate this possibility, in the absence of a transformation system for Wolbachia, host transfection was used. In the Mol line (wPipMol-infected) C. molestus mosquitoes, expression of wtrM was strongly localised to the ovaries in females (Figure 5B), which is where the majority of Wolbachia are located and are the important tissues with respect to CI. The wtrM gene was cloned into an insect expression plasmid and transfected into females of Pel C. quinquefasciatus in order to examine whether CI-like phenotypes could be induced; expression or translocation of WtrM protein in the ovaries would be required for this purpose. The transcription of wtrM was confirmed using RT-PCR in whole females transfected by intra-thoracic injection; however the transcript could not be detected in dissected ovaries, suggesting that the transfected plasmid did not cross the Peritoneal Sheath and enter the ovary (Figure 5B). Using quantitative RT-PCR, levels of transcription of CPIJ005623 were then compared with control females transfected with plasmid that did not contain wtrM. There was more than five-fold up-regulation of CPIJ005623 in wtrM-transfected females compared to the females injected with the plasmid alone (p<0.01, Wilcoxon rank sum test) (figure 5C).
Several lines of evidence support the hypothesis that the differential regulation of a Culex homolog of the D. melanogaster gene grau plays a direct role in CI generation. The presence of Wolbachia induced changes in CPIJ005623 transcription levels and the degree of upregulation of CPIJ005623 varied between incompatible C. pipiens group females. A decrease in embryo hatch rates was detected in a C. molestus line when CPIJ005623 expression was knocked down in females of an otherwise compatible cross. This increase in levels of CI suggests a role for CPIJ005623 in CI rescue in C. molestus females. The knockdown of CPIJ005623 also extended the mean developmental stage reached by embryos in incompatible crosses. When reversal of CI might have been expected (male KDs), more embryos arrested at a later stage and when increased levels of CI might have been expected (female KD), more embryos arrested at an early stage.
It has been hypothesized based on its mutant-sterile phenotype that D. melanogaster grau could be involved in egg activation, act as a negative regulator of the cyclin B complex stabilizing factor involved in meiotic cell cycle regulation, and/or regulate microtubules; these are not mutually exclusive possibilities [26]. Molecular changes that could affect cell cycle regulation fit with the known cell biology of CI [2]–[5]. Differential transcription leading to inappropriate levels of GRAU protein in the presence of Wolbachia during spermatogenesis/oogenesis could therefore be causally linked to the fact that cell cycle events in the male and female pronuclei are asynchronous.
Given that there was no complete reversal of embryo hatch rates in the rescue cross following CPIJ005623 knockdown, it is likely that differential regulation of CPIJ005623 is a component of CI but not the only molecular change that underlies the phenotype. It also seems likely that different Wolbachia strains produce different combinations of molecular manipulations, given the mutual incompatibilities that can exist among Wolbachia strains. The differences in grau homolog upregulation and KD analysis between the two C. pipiens group lines presented in this study reflect that. Further work is needed to determine what other factors along with grau are involved in this complex phenotype and whether other completely independent pathways are targeted in different Wolbachia strain-host associations where CI is induced. In light of the female-sterile phenotype produced by grau mutations, it would also be interesting to search for and examine the expression of grau homologs in Asobara tabida, where Wolbachia has become essential for the completion of this wasp's oogenesis [40]–[42].
The identification of a Wolbachia gene that modifies transcription levels of a non-immune host gene is a significant step forward in understanding the molecular methods by which this endosymbiont manipulates its hosts. This Wolbachia gene is one of only a small number that vary between two incompatible populations. Although it was not possible to directly assay the effects of wtrM on crossing patterns using transfection, since the plasmid did not enter the ovaries, the upregulation of CPIJ005623 in wtrM-transfected females does provide clear support for the hypothesis that wtrM has a role in the generation and/or rescue of CI between these Culex lines. The fact that the natural expression levels of wtrM in C. molestus were highest in ovaries is of interest given that CPIJ005623 knockdown in C. molestus females changed embryo development and hatch rates whereas CPIJ005623 knockdowns in males produced little effect. It should be noted that although wtrM appears to play a role in this particular Wolbachia - host interaction, there may well be a number of other Wolbachia genes beyond this family of homologous transcriptional regulators that interact directly with host genomes.
The small number of differences between the genomes of incompatibility-generating wPip strains makes this a promising system for better elucidating the molecular mechanism of CI. All the genes identified here that vary between mutually incompatible wPip substrains provide new avenues for investigation; their concentration in the wPip prophages also underlines the importance of these hypervariable regions in wPip recent genome evolution [13]. Ultimately, the development of a methodology for Wolbachia transformation may be required in order to produce a direct phenotypic confirmation of implicated Wolbachia genes in generating/rescuing CI. Despite considerable effort this remains a highly challenging goal, since it is an obligately intracellular microbe. Transformation of the host insect to express Wolbachia genes could be carried out, as was performed in Drosophila with several Wolbachia Ankyrin repeat-encoding genes [43]. However, given the possible differences in the activity of proteins when expressed in host cells rather than Wolbachia, for example in post-translational processing, folding and in epistatic interactions with other Wolbachia proteins within the bacteria, failure to generate a sterility phenotype does not rule out the involvement of candidate genes in CI.
Recently, Wolbachia has attracted considerable scientific and public interest as a control tool for mosquito-borne diseases. In the Aedes mosquito vectors of dengue virus, the introduction of several non-native Wolbachia strains produced a strong inhibitory effect on dengue virus transmission. Given its population invasion capacity, Wolbachia-based strategies seem very promising as a new dengue control tool [16], [44]–[47]. A better understanding of the molecular basis of CI has implications for the effectiveness of Wolbachia-based disease control strategies. It would assist in the design of multi-strain superinfections for spread into wild Wolbachia-infected populations or, in the event of the development of pathogen resistance, for achieving successive invasions of different Wolbachia strains [48]. Furthermore, knowledge of the host genes involved in this process could also aid the development of gene drive systems for spreading nuclear genes through pest insect populations [49].
C. pipiens complex laboratory lines Pel (Wolbachia infected C. quinquefasciatus, Sri Lanka) [13], Cpq (Wolbachia infected) and Cpq-T (Wolbachia uninfected C. quinquefasciatus USA) (kind gift from R. Glazer) [50], and Italy (Wolbachia infected) C. molestus line (kind gift from M. Petridis, collected in 1991 in Granarolo-Italy, by A. Medici and G. Rossi) were reared using standard mosquito rearing procedures at low larval densities in standard insectary conditions (27°C, 70% relative humidity) with a 12 h light/dark cycle.
Wolbachia uninfected PelU was generated by treating PelA with rifampicin (1 ml of 2.5 mg/ml added every 3 days in roughly 1 L H20) throughout larval development for 5 consecutive generations. Adults were not treated with antibiotic. The line was subsequently reared in the absence of antibiotic for at least 5 generations more before experiments were performed. Removal of Wolbachia was confirmed by PCR and the line has been kept Wolbachia free since 2011. Cpq-T generation is described in detail in [50]. In brief Cpq adults were treated for one week on 1 mg/ml tetracycline (pH 7) in 10% sucrose. Wolbachia removal was confirmed by PCR and the line has been kept Wolbachia free since 2010.
For crosses conducted to establish patterns of compatibility between lines, mass crosses containing 50 individuals of each sex were set up, blood fed at three to five days post eclosion and embryos collected four days later; hatch of at least 800 embryos were scored per cross [13]. Where there was no embryo hatch from a cross, the experiment was repeated and females separated into individual containers for egg laying, and spermathecae then dissected to ensure that only progeny from inseminated females were scored.
Adult mosquitoes were homogenized in STE buffer (10 mM Tris-HCl, 1 mM EDTA and 100 mM NaCl) and incubated for 10 min at 95–99°C. Samples were centrifuged for 5 minutes at maximum speed at 4°C. Supernatant, containing gDNA, was kept and used for Polymerase Chain Reaction (PCR); all oligonucleotides are listed in table S5.
Adult female and male RNA was extracted from 4–5 adult mosquitoes using TRIzol Reagent (Invitrogen-Life Technologies) following manufacturer's instructions. TRIzol-extracted RNA was DNase I treated for genomic DNA elimination and purified via standard phenol chloroform extraction. RNA from both mosquito ovaries and testes (30 individuals for each sample), was extracted and purified via the RNeasy Mini protocol for animal tissue (Qiagen). All purified RNA samples were quality checked via Nanodrop analysis and only highly pure (A260:230 and A260:280 ratios of >2.00) were kept. Exception was made for testes RNA for which the maximum A260:230 achieved were 1∶50-1.80 even when using a highly optimised protocol.
cDNA synthesis was performed in 10 µl (adults) or 20 µl (ovaries and testes) reaction volumes with 100, 300 or 900 ng of total RNA for the testes, ovaries and adult (male or female) samples respectively, using the iScript cDNA synthesis kit (BioRad). qRT-PCRs were performed on 1∶10, 1∶20 and 1∶30 dilutions of the cDNAs from the testes, ovaries and adult samples respectively, using iQ SYBR Green PCR mastermix (BioRad), a DNA Engine thermocycler (MJ Research) with a Chromo4 real-time PCR detection system (Bio-Rad) and the following cycling conditions: 95°C for 3 minutes, then 41 cycles of 95°C for 10 s, 60°C for 30 s, with fluorescence acquisition at the end of each cycle and a melting curve analysis after the final cycle. The cycle threshold (Ct) values were determined and background fluorescence was subtracted. Transcription levels of target genes were calculated by the standard curve method, as described in technical bulletin #2 of the ABI Prism 7700 Manual (Applied Biosystems-Life Technologies), relative to the endogenous reference genes and RpS7.
T7-tailed primers were used to amplify a PCR template for CPIJ005623 and the control LacZ gene from Culex cDNA and E.coli DNA respectively. DsRNAs were synthesized using the T7 Megascript Kit (Ambion-Life Technologies) following manufacturer's advice. Purified dsRNAs were diluted to a concentration of 3 µg/µl and 69 nl were injected into the thorax of cold anesthetized mosquitoes using a Nanoject microinjector (Drummond Scientific). After dsRNA injections and at the indicated time points, testes and ovaries from 30 male and female mosquitoes were dissected. RNA was extracted by using the RNeasy Mini protocol for animal tissue (Qiagen) in order to verify the successful knockdown of the CPIJ005623 transcript. Amplifications were performed using primers RpL32 (F-R) and CPIJ005623 (F1-R1) for Culex Pel, RpLS7 and CPIJ005623 (F3-R3) for Culex Italy as described above.
Crossing experiments following knockdowns were carried out using 50 virgin individuals of each sex obtained by sexing and isolating pupae for mosquitoes. On the day of pupal eclosion, 50 male or 50 female mosquitoes were injected with dsRNA. Injected individuals were allowed to recover and on the dates represented in the experimental design diagrams mating partners were added to the injected individuals for the designated times indicated (Fig. 3A, 4A and 5A). After egg lay, female spermathecae were examined for the presence of sperm if the hatch rate was zero to confirm insemination and only progeny of inseminated females were scored. The F1 generation progeny from crosses were characterised by i) total number of rafts; ii) total number of eggs, iii) mean proportion of hatched eggs and iv) for unhatched embryos mean proportion of developed embryos.
Transcription levels were analysed using a non-parametric Wilcoxon rank-sum test in the case of 2 sample variables and by a two-way ANOVA (with replication) analysis of variance for 3 sample variables. Embryo developmental results were analysed by using generalized linear models (GLM) with binomial error structures (where reaching SII or beyond is classed as a success). Differences between hatched and unhatched embryos were determined via a Welch two sample t-test. All calculations were performed using the R software (R Development Core Team, 2004).
The purification of Wolbachia wPip DNA from Mol early embryos, library construction, general PCR conditions and crossing experiments were as previously described [10]. The genome was assembled from three independent 454 runs. One half-plate from a standard single-end FLX library generated 14 Mb of data from 80,752 reads, with an average length of 174 bp. Two quarter-plates from a similar library generated 30.9 Mb of data from 80,954 reads with an average length of 264 bp. Finally one quarter-plate from a 3.2 kb paired-end Titanium library generated 48.6 Mb from 157,851 reads with an average length of 232 bp. Together these runs produced a total of 93.5 Mb of data, which was combined and assembled with Newbler (Roche), using default parameters. The final assembly gave 194 contigs >500 bp with a mean coverage of 41× and N50 of 19.5 kb, linked into 24 scaffolds plus 79 unscaffolded contigs, with an N50 of 567 kb, and a total length of 1,435,676 bp. 99.6% of bases had a quality value >Q40. The sequences generated have been submitted to the EMBL/GenBank/DDBJ database with the accession number HG428761.
Phylogenetic trees were constructed based on Maximum Likelihood using PhyML [51], using the Whelan and Goldman substitution model, on an amino acid alignment created with ClustalW; one hundred data sets were generated for bootstrapping.
Contigs and scaffolds were mapped against the wPip genome as a reference using mummer. SNPs were called where the mapped sequences differed from the reference sequence. The contig and scaffold SNPs were mostly at contig boundaries and phage regions, where mapping was difficult. Only SNPs which were agreed upon by both the contig and scaffold mapping were considered strongly supported.
For PCR analysis, genomic DNA was extracted from whole mosquitoes as above. Primers were designed to flank SNP regions and amplified using standard PCR conditions (94°C for 30 sec, 52 to 60°C for 30 sec and 72°C for 30 sec to 60 sec ×38 cycles). PCR products were purified using a Qiagen PCR purification kit and sequenced using GATC Biotech sequencing. The C. pipiens group lines used for these PCR experiments were as previously described [13], [28], [29] with an additional colony assayed from the Italy (Granarolo) collected in 1991.
The transcriptional regulator wtrM was PCR amplified from the strain of Wolbachia found in Culex molestus. DNA was extracted as above. PCR was conducted with Phusion Taq polymerase (Finnzymes) according to manufacturer's instructions. PCR products were digested and ligated into the NotI-SacI site of insect cell expression vector, pIEX8 (Novagen). Plasmids were purified using endotoxin free maxi preps (Qiagen).
Adult C. quinquefasciatus Pel line mosquitoes less than 24 hours post eclosion, infected with Wolbachia wPipPel were intrathoracically injected with 200 ng plasmid DNA and Cellfectin II (Invitrogen) transfection reagent according to a previously published protocol [52]. Using a hand-held Nanoject (Drummond), adults were injected with either pIEX8-wtrM or a control plasmid with no insert and left to recover. Adults were isolated for total RNA extraction in groups of 2 to 4 individuals: 10 replicates of 3–4 individuals (n = 32) in pIEX and 10 replicates of 2–3 individuals (n = 27) for pIEX wtrM. cDNA synthesis was performed in 10 µl reaction volumes with 1000 ng of total RNA using the iScript cDNA synthesis kit (BioRad). qRT-PCRs were performed on 1∶20 dilutions of the cDNA using DyNAmo Colorflash (Fisher), a DNA Engine thermocycler (MJ Research) with a Chromo4 real-time PCR detection system (Bio-Rad) and the cycling conditions as above. Transcription levels were calculated by the standard curve method relative to the endogenous Culex reference gene RpL32.
|
10.1371/journal.ppat.1007050 | Isolation of a natural DNA virus of Drosophila melanogaster, and characterisation of host resistance and immune responses | Drosophila melanogaster has played a key role in our understanding of invertebrate immunity. However, both functional and evolutionary studies of host-virus interaction in Drosophila have been limited by a dearth of native virus isolates. In particular, despite a long history of virus research, DNA viruses of D. melanogaster have only recently been described, and none have been available for experimental study. Here we report the isolation and comprehensive characterisation of Kallithea virus, a large double-stranded DNA virus, and the first DNA virus to have been reported from wild populations of D. melanogaster. We find that Kallithea virus infection is costly for adult flies, reaching high titres in both sexes and disproportionately reducing survival in males, and movement and late fecundity in females. Using the Drosophila Genetic Reference Panel, we quantify host genetic variance for virus-induced mortality and viral titre and identify candidate host genes that may underlie this variation, including Cdc42-interacting protein 4. Using full transcriptome sequencing of infected males and females, we examine the transcriptional response of flies to Kallithea virus infection and describe differential regulation of virus-responsive genes. This work establishes Kallithea virus as a new tractable model to study the natural interaction between D. melanogaster and DNA viruses, and we hope it will serve as a basis for future studies of immune responses to DNA viruses in insects.
| The fruit fly Drosophila melanogaster is a useful model species to study host-virus interaction and innate immunity. However, few natural viruses of Drosophila have been available for experiments, and no natural DNA viruses of Drosophila melanogaster have been available at all. Although infecting flies with viruses from other insects has been useful to uncover general immune mechanisms, viruses that naturally infect wild flies could help us to learn more about the coevolutionary process, and more about the genes that underlie host-virus interactions. Here we present an isolate of a DNA virus (named Kallithea virus) that naturally infects the model species Drosophila melanogaster in the wild. We describe the basic biology of infection by this virus, finding that both male and female flies die from infection. We also find that females are more tolerant of infection than males, but lay fewer eggs than uninfected females. We quantify genetic variation for virus resistance in the flies, and we use RNA sequencing to find which genes are expressed in male and female flies in response to infection. These results will form the basis for further research to understand how insects defend themselves against infection by DNA viruses, and how DNA viruses can overcome antiviral defence.
| Studies of Drosophila melanogaster are central to our understanding of infection and immunity in insects. Moreover, many components of the Drosophila immune response, including parts of the JAK-STAT, IMD, and Toll (and perhaps RNA interference; RNAi) pathways are conserved from flies to mammals [1–8], making Drosophila a valuable model beyond the insects. The experimental dissection of antiviral immune pathways in Drosophila has benefited from both natural infectious agents of Drosophila, such as Drosophila C Virus (DCV) and Sigma virus (DmelSV), and from artificial infections, such as Cricket paralysis virus (isolated from a cricket), Flock House Virus (from a beetle), Sindbis virus (from a mosquito) and Invertebrate Iridescent Virus 6 (from a moth). However, while the availability of experimentally tractable, but non-natural, model viruses has been a boon to studies of infection, it also has two potential disadvantages. First, the coevolutionary process means that pairs of hosts and pathogens that share a history may interact very differently to naive pairs (e.g. [9,10]). For example, the Nora virus of D. immigrans expresses a viral suppressor of RNAi that is functional in the natural host, but not in D. melanogaster [11]. Second, if our aim is to understand the coevolutionary process itself, then the standing diversity in both host and virus populations may be fundamentally altered in coevolving as opposed to naïve pairs. For example, heritable variation for host resistance was detectable for two natural viruses of D. melanogaster, but not for two non-natural viruses [12,13]. This difference was in part due to large-effect segregating polymorphisms for resistance to the natural viruses, which are predicted to result from active coevolutionary dynamics [14–16].
Experimental studies of host-virus interaction using Drosophila have consequently been limited by a lack of diverse natural virus isolates. In particular, no natural DNA viral pathogens of D. melanogaster have previously been isolated ([17,18]; but see [19] for a DNA virus of Drosophila innubila), and all natural (and most artificial) studies of viral infection in D. melanogaster have therefore focussed on the biology of RNA viruses and resistance to them [20,21]. For DNA viruses, our molecular understanding of insect-virus interaction has instead been largely shaped by the response of lepidopterans to their natural baculoviruses. These are often of agronomic and/or ecological importance [22], but lack the genetic toolkit of D. melanogaster. Nevertheless, Lepidopteran studies of the expression response to baculovirus infection have implicated host genes with a diverse array of functions, including cuticle proteins, reverse transcriptases, and apoptotic factors, suggesting previously uncharacterised and/or host-specific antiviral immune mechanisms [23–26].
To date, the only DNA virus studies in D. melanogaster have used Insect Iridescent Virus 6 (IIV6), an enveloped dsDNA moth iridovirus with a broad host range [27]. This work has shown that Drosophila RNAi mutants are hyper-susceptible to IIV6 infection, and that IIV6 encodes a viral suppressor of RNAi, indicating that at least some immune responses to DNA viruses overlap with those to RNA viruses [21,28,29]. However, while IIV6 injections are lethal in D. melanogaster, and IIV6 has provided useful information about the Drosophila response to DNA viruses, for the reasons described above it is hard to interpret the implications of this for our understanding of natural host-virus interaction.
Metagenomic sequencing has recently identified several natural dsDNA nudivirus infections in wild-caught Drosophila, including in D. innubila (D. innubila Nudivirus, DiNV; [19,30]) and in D. melanogaster and D. simulans (‘Kallithea virus’, KV [31]; ‘Esparto virus’ (KY608910.1), and ‘Tomelloso virus’ (KY457233.1)), and also ssDNA densovirus infections in D. melanogaster and D. simulans (‘Vesanto virus’ (KX648534.1), ‘Linvill Road virus’ (KX648536.1), and ‘Viltain virus’ (KX648535.1)) [32]. Like other members of the Nudiviridae, DiNV and KV are enveloped dsDNA viruses of around 120-230Kbp with 100–150 genes. This recently-recognised family forms a clade that is either sister to, or paraphyletic with, the Bracoviruses [33] that have been ‘domesticated’ by Braconid parasitoid wasps following genomic integration, and now provide essential components of the wasp venom [34,35]. Together, the nudiviruses and bracoviruses are sister to the baculoviruses, which are arguably the best-studied dsDNA viruses of insects. They share many of their core genes with baculoviruses, but canonically lack occlusion bodies [36]. PCR surveys of wild flies suggest that DiNV is common in several species in the subgenus Drosophila, and that KV is widespread and common in D. melanogaster and D. simulans, being detectable in 10 of 17 tested populations, with an estimated global prevalence of 2–7% [31]. However, we currently know little about the interaction between these viruses and their hosts. Indeed, although studies of wild-caught D. innubila individuals infected by DiNV suggest that infection is costly [19], in the absence of an experimental D. melanogaster nudivirus isolate, it has not been possible to capitalise the power of D. melanogaster genetics to further elucidate the costs associated with infection, or the genetic basis of resistance.
Here we present the isolation of KV from wild-collected D. melanogaster via passage in laboratory stocks and gradient centrifugation. We use this isolate to characterise the fundamental phenotypic impacts of infection on host longevity and fecundity. We then use the Drosophila Genetic Reference Panel (DGRP; [37]) to quantify and dissect genetic variation in immunity to KV infection in males and females, and use RNA sequencing analyses of an inbred line to quantify host and virus transcriptional response in both sexes. We find that KV causes higher rates of mortality following injection in males, but that males have lower viral titre, suggesting some female tolerance to infection. However, we also find that female movement is decreased following infection, and that infected females have significantly reduced late-life fecundity—highlighting the importance of considering infection phenotypes beyond longevity. We find a genetic correlation in longevity between KV-infected males and females, and a weak negative genetic correlation between mortality and KV titre in females, and we report host loci that have variants significantly associated with each trait. Finally, our expression analysis of infected individuals supports a dramatic cessation of oogenesis following infection, and significant differential regulation of serine proteases and certain immune genes. This work establishes KV as a new natural model for DNA virus infection in D. melanogaster and will enable further dissection of the insect antiviral immune response.
We identified KV-infected flies through a PCR screen for previously published D. melanogaster viruses in 80 previously untested wild-caught flies (see [31] for primers and cycling conditions). We homogenised each fly in 0.1 mL of Ringer's solution, transferred half of the homogenate to Trizol for nucleic acid extraction, and performed RT PCR assays on the resulting RNA for all D. melanogaster viruses reported by Webster et al [31]. We selected a KV-positive sample from Thika, Kenya (Collected by John Pool in 2009; subsequently stored at -80C), removed debris from the remaining fly homogenate by centrifugation for 10 minutes at 1000 × g, and microinjected 50 nL of the supernatant into Dicer-2L811fsX flies, which lack a robust antiviral immune response [38]. After one week, we homogenised 100 KV-injected Dicer-2 L811fsX flies in 10 uL Ringer’s solution per fly, cleared the solution by centrifugation as above, and re-injected this homogenate into further Dicer-2L811fsX flies. This process was then repeated twice more with the aim of increasing viral titres. In the final round of serial passage, we injected 2000 Dicer-2 L811fsX flies, which were homogenised in 5 mL 10 mM Tris-HCl. We cleared the homogenate by centrifuging at 1000 × g for 10 minutes, filtering through cheese cloth, centrifuging twice more at 6000 × g for 10 minutes, and finally filtering through a Millex 0.45 µm polyvinylidene fluoride syringe filter. The resulting crude virus preparation was used as input for gradient ultracentrifugation.
We screened the crude preparation by RT-PCR for other published Drosophila virus sequences, and identified the presence of DAV, Nora virus, DCV, and La Jolla virus. To separate KV from these viruses, we used equilibrium buoyant density centrifugation in iodixanol (“OptiPrep”, Sigma-Aldrich) as enveloped viruses are expected to have lower buoyant densities than most unenveloped viruses. Iodixanol is biologically inert, and gradient fractions can be used directly for downstream infection experiments (avoiding dialysis, which we found greatly reduces KV titres). We concentrated virus particles by centrifuging crude virus solution through a 1 mL 10% iodixanol layer onto a 2 mL 30% iodixanol cushion at 230,000 × g for 4 hours in a Beckman SW40 rotor. Virus particles were taken from the 30%-10% interphase, and layered onto a 40%-10% iodixanol step gradient, with 2% step changes, and centrifuged for 48 hours at 160,000 × g. We fractionated the gradient at 0.5 mL intervals, phenol-chloroform extracted total nucleic acid from aliquots of each fraction, and measured virus concentration by quantitative PCR (qPCR). We pooled all Kallithea-positive, RNA virus-negative fractions and calculated the infectious dose 50 (ID50) by injecting 3 vials of 10 flies with each of a series of 10-fold dilutions and performing qPCR after 5 days. We simultaneously performed the above isolation protocol with uninfected Dicer-2L811fsX flies and extracted the equivalent fractions for use as an injection control solution (hereafter referred to as “gradient control”).
A droplet of viral suspension was allowed to settle on a Formvar/Carbon 200 mesh Copper grid for 10 minutes. We removed excess solution and applied a drop of 1% aqueous uranyl acetate for 1 minute before removing the excess by touching the grid edge with filter paper. The grids were then air dried. Samples were viewed using a JEOL JEM-1400 Plus transmission electron microscope, and representative images were collected on a GATAN OneView camera.
Flies were reared on a standard cornmeal diet until infection, after which they were transferred to a solid sucrose-agar medium. We infected flies by abdominal injection of 50 nL of 105 ID50 KV using a Nanoject II (Drummond Scientific), and these flies were then used to assay changes in viral titre, mortality, fecundity, or daily movement. To test whether the change in viral titre over time was influenced by sex or the presence of Wolbachia endosymbionts, we injected 25 vials of 10 male or female Oregon R flies with KV, with or without Wolbachia (totalling 1000 flies). We phenol-chloroform extracted total nucleic acid at 5 time-points: directly after injection and 3, 5, 10, and 15 days post-infection. We used qPCR to measure viral titre relative to copies of the fly genome with the following (PCR primers: kallithea_126072F CATCAATATCGCGCCATGCC, kallithea_126177R GACCGAGTTAGCGTCAATGC, rpl32_465F CTAAGCTGTCGGTGAGTGCC, rpl32_571R: TGTTGTCGATACCCTTGGGC). We analysed the log-transformed relative expression levels of Kallithea virus as a Gaussian response variable in a linear mixed model using the Bayesian generalised mixed modelling R package MCMCglmm (V2.24; [39]). R code and raw data used to fit all models in this paper is provided on figshare (doi: 10.6084/m9.figshare.c.3936037.v1).
The fixed effects portion of the model included an intercept term and coefficients for the number of days post-inoculation (DPI), sex, and DPI by sex interaction. We estimated random effects for each qPCR plate and assumed random effects and residuals were normally distributed. We initially fitted the model with Wolbachia infection status included as a fixed effect, however this term was not significant and was excluded from the final model.
We also attempted to infect flies with KV by feeding. We anesthetised flies in an agar vial and sprayed 50 uL of 5x103 ID50 KV onto the flies and food. We then collected flies immediately (for the zero time-point) and at 7 DPI and used the primers above to calculate relative KV titre.
We performed mortality assays to test the effect of KV infection on longevity, and to test whether this was affected by sex or Wolbachia infection status. We injected a total of 1200 Oregon R flies with control gradient or KV for each sex with or without Wolbachia (Wolbachia had previously been cleared by 3 generations of Ampicillin treatment and its absence was confirmed by PCR). We maintained flies for each treatment in 10 vials of 10 flies, and recorded mortality daily for three weeks. Mortality that occurred in the first day after infection was assumed to be due to the injection procedure and excluded from further analysis. We analysed mortality using an event-analysis framework as a generalised linear mixed model using MCMCglmm, with per-day mortality in each vial as a binomial response variable. We included fixed effects for DPI, DPI2 (used to capture nonlinear mortality curves), KV infection status, the two-way interaction between DPI and KV infection status, the two-way interaction between DPI and sex, and the three-way interaction between DPI, KV infection status, and sex. We fitted vial as a random effect to account for non-independence among flies within vials, assuming these follow a normal distribution. As in the model for viral titre, we found no evidence for differences associated with Wolbachia infection, and Wolbachia terms were excluded from the final model. The higher rate of male mortality we observed was also confirmed in a second independent experiment using an outbred population derived from the Drosophila Genetic Reference Panel (DGRP; see below).
We measured fecundity during early (1 and 2 DPI) and late (7 and 8 DPI) Kallithea virus infection. Virgin female flies from an outbred population derived from the DGRP ([37]; created from 113 DGRP lines and maintained at a low larval density with non-overlapping generations) were injected with either KV, or with chloroform-inactivated KV as a control, and individually transferred to standard cornmeal vials. The following day we introduced a single male fly into the vial with the virgin female. We transferred the pair to new vials each day and recorded the number of eggs laid. Per-day fecundity was analysed in MCMCglmm as a Poisson response variable using a hurdle model, which models the probability of zeroes in the data and the Poisson process as separate variables. We included fixed effects associated with KV infection status, infection stage (early or late), the interaction between KV infection and infection stage, and random effects associated with each fly pair (vial).
We analysed ovary morphology to examine whether changes in fecundity were detectable in ovaries. Flies were injected with either control virus solution or KV and kept on solid sucrose-agar medium vials. After 8 DPI, flies were transferred to vials with standard cornmeal medium supplemented with yeast. Two days later, we dissected ovaries in phosphate-buffered saline solution, fixed ovaries in 4% paraformaldehyde, and stained nuclei with DAPI. Ovaries were analysed under a Leica fluorescence microscope, and we recorded whether each ovariole within an ovary included egg chambers past stage 8 (i.e. had begun vitellogenesis), and whether any egg chambers within an ovariole exhibited apoptotic nurse cells. The probability of an ovariole containing a post-vitellogenic egg chamber was analysed using a logistic regression in MCMCglmm, with KV infection status as a fixed effect and the ovary from which the ovariole derived as a random effect. We analysed whether apoptotic nurse cells are associated with KV virus-infected ovary in the same way.
We used a Drosophila Activity Monitor (DAM, TriKinetics; [40]) to measure per-day total movement of individual flies. The DAM is composed of multiple hubs, each with 32 tubes containing a single fly, and movement is recorded on each occasion the fly breaks a light beam. We injected 96 female flies from an outbred DGRP population with either chloroform-inactivated KV or KV, randomly assigned these flies within and across 3 hubs, and measured total movement for one week. Movement was binned for each day and this per-day total movement was analysed in a linear mixed model as a Poisson response variable using MCMCglmm. We completely excluded flies that failed to move for a whole day or longer, assuming them to be dead. As before, we included fixed effects associated with DPI, KV infection status, and the interaction between KV and DPI. We included random effects associated with each fly (repeated measures) and each of the DAM hubs, and assumed each of these take values from a normal distribution.
The DGRP is a collection of highly inbred fly lines derived from a D. melanogaster population collected in Raleigh, North Carolina [37], and is widely used to estimate and dissect genetic variation in complex traits in Drosophila. We measured KV titre in females and mortality following KV infection in both sexes for 125 DGRP lines, and estimated genetic (line) variances and covariances among these traits. To measure viral titre in the DGRP, we infected 5 vials of 10 flies for each line across 5 days, with a vial from each line being represented each day. After 8 DPI, living flies were killed and homogenised in Trizol for nucleic acid extraction and qPCR. To measure mortality following KV infection in the DGRP, we injected 3 vials of 10 flies of each sex and recorded mortality on alternate days until half the flies in the vial were dead (i.e. median survival time). Flies were transferred to fresh agar vials every 10 days. Mortality occurring in the first 3 DPI was assumed to be caused by the injection procedure and was removed from the analysis.
We fitted a multi-response linear mixed model in MCMCglmm to estimate heritability and genetic covariances among lines
yiklpqrtrait=βptrait+βpktrait:sex+μpqtrait:date+μlplate+μpkrtrait:sex:line+εiklpqrtrait:sex
[1]
where yiklpqrtrait is the log-transformed relative viral titre or the duration until median mortality (LT50). We only estimated sex-specific fixed effects (βtrait:sex) for LT50, because we did not measure titre in both sexes. The first part of the random effects model accounts for block effects due to date of injection (μtrait:date) and qPCR plate (μplate). We assumed a 2x2 identity matrix as the covariance structure for μtrait:date, with effects associated with each trait from independent normal distributions. Effects for the lth plate were assumed to be normally distributed. The second part of the random effects model (μtrait:sex:line) estimates the variance in each trait across lines and was allowed to vary by sex. We estimated all variance-covariance components of the 3x3 G matrix associated with μtrait:sex:line. Finally, we fitted separate error variances for each trait in each sex (εtrait:sex), where residuals were associated with independent normal distributions.
The diagonal elements of the μtrait:sex:line covariance matrix represent posterior distributions of genetic variances for viral titre in females, LT50 in females, and LT50 in males (VGtitre, VGmortality♀, VGmortality♂). We calculated broad-sense heritability (i.e. line effects) for each trait as H2=VGVG+VR, where VR is the residual variance associated with each trait, estimated in the model as εtrait:sex. However, heritabilities cannot readily be compared because of their dependence on the residual variance, which can be vastly different for different phenotypes [41]. Therefore, we also calculated the coefficient of genetic variation (CVG) as CVG=100*VGμ, where VG is standardised by the phenotypic mean (μ) and is more appropriate for comparisons across phenotypes. All confidence intervals reported are 95% highest posterior density intervals.
We used measurements of viral titre and mortality following KV infection in the DGRP lines to perform a series of genome-wide association studies (GWAS). Although our power to detect small-effect genetic variants with only 125 lines is very low, past studies have demonstrated genetic variation in natural viral resistance in Drosophila is often dominated by few large effect variants ([12,14–16]; but for caveats see [42]). We performed a GWAS on each phenotype separately by fitting an individual linear model for each variant in the genome using the full data. For the titre GWAS, we included focal SNP, qPCR plate, and date of injection as linear predictors. For the mortality GWAS, we included focal SNP, sex, and a sex-by-SNP interaction as linear predictors. Models were fitted using the base R linear model function ‘lm()’. We tested the significance of the SNP and SNP-by-sex predictors with a t-test, and we obtained significance thresholds for each GWAS by permuting genotypes across phenotypes 1000 times and recording the lowest p-value for each pseudo-dataset.
We chose 19 genes identified near significant GWAS hits to further test their involvement in KV infection. For each gene, we crossed a transgenic line containing a homologous foldback hairpin under the control of the Upstream Activating Sequence (UAS) to two GAL4 lines: w*; P{UAS-3xFLAG.dCas9.VPR}attP40, P{tubP-GAL80ts}10; P{tubP-GAL4}LL7/TM6B, Tb1 (Bloomington line #67065; hereafter referred to as tub-GAL4) and w*; P{GawB}Myo31DFNP0001/CyO; P{UAS-3xFLAG.dCas9.VPR}attP2, P{tubP-GAL80ts}2 (Bloomington line #67067; hereafter referred to as myo31DF-GAL4). These lines drive GAL4 expression in the entire fly and in the gut, respectively, and contain a temperature-sensitive Gal80, which is able to inhibit GAL4 at the permissive temperature (18 degrees). RNAi lines included the following genes (BDSC numbers): Pkcdelta (28355), btd (29453), dos (31766), tll (34329), Atg10 (40859), Dgk (41944), Cip4 (53321), hppy (53884), LpR2 (54461), CG5002 (55359), sev (55866), eya (57314), Gprk2 (57316), Sox21b (60120), CG11570 (65014), ATPCL (65175), Pdcd4 (66341), CG7248 (67231), and yin (67334). As a control, we crossed the genetic background of the RNAi lines (Bloomington line #36304) to the two GAL4 lines. All crosses were made at 18 degrees. After eclosion, offspring were transferred to agar vials (10 flies per vial) at the non-permissive temperature (29 degrees) for two days to facilitate silencing of candidate genes, then injected with KV. We measured titre at 5 DPI for 5 vials of each KV-infected genotype for each GAL4 driver. We used a linear mixed model to analyse log-transformed viral titre in each knockdown relative to the genetic background controls, with GAL4 driver as a fixed effect, gene knockdown as a random effect, and with separate error variances for each GAL4 driver. If the random effect associated with a candidate gene was significantly different from zero, we concluded this gene played a role in determining the outcome of infection by KV. The specification of gene as a ‘random effect’ allows comparison of each knockdown to all other knockdowns, accounting for any possible overall effect of overexpressing a dsRNA hairpin. As a proof of principle, we confirmed knock-down of the largest-effect gene (Cip4) using the DRSC FlyPrimerBank qPCR primers Cip4_PP33370F (ATTGCGGGAGTGACGCTTC) and Cip4_PP33370R (CTGTGTGGTGAGGTTCTGCTG). We did not assess knockdown efficiency for the other crosses, and any negative findings should be treated with caution.
We next aimed to characterise the host expression response to KV infection, and whether this differed between males and females. We injected 6 vials of 10 flies for each sex with either the control gradient solution or with KV. After 3 DPI, we homogenised flies in Trizol, extracted total nucleic acid, and enriched the sample for mRNA through DNAse treatment and poly-A selection. We used the NEB Next Ultra Directional RNA Library Prep Kit to make strand-specific paired-end libraries for each sample, following manufacturer’s instructions. Libraries were pooled and sequenced by Edinburgh Genomics (Edinburgh, UK) using three lanes of an Illumina HiSeq 4000 platform with strand-specific 75 nucleotide paired end reads. We subsequently identified a low level of Drosophila A Virus (DAV) contamination in both KV treated and untreated flies, reflecting the widespread occurrence of this virus in fly stocks and cell cultures. All reads have been submitted to the European Nucleotide Archive under project accession ERP023609.
We used SPAdes genome assembler (v3.11.1; [43]) to assemble the KV genome from RNA-sequencing reads, using the previously published genome (NC_033829.1) as an ‘untrusted contig’ (File S1).
We removed known sequence contaminants (primer and adapter sequences) from the paired end reads with cutadapt (V1.8.1; [44]) and mapped remaining reads to the D. melanogaster genome (FlyBase release r6.15) and to all known Drosophila virus genomes using STAR (V2.5.3a; [45]), with a maximum intron size of 100 KB, but otherwise default settings. We counted the number of reads mapping to each gene using the featurecounts command in the subread package (V1.5.2; [46]) and used these raw count data as input to DESeq2 (V1.16.0; [47]) for differential expression analysis. DESeq2 fits a generalised linear model for each gene, where read counts are modelled as a negative binomially distributed variable [47,48] and includes a sample-specific size factor and a dispersion parameter that depends on the shared variance of read counts for genes expressed at similar levels [47,48]. Our design matrix included sex, KV infection status, and the interaction between the two, allowing us to test for expression changes following KV infection and how these changes differ between the sexes. To account for the unintended presence of DAV, and differences in the level of DAV within and between the treatments, we also include the relative titre of DAV as a continuous predictor. Using this model, we calculated log2 fold changes in DESeq2, and tested for significance using Wald tests. We used the ‘plotPCA’ function implemented in DESeq2 to perform principal component analysis of the rlog-transformed read count data [47].
We performed five independent gene ontology (GO) term enrichment analysis, using: (1) genes with significant SNPs in the GWAS for titre; (2) genes with significant SNPs in the GWAS for mortality, (3) genes upregulated in either sex (p < 0.001); (4) genes downregulated in either sex (p < 0.001); and (5), genes significantly different between males and females (p < 0.05). For each of these gene lists, we tested for GO term enrichment using the ‘goseq’ R package (V1.26.0; [49]), which accounts for the difference in power for detecting differential expression caused by gene length, and tests for significant over-representation of genes in a GO term.
We performed a network analysis on genes identified in GWAS or RNA-sequencing studies at a liberal significance threshold (p < 0.10) to infer broadly acting pathways involved in KV infection that may have been overlooked in individual gene analyses. We used the PPI-spider tool [50] available on the bioprofiling webserver [51], which uses the IntAct database to find enriched subpathways within a provided gene list, allowing one gene absent from the provided list to mediate an interaction. Enriched pathways in the given gene list are then compared to random gene lists of the same length to assess significance.
We isolated Kallithea Virus (KV) by gradient centrifugation following 4 rounds of serial passage in flies. Many laboratory fly stocks and cell culture lines are persistently infected with RNA viruses [17; 31], and following serial passage we identified co-infections of DAV, Nora Virus, and Drosophila C Virus (DCV) by PCR. The high prevalence of these viruses in laboratory stocks presents a substantial hurdle in the isolation of new Drosophila viruses, requiring the separation of the new viruses of interest. Although this can be relatively simple (e.g. separating enveloped from non-enveloped viruses), most of the recently identified Drosophila viruses [31,52,53] are from ssRNA virus families with buoyant densities similar to common laboratory infections. To exclude these from our isolate, we concentrated KV using a 1.18 g/mL cushion, retaining KV at the interphase, but excluding most of the contaminating RNA viruses. Subsequent equilibrium density gradient centrifugation produced a KV band at 1.17 g/mL, and with some DAV contamination at approximately 1.20 g/mL (Fig 1A). Although nudiviruses have not previously been prepared using an iodixanol gradient, the equilibrium buoyant density was consistent with the lower buoyant densities of enveloped particles [54] and similar to other enveloped dsDNA viruses (e.g. Herpesviruses: 1.15 g/mL). KV was estimated to be an approximately 650-fold higher concentration than DAV at 1.17 g/mL, and we were unable to identify intact DAV particles by electron microscopy (KV shown in Fig 1B). KV is morphologically similar to Oryctes rhinoceros nudivirus [55], with an enveloped rod-shaped virion approximately 200 nm long and 50 nm wide.
We injected the KV isolate into Drosophila Oregon R males and females, with and without Wolbachia, and measured viral titre at four time-points by qPCR. In females, KV increased approximately 45,000-fold by day 10, and then began to decrease (Fig 1C). In males, the KV growth pattern was altered, growing more slowly (or possibly peaking at an earlier un-sampled time point), resulting in a 7-fold lower titre than in females after 10–15 days, (nominal MCMC p-value derived from posterior samples, MCMCp = 0.002). Wolbachia did not affect virus growth rate in either sex (MCMCp = 0.552, S1 Fig), reaffirming previous findings that Wolbachia do not offer the same protection against DNA viruses in Drosophila as they do against RNA viruses [56].
Nudiviruses have previously been reported to spread through sexual and faecal-oral transmission routes. The Drosophila innubila Nudivirus (DiNV), a close relative of KV, is thought to spread faecal-orally, so we tested whether KV can spread through infected food. We found that although oral transmission occurred, it was relatively inefficient (Fig 1D). However, the concentration of DiNV found in D. innubila faeces is broadly similar to our KV isolate after gradient centrifugation ([19]; Fig 1D), but the administered suspension had been diluted 50-fold and may consequently provide a lower dose than flies encounter naturally. To explore the potential for transovarial vertical transmission or gonad-specific infections following sexual transmission (as reported for Helicoverpa nudivirus 2; [57]), we also performed qPCR on dissected ovaries and the remaining carcasses at 3 DPI (S2 Fig). We found that KV was highly enriched in the carcass relative to the ovaries. Although intra-abdominal injection could influence KV tissue-specificity, there were still substantial levels of KV in the ovaries, indicating there is not a complete barrier to infection. These results imply that KV is likely transmitted faecal-orally, as are closely related nudiviruses, but explicit tests for transovarial or sexual transmission are required.
Drosophila innubila infected with DiNV suffer fitness costs including increased mortality and decreased fecundity [19]. We investigated KV-induced mortality in D. melanogaster by injecting males and females, with and without Wolbachia, with either control gradient solution or KV. We found that KV caused slightly, but significantly, increased mortality in females compared with controls (21% dead by day 20, vs. 11% in controls, MCMCp = 0.001), but caused a dramatically increased mortality in males compared to females (63% dead by day 20, vs. 14% in controls, sex:virus interaction MCMCp < 0.0001; Fig 2A). Therefore, males appear less tolerant of infection by KV, displaying increased mortality and a lower titre than females. We confirmed the KV-induced male death was not caused by DAV or other unknown small unenveloped RNA viruses present in our initial isolate, as chloroform treatment of the KV isolate eliminated treatment associated mortality (S3 Fig). Male-specific costs of infection are widespread across animal hosts and their pathogens (e.g. [58]), and reduced male tolerance has been found in flies infected with DCV [59]. We found that Wolbachia infection had no detectable effect on KV-induced mortality in males or females, and thus does not affect tolerance (MCMCp = 0.20; S1 Fig). This is consistent with previous studies showing that Wolbachia infection affects resistance and tolerance to RNA viruses but not a DNA virus [56].
We next tested whether female flies suffer sub-lethal fitness costs, by monitoring fly movement for a week following infection. KV-infected female flies showed similar movement patterns to chloroform-treated KV-injected flies for two days post-infection, but from three days post-infection moved significantly less (~70% reduction relative to controls; MCMCp < 0.001; Fig 2B). We conclude that females suffer from increased lethargy resulting from KV infection. In a natural setting, this could translate into fitness costs associated with increased predation, and reduced egg dispersal, mating, and foraging.
Finally, we tested whether KV infection resulted in decreased fecundity by monitoring the number of eggs laid for 8 days post-infection. We found that infected females exhibited markedly different egg laying patterns (MCMCp < 0.001; Fig 2), with KV-infected flies consistently laying fewer eggs between 7 and 8 days post-inoculation. This reduction in egg-laying during late infection could be due to a behavioural response or a cessation of oogenesis. To differentiate between these possibilities, we dissected ovaries, and determined the proportion of ovarioles that contained mature egg chambers. We found that ovaries from KV-infected flies halt oogenesis around stage 8 (MCMCp < 0.001), before vitellogenesis, and house an increased number of apoptotic egg chambers (MCMCp < 0.001) (Fig 2). This phenotype is similar to that seen upon starvation [60], and could be the manifestation of a trade-off to reroute resources to fighting infection, or of sickness-induced anorexia (e.g. [61]). Alternatively, this could be a direct consequence of viral infection, consistent with the gonadal atrophy reported for HzNV-2 [57]. Future studies should address whether this phenotype is a direct or indirect consequence of infection, and if the latter, whether it is orchestrated by the host or the virus.
The DGRP [37] have previously been used to dissect genetic variation underlying resistance and tolerance to bacterial, fungal, and viral pathogens [12,13,62,63]. We infected 125 DGRP lines with KV and estimated broad-sense heritabilities (H2: the proportion of phenotypic variance attributable to genetic line) and coefficients of genetic variation (CVG: a mean-standardised measure of genetic variation) in viral titre and LT50 values in females, and LT50 values in males (Table 1). Our estimates of H2 and CVG fall within the range found for resistance to other pathogens in the DGRP, although direct comparison is difficult as studies are inconsistent in the statistics used to report genetic variation. H2 in survival following infection with an opportunistic bacterium or fungus was similar to our estimate for survival following KV infection (Pseudomonas aeruginosa: H2 in males = 0.47, H2 in females = 0.38; Metarhizium anisopliae: H2 in males = 0.23, H2 in females = 0.27; 13), although comparing heritability can be easily confounded by differences in environmental (residual) variance [41]. Genetic variation in resistance has also been measured in response to two non-native D. melanogaster viruses (Flock House Virus and Drosophila affinis Sigma Virus) and two native viruses (DCV and DmelSV) in females of the DGRP. Of these, the lowest heritabilities are those associated with resistance to non-native fly viruses (FHV: narrow sense heritability h2 = 0.07, CVG = 7; D. affinis sigma virus: h2 = 0.13), and the highest are associated with native fly viruses (DCV: h2 = 0.34, CVG = 20; DmelSV: h2 = 0.29). Although Magwire et al [12] inferred h2 as half VG and accounted for the homozygosity of inbred lines when inferring CVG, it is clear that VG for resistance to KV is closer to the VG for resistance to other native fly viruses than to non-native ones, at least for survival. It is also notable that CVG estimates for survival are higher than estimates for titre, consistent with the observation that traits more closely related to fitness are expected to have higher CVG values [41].
We calculated genetic correlations between male and female mortality, and between viral titre and mortality in females (Fig 3). Note that we found no correlation between survival time following KV infection and published estimates of longevity in the absence of infection, nor to resistance to any other RNA viruses [12,64]. We found a strong positive correlation between males and females in median survival time following KV infection (0.57 [0.34–0.78]; MCMCp <0.001), such that lines in which infected males die quickly are also lines in which infected females die quickly, suggesting a shared genetic basis for early lethality following infection. We also surprisingly find a positive genetic correlation between viral titre and LT50 values (r = 0.32 [0.05–0.59], MCMCp = 0.017), such that fly lines that achieved higher titres on day 8 tended to live slightly longer. However, the effect size is small (a doubling of viral titre led to a half-day increase in median survival time) and the result is only marginally significant. The absence of a negative correlation is counter-intuitive, and contrasts with infection of the DGRP with Providencia rettgeri and Metarhizium anisopliae, and infection across Drosophila species with DCV, where fly lines or species with higher parasite loads suffer increased mortality [13,63,65]. This apparent decoupling of titre and mortality could result from inherent costs associated with the induction of an immune response, whereby flies that raise a more potent immune response keep KV at lower titres but induce greater tissue damage.
Using the phenotypes in the DGRP lines measured above, we performed a genome-wide association study to identify candidate genes underlying variation in titre, LT50, and differences in LT50 between the sexes. We found 10 SNPs (9 near genes) that were significantly associated with viral titre (prand < 0.05, based on 1000 random permutations of phenotypes across lines; Fig 4). The SNP with the smallest p-value appeared in Lipophorin receptor 2 (LpR2), which encodes a low-density lipoprotein receptor, previously found to be broadly required for flavivirus and rhabdovirus cell entry (e.g. [66–68]).
We tested whether these candidate polymorphisms were enriched in any molecular, biological, or cellular processes using a GO enrichment analysis, and found the top hit to be the torso signalling pathway with 2 genes of 34 in the category (p = 0.0004), tailless and daughter of sevenless (dos). Torso signalling is upstream of extracellular-signal-regulated kinase (ERK) pathway activation in some tissues, and human orthologues of dos (GAB1/GAB2/GAB4) are cleaved by an enterovirus-encoded protease, thereby activating ERK signalling and promoting viral replication [69,70]. ERK signalling is also an important regulator of virus replication in the fly midgut, where it couples nutrient availability with antiviral activity [71,72]. See S1 Table for a list of all nominally significant SNPs with associated locations, mutation types (e.g. intronic, synonymous coding, etc), nearby genes, p-values, effect sizes, and GO terms.
We found 86 SNPs (65 near genes) that were significantly associated with LT50 following KV infection in the DGRP (prand < 0.05; Fig 4 and S1 Table), none of which were identified in the GWAS for viral titre. We performed a GO enrichment analysis, and found genes associated with these SNPs were enriched for hydrolase activity (top molecular function GO term, p = 0.0004), stem cell fate determination (top biological process GO term, p = 0.002), and in the plasma membrane (top cell component GO term, p = 0.004), among others (S2 Table). Of these 86 SNPs, we found 34 (26 near genes) that were highly significant, and selected these for further analysis and confirmation (prand < 0.01; see S1 Table for all significant SNPs). The polymorphism with the most confident association was located in Cdc42-interacting protein 4 (Cip4), a gene involved in membrane remodelling and endocytosis [73,74]. This SNP is intronic in the majority of Cip4 transcripts, but represents a nonsynonymous polymorphism segregating leucine and proline in the first exon of Cip4-PB and Cip4-PD isoforms, perhaps indicating spliceform-specific effects on KV-induced mortality. Of particular interest from the remaining 33 highly significant SNPs was a synonymous SNP in the receptor tyrosine kinase, sevenless, known to interact with dos (above), and seven genes (Dgk, Atg10, ATPCL, Hppy, Pkcdelta, Gprk2, Pdcd4) previously implicated in viral pathogenesis or general immune processes. Of these, three (Gprk4, hppy, Pkcdelta) are involved in NF-κB signalling [75–77]. ATPCL was identified in an RNAi screen for factors regulating Chikungunya virus replication in humans [78] and is involved in the late replication complexes of Semliki Forest Virus [79]. Finally, Atg10 and Pdcd4 are involved in autophagy and apoptosis, respectively, both broadly antiviral cellular functions known to have a role in antiviral immunity in Drosophila [80,81]. We found no SNPs significantly associated with sex-specific KV-induced mortality (S4 Fig).
We chose 18 GWAS-candidate genes with available UAS-driven RNAi constructs to verify their involvement in KV infection. We found that knockdown of Cip4 and CG12821 caused significantly increased viral titre, and knockdown of sev and dos resulted in significantly decreased viral titre, relative to other knockdown lines (Figs 5 and S5). We confirmed tub-GAL4>Cip4IR flies had reduced (26% of wild-type) Cip4 RNA levels and a concomitant increase in viral titre relative to the genetic background control (3.4-fold increase, 95% C.I. 1.3–9.6 fold) (Fig 5). This strongly suggests that Cip4 is a KV restriction factor that likely segregates for functional polymorphism affecting survival following KV infection (Fig 5). It is known that baculovirus budded virions enter cells through clathrin-mediated endocytosis or micropinocytosis [82,83], and gain their envelope at the cell membrane upon exit [84]. Cip4 could therefore plausibly enact an antiviral effect by limiting KV cell entry or spread, perhaps through its known function in cell membrane remodelling and trafficking.
Previous transcriptional profiling in response to RNA virus infection has shown upregulation of heatshock proteins, JAK-STAT, JNK, and Imd pathways [4,28,85,86]. However, the D. melanogaster expression response to a DNA virus has not previously been investigated. We separately injected male and female flies with control gradient solution or KV and extracted mRNA for sequencing 3 days post-infection. KV gene expression increased dramatically 3 days post-inoculation, consistent with our qPCR analysis of genome copy-number (Figs 1 and S10). Although not previously detectable by PCR, RNAseq read mapping identified a low level of DAV in both control and KV-infected flies, with an overall higher level in KV-infected flies. To account for this potentially confounding contaminant, we fitted the number of DAV-mapped reads as a covariate in the differential expression analysis, and used a stringent Benjamini–Hochberg adjusted significance threshold of p < 0.001 to infer nominal significance. We found 54 genes upregulated and 79 genes downregulated in response to KV in either males or females (Fig 6 and S3 Table). There was no enrichment for GWAS hits among the KV-responsive genes (S6 Fig). Principal components analysis on depth-normalised read counts separated males and females along PC1 and partially separated KV-infected and control-injected libraries along PC3 (S7 Fig). GO term analysis identified ‘defense response to virus’ (p = 3.1x10-4), ‘serine peptidase activity’ (p = 1.2x10-7, identified in part due to downregulation of Jonah family serine proteases), and ‘chorion’ (p < 1x10-8) as the most highly enriched biological process, molecular function, and cellular component, respectively (Fig 6 and S4 Table). Subsequent network analysis identified a large pathway of interactions enriched for genes either differentially expressed or associated with variation in KV infection, including known defense response genes (S8 Fig).
There are few described induced antiviral immune effectors in Drosophila (e.g. [81]). In line with this we observe 57% of differentially expressed genes have not yet been named (i.e. “CG” genes), significantly greater than the genome-wide rate of 41% (p = 3x10-4), and the most highly induced genes have not been implicated in viral pathogenesis. The cytochrome P450 family gene Cyp304a1 was most highly upregulated, concomitant with the upregulation of four other genes in this family (Cyp309a1, Cyp309a2, Cyp4p3, and Cyp6a20). The next most highly induced genes include the hemocyanin Larval serum protein 2, the cytidine deaminase CG8353, four genes without functional annotation or recognisable domains (CG33926, CG31955, CG32368, CG13641), and an additional six genes without functional annotation (CG43064, CG42825, Gagr, CG10211, CG17264, and CG17224 –the last two of which are adjacent on chromosome arm 2L). We also note the striking but variable upregulation of 11 of the 24 Tweedle genes (S9 Fig) in some (but not all) of the infected samples. These are secreted, insect-specific cuticle proteins that regulate body shape [87], and are also upregulated in response to Sindbis virus infection in cell culture [88], perhaps suggesting a general role in viral pathogenesis.
Genes with known involvement in viral pathogenesis were also found to be induced following KV infection. The RNAi effector AGO2 was upregulated, consistent with the previous results that DNA viruses are a target of the RNAi pathway [21,28,29]. Vago, an antiviral factor downstream of Dicer-2 [89], was upregulated and was also adjacent to a SNP found in the mortality GWAS (dos; Figs 4 and 6), as were pastrel and ref(2)P, identified in previous genome wide association analyses for resistance to DCV and DMelSV, respectively. Finally, we found that KV induced expression of CG1667, the Drosophila homologue of STING. The vertebrate cGAS-STING pathway is involved in cytosolic DNA sensing and activation of immune factors in response to DNA virus infection [90]. This upregulation of CG1667 may suggest that this is another pathway conserved between Drosophila and vertebrates.
As we had observed male and female differences in KV-induced mortality and titre (Figs 1 and 2), we tested for sex-specific transcriptional regulation in response to KV infection. We found that females and males had similar patterns of differential expression following KV infection (spearman rank correlation coefficient, ρ = 0.57, p = 2.2x10-16), although the male response was often less potent (Fig 6). Nine genes were significantly differentially expressed (p < 0.05) between the sexes specifically in response to KV (S3 Table), and these were all downregulated in females and highly enriched for genes associated with the chorion (Fig 6 and S4 Table). Strikingly, all but three genes classified with the GO term ‘chorion’ were downregulated in females (Fig 6), consistent with the observed reduction in mature ovarioles and eggs during late infection, and implying a substantial reorganization of oogenesis (Fig 2). We did not identify any previously described immune genes with significant sex-specific regulation during KV infection.
We have isolated Kallithea virus, a dsDNA nudivirus that naturally infects D. melanogaster, and find it to be experimentally tractable. KV infection leads to reduced fertility and movement in females, highlighting the importance of measuring fitness associated traits besides longevity. Although males suffered greater mortality than females, they achieved lower titres, consistent with increased resistance and/or reduced tolerance in males. Similar to RNA viruses, we identified moderate host genetic variation in resistance to KV infection, however, we found that the underlying genetic architecture of this variation is unlike previously studied RNA viruses of D. melanogaster, in which a high proportion of genetic variation was apparently determined by a small number of loci. This could reflect a difference in the co-evolutionary dynamics between D. melanogaster and KV, versus other RNA viruses such as DCV and DMelSV. The D. melanogaster transcriptional response to KV included genes with known involvement in viral pathogenesis, but also genes that could represent infection responses distinctive to DNA viruses or KV, including downregulation of chorion genes. Upregulation of widely conserved immune factors, such as STING, represent promising candidates involved in fly antiviral immunity, and demonstrate the continued utility of the Drosophila system for understanding host-virus interactions.
|
10.1371/journal.pntd.0006219 | Caregivers’ views on stigmatization and discrimination of people affected by leprosy in Ghana | Leprosy is a condition that has long been associated with stigma and discrimination, even when infected persons have been cured. This paper describes stigma and discrimination as viewed by caregivers who are associated with people affected by leprosy in Ghana.
A qualitative interview with semi-structured interviews were conducted for twenty caregivers.
Findings indicated that caregivers were of the view that people affected by leprosy in Ghana are stigmatized and discriminated against by the larger society thus making their movements and interactions restricted to the Leprosarium. Besides, employments opportunities are unavailable to them thus making them exposed to financial challenges. The livelihood Empowerment Against poverty (LEAP) money given them is not sufficient for their daily upkeep.
People affected by leprosy in Ghana are stigmatized and therefore find it difficult to interact freely with the public. The associated physical deformities with the disease also tend to impede their ability to relate to the general public. The LEAP cash given to people affected by leprosy is helpful however, it could be enhanced to keep pace with prevailing economic conditions in the country.
| In Ghana, the social interpretation of leprosy regardless of the language, culture and tradition engenders stigmatization and discrimination that leads to social rejection and exclusion of persons who have been cured of the disease. Often, these persons are cared for by relatives who happen to live with them in a confined place. From the views of these caregivers, this paper identifies areas of stigmatizing and discriminatory tendencies against people affected by leprosy who reside in a Leprosarium in Accra, the capital city. It is expected that persons suffering from such neglected tropical disease would be empowered to enable them go about their daily routines without the fear of being rejected. Besides, the intensification of public education about curing tendencies of stigma becomes paramount.
| The understanding and management of leprosy has seen tremendous improvement over the years. This accomplishment can be attributed to the increasing availability of the Multi-Drug Therapy (MDT) that has enabled many sufferers globally to be cured of the disease. Over the years in Ghana, efforts put in place have realized a reduction of 40,000 registered leprosy cases in 1948 [1] to 345 registered cases in 2014. Besides, new cases detected was put at 366 [2] an indication that the disease is still prevalent in Ghana albeit in relatively small numbers. For instance, there is report of the prevalence of leprosy in the Sene Distirct in the Brong Ahafo Region [3], and some other districts in the Volta and Northern Regions [4].
The disease can be cured as long as the precise diagnosis is made and treatment initiated with the appropriate medication [3]. Consequently, infected persons can receive medical attention and be cured in a year and get back to normal life, especially when the disease does not reach the disabling stage [4]. Unfortunately, for many sufferers of leprosy, medical attention is sought when the disease has reached an advanced stage and caused severe deformities and disabilities [5]. For example, persons who get treatment may end up having some physical deformities such as scarring on parts of their body [6]. These physical impairments result in exacerbating the stigma associated with the disease thus sustaining the cycle of stigmatization [4]. In Ghana, the social interpretation of the disease regardless of the language, culture and tradition engenders stigmatization. The Akan language for instance, that is spoken mostly from the south to the north of the country, the condition is commonly called “KWATA” a term that is associated with apprehension for dealing with leprosy-affected people. Again, in the capital city where the Ga language is spoken by the indigenes, it is called “KPITI” which also elicits similar apprehension. This tendency tends to entrench the stigma associated with the disease. Leprosy has therefore become a disease of public health concern because of its association with stigmatization of persons suffering from the disease [7].
The concept of stigma is defined as an attribute that is deeply discrediting within a particular social interaction [8]. Following from that, stigma has been described as “the social devaluation of a person” [9]. As long as individuals who have leprosy are stigmatized, tendencies of discrimination are exhibited towards them. These discriminatory tendencies lead to disadvantages in many areas of life that include personal relationships and work. These individuals tend to accept the situation thus culminating in an acceptance of the discrediting prejudices held against them, which tend to diminish their self-esteem, which also leads to feelings of shame, a sense of alienation and social withdrawal [10, 11].
Over the years, advancements have been made in the treatment of the disease and in Ghana, a number of people who once had the disease have been cured of it. That notwithstanding, stigma is still high for these people who are mostly unable to integrate back into their previous communities for fear of rejection. Some studies have shown that the stigma associated with the disease is heightened by the physical deformities associated with the disease [12]. Clearly, the lack of understanding and knowledge about leprosy increases the misconceptions about the transmission and treatment of the disease [13]. In effect, people affected by leprosy continue to experience negative connotations such that they are perceived to be lepers, a terminology that in itself is synonymous to stigma [4].
In spite of the cured status of persons living at the Weija Leprosarium in Ghana, they continue to experience stigma and discrimination because the social pathology of the disease continues to be associated with them. For example, people affected by leprosy at the Leprosarium reported experiences of stigma and discrimination from their families, friends, healthcare providers and community members [14]. In the lives of people affected by leprosy are caregivers who take on different roles in the daily upkeep of the latter [15]. These caregivers are among those who come into close contact with people affected by leprosy and also interact freely with the general public. Per their caregiving role, they observe at firsthand stigmatizing and discriminatory tendencies that are exhibited by the public towards persons who have been cured of leprosy. It is of paramount significance to be aware of the views of these caregivers so as to help develop an appropriate and comprehensive intervention to address stigma and discrimination tendencies by the public towards persons affected by leprosy. This study therefore sought to understand those experiences from the perspectives of the caregivers at the Weija Leprosarium in Accra, so as to confirm that people affected by leprosy in Ghana do experience stigma and discrimination.
This study draws on the Modified Labeling Theory (MLT) developed by Link and colleagues [16, 17], and the conceptualized four dimensional mechanisms of perceived stigma [18]. Per the MLT, stigma is an internal process that inherently involves the negative responses of persons in the environment, which is defined as the “labeling” behaviors of others.
People in the society tend to be apprehensive coming into contact with people affected by leprosy. Besides, there is hesitation whenever people have to relate freely with people affected by leprosy. The theory proposes that labeled individuals will respond behaviorally to anticipate social rejection. In effect, people affected by leprosy in anticipation of how the society will react towards them would rather keep to themselves and mostly remain in the leprosarium. Further, harmful effects may arise from internalized conceptions of anticipated stigma or from the stigma coping response enacted. Labeling thus may negatively affect one’s psychological state.
The four dimensions of perceived stigma include social rejection (e.g., friends, family, colleagues abandoning not wanting to get close to people affected by leprosy), financial insecurity (e.g., feeling financially inadequate), internalized shame (e.g., feelings of embarrassment about deformities of people affected by leprosy), and social isolation (e.g., limiting social contact due to societal behaviours towards people affected by leprosy). People affected by leprosy at the Weija Leprosarium tend to experience all these dimensions. For people who experience internalized stigma, they may suffer poor psychological wellbeing [19, 20].
The ethical review and approval of the study was sought from the Ethics Committee for the Humanities (ECH) at the University of Ghana. Permission was also sought from the authorities of the Weija Leprosarium in the Greater Accra Region of Ghana. Furthermore written and informed consent forms were signed or thumb-printed by study participants before they participated in the study. The objectives of the study and study procedures were explained to all participants. In addition, anonymity and confidentially was assured to the participants prior to each interview. Participants were made aware that their participation was entirely voluntary and they had the right to refuse to participate or to withdraw from the study at any time if they so desired. It was explained to the participants that their participation in the study would not pose any risks to them and also refusal to participate will not affect any services provided to their respective cured leper. All information received were anonymized and can therefore not be traced to any particular participant.
A qualitative research design was employed for this study because it provides complex social processes that capture important aspects of a phenomenon from the perspective of study participants [21]. Semi-structured in-depth-interviews [22] were carried out with purposively selected caregivers at the Weija Leprosarium. This technique was used because we were interested in informants who have the best knowledge or experience concerning the research topic and with the expectation that each participant will provide unique and rich information of value to the study [23].
The selection criterion was caregivers who had at least one year continuous experience in providing care to the person affected by leprosy. Subsequently, appointments were scheduled with caregivers who were willing to participate in the study at their convenience.
The interviews were conducted and recorded in English, and Twi (a local dialect) that was mostly understood and spoken by participants who could not speak English. The interviews were conducted in a secluded environment within the leprosarium so as to avoid interruptions by other people.
The interviews were conducted in person, with two interviewers and a participant. The two interviewers were linguistically competent in both English and Twi (one of the local dialects). These interviewers took turns to moderate all the interviews. The interviews focused on issues such as stigma and discrimination among people affected by leprosy, challenges they encounter in accessing health care and employment opportunities available as well as support systems available to them. In all, twenty in-depth interviews were conducted. When it was realized that no new further information was obtained, saturation was attained and therefore the interviews were ended [24, 25]. All the interviews were audio-recorded and field notes taken by two research assistants who were fluent in both English and Twi.
The qualitative responses recorded during the In-depth Interviews (IDIs) were translated verbatim and transcribed by two interviewers into English. Further, the notes taken during the interview sessions were expanded. In a situation where there was disagreement between the two interviewers, the transcripts and the original recordings were reviewed until consensus was reached. The transcripts and expanded notes were stored as files and coded manually for textual analysis in accordance with the principles of grounded theory [24]. Coding was specifically done by placing blocks of text into various nodes based on the categories and subcategories. Using the categories, information was compared across the transcripts based on similar and contrasting views of caregivers’ on stigma and discrimination among people affected by leprosy. The themes were illustrated with verbatim quotes and interpreted based on existing literature.
A total of 20 caregivers took part in the study. They were made up of 8 males and 12 females whose ages ranged from 18 to 70 years. The males were made up of a husband, two sons, and five grandchildren. On the other hand, the females were made up of two wives, three daughters and seven grandchildren. Of the 20 participants, only six had no formal education, the rest of them had some education ranging from primary through to post-secondary. Majority of them had some form of employment. Whereas 16 of them indicated that they were Christians, 3 were Muslims and only 1 traditionalist.
The caregivers demonstrated knowledge of stigma and discrimination through their own behaviours prior to the care giving role and their observations of the general public. As caregivers of people affected by leprosy, participants perceived that they themselves at the initial stage, had strong stigmatizing and discriminatory tendencies about the leprosarium and the leprosy disease.
For many people who come into contact with the facility, their apprehensions are expressed in diverse ways. When the contact is with the residents of the facility, then there is the likelihood of heightened apprehension in spite of the cured status of the residents at the leprosarium. The participants in this study indicated these apprehensions as they noted:
Similarly, other participants made the following statements:
The above statements attest to the initial apprehensions that are exhibited by people who come into contact with people affected by leprosy. Over time however, these apprehensions tend to dissipate suggesting that continuous interaction with the leprosarium would get rid of the fear that is otherwise associated with the leprosarium and persons who have been cured of the disease.
Whereas caregivers over time get used to the leprosarium and the inmates, the public on the other hand continue to exhibit stigmatizing and discriminatory behaviours towards people affected by leprosy. These behaviours reinforce the social unacceptability of people affected by leprosy. Caregivers were of the view that the public through their actions make it obvious that people affected by leprosy are stigmatized. Participants expressed sentiments that indicate that people affected by leprosy and anything associated with them tend to be devalued. The following statements made by participants emphasize that notion:
It is clear from the statements above that stigma and discrimination exhibited towards people affected by leprosy is extended to products they sold thus curbing their desire to engage in any income generating activities. The fore-going expressions of the caregivers are manifestations of the theoretical dimensions of social rejection, financial insecurity, and social isolation that people affected by leprosy experience in their lives.
Caregivers noted that people affected by leprosy could access health care services from the clinic within the Leprosarium. This clinic was purposely built with the support and funds from a Philanthropist to attend solely to people affected by leprosy when it had to do with very minor cases. Besides, there is a Municipal Hospital that is adjacent to the Leprosarium where people affected by leprosy are sent when their health condition is beyond the capacity of the clinic. In instances where their conditions require specialized attention, they are sent to major health facilities in the city. Apart from the clinic within the Leprosarium that provided care without any show of stigma, caregivers were of the view that people affected by leprosy experienced stigma and discrimination when they attend to other health care facilities, thus reinforcing the social rejection experienced by people affected by leprosy A caregiver for example in a statement said:
In a similar way, another caregiver also recounted an experience she witnessed while she had taken a person affected by leprosy to a major hospital for a surgical operation.
For all the people affected by leprosy who participated in the study, they had been enrolled onto the National Health Insurance Scheme (NHIS) which they found to be helpful. That notwithstanding, there were other additional health costs that NHIS did not cover. In instances like that, people affected by leprosy themselves had to bear the cost. Incidentally, people affected by leprosy were mostly not financially empowered, thereby emphasizing their financial insecurity. As a result, those additional costs had to be borne by the Lepers’ Aid Committee. The following statements by caregivers attest to that arrangement:
It has always been the right of people to be employed as long as they possess the requisite skills and qualifications needed for a particular job. The expectation therefore is that employment rights of people affected by leprosy be respected. This study revealed that these rights are highly disregarded by most if not all employers. Participants in this study for example mentioned that many employers were unwilling to employ them because of their negative perceptions and beliefs about the disease. When caregivers in this study were asked questions in respect of employment opportunities for people affected by leprosy, their responses suggested that people affected by leprosy were socially rejected in respect of employment that ultimately leads to their financial insecurities. The caregivers made the following statements in their responses:
As to why people affected by leprosy had difficulty getting employed, caregivers enumerated some concerns that could pass for reasons why employers do not want to engage people affected by leprosy. In the view of the caregivers, employers will do then a lot of good if they were to outline the reasons why they refuse to engage people affected by leprosy. That would have provided some clarity as to who can apply for a job. Participants expressed these sentiments in the following quotes:
People who have been cured of leprosy may benefit from social services such as counseling and programmes that seek to empower them in their daily lives. Caregivers interviewed indicated that there were no counseling programmes in place for people affected by leprosy except for the LEAP cash benefits that is given to them periodically. Even that, caregivers were of the view that the amount of money given to people affected by leprosy was inadequate considering the fact that the cost of living lately has increased, more so when people affected by leprosy do not have any other source of income basically because they do not have any job. Caregivers expressed their sentiments in the following statements:
It is clear from the foregoing that the only formal support available to the people affected by leprosy is the LEAP programme. In respect of any formal social services instituted for the people affected by leprosy, no such programme exists. This financial support is provided obviously because people affected by leprosy are financially insecured.
The results of this study were specifically on caregivers’ perspectives of stigma and discrimination experienced by people affected by leprosy who live in one of the five known leprosaria located in the capital city Accra of Ghana, thus the results may not be generalised to people affected by leprosy in the other leprosaria. Besides, the small numbers of caregivers who participated in the study puts limitation on generalisation of findings to all caregivers of people affected by leprosy in Ghana. It is therefore recommended that future studies will consider caregivers in the other leprosaria in Ghana.
The purpose of this study was to explore caregivers’ perspectives of stigma and discrimination among people affected by leprosy in Ghana. The findings indicate that caregivers are witnesses of stigma and discrimination that people affected by leprosy’ experience. These experiences demonstrate the Modified Labeling Theory [17], that stigma ultimately is an outcome of inherent negative responses in the environment that labels behaviours of others. The four dimensional provision of the theory is a manifestation of experiences that are encountered by people affected by leprosy in Ghana. These are first, the perceived stigma that includes social rejection from relations and the general society. Findings in this study corroborates the long held construction of stigma as a deep rooted phenomenon in societies [26]. Indeed certain entrenched beliefs as well as the lack of knowledge, fear and shame associated with the disease result in the stigmatization of people affected by leprosy [26]. Consequently, irrational behaviours are exhibited towards them. Advances in medicine over the years have turned leprosy into a completely curable disease that can be rendered non-infectious. This is attributed to the introduction and subsequent implementation of the Multi-Drug Therapy (MDT) that has demonstrated to be effective, in curing leprosy [27]. That notwithstanding, many people are unaware of this development thus maintain a certain distance even with people affected by leprosy. The belief therefore continues to be held that one could be infected by getting close to people affected by leprosy. This fear of infection culminates in apprehension, a situation that is similar to a study in Nepal that found the fear of transmission makes people attempt to keep distance with affected persons [28].
Additionally, caregivers in our study identified some hesitation and reluctance on the part of medical practitioners to provide care to people affected by leprosy who are sent to health facilities. These discriminatory tendencies exhibited by health personnel notably some medical Doctors who are assumed to know and understand aspects of the disease including its etiology, causation, means of transmission and curability do not help the course of people affected by leprosy who may require health attention when the need arises. This tendency reinforces stigma associated with the disease [29]. Besides, it is an affront to the World Health Organisation’s (WHO) encouragement for the integration of leprosy into the general health service where leprosy patients should be treated in the same outpatient department as those with any other disease [30]. This can signal to the patients and their communities that leprosy is not a 'different' disease. It has been found that positive attitude of health professionals can contribute significantly in the reduction of stigma due to leprosy [31].
It was evident from our study that people affected by leprosy are financially vulnerable basically because they could not engage in any income generating activities. For those who tried to undertake some economic venture, it could not prevail as soon as patrons realized that persons behind were people affected by leprosy. It is also the case that people affected by leprosy end up having some deformities which in most cases disable them from doing any meaningful work. The consequences of this is loss of earning capacity [32]. To help fill this gap is the LEAP, a social intervention programme instituted in Ghana for vulnerable people. People affected by leprosy have been enrolled unto the programme alright but in the view of caregivers, the amount of money given to them is not sufficient enough. For instance, every cured leper receives and amount of forty eight cedis for two months (equivalent of about fourteen dollars). It is perhaps for this reason that some people affected by leprosy take to begging on the streets as a means to supplement their earnings just as it happens in Ghana.
In the view of caregivers, people affected by leprosy themselves do not express shame in respect of their deformities as long as they remain in the leprosarium where there is some feeling of belonging. This is so because individuals tend to see their personal deformities in other persons as well. It therefore becomes a situation where it is not just one person with some physical deformities. It is rather when they have to interact with the public that they encounter mocking and rude behavior.
Knowing that the deformities that come up due to the disease cannot be reversed, the United Nations General Assembly in the year 2010, unanimously adopted a resolution on the elimination of discrimination against persons affected by leprosy. That resolution was accompanied by principles and guidelines that listed measures to improve the living conditions of such persons. Further, the United Nations Convention on the Rights of Persons with Disabilities (UNCRPD) adopted in the December of 2006 seeks to promote, protect and ensure the full and equal enjoyment of all human rights and fundamental freedoms by all persons with disabilities [33]. People affected by leprosy in Ghana do not seem to enjoy these provisions outlined by the World Body. Their situation is worsened by the secluded and confined environment where they find themselves. About a decade ago for instance, the current location of the leprosarium in Accra appeared remote. In recent times however, population increase and infrastructural developments around the area have brought other communities closer to the leprosarium. In spite of that, there is still the feeling of isolation because walls have been constructed all around the facility, buttressing the phenomenon of social isolation that is still experienced by people affected by leprosy [34].
Following from that, there is limited social contact between people affected by leprosy and the generality of the populace. This social isolation results in some feeling of internal stigma among individual people affected by leprosy. Studies have shown that people who experience internalized stigma tend to suffer poor psychological wellbeing [19, 20]. Psychological well-being connotes lives going well, a combination of feeling good and functioning effectively. In effect, the expected flexibility and creative thinking as well as pro-social behavior and good physical health associated with psychological wellbeing will be lost among people affected by leprosy [35]. Reversing this situation may require an intervention that seeks active participation of people affected by leprosy in programmes that aim at addressing the stigma associated with the disease. This has proven successful in other countries. A Nepalese study for example showed that people affected by leprosy who took part in the Stigma Elimination Programme (STEP) [36] were less stigmatized and participated more in the community than those who did not take part. Besides, STEP participants were more empowered and became change agents in their own communities. This kind of intervention has been proven to be effective, as demonstrated in Ethiopia [37].
Caregivers interviewed in this study had an understanding of the extent of stigma and discrimination experienced by people affected by leprosy in the Weija Leprosarium. These experiences were generally in the areas of social rejection whereby people see people affected by leprosy as outcasts and therefore, would not want to be close with them. The manifestation of this rejection is the area of social isolation that people affected by leprosy find themselves. The result of this is the limited or no interaction that take place between them and the society in general. People affected by leprosy also had physical deformities which they are not comfortable with especially when they come into contact with the public. In addition, people affected by leprosy receive some amount of money periodically. In their estimation, those amounts could be enhanced to enable them live meaningfully.
Knowing that caregivers play an important role especially for persons with stigmatized conditions, this study brings to the fore the need for concerted efforts to reduce stigma and discrimination at the community level with subsequent extensions to generality of the population. This would mean consciously involving caregivers of people affected by leprosy in such activities where they can share their experiences as in getting close to people affected by leprosy will not make one to be infected with the disease. This will help disabuse the minds of people in respect of the negative perceptions about leprosy. There is equally the need for psychosocial interventions to be adopted for people affected by leprosy. This should include activities that seek to empower them to get over feelings of alienation, and their ability to deal with negative reactions from society. Finally, the social benefits in the form of the LEAP that is provided to people affected by leprosy can be leveraged in line with the minimum wage that prevails in Ghana. That way, whenever increments are announced, people affected by leprosy can also be assured that they will get an amount of money that could improve their barest standard of living.
|
10.1371/journal.pgen.1006183 | Overexpression of Mitochondria Mediator Gene TRIAP1 by miR-320b Loss Is Associated with Progression in Nasopharyngeal Carcinoma | The therapeutic strategy for advanced nasopharyngeal carcinoma (NPC) is still challenging. It is an urgent need to uncover novel treatment targets for NPC. Therefore, understanding the mechanisms underlying NPC tumorigenesis and progression is essential for the development of new therapeutic strategies. Here, we showed that TP53-regulated inhibitor of apoptosis (TRIAP1) was aberrantly overexpressed and associated with poor survival in NPC patients. TRIAP1 overexpression promoted NPC cell proliferation and suppressed cell death in vitro and in vivo, whereas TRIAP1 knockdown inhibited cell tumorigenesis and enhanced apoptosis through the induction of mitochondrial fragmentation, membrane potential alteration and release of cytochrome c from mitochondria into the cytosol. Intersecting with our previous miRNA data and available bioinformatic algorithms, miR-320b was identified and validated as a negative regulator of TRIAP1. Further studies showed that overexpression of miR-320b suppressed NPC cell proliferation and enhanced mitochondrial fragmentation and apoptosis both in vitro and in vivo, while silencing of miR-320b promoted tumor growth and suppressed apoptosis. Additionally, TRIAP1 restoration abrogated the proliferation inhibition and apoptosis induced by miR-320b. Moreover, the loss of miR-320b expression was inversely correlated with TRIAP1 overexpression in NPC patients. This newly identified miR-320b/TRIAP1 pathway provides insights into the mechanisms leading to NPC tumorigenesis and unfavorable clinical outcomes, which may represent prognostic markers and potential therapeutic targets for NPC treatment.
| The therapeutic strategy for advanced nasopharyngeal carcinoma (NPC) is still challenging. The most urgent need for NPC is novel treatment targets. Therefore, understanding the mechanisms underlying NPC tumorigenesis and progression is essential for the development of new therapeutic strategies. Here, we identified TRIAP1 could serve as a prognostic biomarker in NPC, and function as an oncogene in NPC tumorigenesis and mitochondrial apoptosis through inhibiting the release of cytochrome c. Moreover, miR-320b post-transcriptionally regulated TRIAP1 expression, and exhibited inhibitory effects on proliferation and promoted apoptosis through targeting TRIAP1. Thus, our study provides new insights into the mechanisms of NPC tumorigenesis and progression and identifies novel therapeutic targets for NPC treatment.
| Nasopharyngeal carcinoma (NPC) is the most prevalent head and neck malignancy in Southeast Asia, especially in Southern China [1]. A majority of NPC patients are diagnosed at advanced stages, leading to approximately 30% of NPC patients developing treatment failure [2]. Although NPC is a heterogeneous disease, a combination of radiotherapy and platinum-based chemotherapy remains the standard treatment method [3,4]. Therefore, identification of effective molecules regulating NPC development and progression is essential for developing novel therapeutic strategies.
Sustaining proliferative signaling and resisting apoptosis are typical hallmarks of cancer [5]. Mitochondria are at the core of programmed cell death or apoptosis [6,7]. Proteins involved in mitochondrial network could regulate the apoptotic pathway [8–10]. Thus, it is crucial to elucidate the molecular mechanisms of proliferation and mitochondrial apoptosis to excavate potential therapeutic targets for NPC therapy. TP53-regulated inhibitor of apoptosis (TRIAP1) is a small ~9-kDa protein transcriptionally activated by TP53 [11,12]. It has been reported that TRIAP1 protects cancer cells from apoptosis through interaction with hear shock protein 70–4 (HSP70) or the repression of cyclin-dependent kinase inhibitor 1 (p21) [12,13]. Recent evidence has also revealed that TRIAP1 contributes to the resistance of apoptosis in a mitochondria-dependent manner [14,15]. However, the function and clinical value of TRIAP1 remain unknown in NPC. In addition, TP53 is commonly inactivated in tumor cells to escape apoptosis, indicating there may be other mechanisms regulating TRIAP1 expression and extensive investigation is warrant.
MicroRNAs (miRNAs) are a class of small non-coding RNAs that negatively regulate gene expression by provoking mRNA degradation or suppressing mRNA translation [16–18]. Importantly, miRNAs have important roles in a wide range of biological processes, including cell proliferation, cell death and motility [19–21]. Accumulating evidences have shown that miRNAs are dysregulated and function as either oncogenes or tumor suppressors in different cancer types [22–24]. In our previous microarray study, a profile of deregulated miRNAs is identified in NPC [25,26], and some miRNAs affect cell growth, proliferation and metastasis in NPC [27–29]. However, it is yet unclear whether these miRNAs maintain their apoptotic effects in NPC. Therefore, understanding the role of miRNAs in apoptosis may provide insight into the mechanisms underlying carcinogenesis and aggressiveness in NPC.
In the present study, we demonstrated that TRIAP1 functioned as an oncogene in proliferation and apoptosis through preventing mitochondrial fragmentation and cytochrome c release and its overexpression was correlated with poor survival in NPC patients. miR-320b was revealed to negatively regulate TRIAP1 and exhibited proliferative inhibition and apoptotic promotion, which could be rescued by TRIAP1 overexpression. Thus, the altered miR-320b/TRIAP1 pathway contributes to the proliferation and apoptosis of NPC and may provide novel therapeutic targets for NPC treatment.
To investigate the clinical significance of TRIAP1 in NPC, we first examined TRIAP1 mRNA expression in 16 fresh-frozen NPC and 8 normal nasopharyngeal epithelial tissues. The mRNA expression level of TRIAP1 was significantly upregulated in NPC tissues (Fig 1A, P < 0.01) and in 6 NPC cell lines compared with the normal nasopharyngeal epithelial cell line NP69 (Fig 1B). In addition, protein immunoblotting analysis confirmed high TRIAP1 expression in various NPC cell lines (Fig 1C).
To further evaluate the expression status of TRIAP1 in NPC, we performed immunohistochemistry (IHC) for TRIAP1 in 204 NPC specimens. The results showed that TRIAP1 was overexpressed in 47.1% (96/204) of the NPC specimens (Fig 1D). Importantly, the level of TRIAP1 expression was strongly correlated with distant metastasis (P < 0.001) and death (P = 0.003; S1 and S2 Tables). Patients with high TRIAP1 expression showed significantly shorter 5-year overall survival (OS; 92.5% vs. 71.5%, P = 0.002) and disease-free survival (DFS; 83.1% vs. 71.5%, P = 0.001; Fig 1E) rates than those with low TRIAP1 expression. Moreover, multivariate analysis revealed TRIAP1 overexpression was an independent prognostic factor for OS (HR, 2.75; 95% CI, 1.50–5.03; P = 0.001) and DFS (HR, 2.54; 95% CI, 1.47–4.38; P < 0.001; S3 Table). Taken together, these data demonstrate that TRIAP1 overexpression is a risk factor for a poor prognosis in NPC patients.
To explore the biological role of TRIAP1 in NPC, we transiently overexpressed or knocked down TRIAP1 in CNE-2 and SUNE-1 cells (S1A–S1D Fig). While cell proliferation was significantly promoted following ectopic TRIAP1 overexpression, TRIAP1 knockdown remarkably inhibited cell proliferation (Fig 2A, P < 0.001). Furthermore, TRIAP1 overexpression significantly increased the colony-formation rate and anchorage-independent growth ability, which were impaired by TRIAP1 silencing (Fig 2B–2D, P < 0.01). These data suggest that TRIAP1 promotes NPC cell proliferation.
Furthermore, we investigated the effect of TRIAP1 on apoptosis through flow cytometric analysis and found that knockdown of TRIAP1 expression induced a significantly higher rate of apoptotic cells compared with the control group (Fig 2E, P < 0.01). These findings demonstrate that TRIAP1 participates in regulating NPC cell apoptosis.
Next, we investigated the effect of TRIAP1 on tumorigenesis in vivo through xenograft tumor models. As shown in Fig 3A–3C, TRIAP1 overexpression significantly enhanced tumor growth, with regard to both tumor volume and tumor weight, compared with the control LV-Vector group. TRIAP1 expression in dissected specimens was confirmed by IHC (Fig 3D). In addition, TRIAP1 overexpression displayed a higher proportion of Ki67-positive cells and a lower percentage of TdT-mediated dUTP Nick-End Labeling (TUNEL)-positive cells (Fig 3D, P < 0.001), suggesting that cells with ectopic TRIAP1 expression were actively proliferating. Conversely, tumor growth, tumor size and tumor weight were significantly inhibited by TRIAP1 knockdown (Fig 3E–3G, P < 0.05). Meanwhile, the TRIAP1-knockdown group showed a decreased proliferation index and increased apoptotic index (Fig 3H, P < 0.01). All together, these results support that TRIAP1 promotes NPC tumor growth and inhibits cell apoptosis in vivo.
To explore the underlying mechanism of TRIAP1 on NPC cell proliferation and apoptosis, we investigated the subcellular location of TRIAP1. The observation showed that ectopically expressed TRIAP1 accumulated in the mitochondria, indicating that TRIAP1 co-localized with mitochondria (Fig 4A). Furthermore, TRIAP1 knockdown led to marked mitochondrial fragmentation in both live mitochondrial images (S2 Fig) and fixed mitochondrial views (Fig 4B). We next investigated the status of mitochondrial membrane potential (△Ψm). TRIAP1 knockdown induced a significantly increased number of depolarized mitochondria (Fig 4C). Taken together, these findings indicate that TRIAP1 participates in the regulation of mitochondrial fragmentation and is required for normal polarized mitochondrial membrane potential.
Furthermore, we examined the potential role of TRIAP1 in mitochondria-dependent apoptosis. Immunofluorescent staining displayed that cytochrome c co-localized with mitochondria and TRIAP1 (Fig 5A and 5B). Interestingly, the loss of TRIAP1 induced the release of cytochrome c from mitochondria into the cytosol accompanied by mitochondrial fragmentation (Fig 5A and 5B). Subsequently, the activity of caspase-3 and -7 was significantly increased by TRIAP1 knockdown, revealing a significant induction of apoptosis (Fig 5C). Together, these results suggest that knockdown of TRIAP1 led to apoptosis through mitochondrial fragmentation and the subsequent release of cytochrome c from mitochondria.
To investigate the mechanisms of TRIAP1 expression aberration, we used available bioinformatic algorithms as filters to screen miRNAs targeting TRIAP1. A total of 98 miRNAs were identified as candidates and subsequently intersected with the 33 downregulated miRNAs identified in our published data set (NCBI/GEO/GSE32960, n = 330, including 312 NPC tissues and 18 normal nasopharyngeal tissues; Fig 6A and 6B and S3 Fig) [25]. Finally, miR-320b was identified as the sole candidate (Fig 6B, P < 0.01). In determining whether miR-320b negatively regulates TRIAP1 expression, we found that miR-320b mimics significantly inhibited TRIAP1 expression at both the mRNA and protein levels, whereas miR-320b inhibitor increased its expression in NPC cells (Fig 6C–6E, P < 0.05). To further confirm the site-specific repression of miR-320b on TRIAP1, we constructed wild-type and mutant TRIAP1 3′ UTR luciferase reporter vectors (Fig 6F). miR-320b overexpression or inhibition suppressed or increased the luciferase activity of the wild-type TRIAP1 3′ UTR reporter gene but had no inhibitory effect on the mutant reporter (Fig 6G, P < 0.05). Taken together, these data demonstrate that TRIAP1 is a novel direct target of miR-320b in NPC cells.
Subsequently, we investigated whether miR-320b has biological roles in NPC progression. Similar to the effect induced by loss of TRIAP1, overexpression of miR-320b significantly suppressed cell proliferation (Fig 7A, P < 0.01), but led to mitochondrial membrane depolarization, mitochondrial fragmentation and apoptosis (Fig 7B–7E and S4A and S4B Fig, P < 0.05). Inversely, miR-320b inhibition increased cell proliferation, but decreased mitochondrial membrane depolarization and apoptosis (S5A–S5F Fig, P < 0.05). Furthermore, we explored how miR-320b exerts its functional effects. Restoration of TRIAP1 remarkably abrogated the proliferation inhibition, mitochondrial membrane depolarization, fragmentation and apoptosis induced by miR-320b (Fig 7A–7E and S4A and S4B Fig, P < 0.05), while inhibition of TRIAP1 expression significantly abrogated the induction of proliferation, and the suppression of mitochondrial membrane depolarization and apoptosis induced by miR-320b knockdown (S5A–S5F Fig, P < 0.05). In addition, enforced TRIAP1 overexpression prevented cytochrome c release from mitochondria to the cytoplasm (Fig 7F and S6A and S6B Fig). These results suggest that TRIAP1 is a functional mediator of miR-320b on cell proliferation and mitochondria-dependent apoptosis in NPC.
Consistent with our previous miRNA microarray data (NCBI/GEO/GSE32960; S7 Fig), miR-320b was significantly downregulated in 16 fresh-frozen NPC compared with 8 normal nasopharyngeal epithelial tissues, as well as in 6 NPC cell lines compared with the normal cell line NP69 (Fig 8A and 8B). Quantitative RT-PCR revealed that miR-320b expression was inversely correlated with TRIAP1 levels in NPC tissues (n = 204; Fig 8C; P < 0.01). When combining the expression of miR-320b and TRIAP1, patients in group III with low miR-320b expression and high TRIAP1 expression displayed worse OS and DFS than those in groups I and II with low TRIAP1 expression (n = 204; S8A and S8B Fig; P < 0.001). Furthermore, miR-320b overexpression significantly suppressed tumor growth and displayed a lower proliferation index and a higher apoptotic index, while miR-320b inhibition promoted tumorigenesis and inhibited apoptosis in vivo (Fig 8D–8G, S9 Fig; P < 0.05). Taken together, these results support that the miR-320b/TRIAP1 pathway regulates the proliferation and apoptosis by repressing the release of cytochrome c from mitochondria, leading to NPC tumorigenesis and poor clinical outcomes (Fig 8H).
In our current study, we found that TRIAP1 was upregulated and associated with poor clinical outcomes in NPC. Moreover, we firstly reported that TRIAP1 could be post-transcriptionally regulated by miR-320b, and TRIAP1 expression was inversely correlated with miR-320b expression in clinical NPC samples. Furthermore, miR-320b inhibited cell proliferation and increased apoptosis through the release of cytochrome c from mitochondria in a TRIAP1-dependent manner. Therefore, our findings uncovered a novel mechanism post-transcriptionally regulating TRIAP1 expression by miR-320b and its role in tumorigenesis and unfavorable survival in NPC.
Sustaining proliferation and resisting apoptosis are hallmarks of cancer [5]. Apoptosis is programmed cell death regulated by intrinsic and extrinsic pathways centralized in the mitochondria [6,7]. However, cell death is commonly resisted in cancer cells. Mitochondria constantly undergo fusion and fission, which are required for cells to maintain mitochondrial integrity and respond to intrinsic apoptotic stimuli [30–33]. A low fusion-to-fission ratio has been reported to result in the loss of mitochondrial fusion, the generation of mitochondrial fragmentation and the release of cytochrome c to trigger cell death (apoptosis) [8–10]. Proteins involved in mitochondrial fusion and fission may participate in cancer cell resistance to apoptotic stimuli and serve as new therapeutic targets. A number of studies have observed mitochondrial-mediated apoptosis in treating NPC cells [34,35]. However, the regulation of mitochondrial network dynamic and apoptosis in NPC remains undefined.
Emerging evidence indicates that TRIAP1 promotes cell survival and prevents apoptosis [11–14,36]. In our present study, we found that TRIAP1 promoted NPC cell proliferation and suppressed apoptosis in vitro and in vivo, supporting the contribution of TRIAP1 in NPC development and progression. Moreover, we demonstrated that TRIAP1 overexpression was associated with poor survival and was an independent risk factor in NPC, indicating a significant therapeutic implication of TRIAP1 in NPC. As we known, mitochondrial fragmentation is required for apoptosis induction. In this study, we found that knockdown of TRIAP1 induced mitochondrial fragmentation, membrane potential depolarization and the subsequent release of cytochrome c, and enhanced apoptosis in NPC cells, which is consistent with a previous study in colon cancer [14]. Although another study reports that TRIAP1 exerts its function through repressing p21 [13], we found that knockdown of TRIAP1 did not increased, but decreases p21 expression, suggesting that TRIAP1did not function through repressing p21 in NPC (S10A Fig). Our study elucidates the mechanisms of TRIAP1 regulating mitochondrial fragmentation and apoptosis in NPC, suggesting that TRIAP1 modulation can be a promising therapy for NPC apoptotic resistance.
As we known, TRIAP1 is transcriptionally upregulated by TP53 [7,8] and it has been reported that TP53 can be activated by EBV encoded protein LMP1 in NPC [37–39]. However, no obvious upregulation of TP53 and TRIAP1 was observed after LMP1 overexpression in NPC cells (S10B Fig), suggesting that TRIAP1 is not regulated though LMP1/TP53 pathway in NPC and there may be other regulatory mechanisms involved in TRIAP1 overexpression. In this study, we provided evidence that miR-320b negatively regulated TRIAP1 expression and exerted its function on mitochondrial fragmentation and apoptosis by targeting TRIAP1. The inhibitory effect of miR-320b is consistent with the previous study in other cancer types and cardiomyopathy [40–42]. Here, the miR-320b level was inversely correlated with TRIAP1 expression in NPC patients, revealing that loss of miR-320b determines TRIAP1 overexpression and function in NPC. We also acknowledged that the correlation between miR-320b and TRIAP1 expression was modest in NPC patients, which indicating that there may be some other mechanisms involved in regulating TRIAP1 expression.
In conclusion, our study revealed TRIAP1 as an oncogene in tumor progression and unfavorable survival, and miR-320b as a novel post-transcriptional regulator of TRIAP1 expression in NPC. The newly identified miR-320b/TRIAP1 pathway uncovers the molecular mechanisms underlying tumorigenesis and poor clinical outcomes of NPC and may facilitate the development of novel therapeutic strategies against NPC.
For Human Subject Research, this study was approved by the Institutional Ethical Review Boards of Sun Yat-sen University Cancer Center (approval number: L20150201). Written informed consent was obtained from each patient before the study. For Animal Research, all experiments were performed according to the guidelines approved by the Institutional Animal Care and Use Ethics Committee of Sun Yat-sen University Cancer Center (approval number: 00111032).
A total of 204 consecutive patients diagnosed with non-distant metastatic NPC were recruited from Sun Yat-sen University Cancer Center between January 2004 and January 2007. Paraffin-embedded biopsy specimens of individual patients were histologically-confirmed and collected for immunohistochemistry and quantitative RT-PCR. Written informed consent was obtained from each patient before the study. This study was approved by the Institutional Ethical Review Boards of Sun Yat-sen University Cancer Center. No patient had received radiotherapy or chemotherapy before biopsy. The TNM stage was reclassified according to the 7th edition of the AJCC Cancer Staging Manual. All patients were treated with radiotherapy, as previously described [43]. Patients with stage III-IV NPC received concurrent platinum-based chemotherapy [3,44]. The median follow-up time was 81.7 months (range, 8.2 to 113.9 months). The detailed clinicopathological characteristics are listed in S1 Table. Sixteen fresh-frozen NPC samples with histological diagnosis and eight normal nasopharyngeal epithelium samples were collected and stored in liquid nitrogen until required.
IHC analysis was performed on individual sections of 204 specimens, using the polyclonal anti-TRIAP1 antibody (1:200, Sigma-Aldrich, Ronkonkoma, NY, USA). The degree of immunostaining was independently evaluated by two pathologists blinded to the clinicopathological characteristics of the patients. The scores were determined on the basis of the staining intensity and the percentage of positively stained cells. The staining intensity was graded as follows: 0, no staining; 1, weak staining, light yellow; 2, moderate staining, yellow brown; and 3, strong staining, brown. The percentages were scored according to the following standard: 1, < 10% positive cells; 2, 10–35% positive cells; 3, 35–70% positive cells; and 4, > 70% positive cells, as previously described [45]. The staining index was used to evaluate TRIAP1 expression in NPC sections, with possible scores of 0, 1, 2, 3, 4, 6, 8, 9 and 12. Cutoff values were determined on the basis of Receiver operating characteristic (ROC) curve analysis, and classified as follow: low TRIAP1 expression, staining index < 6; and high TRIAP1 expression, staining index ≥ 6 [46,47].
The human immortalized nasopharyngeal epithelial cell line NP69 and human NPC cell lines CNE-2, SUNE-1, CNE-1, HNE-1, HONE-1 and C666-1 were obtained from Professor Musheng Zeng within 6 months and authenticated by short tandem repeat profiling (Sun Yat-sen University, Guangzhou, China). NP69 was cultured in keratinocyte/serum-free medium (Invitrogen, Grand Island, NY, USA) supplemented with bovine pituitary extract (BD Biosciences, San Diego, CA, USA). CNE-2, SUNE-1, CNE-1, HNE-1, HONE-1 and C666-1 were grown in RPMI-1640 (Invitrogen) supplemented with 10% FBS (Gibco, Grand Island, NY, USA); 293FT cells were maintained in DMEM (Invitrogen) supplemented with 10% FBS.
Total RNA from cultured cells and NPC specimens was extracted using TRIzol reagent (Invitrogen) as previously described [48]. RNA was reverse transcribed using reverse transcriptase (Promega, Madison, WI, USA) with random primers (Promega) for TRIAP1 or Bulge-Loop miRNA specific RT-primers (RiboBio, Guangzhou, China) for miR-320b. Quantitative RT-PCR reactions were performed on a CFX96 Touch sequence detection system (Bio-Rad, Hercules, CA, USA). Using GAPDH or U6 as internal controls for TRIAP1 and miR-320b, respectively, the relative expression levels were calculated by the 2-ΔΔCT method [49].
Cell lysis was performed at 4°C using RIPA buffer containing a protease inhibitor cocktail (Fdbio Science, Hangzhou, China). Equal amounts of protein were separated on 12% SDS-PAGE gels and transferred to polyvinylidene fluoride membranes (Merck Millipore, Billerica, MA, USA). The membranes were incubated with rabbit polyclonal anti-TRIAP1 antibody (1:200; Santa Cruz Biotechnology, Beverly, MA, USA), followed by incubation with anti-rabbit IgG secondary antibody (1:5000; Epitomics, Burlingame, CA, USA). Anti-α-tubulin antibody (1:1000; Sigma-Aldrich) was used as a protein loading control. Detection was visualized by enhanced chemiluminescence.
Small interfering RNAs targeting TRIAP1 (siTRIAP1-1 5’-AGGCAUGCACGGACAUGAATT-3’; siTRIAP1-2 5’-GAAAGAGAUUCCUAUUGAATT-3’), miR-320b mimic (5’-AAAAGCUGGGUUGAGAGGGCAA-3’) and miR-320b inhibitor (5’-UUGCCCUCUCAACCCAGCUUUU-3’) were purchased from GenePharma company (Suzhou, China). The human TRIAP1 gene, EBV encoded LMP1 gene and synthesized short hairpin RNA targeting TRIAP1 (shTRIAP1) were cloned into the pSin-EF2- puromycin and pSuper-retro-puromycin vectors, respectively (Addgene, Cambridge, MA, USA). CNE-2 and SUNE-1 cells were transfected with oligonucleotides (100 nM) or plasmids (2 μg) using Lipofectamine 2000 reagent (Invitrogen), and then harvested for assays 48 h after transfection. Stable SUNE-1 cell lines expressing TRIAP1 and shTRIAP1 were generated by lentiviral infection using 293FT cells and selected using 0.5 μg ml-1 puromycin.
For the MTT assay, 1,000 transfected cells were seeded in 96-well plates and exposed to MTT (BD Biosciences) for 4 h at 1, 2, 3, 4, and 5 days. The absorbance values were measured at 490 nm. For the colony-formation assay, 500 transfected cells were plated in six-well plates and cultured for 7 or 12 days. The colonies were stained with 0.5% crystal violet for quantification after fixation with 4% paraformaldehyde. The anchorage-independent growth assays were performed by soft agar culture of 2.5 × 104 transfected cells in six-well plates for 7 or 12 days. The colony numbers were counted using an inverted microscope.
The apoptosis assay was performed using the Annexin V-FITC/PI Apoptosis Detection Kit (KeyGEN BioTECH, Nanjing, China). Briefly, 2 to 5 × 105 transfected cells were rinsed twice with PBS and then resuspended in 500 μl of binding buffer, followed by staining with 5 μl of Annexin V-FITC and propidium iodide (PI) for 15 minutes at room temperature in the dark. The detection was performed using a flow cytometer on a Beckman Gallios detection system (Beckman Coulter Inc., CA, USA).
For mitochondrial staining, transfected cells were grown on coverslips inside a Petri dish (Nest Biotechnology, Wuxi, China) for 24 h and stained with MitoTracker Red CMXRos (0.04 μM for live mitochondrial imaging; 0.4 μM for fixation and permeabilization after mitochondrial staining; ThermoFisher Scientific, Waltham, MA) for 30 minutes at 37°C. After staining, the staining solution was replaced with pre-warmed PBS and live mitochondrial were either observed using a confocal laser-scanning microscope (Olympus FV1000, Tokyo, Japan) or allowed continue to fix and permeabilize. After permeabilization, coverslips were incubated with either rabbit polyclonal anti-TRIAP1 antibody (1:100; Santa Cruz Biotechnology) or mouse monoclonal anti-cytochrome c antibody (1:300; Cell Signaling Technology, Danvers, MA) and then incubated with species-matched Alexa Fluor 488 or 594 goat IgG secondary antibody (Life Technologies, Carlsbad, CA, USA). After being counterstained with 4′, 6-diamidino-2-phenylindole (DAPI), cells were imaged using the confocal laser-scanning microscope (Olympus FV1000).
Mitochondrial membrane potentials (△Ψm) were detected using a JC-1 Apoptosis Detection Kit (KeyGEN BioTECH). A total of 2 to 5 × 105 transfected cells were collected, washed twice with PBS, and incubated with 500 μl of prewarmed JC-1 incubation buffer at 37°C for 20 minutes. After incubation, cells were centrifuged, rinsed twice with incubation buffer, and resuspended in 500 μl of incubation buffer. The △Ψm analysis was performed using a flow cytometer on a Beckman Gallios detection system (Beckman Coulter).
The activity of caspase-3 and caspase-7 was determined using a Caspase-Glo 3/7 Assay Kit (Promega) according to the manufacturer’s instructions. Transfected cells were grown in 100 μl of cultured medium in 96-well plates and equilibrated to room temperature before the assay. A total of 100 μl of Caspase-Glo 3/7 reagent was added to each well and incubated for 1 h in the dark. Luminescence was measured using the luminometer (Promega).
Six-week-old male BALB/c nude mice were purchased from the Medical Experimental Animal Center of Guangdong Province (Guangzhou, China). The nude mice were implanted with 1 × 106 SUNE-1 cells stably overexpressing TRIAP1, shTRIAP1 or the corresponding negative control in the dorsal flank. For miR-320b overexpression and inhibition experiments, 1 × 106 SUNE-1 cells were subcutaneously injected into the dorsal flank of nude mice after pre-transfected with agomir-320b, antagomir-320b or NC control (200nM, RiboBio). After 7 days, when the tumor volume reached 100mm3, intratumoral injection of agomir-320b (5nM), antagomir-320b (5nM) or NC control (5nM) was performed twice a week for 3 weeks. The weights and tumor volumes were measured twice weekly. The mice were sacrificed 28–35 days after implantation, and the tumors were dissected, weighted and paraffin embedded. Serial sections were subjected to IHC analysis using anti-TRIAP1 antibody or anti-Ki67 antibody (ZSGB-Bio, Beijing, China). A proliferation index was measured using the percentage of positive Ki67 cells. A TUNEL assay was performed on sections from paraffin-embedded mouse specimens using the TUNEL In situ Cell Death Detection Kit, Biotin POD (KeyGEN BioTECH) according to manufacturer’s instructions. The apoptotic index was quantified by the proportion of positive TUNEL cells. All animal experiments were approved by the Institutional Animal Care and Use Ethics Committee.
Both the conserved (position 265–272) and poor conserved (position 550–556) binding sites were mutated. The mutation (Mt) and wild-type (Wt) versions of the TRIAP1 3′ UTR were generated and cloned into the psiCHECK-2 luciferase reporter plasmid (Promega). Cells were seeded into 6-well plates and co-transfected with the TRIAP1 Wt or Mt 3′ UTR reporter plasmids (2 μg), along with the miR-320b mimics (100 nM) or miR-320b inhibitor (100nM) or miRNA negative control (miR-Ctrl, 100 nM) using Lipofectamine 2000 reagent (Invitrogen). Renilla and firefly luciferase activities were measured 24 h after transfection using the Dual-Luciferase Reporter Assay System (Promega).
All statistical analyses were performed using SPSS 16.0 software (IBM, Armonk, NY, USA). The data are represented as the mean ± SD resulting from at least three independent experiments. The χ2 and Fisher’s exact tests were used to compare Categorical variables. Survival curves were calculated using the Kaplan-Meier method, and the differences were compared using a log-rank test. A multivariate analysis using a Cox proportional hazards model was performed to assess independent prognostic factors. The correlation of TRIAP1 and miR-320b mRNA expression was compared using a Pearson’s χ2 test. Comparisons between groups were evaluated by two-tailed Student’s t-tests. P < 0.05 was considered significant.
|
10.1371/journal.ppat.1004040 | Regulation of Human T-Lymphotropic Virus Type I Latency and Reactivation by HBZ and Rex | Human T lymphotropic virus type I (HTLV-I) infection is largely latent in infected persons. How HTLV-1 establishes latency and reactivates is unclear. Here we show that most HTLV-1-infected HeLa cells become senescent. By contrast, when NF-κB activity is blocked, senescence is averted, and infected cells continue to divide and chronically produce viral proteins. A small population of infected NF-κB-normal HeLa cells expresses low but detectable levels of Tax and Rex, albeit not Gag or Env. In these “latently” infected cells, HTLV-1 LTR trans-activation by Tax persists, but NF-κB trans-activation is attenuated due to inhibition by HBZ, the HTLV-1 antisense protein. Furthermore, Gag-Pol mRNA localizes primarily in the nuclei of these cells. Importantly, HBZ was found to inhibit Rex-mediated export of intron-containing mRNAs. Over-expression of Rex or shRNA-mediated silencing of HBZ led to viral reactivation. Importantly, strong NF-κB inhibition also reactivates HTLV-1. Hence, during HTLV-1 infection, when Tax/Rex expression is robust and dominant over HBZ, productive infection ensues with expression of structural proteins and NF-κB hyper-activation, which induces senescence. When Tax/Rex expression is muted and HBZ is dominant, latent infection is established with expression of regulatory (Tax/Rex/HBZ) but not structural proteins. HBZ maintains viral latency by down-regulating Tax-induced NF-κB activation and senescence, and by inhibiting Rex-mediated expression of viral structural proteins.
| Most HTLV-1-infected individuals are asymptomatic. It is thought that the proviral DNA is transcriptionally inert and HTLV-1 replicates through mitotic expansion of host cells. The evolving provirus integration patterns in HTLV-1 carriers, however, suggest new infection occurs continuously. Whether or how HTLV-1 establishes latency and reactivates is unclear. We show that HTLV-1 infection in culture can lead to two alternative outcomes — productive infection accompanied by senescence or latent infection followed by clonal expansion — based on the relative expression of regulatory proteins: Tax, Rex, and HBZ. HTLV-1 latency is established by HBZ, and reactivation is achieved by Rex through regulating nuclear export of viral mRNAs. Elucidating mechanisms underlying HTLV-1 latency and reactivation can facilitate virus control to prevent progression to disease.
| Human T-lymphotropic virus type 1 (HTLV-1) is a complex human retrovirus that infect approximately 10–20 million people worldwide [1]. In 3–5% of infected individuals a malignancy of CD4+ T cells known as adult T-cell leukemia/lymphoma (ATL) develops over a course of several decades [2], [3]. Other diseases caused by HTLV-1 include HTLV-1-associated myelopathy/tropical spastic paraparesis (HAM/TSP), HTLV-1 uveitis, and other inflammatory diseases.
Most HTLV-1-infected individuals become asymptomatic virus carriers. The prevailing view of HTLV-1 infection is that it is rather inactive and integrated HTLV-1 proviral DNA replicates largely through mitotic expansion of host cells. This view is based on three lines of evidence (see [4] for comments): (i) undetectable viral structural mRNA or protein expression in most infected PBMCs; (ii) undetectable cell-free viral particles in the plasma; and (iii) genetically stable viral genome due to very limited de novo infection through error-prone reverse transcription. However, longitudinal studies of HTLV-1 carriers indicate that the patterns of proviral DNA integration in PBMCs continue to evolve over time [5], suggesting that de novo infection of naïve cells occurs constantly in virus carriers (see [4] for a review). In infected individuals, there is also a robust CTL response against Tax and HBZ [6]–[8], implicating immune activation via persistent expression of viral antigens. Whether and how HTLV-1 establishes latency and reactivates is not understood.
HTLV-1 viral trans-activator Tax is a potent activator of viral mRNA transcription and the NF-κB pathway [3], [9]. We have shown previously that hyper-activation of NF-κB by Tax induces cellular senescence [10]. Remarkably, HBZ, a regulatory protein encoded by the HTLV-1 anti-sense transcript [11], dampens NF-κB activation [12] and thereby mitigates Tax-induced senescence [10]. These results raise the possibility that HTLV-1 infection may lead to two alternative outcomes dictated by the levels of Tax and HBZ [10]. When expressed at high levels, Tax drives robust viral replication, hyper-activates NF-κB, and triggers a senescence checkpoint response. Low levels of Tax and higher levels of HBZ, by contrast, result in moderation of NF-κB activation, prevention of senescence, and survival and persistence of HTLV-1-infected cells.
Our previous studies have shown that most HeLa and SupT1 cells infected by HTLV-1 in culture become senescent or arrested in cell cycle progression [13]. Here we demonstrate that HTLV-1 infection indeed can lead to productive infection with expression of all viral proteins, NF-κB activation, and senescence; or latent infection with expression of regulatory but not structural proteins. HTLV-1 latency is regulated by HBZ, which dampens LTR and NF-κB activation by Tax. The latter activity of HBZ prevents senescence induction and allows latently infected cells to proliferate. Interestingly, HBZ further inhibits Rex-mediated export of intron-containing viral mRNAs, thereby shutting off Gag, Gag-Pol, and Env expression, and virus production. The latent provirus can be reactivated by over-expressing Rex or down-regulating HBZ. Thus, the “latency state” of HTLV-1 in the cell culture system resembles that established by γ-herpesviruses such as EBV and KSHV, which express a handful of potentially oncogenic latency-associated viral proteins and RNAs that stimulate mitotic expansion of latently infected cells. We speculate that the persistent expression of Tax and HBZ during the early stage of HTLV-1 latency propels infected cells to proliferate.
We have previously derived a cell line known as HeLa-G that harbors a reporter cassette, 18×21-EGFP, consisting of the enhanced green fluorescence protein (EGFP) gene under the transcriptional control of 18 copies of the Tax-responsive 21-base-pair enhancer element [14]. HeLa-G cells express abundant GFP as a function of Tax expression either after HTLV-1 infection or transduction of the tax gene. We infected them by co-culture with mitotically inactivated HTLV-1-producing T-cell line, MT2 (Fig. 1A Top, see Materials and Methods), and found that most infected cells (98%) became senescent (Fig. 1A left panel). However, a careful examination of infected HeLa-G cells revealed that a small population (2%) continued to proliferate (Fig. 1A middle panel). Since our recent data indicate that hyper-activation of NF-κB by Tax is responsible for inducing cellular senescence [10], [15], we also tested the effect of HTLV-1 infection on a HeLa-G-derived cell line, HeLa-G/ΔN-IκBα in which the transcriptional activity of NF-κB is shut off by the stable expression of a degradation-resistant form of IκBα(ΔN-IκBα). As anticipated, HTLV-1-infected HeLa-G/ΔN-IκBα cells continued to proliferate after infection (Fig. 1A right panel). We also monitored infected HeLa-G and HeLa-G/ΔN-IκBα cells by immunofluorescence for p65/RelA, the capsid protein p24, and Rex at 48 hours after infection. As shown in Fig. 1B, NF-κB is activated in HTLV-1 infected HeLa-G cells as revealed by the localization of p65/RelA to the nucleus. Most of these cells became senescent as described above. By contrast, in HTLV-1 infected HeLa-G/ΔN-IκBα cells, p65/RelA is localized in the cytoplasm as a result of inhibition by ΔN-IκBα. Doublets of infected GFP-positive HeLa-G/ΔN-IκBα cells could be seen, indicative of cell proliferation. Both types of infected cells express Tax, Rex, and p24 as might be expected (Fig. 1B). Proliferating HeLa-G/HTLV-1 and HeLa-G/ΔN-IκBα/HTLV-1 clones were isolated by cell sorting based on GFP expression, expanded in cell culture, and characterized. Intriguingly, while the HeLa-G/ΔAN-IκBα/HTLV-1 clones expressed robust levels of p24 (Fig. 1B ΔN-IκBα/HTLV-1 clones 1–3), no p24 expression was detectable in HeLa-G/HTLV-1 clones (Fig. 1B HeLa-G/HTLV-1 clones 1 to 5). PCR analyses showed the chromosomal DNA of isolated clones to be positive for integrated HTLV-1 proviral DNA. Results from 3 (G1-3) and 2 (ΔN1-2) representative clones of each group are shown (supplementary Fig. S1). While all PCR products of ΔN1 and ΔN2 clones were of correct sizes as might be expected, G1-3 clones lacked (G1 and G2) or yielded a smaller env PCR product (G3) in the region that spans nucleotides 5318-5784 of the HTLV-1 genome (see supplementary Fig. S1 and Table S1), indicating gene deletions in this env region.
As indicated by immunoblotting, most if not all HeLa-G/ΔN-IκBα/HTLV-1 cells were productively infected by HTLV-1 and abundantly expressed Gag (p24 and p19 matrix protein, abbreviated as p19), Env (gp46), Tax, and Rex (Fig. 2A). Together with the results described above (Fig. 1A left panel), these data indicate that cells productively infected by HTLV-1 usually undergo senescence as a result of chronic NF-κB activation by Tax. However, when NF-κB activity is blocked, the senescence response is prevented and the productively infected cell (PIC) population can grow and divide, and be established as individual virus-producing clones.
We next examined in depth the 3 cell lines derived from the minor population of proliferating HTLV-1-infected HeLa-G (HeLa-G/HTLV-1) cells whose NF-κB activity was unaltered. Interestingly, all three expressed Tax and Rex, albeit at lower levels (Fig. 2A right panels), but showed no detectable expression of p19, p24, or gp46 (Env) (Fig. 2A). The absence of Env from these clones correlated with env mutations detected by PCR (Fig. S1). Since only GFP+ cells were sorted and isolated, the positive detection of Tax expression is perhaps not surprising. These clones are designated as latently infected cells (LICs) for reasons that will become obvious later. It should be pointed out that Tax/Rex and HBZ expression is very low for LIC clone 3 (LIC3), where Tax expression is only detectable by the 18×21-EGFP reporter and Rex can only be seen by immunoblotting occasionally (see below).
Activation of viral transcription by Tax is intact in LICs and PICs (albeit at a very low level for LIC3) as revealed by significant luciferase activities after transfection with an LTR-Luc reporter (Fig. 2B). This suggests that for LTR activation, the levels of Tax expressed in LICs and PICs are not limiting, with the exception of LIC3. This is as might be expected since each LTR has only 3 Tax-responsive 21-bp repeat elements (TxREs), and as the Tax/CREB complex recruited to the TxREs is known to have a high affinity for them, the effective concentrations of Tax necessary to drive LTR transcription need not be high. The lower levels of viral expression in LICs may be related to the chromosomal environments of proviral DNAs that dampen Tax-mediated trans-activation. As anticipated, the NF-κB activity in PICs is profoundly inhibited by the IκBα super-repressor (ΔN-IκBα) despite Tax expression. This is confirmed by the absence of detectable luciferase activity in them after transfection of an NF-κB reporter plasmid, E-selectin-Luc (Fig. 2C PIC lanes). Despite detectable Tax expression, NF-κB activity was not significantly induced in LICs (Fig. 2C). We think this is due to the expression of HBZ, which is known to down-modulate NF-κB activity, albeit not as drastically as ΔN-IκBα [10], [12]. Indeed, unspliced, but not spliced HBZ mRNA was readily detected and its levels are similar in both PICs and LICs (again, with the exception of LIC3) as determined by RT-PCR (Fig. 3A), consistent with the notion that its expression is independently regulated. The abundance of unspliced versus spliced HBZ mRNA most likely depends on the availability of splicing factors and can vary from cells to cells. Finally, it should be pointed out that although existing HBZ antibody can detect HBZ after DNA transfection, its sensitivity is insufficient for detecting HBZ during infection.
While the Tax/Rex mRNA (pXIII) levels in LICs were lower than those in PICs as indicated by mRNA quantitation (Fig. 3B, LIC1/PIC1 and LIC1/PIC2 about 1/3 and 1/4 respectively), greater differences were seen for Gag-Pol mRNAs (Fig. 3B, LIC1/PIC1 and LIC1/PIC2 about 1/5 and 1/8 respectively), Thus, more viral mRNAs of PICs are in the unspliced (Gag-Pol) form, and less so for LICs. The Env mRNAs in LICs were much lower (Fig. 3B). We think this is due to nonsense-mediated degradation caused by env mutations (Fig. S1). As mentioned earlier, the reduced Gag-Pol and Tax/Rex mRNA expression in LICs is most likely associated with the chromosomal sites of integration. The mRNA stabilization by higher levels of Rex in PICs [16], [17] likely also influences viral mRNA expression. Finally, inhibition of Rex-mediated nuclear export of intron-containing viral mRNAs in LICs contributes to the altered gene expression profile as elaborated below.
Since the level of Rex is lower in LICs, we tested the possibility that their lack of Gag expression might be caused by a block in the nuclear export of unspliced viral mRNAs. Nuclear and cytoplasmic RNAs were fractionated for the LIC clones 1–3 and PIC clones 1 and 2, and subjected to qRT-PCR to quantify the nuclear and cytoplasmic levels of Gag-Pol, pX-III, and the control β-actin mRNAs. As indicated in Fig. 4A, there is a block in nuclear export of Gag-Pol mRNA in LICs with nuclear to cytoplasmic (N/C) ratios of approximately 30–40. By contrast, N/C ratios of Gag-Pol mRNA in PICs were approximately 1–2. The N/C ratios of the doubly spliced pX-III mRNA range from 0.6 to 2 in both cell types. Importantly, even though the levels of Rex in LICs were modest compared to those in PICs (Fig. 2A Rex panel on the right), it was detectable by immunoblotting, and was expected to export at least some unspliced Gag-Pol mRNAs. Intriguingly, Rex appeared altogether inactive in LICs.
We next asked if p24 expression could be reactivated by Tax or Rex in the LICs. Contrary to conventional wisdom, over-expression of Tax in LICs via transfection of an expression vector, Bc12-Tax, had very little impact on inducing p24 expression except in LIC clone 3 (Fig. 4B right 3 lanes). By contrast, when a Rex-expression plasmid was transfected, p24 expression was readily induced in all three LIC lines (Fig. 4B compare middle 3 and left 3 lanes). As expected, more robust reactivation of LIC3 could be achieved with co-transfection of both Rex and Tax (supplemental Fig. S2). These results again suggest that Tax is not a limiting factor for viral gene expression in many LICs. Importantly, they also suggest that the endogenous Rex in LICs may be defective or inhibited; and the defect or block can be complemented or overcome by over-expressing exogenous Rex.
To determine if the activity of Rex was blocked or defective in LICs, we transfected both LICs and PICs with an HTLV-1 Rex reporter plasmid, pRxRE1-RLuc (Fig. 5A upper panel; [18]). This reporter encodes an mRNA that contains the HTLV-1 Rex-response element (RxRE1) in the 3′ end, and an intron that harbors the coding sequence for the Renilla luciferase (RLuc). In the absence of Rex, the RLuc sequence is removed by splicing. The mRNA exported to the cytoplasm is therefore without Rluc, hence no luciferase activity is expressed. In the presence of Rex, however, the RxRE1-RLuc-intron-containing mRNA is exported to the cytoplasm and translated to yield Renilla luciferase. Indeed, pRxRE1-RLuc-transfected PICs readily produced Renilla luciferase activity (Fig. 5B PIC 1 and 2), consistent with their chronic production of viral structural proteins facilitated by higher levels of Rex. By contrast, RxR1E-RLuc-transfected LICs expressed little luciferase activity (Fig. 5B LIC 1–3), in agreement with the notion that Rex is either defective or inhibited in LICs. No luciferase activity is detectable in transfected control HeLa-G cells (Fig. 5B leftmost lane).
We next asked if the lack of Rex activity in LICs might be due to inhibition by a trans-acting viral factor. Of all viral proteins, we thought HBZ to be the most likely to have a role in inhibiting the nuclear export function of Rex. This is because low or no NF-κB activation was detected in LICs despite Tax expression, suggesting that HBZ was expressed in LICs and was inhibiting NF-κB activation and senescence induction as previously proposed [10], [12]. Since HBZ is already playing a critical role in rendering possible the continuous proliferation of Tax-expressing cells [10], it is logical that it might additionally prevent virus production so as to establish latency.
To test if HBZ could block nuclear export of unspliced mRNA by Rex, we titrated Rex and pRxRE1-RLuc plasmids to determine the minimal amount of Rex DNA needed to achieve maximal reporter activity (Fig. 5C lanes 1–5). That amount of Rex (50 ng) was then used in co-transfection with increasing amounts of an HBZ-expressing plasmid. Indeed, a dose-dependent reduction in Renilla luciferase activity was observed when Rex and pRxRE1-RLuc reporter were co-transfected with HBZ (Fig. 5C lanes 6–8), suggesting that HBZ blocked Rex-mediated export of RexRE1-RLuc-intron mRNA. This effect of HBZ is mediated by the HBZ protein and not mRNA, because an HBZ mutant with the ATG translational start codon mutated to TTG failed to block the activity of Rex (Fig. 5C lanes 9 and 10).
To confirm that HBZ is indeed responsible for preventing Gag expression in LICs, we derived a puromycin-selectable lentiviral vector encoding an HBZ-targeting shRNA under the transcriptional control of the Pol III-dependent snRNA U6 promoter. LICs were transduced with the LV-HBZ-shRNA-SV-puro (Fig. 6A & B, HBZ shRNA) or the empty vector, and selected in puromycin-containing medium for seven days. Down-regulation of HBZ mRNA in the HBZ-shRNA-treated cells was confirmed by qRT-PCR (Fig. 6A). As expected, after HBZ knockdown, an increase in Tax/Rex expression was observed (Fig. 6B), consistent with the notion that HBZ down-regulates viral gene expression [19]. Importantly, p24 expression was significantly induced and readily detected for LIC clones 1 and 2, indicating viral reactivation (Fig. 6B). LIC clone 3 did not show appreciable p24 expression, most likely because its level of sense mRNA transcription was too low. These results demonstrate that HBZ is responsible for down-regulating Tax-mediated viral sense mRNA transcription, and blocking Rex-mediated nuclear export of intron-containing HTLV-1 mRNAs and expression of viral structural proteins in LICs. It is important to note that in order for HBZ to exert effective control over viral replication, low levels of Tax/Rex expression are needed.
The levels of Rex and Tax in PICs are 3- to 4-fold higher than those in LICs (Fig. 2A right panels). While one cause for this difference may be the sites of proviral integration, since the major difference between LICs and PICs is the profound NF-κB inhibition by ΔN-IκBα in the latter, we thought strong NF-κB repression might contribute to the increased expression of Rex and Tax, and thereby reactivated latent HTLV-1 genome. Indeed, stable expression of the IκBα super-repressor, ΔN-IκBα, in LIC clones up-regulated Rex and Tax expression, and reactivated the latent proviruses as indicated by the induction of p24 expression, especially for LIC clones 1 and 2 (Fig. 6C). The mechanism by which NF-κB inhibition induces Rex and Tax expression is currently under investigation.
In this paper, we present evidence to demonstrate that HTLV-1 infection can lead to two alternative outcomes based on the relative expression of Tax/Rex and HBZ (summarized in Fig. 7). In most HTLV-1 infected cultured cells, Tax/Rex expression is robust and viral structural proteins are abundantly expressed. In this condition, IKK/NF-κB is hyper-activated by Tax, triggering a host senescence response. When senescence induction is prevented by inhibiting NF-κB, cell clones productively infected by HTLV-1 can be readily established (Fig. 1). On the relatively rare occasions when Tax/Rex expression is weak, HBZ moderates NF-κB activation by Tax [10], [12], thus averting the host senescence response and allowing infected cells to continue to proliferate. Importantly, HBZ additionally inhibits Rex-mediated nuclear export of intron-containing mRNAs, thereby shutting off Gag, Gag-Pol, and Env production and committing infected cells into latency (Fig. 7). In latently infected cells, viral regulatory proteins, Tax, Rex, and HBZ, but not structural proteins are persistently expressed. Reactivation of Gag, Gag-Pol, and Env expression is achieved through up-regulation of Rex or down-regulation of HBZ (Fig. 7). Most interestingly, strong inhibition of NF-κB increases Rex and Tax expression and reactivates HTLV-1 replication.
We did not investigate the involvement of other HTLV-1 accessory proteins including p21Rex, p12I, p13II, and p30II in the present model. P30II is a nuclear and nucleolar protein thought to be a post-transcriptional modulator of viral replication [20]. Published data suggest that p30II retains the doubly-spliced Tax/Rex mRNA in the nucleus and thereby down-modulates viral gene expression by reducing the levels of Tax and Rex [20]. It has also been shown to interact with CBP/p300 and interfere with LTR trans-activation by Tax [21]. The continuous expression of Tax and Rex in LICs, albeit at low levels, indicates that the Tax/Rex mRNA is not sequestered. Importantly, the LICs described here were identified by virtue of Tax-driven GFP expression via the 18×21-EGFP reporter cassette. Cells that had no Tax/Rex expression would not have been scored in this system. Thus, a study of the latency state where Tax/Rex expression is completely silenced by p30II requires other approaches.
An unexpected finding from the present study is the high frequency with which LIC clones were found to harbor env mutations (Fig. S1). Additional attempts to isolate LIC clones that contain fully functional proviral DNA were not successful. This contrasts with the PIC clones, most if not all of which readily express Tax/Rex, Gag and Env. Whether the profound NF-κB inhibition in HeLa-G/ΔN-IκBα cells is responsible for shutting off innate host defense mechanism(s) that target mutations to retroviral genomes remains to be investigated.
The outcomes of HTLV-1 infection reported herein can adequately explain data from clinical and in vivo studies [4], [22]. In this model, cells productively infected by HTLV-1 immediately enter into senescence (Fig. 7; and ref. [10]) and most likely become eliminated by cytotoxic T lymphocytes [22]–[24]. Removal of senescent cells by NK cells is also a likely possibility [25], [26]. The latently infected cells, however, continue to express Tax, Rex and HBZ and can rely on the mitogenic activities of Tax, and HBZ protein and/or mRNA to drive cell proliferation. This is in accordance with the detection of Tax/Rex mRNA in a small population of infected cells reported previously [6], and the robust CTL response against Tax and HBZ seen in infected individuals [6], [8], [27]. The dynamically evolving proviral integration patterns in asymptomatic HTLV-1 carriers can now be explained by viral reactivation and de novo infection of naïve cells brought about via up-regulation of Rex and/or down-regulation of HBZ expression. In this model, it is necessary for latently infected cells to express only muted levels of Tax/Rex such that their activities can be controlled by HBZ. Indeed, proviral DNA integration sites in asymptomatic carriers were found mostly to locate in transcriptionally silent regions of chromosomes [28]. Finally, the observation that strong NF-κB inhibition can increase Rex expression and viral reactivation has clinical implications. Bortezomib, a proteasome inhibitor that inhibits NF-κB by stabilizing IκBα, has been entered into clinical phaseI/II trials for ATLL. Although ATLL cells in general no longer replicate HTLV-1, latently infected cells most likely persist in patients and may be reactivated by NF-κB inhibitors so as to influence the course of the disease. Based on present data, administration of antivirals to prevent HTLV-1 reactivation and spread may be advisable when bortezomib is used for ATLL treatment.
Given the alternative outcomes of HTLV-1 infection, do ATL cells emerge from productively or latently infected T lymphocytes? We think productive HTLV-1 infection of T lymphocytes whose senescence checkpoint response has been impaired is most likely the first step in ATL development. Such precancerous lymphocytes can express sufficient levels of Tax to overcome HBZ inhibition and achieve persistent IKK/NF-κB activation without inducing senescence. Through the loss of senescence response, the proliferative and survival advantages conferred by Tax-driven NF-κB activation, and the mitogenic activity of HBZ, such lymphocytes can readily undergo clonal expansion. This view agrees with recent high-throughput DNA sequencing data showing that most proviral integrations in ATL cells occur in transcriptionally active regions in the sense orientation [28]. As Tax is a primary CTL target, the loss of Tax and forward (sense) viral gene expression from precancerous T lymphocytes is selected and occurs through 5′ LTR DNA methylation, nonsense mutations, and deletions [29]. However, after Tax and Tax-dependent NF-κB activation are lost from pre-cancerous T lymphocytes, the impairment to the senescence checkpoint response remains, and can facilitate the evolution of Tax-independent NF-κB activation. As the expression of HBZ mRNA and protein is independently regulated, they can continue to exert mitogenic effect to propel the proliferation of cancerous cells [30]–[32]. Understanding how Rex, Tax, and HBZ expression is altered by cell signaling and cellular physiology to affect latency establishment and viral reactivation can facilitate the control of viral infection to prevent progression to disease.
HTLV-1 infections were performed in a 10 cm dish by co-culturing HeLa-G or HeLa-G/ΔN-IκBα cells (1–2×106) with HTLV-1-producing MT2 cells (3×106) that have been mitotically inactivated by mitomycin C (MMC) treatment (10 µg/ml for 2 hours). The co-culture was carried out in the presence of polybrene (8 µg/ml) for 16 hours. MT2 cells were then removed by washing with phosphate buffered saline (PBS). Fresh media was added and cells were grown for an additional 24 hours, and then harvested. GFP-positive cells were isolated using a cell sorter (BD FACSAria) housed in a lamella flow hood under aerosol-protection condition. Sorted cells were plated at low density on a 15-cm dish. Individual proliferating colonies were picked after a week into 96-well plates and further screened for the integrated HTLV-1 genome by PCR.
HeLa-G or HeLa-G/ΔN-IκBα cells grown on chamber glass slides were infected with HTLV-1 as described above. Forty eight hours after infection, cells were fixed with 4% paraformaldehyde, permeabilized with 0.1% Triton X-100, and immunostained overnight with the indicated primary antibodies followed by Alexa Fluor 568 secondary antibodies (Invitrogen, Carlsbad, CA.) Nuclei were counterstained using 4′,6′-diamidino-2-phenylindole (DAPI). The slides were then mounted in a fluorescence mounting medium (Dako). Images were captured using a Zeiss Pascal inverted confocal microscope.
Standard procedures were used for immunoblotting. Typically, 30–50 µg of total proteins was used in each sample. The HTLV-1 Tax hybridoma monoclonal antibody 4C5 had been described previously [15]. The rabbit polyclonal antibody against Rex was a generous gift of Dr. Gisela Heidecker. HTLV-1 p24 antibody was purchased from Advanced Bioscience, HTLV-1 p19 and HTLV-1-gp46 (Env) antibodies were from Zeptometrix, IκBα, β-actin, goat anti-mouse, and goat anti-rabbit HRP conjugated secondary antibodies were from Santa Cruz.
HTLV-1-infected cells generated as above were harvested, washed, and dissolved in lysis buffer (50 mM Tris, pH 8.0, 10 mM EDTA, 100 mM NaCl, 0.5% sarcosyl, 0.5 mg/ml proteinase K). DNA was then precipitated with isopropanol. To screen for the full-length HTLV-1 provirus, sequence-tagged site polymerase chain reaction was carried out as described [33] using primers specified for distinct regions/genes of the HTLV-1 genome (supplementary Table S1). PCR products were resolved on 2% agarose gels.
Total mRNA, nuclear and cytoplasmic mRNAs from HTLV-1-infected cell clones were isolated using the PARIS kit (Ambion) according to manufacturer's instructions. Contaminating genomic DNA was removed using the turbo DNA-free kit (Ambion). Complementary DNA (cDNA) was synthesized from 500 ng of RNA in a total volume of 10 µl with iScript reverse transcription super mix (Biorad). The cDNA used for HBZ mRNA quantitation was prepared using a gene-specific antisense primer, HBZ-R2 (see supplementary Table S 2), to avoid contamination from HTLV-1 sense strand cDNA. Real-time PCR was performed using 2 µl of the cDNA as template in a 20 µl reaction, using gene-specific primers (see supplementary Table S2 for sequences) and LightCycler DNA SYBR Green I master mix (Roche applied science) in a LightCycler thermal cycler (Roche Diagnostics). The mRNA level in each sample was normalized to that of the β-actin mRNA. Relative mRNA levels were calculated using the 2−ΔCt method [34]. To determine the nuclear-to-cytoplasmic ratio of a given viral mRNA species, 150 ng of nuclear or 300 ng of cytoplasmic mRNA was used for cDNA synthesis. Complementary DNA from each fraction was quantified for the level of Gag-Pol, pXIII, and β-actin mRNA transcripts respectively by PCR, the relative abundance of each viral mRNA in the nuclear or cytoplasmic compartment was determined by normalizing the level of a given viral mRNA against that of the β-actin mRNA in the same mRNA preparation. The nuclear-to-cytoplasmic ratio of viral mRNAs was calculated based on the relative abundance measurements, and then plotted.
Cells (3×105) were seeded into a 24-well plate overnight. After 16 hours, DNA transfections were performed using Fugene HD reagent (Promega). Two hundred nanograms of each reporter plasmid HTLV-1 LTR-Luc, E-selectin-Luc, or RxRE-RLuc were used in the respective luciferase reporter assay. The RxRE1-RLuc contains the HTLV-1 Rex-response element downstream of the Renilla luciferase reporter (RLuc) gene located within an intron (kindly provided by Dr. Jaqueline Dudley). The amounts of Rex or HBZ expression plasmid used range from 0 to 200 ng. All transfections were performed in duplicates. The total DNA amount (500 ng) was kept constant in all transfections using an empty vector plasmid, pcDNA3.1. Twenty nanograms per well of control luciferase plasmid pGL3-Luc (firefly) or pRL-TK (renilla) were also included in each transfection. After 48 hours, cells are harvested, and luciferase activity was measured using the Dual-Luciferase Reporter Assay System (Promega) according to the manufacturer's instructions. Transfection efficiencies were normalized using either TK-renilla or PGL3-Luc. Data are mean ± s.d. from at least three independent experiments.
The lentivirus expressing the degradation resistant mutant of IκBα, LV-ΔN-IκBα-SV-puro, has been described earlier [10]. For delivery of the anti-HBZ short hairpin RNA (shRNA), A small hairpin RNA (shRNA) expression cassette containing the HBZ shRNA sequence [30] downstream of the mouse U6 promoter was amplified by PCR and cloned into a self-inactivating lentiviral vector, SMPU [14], engineered to contain the SV-puro (the puromycin resistance gene placed under the control of the SV40 early promoter). Lentiviral vectors were prepared as previously reported [15]. HeLa-G cells were transduced with the lentiviral vector in DMEM supplemented with 10% fetal bovine serum and selected in the same medium containing puromycin (1 µg/ml).
|
10.1371/journal.pgen.1004177 | miR-100 Induces Epithelial-Mesenchymal Transition but Suppresses Tumorigenesis, Migration and Invasion | Whether epithelial-mesenchymal transition (EMT) is always linked to increased tumorigenicity is controversial. Through microRNA (miRNA) expression profiling of mammary epithelial cells overexpressing Twist, Snail or ZEB1, we identified miR-100 as a novel EMT inducer. Surprisingly, miR-100 inhibits the tumorigenicity, motility and invasiveness of mammary tumor cells, and is commonly downregulated in human breast cancer due to hypermethylation of its host gene MIR100HG. The EMT-inducing and tumor-suppressing effects of miR-100 are mediated by distinct targets. While miR-100 downregulates E-cadherin by targeting SMARCA5, a regulator of CDH1 promoter methylation, this miRNA suppresses tumorigenesis, cell movement and invasion in vitro and in vivo through direct targeting of HOXA1, a gene that is both oncogenic and pro-invasive, leading to repression of multiple HOXA1 downstream targets involved in oncogenesis and invasiveness. These findings provide a proof-of-principle that EMT and tumorigenicity are not always associated and that certain EMT inducers can inhibit tumorigenesis, migration and invasion.
| Induction of epithelial-mesenchymal transition (EMT) in epithelial tumor cells has been shown to enhance migration, invasion and cancer ‘stemness’. Here we demonstrate that a miRNA downregulated in human breast tumors, miR-100, can simultaneously induce EMT and inhibit tumorigenesis, migration and invasion through direct targeting of distinct genes. This is the first report of an EMT inducer that suppresses cell movement and tumor invasion, which indicates that EMT is not always associated with increased tumorigenesis, migration and invasion, and that all EMT inducers are not equal: while some of them can promote tumorigenicity, motility and invasiveness, others inhibit these properties owing to their ability to concurrently target both EMT-repressing genes and oncogenic/pro-invasive genes. These findings provide new insights into the complex roles of EMT inducers.
| Epithelial-mesenchymal transition (EMT) is regulated by transcription factors [1], [2], extracellular ligands [3] and microRNAs (miRNAs) [4]–[9]. It has been proposed that inducing EMT in epithelial tumor cells enhances migration, invasion and dissemination, whereas the mesenchymal-epithelial transition (MET) process facilitates metastatic colonization [1], [2], [10]–[12]. In addition, induction of EMT in differentiated tumor cells has been shown to generate cells with properties of tumor-initiating cells, or cancer stem cells [13], [14]. However, whether EMT and tumorigenicity are always linked is debated. Recently, analysis of clonal populations derived from the PC-3 prostate cancer cell line demonstrated that a metastatic clone was highly proliferative and expressed genes associated with an epithelial phenotype, whereas a non-metastatic clone was poorly proliferative and expressed genes associated with EMT [15]. Whether this finding is attributed to clonal bias or holds true in general is unknown. Moreover, whether there exist specific gene products that concurrently induce EMT and inhibit tumorigenesis remains elusive.
To systematically identify miRNAs differentially expressed in EMT, we overexpressed EMT-inducing transcription factors, Twist, Snail or ZEB1, in the experimentally immortalized, non-transformed human mammary epithelial cells [16], termed HMLE cells. Each of these transcription factors was capable of inducing EMT, as evidenced by changes in morphology (Figure S1A), downregulation of E-cadherin (CDH1), and upregulation of N-cadherin (CDH2), vimentin (VIM) and multiple EMT-inducing transcription factors (Figure S1B).
Next, we performed miRNA microarray profiling analysis (Table S1) of these HMLE cells that had been induced to undergo EMT and identified a set of 13 EMT-associated miRNAs (Figure 1A and S2A; Table S2). Using TaqMan qPCR assays, we confirmed that four miRNAs, miR-100, miR-125b, miR-22 and miR-720, were commonly upregulated miRNAs in EMT; five miRNAs, miR-200c, miR-141, miR-205, miR-663 and miR-638, were commonly downregulated miRNAs in EMT (Figure 1B and Table S2). The most dramatically deregulated miRNAs were miR-205 and two clustered miR-200 family members – miR-200c and miR-141 (Table S2), which are the first EMT-regulating miRNAs discovered through other approaches [4], [5]. This result validated the efficacy of our experimental system.
Differentially expressed miRNAs could be either causes or consequences of EMT. We cloned the four upregulated miRNAs into puromycin resistance cassette-containing retroviral vectors (MSCV-PIG and pBabe-puro) and expressed them individually in HMLE cells. While miR-125b and miR-720 did not cause any changes in cell morphology or EMT markers (Figure 1C and data not shown), expression of either miR-100 or miR-22 (Figure S2B) was sufficient to induce EMT: upon expression of either miRNA, epithelial cells became scattered and assumed fibroblastic morphology (Figure 1C); E-cadherin expression was undetectable and the mesenchymal marker vimentin was dramatically induced (Figure 1D). Similarly, expression of miR-100 in the MCF7 human epithelial breast cancer cell line (Figure S2C) also markedly downregulated E-cadherin and upregulated vimentin (Figure 1D), although we did not observe a clear morphological change.
We examined miR-100 expression levels in a series of human breast cancer cell lines. Relative to HMLE cells, epithelial-like tumor cell lines exhibited either comparable or much lower miR-100 expression, whereas mesenchymal-like tumor cell lines showed higher levels of miR-100 (Figure 1E). The association between miR-100 and EMT markers was further validated in human tumors: from the Cancer Genome Atlas (TCGA) breast cancer data [17], we observed a moderate but significant inverse correlation between miR-100 and E-cadherin expression levels (Rs = −0.1, P = 0.006, Figure 1F) and a highly significant positive correlation between miR-100 and vimentin expression levels (Rs = 0.43, P<2×10−16, Figure 1G).
We performed TCGA data analysis to determine the expression levels of miR-100 and miR-22 in human breast cancer. Surprisingly, miR-100 was found to be downregulated in all subtypes of human breast tumors, including luminal A (P = 1×10−11), luminal B (P = 0.008), basal-like (P = 0.006) and HER2 (P = 0.001) subtypes, compared with paired normal breast tissues (Figure 2A). Consistent with the correlation of miR-100 with EMT markers (Figure 1F and 1G), the luminal A subtype of primary breast tumors (which are known to be E-cadherin-positive and vimentin-negative) exhibited the most significant downregulation of miR-100 (Figure 2A). In contrast, miR-22 expression showed no significant difference between cancer and paired normal tissues (Figure S3). To determine the cellular origin of miR-100 expression, we performed in situ hybridization on human normal and cancer tissues, and found that miR-100 was indeed highly expressed in normal human mammary epithelium as opposed to barely detectable expression in the stroma, whereas human breast tumors exhibited reduced miR-100 expression (Figure 2B and 2C). Therefore, downregulation of miR-100 reflects the difference between normal mammary epithelium and breast tumor cells, but is not due to the difference in the stroma.
This observed downregulation of miR-100 in human breast tumors prompted us to determine whether it could be a tumor suppressor. Indeed, expression of miR-100 significantly inhibited the proliferation of HMLE cells in vitro, either in the presence or absence of ectopic expression of the Erbb2 mammary oncogene (Figure 2D and S4A). To validate this effect in vivo, we subcutaneously implanted Erbb2-expressing HMLE cells (HMLE-Erbb2) with or without miR-100 overexpression into nude mice. Strikingly, miR-100 expression dramatically suppressed tumor formation and growth (Figure 2E–2G), as it not only delayed initial tumor onset by one week (Figure 2E), but also caused a 83% reduction in tumor volume (683.3 mm3 vs. 117.2 mm3, Figure 2E) and a 84% reduction in tumor weight (0.62 g vs. 0.098 g, Figure 2F and 2G) at the late stage. Western blot analysis of E-cadherin and vimentin in tumor lysates (Figure 2H) and E-cadherin immunohistochemical staining of the tumors (Figure S4B) confirmed that the EMT status was retained in tumors formed by miR-100-expressing HMLE-Erbb2 cells. Furthermore, a 91% decrease in tumor weight was observed in mice implanted with miR-100-overexpressing MCF7 human breast cancer cells, compared with hosts of mock-infected MCF7 cells (Figure 2I and 2J).
We hypothesized that different target genes of miR-100 mediate the two distinct functions of this miRNA. Four miR-100 targets, SMARCA5, SMARCD1, MTOR (mammalian target of rapamycin) and BMPR2, have been identified by reporter assays previously [18], [19]. In addition, among all predicted targets of miR-100, HOXA1 is a mammary oncogene [20] and is upregulated in human breast cancer [21]; overexpression of HOXA1 in immortalized human mammary epithelial cells was sufficient to induce aggressive tumor formation in vivo [20]. While miR-100 did not substantially alter expression levels of SMARCD1, mTOR and BMPR2 in HMLE cells (Figure S5A), overexpression of this miRNA in both HMLE and MCF7 cells resulted in a pronounced decrease in SMARCA5 and HOXA1 protein levels (Figure 3A). Moreover, the activity of a luciferase reporter fused to a wild-type HOXA1 3′ UTR, but not that of a reporter fused to a mutant HOXA1 3′ UTR with mutations in the miR-100 binding site (Figure S5B), was reduced by 80% upon expression of miR-100 (Figure 3B), which validated HOXA1 as a direct target of this miRNA.
We silenced SMARCA5 in HMLE cells. This markedly reduced E-cadherin protein expression (Figure 3C) but did not alter cell proliferation (Figure S5C), suggesting that downregulation of SMARCA5 partially mediates the EMT-inducing effect of miR-100 but not its growth-inhibitory function. Conversely, re-expression of SMARCA5 in miR-100-overexpressing HMLE cells restored the expression of E-cadherin at both mRNA and protein levels (Figure 3D and 3E), although the mesenchymal morphology was not reversed. SMARCA5 (also named hSNF2H) is a chromatin-remodeling protein that physically interacts with the DNA methyltransferase DNMT3B [22]. Although it is not clear how this interaction modulates DNMT3B activity, we speculated that miR-100 might promote CDH1 (encoding E-cadherin) gene methylation by targeting SMARCA5. Indeed, bisulfite sequencing assays of the 27 CpG sites in the CDH1 promoter region revealed 29.6% methylation in the control HMLE cells and 55.1% methylation in miR-100-overexpressing HMLE cells, while re-expression of SMARCA5 reversed the effect of miR-100 on CDH1 promoter methylation (Figure 3F).
In contrast to the effect of SMARCA5, restoring HOXA1 expression in miR-100-overexpressing HMLE-Erbb2 cells to the same level as the control HMLE-Erbb2 cells (Figure 4A) did not affect expression levels of EMT-associated markers (Figure S5D), but instead fully rescued tumor onset and partially rescued tumor volume (51% rescue, Figure 4B) and tumor weight (40% rescue, Figure 4C and 4D). Consistent with the in vitro effect of miR-100 on EMT induction (Figure 1C and 1D) and cell proliferation (Figure S4A), the control HMLE-Erbb2 tumors were epithelial and had 80% Ki-67-positive cells, miR-100-expressing HMLE-Erbb2 tumors exhibited mesenchymal morphology and 8% Ki-67-positive cells, whereas HMLE-Erbb2 tumors with co-expression of miR-100 and HOXA1 were mesenchymal but showed 63% Ki-67-positive cells (Figure 4E). Taken together, downregulation of HOXA1 mediates, at least in part, the tumor-suppressing effect of miR-100 but not its EMT-inducing function.
Unexpectedly, despite strong EMT induction in both HMLE-Erbb2 and MCF7 cells, expression of miR-100 suppressed their migration and invasion in vitro, as gauged by Transwell assays (Figure 5A and 5B; Figure S6A and S6B). To further confirm the inhibitory effect of miR-100 on cell motility, we tracked the movement of individual cells cultured on top of collagen over a 24-hour period. Using time-lapse video microscopy, we observed a 53% decrease in the speed of movement of miR-100-expressing HMLE-Erbb2 cells compared with HMLE-Erbb2 cells (Figure 5C; Video S1 and S2). It should be noted that in order to permit the space for cell movement, the condition used for this experiment was low density and did not allow the majority of HMLE-Erbb2 cells to form epithelial clusters; however, we did observe HMLE-Erbb2 cell clusters with epithelial island structure that exhibited a surprisingly rapid collective movement and long trajectories without cell dissociation (Video S1 – note that an epithelial cell cluster initially appeared in the upper left corner and then moved to the lower part of the field), whereas all miR-100-expressing HMLE-Erbb2 cells had highly limited area of movement and reduced speed (Video S2).
To our knowledge, this is the first time that conversion from an epithelial state to a mesenchymal state has been found to be accompanied by reduced motility and invasiveness, which indicates that miR-100 may concurrently target EMT-repressing genes (SMARCA5) and pro-invasive genes. Indeed, HOXA1 has been identified as a driver of both oncogenesis and the invasion-metastasis cascade in human melanoma [23]. Consistent with this finding, restoration of HOXA1 in miR-100-overexpressing HMLE-Erbb2 cells (Figure 4A) rescued cell migration and invasion (Figure 5A and 5C; Figure S6A; Video S3). In contrast, neither re-expression of SMARCA5 in miR-100-overexpressing HMLE cells nor knockdown of SMARCA5 in HMLE cells affected cell motility (Figure S6C and S6D). To determine the loss-of-function effect, we used a miR-Zip method to achieve lentiviral inhibition of miR-100 in MDA-MB-231 breast cancer cells. Compared with cells infected with a scrambled hairpin control (Zip-scr), cells with approximately 60% knockdown of miR-100 (Zip-100, Figure 5D) displayed a significant increase in their migratory and invasive capacity (Figure 5E), while their mesenchymal status was not altered (data not shown).
We further validated the effect on tumor invasion in vivo: tumors formed by miR-100-overexpressing HMLE-Erbb2 cells were well demarcated and did not show overt invasion to their surrounding tissues (Figure 5F); in contrast, tumors formed by either the control HMLE-Erbb2 cells (mock) or HMLE-Erbb2 cells with simultaneous expression of miR-100 and HOXA1 were invasive and infiltrated muscular, adipose and stromal tissues (Figure 5F). We conclude from these experiments that miR-100 suppresses migration and invasion, at least in part, through direct targeting of HOXA1 but not SMARCA5.
HOXA1 is required for the development of the hindbrain, inner ear and neural crest in mammals [24]–[26]. Genome-wide expression profiling analysis of Hoxa1-null mouse embryos identified a number of Hoxa1 downstream targets involved in developmental processes [26]; three of the genes downregulated in Hoxa1 null embryos, Met, Smo (smoothened) and Sema3c (semaphorin 3c), are positive regulators of tumor cell migration, invasion and/or growth. MET, the receptor for hepatocyte growth factor, has been identified as a driver of tumorigenesis, motility and metastasis [27]. SMO is a central mediator of Hedgehog signaling, whereby Hedgehog binds to the twelve-pass transmembrane protein patched, alleviating patched-mediated inhibition of SMO [28]. It has been shown that the SMO inhibitor cyclopamine can lead to regression of medulloblastoma deficient in patched [29]. SEMA3C is a secreted protein that can induce migratory and invasive properties of breast cancer and prostate cancer cells [30], [31]. In addition, ectopic expression of HOXA1 in MCF7 breast cancer cells upregulated cyclin D1 [20], a cyclin that is required for steroid-induced proliferation of mammary epithelium during pregnancy [32] and promotes the development of mammary adenocarcinomas when overexpressed [33].
In the present study, ectopic expression of miR-100 markedly reduced the mRNA levels of MET, SMO, SEMA3C and CCND1, either in the presence or absence of Erbb2 expression (Figure 6A and 6B), while restoration of HOXA1 rescued the expression of each of these four genes (Figure 6B). A similar effect was observed on cyclin D1 protein expression levels (Figure 6C). Therefore, miR-100 downregulates multiple HOXA1 downstream targets involved in oncogenesis and invasiveness.
We sought to understand how miR-100 expression is regulated. Examination of the 2.5 kb genomic sequence upstream of the human mir-100 stem-loop identified two putative ZEB1-binding sites at −400 bp (Z-box, CAGGTA) and −2.2 kb (E-box, CAGCTG), respectively (Figure S7A). We designed PCR amplicons to assay for the presence of these putative binding sites in chromatin immunoprecipitates. This experiment revealed that ZEB1 bound to the E-box but not to the Z-box (Figure 7A and 7B). Moreover, luciferase assays demonstrated that ZEB1 significantly increased the activity of the putative mir-100 promoter (Figure 7C), suggesting that mir-100 is likely to be a transcriptional target of ZEB1. Interestingly, overexpression of either Twist or Snail increased ZEB1 expression to the level as high as that of ZEB1-overexpressing cells (Figure S1B), which could explain why miR-100 was identified as a commonly upregulated miRNA in HMLE cells overexpressing Twist, Snail or ZEB1. Consistently, miR-100 exhibited a strong positive correlation with Twist (Rs = 0.3, P = 5×10−19), Snail (Rs = 0.2, P = 4×10−7) and ZEB1 (Rs = 0.5, P<2×10−16) expression levels in human breast tumors (Figure S7B–S7D)
Upregulation of ZEB1 has been observed in triple-negative and basal-like breast tumors [34], [35]. Paradoxically, miR-100 is commonly downregulated in all subtypes of human breast cancers (Figure 2A), which indicates that other mechanisms lead to downregulation of miR-100. The mir-100 gene is embedded in a non-coding host gene, MIR100HG. Analysis of TCGA data revealed that 1.2% of the breast tumors (11 out of a total of 913 samples with copy number data available) had homozygous deletion of both mir-100 and MIR100HG, which could explain loss of miR-100 in these samples. Besides genetic alterations, a second common cause of loss of a tumor suppressor is DNA hypermethylation. From TCGA data, the majority of breast tumors (consisting of luminal A, luminal B, basal-like and HER2 subtypes) had much higher levels of MIR100HG gene methylation compared with paired normal mammary tissues (P = 2×10−12, n = 90, Figure 7D). Moreover, we observed a significant inverse correlation between MIR100HG gene methylation and miR-100 expression levels in breast cancer patients (Rs = −0.3, P = 7×10−17, n = 522, Figure 7E). Consistently, treatment of MCF7 and SUM149 human breast cancer cell lines with the DNA demethylating agent 5-azacytidine led to significant upregulation of miR-100 expression (Figure 7F and 7G). Taken together, these data suggest that miR-100 expression is regulated by both transcriptional activation and epigenetic silencing.
In summary, we identified miR-100 as a novel EMT inducer and a tumor suppressor, and validated in human tumors that miR-100 is downregulated in clinical breast cancer and correlates with EMT-associated markers. Notably, our results indicate the following: on one hand, both DNA hypermethylation and genetic deletion could contribute to miR-100 downregulation or loss in all subtypes of human breast tumors independently of EMT. On the other hand, induction of miR-100 may serve as a negative feedback mechanism to counteract the tumor-promoting and pro-invasive effect of EMT-inducing transcription factors. However, these transcription factors also regulate many other genes involved in cancer stemness, invasion and metastasis; for example, ZEB1 represses miR-200 [7] and Twist transactivates miR-10b [36]. This appears to be similar to other pleiotropically acting transcription factors: for instance, MYC is a cancer-driving oncoprotein and it is known to transcriptionally activate both pro-survival and pro-apoptotic genes [37].
Because induction of the EMT program can generate stem-like cells [13], [14], we examined the ability of miR-100 to regulate stem cell properties, as gauged by the stem cell marker ALDH1 (aldehyde dehydrogenase 1) [38] and mammosphere-forming ability [39]. Indeed, we observed induction of both ALDH1 expression (Figure S8A) and mammosphere formation (Figure S8B and S8C) by miR-100 in HMLE and HMLE-Erbb2 cells. Thus, the anti-tumor function of miR-100 is not due to depletion of the stem-like cell population, but instead results from inhibition of cell proliferation. In support of this notion, miR-100-expressing HMLE-Erbb2 tumors displayed a 90% reduction in the percentage of Ki-67-positive cells compared with the control HMLE-Erbb2 tumors (Figure 4E).
Our work is consistent with the anti-proliferative function of miR-100 described in several recent studies [18], [40], and is the first report of an EMT inducer that suppresses cell movement and invasion. Mechanistically, miR-100 induces EMT by targeting SMARCA5, an epigenetic regulator of E-cadherin, and inhibits tumorigenesis, migration and invasion by targeting HOXA1, leading to downregulation of multiple HOXA1 downstream targets involved in oncogenesis and invasiveness, including CCND1, MET, SMO and SEMA3C (Figure 7H). It should be noted that miR-100 has been reported to target IGF2 in 4T1 mouse mammary tumor cells [40]; however, IGF2 expression is undetectable in the human mammary epithelial cells (HMLE) used in this study (data not shown), although it is possible that IGF2 mediates the function of miR-100 in cells that express IGF2.
Another EMT-inducing miRNA identified in our study is miR-22. Consistent with our results, a recent report also demonstrated that miR-22 is an EMT inducer [41]. However, in stark contrast to miR-100, miR-22 functions to promote tumorigenesis, invasion and metastasis, ostensibly through direct targeting of the TET family of methylcytosine dioxygenases [41]. Although miR-22 expression showed no significant difference between breast tumors and paired normal mammary tissues based on TCGA data analysis (Figure S3), patients with high levels of miR-22 had worse survival rates than patients with low levels of miR-22 [41].
Taken together, these results do not argue that EMT itself suppresses cancer, but instead demonstrate that EMT is not always associated with increased tumorigenesis, migration and invasion, and that all EMT inducers are not equal: while some of them (such as miR-22) can promote tumorigenicity, motility and invasiveness, others (such as miR-100) inhibit these properties owing to their ability to target both EMT-repressing genes and oncogenic/pro-invasive genes (Figure 7H). Our findings raise the caution that the validity of using EMT-associated gene products as cancer biomarkers should be carefully assessed.
The HMLE cell line was from R. A. Weinberg's lab stock and cultured in complete Mammary Epithelial Cell Growth Medium (MEGM from Lonza). The MCF7, T47D, MDA-MB-231 and 293T cell lines were purchased from American Type Culture Collection and were cultured under conditions specified by the manufacturer. The SUM149, SUM159 and SUM229 cell lines were from S. Ethier and cultured as described (http://www.asterand.com/Asterand/human_tissues/149PT.htm). For demethylating studies, the MCF7 and SUM149 cells were treated with 2 µM 5-azacytidine (Sigma) for 12 hours.
The human mir-100, mir-22, mir-125b and mir-720 genomic sequences were PCR amplified from normal genomic DNA and cloned into the MSCV-PIG or pBabe-puro retroviral vector. A 1.5 kb putative human mir-100 promoter sequence containing the ZEB1-binding site (E-box) was PCR amplified from normal genomic DNA and cloned into the pGL3-Basic vector. A HOXA1 3′ UTR fragment was cloned into the pMIR-REPORT luciferase construct, using the following cloning primers: forward, 5′-ATCTTAGCTGGATATAATGTA-3′; reverse, 5′-TGCTTCATAAATTTCTTCATC-3′. A rat oncogenic (activated) form of Erbb2/NeuNT was from W. Guo. The Twist, Snail and ZEB1 expression constructs were from R. A. Weinberg. The human HOXA1 ORF was from Open Biosystems through MD Anderson's shRNA and ORFeome Core (PLOHS_100003514). The human SMARCA5 shRNA was from Sigma (TRCN0000013215). The human SMARCA5 expression vector was from GeneCopoeia (EX-E2767-Lv105). The miR-Zip construct expressing a short hairpin inhibiting miR-100 was from System Biosciences. The HOXA1 3′ UTR mutant was generated using a QuikChange Site-Directed Mutagenesis Kit (Stratagene). The vectors used in this study are listed in Table S3.
Total RNA, inclusive of small RNAs, was isolated using the mirVana miRNA Isolation Kit (Ambion) and was then reverse transcribed with an iScript cDNA Synthesis Kit (Bio-Rad). The resulting cDNA was used for qPCR using the TaqMan Gene Expression Assays (Applied Biosystems), and data were normalized to an endogenous control β-actin. Quantification of the mature form of miRNAs was performed using the TaqMan MicroRNA Assay Kit (Applied Biosystems) according to the manufacturer's instructions, and the U6 small nuclear RNA was used as an internal control. Real-time PCR and data collection were performed on a CFX96 instrument (Bio-Rad).
The production of lentivirus and amphotropic retrovirus and infection of target cells were performed as described previously [42].
Genes that contain the miR-100-binding site(s) in the 3′ UTR were obtained using the TargetScan program [43] (www.targetscan.org; version 6.2). The RNAhybrid program [44] was used to predict duplex formation between miR-100 and human HOXA1 3′ UTR.
To determine growth curves, we plated equal numbers of cells in 6-cm dishes. Starting from the next day, cells were trypsinized and counted every day. Cell counts were obtained from a TC10 Automated Cell Counter (Bio-Rad).
Transwell migration and Matrigel invasion assays were performed as described previously [36].
Mammosphere assay was performed according to the vendor (Stemcell Technologies)'s protocol. Briefly, single cell suspensions were seeded in the 6-well ultra-low attachment plate (Corning, 3471) at a density of 3.5–4.0×104 cells in 2 ml of freshly prepared Complete MammoCult Medium (Stemcell Technologies, 05620) per well. After incubation for 7 days, the number of mammospheres that were larger than 40 µm in diameter was counted.
Dual luciferase reporter assays were performed as described previously [36].
Western blot analysis was performed with precast gradient gels (Bio-Rad) using standard methods. Briefly, cells were lysed in the RIPA buffer containing protease inhibitors (Roche) and phosphatase inhibitors (Sigma). Proteins were separated by SDS-PAGE and blotted onto a nitrocellulose membrane (Bio-Rad). Membranes were probed with the specific primary antibodies, followed by peroxidase-conjugated secondary antibodies. The bands were visualized by chemiluminescence (Denville Scientific). The following antibodies were used: antibodies to E-cadherin (1∶1000, BD Transduction Laboratories, 610182), vimentin (1∶2000, NeoMarkers, MS-129-P), Erbb2 (1∶500, Cell signaling Technology, 2242), HOXA1 (1∶1000, Santa Cruz Biotechnology, sc-17146), SMARCA5 (1∶500, sc-8760 from Santa Cruz Biotechnology and ab3749 from Abcam), SMARCD1 (1∶500, Abcam, ab86029), mTOR (1∶1000, Cell signaling Technology, 2972), BMPR2 (1∶1000, Cell signaling Technology, 69679), cyclin D1 (1∶1000, Cell signaling Technology, 2922), ALDH1A1 (1∶1000, Santa Cruz Biotechnology, sc-22589), HSP90 (1∶3000, BD Transduction Laboratories, 610419) and cyclophilin B (1∶2000, Thermo, PA1-027A).
ChIP was performed with 293T cells transfected with SFB-tagged GFP or ZEB1, by using a Chromatin Immunoprecipitation (ChIP) Assay Kit (Millipore, 17–295) according to the manufacturer's instructions. After immunoprecipitation with FLAG antibody-conjugated beads (Sigma, M8823), protein-DNA crosslinks were reversed and DNA was purified to remove the chromatin proteins and used for PCR and qPCR. The PCR primers are: E-box, 5′-TACTAGGTCAGTATTTGATTT-3′ (forward) and 5′-GTTAGCGATAGACTAAGATCTAT-3′ (reverse); Z-box, 5′-ACCTATAAATCCGTTGGTAG-3′ (forward) and 5′-AATCTGGGCAAAGTGATACC-3′ (reverse). The qPCR primers are 5′-ACTTTGGATTGTTTGGAGGTTAAC-3′ (forward) and 5′-AATTTGCATGGCGCTCTTG-3′ (reverse).
Genomic DNA was extracted using the DNeasy Kit (Qiagen, 69504). The MethylDetector kit (Active Motif, 55001) was used to generate bisulfite-modified DNA. The modified DNA was purified and used as the template for nested PCR reactions with the following primers: outer primers, 5′-ATTCGAATTTAGTGGAATTAGAATC-3′ (forward) and 5′-AACCTACAACAACAACAACAACG-3′ (reverse); nested primers, 5′-TTAGTAATTTTAGGTTAGAGGGTTATCG-3′ (forward) and 5′-ACTCCAAAAACCCATAACTAACCG-3′ (reverse). The second-round PCR products were subcloned into the TOPO cloning vector (Invitrogen, K4600-01) and clones were randomly picked for DNA sequencing.
The double (5′ and 3′) digoxigenin (DIG)-labeled miR-100 probe and U6 probe were purchased from Exiqon. The normal mammary tissue and breast tumor sections were purchased from Origene (normal: CS807851; tumor: CS704488 and CS 711714). The tissue microarray (TMA) slide was purchased from Biomax (BR1006). In situ hybridization was performed according to the protocol of the miRCURY LNA microRNA ISH Optimization Kit (FFPE) (Exiqon). The stained slide was scanned on the Automated Cellular Image System III (ACIS III, Dako, Denmark) for quantification by digital image analysis. The color threshold was set up and standardized for all samples, and the color intensity was automatically scored for all individual cores on the TMA slide. The expression level was calculated from the score of color intensity and normalized to the internal control U6.
The 3.5-cm glass bottom multi-well plates (MatTek Corporation) were covered with 1 ml of 1.7 mg/ml collagen solution. After collagen solidified, we seeded 1×105 cells in serum-free and growth factor-free medium on top of the collagen. The cells were incubated overnight, and then were observed for 24 hours in a humidified, CO2-equilibrated chamber mounted on a Zeiss Axio Observer Z1 microscope. To quantitate the speed, we tracked the distance of individual cell movement by using Axio Vision software (Zeiss) in randomly selected fields. The speed of movement was calculated and presented as micrometers per minute.
Agilent human miRNA 8×15K microarray was used to profile global miRNA expression with standard procedures. Arrays were scanned using an Agilent scanner and data were extracted using Agilent's Feature Extraction software set to the default miRNA analysis protocol. The raw data were normalized and quantified by the LIMMA (linear models for microarray data) library, part of the Bioconductor project, using the R statistical environment. The raw data from all arrays were first background-corrected and then normalized using quantile normalization. The difference in miRNA expression between different groups was analyzed using empirical Bayes method implemented in the LIMMA package. P values obtained from the multiple comparison tests were corrected by false discovery rates.
Six- to eight-week-old athymic female nude mice were used for tumor cell implantation. Cells were injected subcutaneously into the left back of recipient animals. For recipients of MCF7 cells, Depo-Estradiol (Pfizer) was given to the mice two days before tumor cell implantation (1.5 mg/kg body weight), and the same dose was given once a week after implantation. Tumor size was measured weekly using a caliper, and tumor volume was calculated using the standard formula: 0.5×L×W2, where L is the longest diameter and W is the shortest diameter. Mice were euthanized when they met the institutional euthanasia criteria for tumor size and overall health condition. The tumors were removed and weighed. The harvested tumor samples were fixed in 10% buffered formalin for 12 h, washed with PBS, transferred to 70% ethanol, embedded in paraffin, sectioned and stained with hematoxylin and eosin (H & E).
Samples were deparaffinized and rehydrated. Antigen retrieval was done using 0.01 M sodium-citrate buffer (pH 6.0) at a sub-boiling temperature for 10 min after boiling in a microwave oven. To block endogenous peroxidase activity, the sections were incubated with 3% hydrogen peroxide for 10 min. After 1 h of preincubation in 5% normal goat serum to prevent nonspecific staining, the samples were incubated with the antibody to Ki-67 (1∶50, BD Biosciences, 550609) or E-cadherin (1∶500, Cell signaling Technology, 3195) at 4°C overnight. The sections were incubated with a biotinylated secondary antibody (1∶500, Vector Laboratories, BA-9200) and then incubated with avidin-biotin peroxidase complex solution (Vector Laboratories, PK-6100) for 30 min at room temperature. Color was developed using the Diaminobenzidine (DAB) substrate kit (BD Biosciences, 550880). Counterstaining was carried out using Harris modified hematoxylin.
We obtained level 3 data of mRNA expression, miRNA expression and gene methylation of human breast tumors from Synapse (http://synapse.org) (syn1461151). The mRNA expression levels (RNA-Seq by Expectation Maximization, RSEM) were measured by Illumina HiSeq (V2). The miRNA expression levels (normalized read counts) were measured by Illumina HiSeq and Illumina Genome Analyzer. The DNA methylation level (β value) was measured by the Illumina Infinium Human DNA Methylation 450 platform. The breast cancer subtype information (luminal A, luminal B, basal-like and HER2 subtypes) was described previously [17]. Paired t test was used to compare miRNA expression levels in all cases with miRNA expression data available from paired normal and cancer tissues (n = 56). Wilcoxon signed-rank test was used to compare MIR100HG methylation levels in all cases with gene methylation data available from paired normal and cancer tissues (n = 90). Spearman rank correlation test was used to assess the correlation between miR-100 expression level and MIR100HG gene methylation level (n = 522), and the correlation between miRNA expression levels and mRNA expression levels in all breast cancer samples with both miRNA and mRNA expression data available (n = 777).
Each experiment was repeated three times or more. Unless otherwise noted, data are presented as mean ± s.e.m., and two-tailed Student's t test was used to compare two groups for independent samples. Statistical methods used for TCGA data analysis are described above. P<0.05 was considered statistically significant.
All animal experiments were performed in accordance with a protocol approved by the Institutional Animal Care and Use Committee of MD Anderson Cancer Center.
|
10.1371/journal.pmed.1002322 | Long-term health status and trajectories of seriously injured patients: A population-based longitudinal study | Improved understanding of the quality of survival of patients is crucial in evaluating trauma care, understanding recovery patterns and timeframes, and informing healthcare, social, and disability service provision. We aimed to describe the longer-term health status of seriously injured patients, identify predictors of outcome, and establish recovery trajectories by population characteristics.
A population-based, prospective cohort study using the Victorian State Trauma Registry (VSTR) was undertaken. We followed up 2,757 adult patients, injured between July 2011 and June 2012, through deaths registry linkage and telephone interview at 6-, 12-, 24-, and 36-months postinjury. The 3-level EuroQol 5 dimensions questionnaire (EQ-5D-3L) was collected, and mixed-effects regression modelling was used to identify predictors of outcome, and recovery trajectories, for the EQ-5D-3L items and summary score. Mean (SD) age of participants was 50.8 (21.6) years, and 72% were male. Twelve percent (n = 333) died during their hospital stay, 8.1% (n = 222) of patients died postdischarge, and 155 (7.0%) were known to have survived to 36-months postinjury but were lost to follow-up at all time points. The prevalence of reporting problems at 36-months postinjury was 37% for mobility, 21% for self-care, 47% for usual activities, 50% for pain/discomfort, and 41% for anxiety/depression. Continued improvement to 36-months postinjury was only present for the usual activities item; the adjusted relative risk (ARR) of reporting problems decreased from 6 to 12 (ARR 0.87, 95% CI: 0.83–0.90), 12 to 24 (ARR 0.94, 95% CI: 0.90–0.98), and 24 to 36 months (ARR 0.95, 95% CI: 0.95–0.99). The risk of reporting problems with pain or discomfort increased from 24- to 36-months postinjury (ARR 1.06, 95% CI: 1.01, 1.12). While loss to follow-up was low, there was responder bias with patients injured in intentional events, younger, and less seriously injured patients less likely to participate; therefore, these patient subgroups were underrepresented in the study findings.
The prevalence of ongoing problems at 3-years postinjury is high, confirming that serious injury is frequently a chronic disorder. These findings have implications for trauma system design. Investment in interventions to reduce the longer-term impact of injuries is needed, and greater investment in primary prevention is needed.
| Improvements in trauma care have improved the chances of surviving serious injury, requiring a shift in focus to better understanding how well people recover from injury and how long this takes.
Longitudinal studies of the long-term health outcomes of seriously injured people are few. This study was undertaken to close this knowledge gap and provide valuable data necessary to inform trauma system design, injury rehabilitation programs, compensation schemes, and estimates of injury burden.
We followed a cohort of 2,757 major trauma patients in Victoria, Australia at 6 months, 12 months, 24 months, and 36 months after injury to collect health outcomes using the 3-level EuroQol 5 dimensions questionnaire (EQ-5D-3L).
We found that 20% of patients had died by 36-months postinjury. The proportion of survivors reporting persistent problems was high for each of the EQ-5D-3L items, although improvement was continuing at 36 months after injury for the usual activities item.
After adjusting our analyses to account for possible confounding factors, we found that lower levels of education, claiming compensation for injury, and age were consistent predictors of reporting problems at follow-up. The nature of injuries sustained, gender, preinjury employment, and level of socioeconomic disadvantage were also important predictors of problems on many of the 5 EQ-5D-3L items.
The prevalence of ongoing problems following serious injury was high at 36 months, though continued improvement was evident.
Investment in interventions designed to prevent major trauma overall, and to reduce the impact of injury, is clearly needed.
| The implementation of organised trauma systems has enabled considerable reductions in injury-related mortality in high-income countries [1–3]. With improving survival rates comes the potential for greater numbers of people living with long-term injury impacts, including reduced health status or health-related quality of life. An expert consensus group identified the capture of functional and quality-of-life outcomes following trauma as critically important for improving healthcare quality more than 2 decades ago [4]. Despite this, integration of these outcomes into trauma registries and system monitoring has been largely absent, and calls for their inclusion continue [5,6].
Improved understanding of the quality of survival is crucial in evaluating the quality of care provided to trauma patients. This information is important for understanding the patterns and timeframe of recovery and for best informing provision of healthcare, social, and disability services to those with ongoing issues following injury [6–8]. Establishing rates and patterns of recovery, and how these change over time, requires longitudinal data to inform what to measure, when, and for how long. To date, longitudinal studies of the health status of trauma patients have typically limited follow-up to 12- or 24-months postinjury [8–15]. Many have focused on less severely injured patients [10,11,13,14] and found little improvement beyond 9- to 12-months postinjury [10,12,14]. Previous studies of major trauma patients have shown continued improvement in function, return to work, and pain outcomes to 2 years after injury [8,9,16,17], and recent injury studies have shown variable trajectories of recovery in the first 2 years after injury [8,11]. However, studies extending beyond 2 years after injury are few and have involved very small cohorts [18,19], with 1 study showing improvement in physical health and function 10 years following injury when compared to 5 years in a cohort of 58 patients [18]. Our understanding of how long it takes to recover, and the factors that influence the degree and time course of recovery, after serious injury remains incomplete.
The aims of this population-based study were to: (i) describe the health status of seriously injured patients over a 3-year follow-up period, (ii) identify predictors of health status, and (iii) establish whether recovery trajectories differ by key patient and injury characteristics.
The Victorian State Trauma Registry (VSTR) and the REcovery after Serious Trauma—Outcomes, Resource use, and patient Experiences (RESTORE) project have been approved by the Human Research Ethics Committee of each participating hospital and Monash University.
The state of Victoria has a population 5.8 million and represents approximately 25% of the Australian population. An integrated trauma system was established in 2000 and is monitored using the VSTR, which captures data about all hospitalised major trauma cases. Major trauma is defined as any of the following: death following injury, an Injury Severity Score >12, urgent surgery, or admission to intensive care for >24h [20]. The registry data are regularly linked with the Victorian deaths registry to identify postdischarge deaths. Survivors to hospital discharge are routinely followed-up by telephone at 6, 12, and 24 months after injury to collect data about return to work, function, pain, and health status. The RESTORE study is extending the timeframe for follow-up to 36-, 48-, and 60-months postinjury for all patients with a date of injury from July 2011 to June 2012 [21]. Adult (aged 18 years and over) patients from the RESTORE project were included in this study.
Eligible patients were provided with a letter and brochure about the registry, explaining what is collected, how the data are used, and how to have their details removed. At each telephone interview, verbal consent to complete the interview was obtained.
While Australia’s publicly funded healthcare system (Medicare) provides healthcare coverage for all Australian citizens and permanent residents, 57% of the adult population purchase private health insurance. Additionally, Victoria has no-fault third party insurers for road (Transport Accident Commission [TAC]) and work-related (WorkSafe Victoria) injury, which provide compensation for treatment, rehabilitation, income replacement, and long-term support services.
The protocol for the RESTORE study is described elsewhere but summarised here [21]. Survivors to hospital discharge were telephoned at 6-, 12-, 24-, and 36-months postinjury using a standardised interview, which included validated patient-reported outcome measures. The collection of outcomes is also planned at 48- and 60-months postinjury. The health-status measure used was the 3-level EuroQol 5 dimensions questionnaire (EQ-5D-3L), comprising 5 items, including mobility, usual activities (e.g., work, study, housework, family, or leisure activities), self-care, pain or discomfort, and anxiety or depression [22]. For each item, the level of problems experienced is measured on a 3-point scale: no problems, some problems, severe problems. Age- and gender-specific population weights (tariffs) are applied to the item responses to generate preference weights, resulting in a utility score ranging from −0.594 to 1, in which 0 represents a health state equivalent to death, 1 represents perfect health, and a score <0 represents a health state considered worse than death. The United Kingdom value sets (or tariffs) were used for this study to calculate EQ-5D-3L preference weights, as these are most commonly used [23].
Demographic factors, injury event, injury type and severity, and other relevant factors were extracted from the registry and RESTORE for analysis. The registry receives the 10th Revision of the International Classification of Diseases—Australian Modification (ICD-10-AM) codes for each trauma patient. The ICD-10-AM coding is a requirement for all hospital admissions in Australia. The Charlson Comorbidity Index (CCI) and presence of preexisting mental health, drug, or alcohol conditions were mapped from the ICD-10-AM codes for each patient using published algorithms [24,25]. The CCI weight was used for analysis: 0 representing no CCI conditions, 1 representing at least 1 CCI condition with a weight of 1, and 2+ representing patients with at least 1 CCI condition with a weight of 2 or greater [24]. Socioeconomic status was obtained by applying the Index of Relative Socioeconomic Advantage and Disadvantage (IRSAD) to the patient’s postcode of residence to obtain quintiles of socioeconomic advantage ranging from 1 (most disadvantaged) to 5 (most advantage). Geographic remoteness was assessed by applying the Accessibility/Remoteness Index of Australia (ARIA), which classifies the patient’s postcode of residence into 1 of 5 categories: major city, inner regional, outer regional, remote, or very remote. These categories were collapsed into 2 categories, major city and all others combined, as the number of cases in regional and remote areas was low. Abbreviated Injury Scale 2005 (2008 revision) diagnosis codes were used to categorise each patient’s injuries into 1 of 7 nature-of-injury groups. These groups represent the 6 most common nature-of-injury groups and 1 residual group representing patients with burns or multiple injuries but without serious neurotrauma. The cause of injury categories was collapsed for analysis into the 6 most common causes of injury (motor vehicle occupant, motorcyclist, pedestrian or pedal cyclist, low fall, high fall, and struck by or collision with an object or person) and 1 residual category. Age was categorised into 7 groups for ease of interpretation of the findings and because the relationship with outcome was not linear. Self-reported preinjury disability was determined using a validated question asking the patient’s level of disability in the week prior to injury with options of none, mild, moderate, marked, or severe [26]. The marked and severe categories were combined for analysis due to the low number of patients reporting preinjury disability in these categories.
Frequencies, percentages, and the 95% confidence intervals (CI) of the percentages were used for categorical variables, and mean and SD were used for continuous variables. Patients were considered lost to follow-up if the EQ-5D-3L was missing at every follow-up time point (6-, 12-, 24-, and 36-months postinjury), and the patient had not died since hospital discharge.
Predictors of outcomes were assessed using mixed effects regression models [27]. All models included covariates previously identified as predictors of outcome in published studies and were adjusted for self-reported levels of preinjury disability. Both the CCI and self-reported preinjury level of disability were included in the model, as these represent separate constructs of preinjury health and function. Linear regression was used for the EQ-5D-3L preference or utility score, based on the age- and gender-specific population weights (tariffs), and a modified Poisson model was used for each of the 5 items, in which responses were dichotomised into no problems versus some/severe problems. Whilst the protocol paper was closely followed, a modified Poisson model was used to allow the estimation of relative risks (RRs) rather than odds ratios for the binary outcomes, as RRs are the preferred estimate for prospective studies [28]. The modified Poisson method has been shown to provide consistent results with a log-binomial model in which average risk levels are low to moderate, as observed in our data [29]. Similarly, the relative efficiency of the log-Poisson model is high compared to the log-binomial model at the average risk levels observed in our study [29]. Adjusted mean differences and 95% CI were calculated for linear models. Adjusted RR and 95% CI were calculated for binary outcome models. Differences in change over time in the prevalence of each outcome between patient subgroups (e.g., males versus females) were assessed using an interaction term between each covariate and time postinjury, with the adjusted mean difference and RR representing the difference in mean scores, or risk of improvement in outcome, in that group relative to the previous time point, respectively. These tests for interactions were performed to ascertain whether outcome trajectories differed between categories of the covariates—for example, whether the change in prevalence of the outcome differed over time for men compared to women.
Data were missing for some covariates: compensable status (n = 17, 0.7%); injury intent (n = 21, 0.9%); presence of a preexisting mental health, drug, or alcohol condition (n = 65, 2.9%); geographic remoteness and socioeconomic status (n = 65, 2.9%); preinjury work status (n = 107, 4.7%); preinjury level of disability (n = 115, 5.1%); and highest level of education (n = 315, 13.9%). Missing covariates were imputed using multiple imputation by chained equations. All covariates, and each outcome from the study, were used in the process to impute the missing covariate data. However, the data points with missing outcomes were not used in the modelling procedures [30]. Twenty datasets were produced, which were then combined using Rubin’s rules, a multiple-imputation algorithm [31]. A P value less than 0.05 was considered significant, and all analyses were performed using Stata Version 13.
There were 2,757 hospitalised major trauma patients in Victoria with a date of injury from July 2011 to June 2012; 12.1% (n = 333) died during their hospital stay, and a further 222 (8.1%) patients died postdischarge (Fig 1). The follow-up rates were 84% at 6-months, 85% at 12-months, 84% at 24-months and 74% at 36-months postinjury (Fig 1). The mean (SD) age of participants was 50.8 (21.6) years, and 72% were male. Ninety-two percent of cases were blunt trauma, with falls and road transport being the most common causes of injury. Over the study timeframe, 155 (7.0%) survivors to 36-months postinjury did not complete the 3-level EuroQol 5 dimensions questionnaire (EQ-5D-3L) at any time point and were considered lost to follow-up. Patients lost to follow-up were younger; less severely injured; more commonly injured through self-harm or assault; more commonly had a penetrating trauma type; had a higher prevalence of mental health, drug, or alcohol conditions; and more commonly lived in major cities (Table 1).
The pattern of change in the prevalence of patients reporting problems at each time point postinjury differed by EQ-5D-3L item (Fig 2 and Table 2). The usual activities item was the only outcome where improvement (compared with the previous time point) was observed at every time point after injury. There was a 13% (ARR 0.87, 95% CI 0.83–0.90) reduction in the prevalence and risk of reporting problems with usual activities at 12 months compared to 6 months, a 6% (ARR 0.94, 95% CI 0.90–0.98) reduction from 12 months to 24 months, and a 5% (ARR 0.95, 95% CI: 0.90, 0.99) reduction from 24 months to 36 months. The risk of reporting problems with self-care decreased by 14% (ARR 0.86, 95% CI 0.80–0.93) from 6- to 12-months postinjury but showed no improvement after 12-months postinjury, while the ARR of reporting mobility problems declined 7% (ARR 0.93, 95% CI 0.89–0.97) from 6 to 12 months after injury and a further 6% (ARR 0.94, 95% CI 0.90–0.99) from 12 to 24 months. Pain or discomfort outcomes improved 8% (ARR 0.92, 95% CI 0.88–0.96) from 6 to 12 months and to 6% (ARR 0.94, 95% CI 0.89–0.98) from 12 to 24 months, but there was a 6% (ARR 1.06, 95% CI 1.01–1.12) increase in the ARR of reporting problems with pain or discomfort from 24- to 36-months postinjury (Table 2). Improvement in anxiety or depression outcomes was only observed between 12- and 24-months postinjury. Overall health status, as measured by the EQ-5D-3L summary score, showed significant improvement from 6- to 12- and 12- to 24-months postinjury, before declining from 24-months to 36-months postinjury, driven by the poorer pain outcomes at 36 months (Table 2). The mean (SD) summary score at 6-months postinjury was 0.67 (0.31), 0.68 (0.32) at 12-months postinjury, 0.71 (0.31) at 24-months postinjury, and 0.70 (0.32) at 36-months postinjury.
Fig 3 provides a summary of the key findings from the multivariable mixed-effects regression models investigating the predictors of reporting problems for each EQ-5D-3L item. The frequency and percentage of patients experiencing problems on each item, and the full results from the models, including the ARR and 95% CI are detailed in the supplementary material (S1 Table to S5 Table). Significant differences in change over time in the prevalence of each outcome between patient subgroups are noted in each section, and the full tables of these analyses are available on request from the authors.
To our knowledge, we have conducted the largest, longitudinal study of health outcomes in seriously injured patients to date. Of the 2,757 hospitalised, adult major trauma patients, approximately 1 in 5 had died by 36-months postinjury, and the prevalence of persistent problems in survivors remained high for each of the EQ-5D-3L items. Forty percent of deaths in the cohort occurred after hospital discharge. Improvement to 24-months postinjury was evident for most items of the EQ-5D-3L, while improvement from 24- to 36-months postinjury was observed only for the usual activities item. Notably, the EQ-5D-3L summary score declined from 24- to 36-months postinjury, driven by an increase in reporting of pain or discomfort. Three factors were important predictors of outcome for all items of the EQ-5D-3L, age, compensable status, and level of education, while the nature of injuries sustained, gender, preinjury employment, and level of socioeconomic disadvantage were important predictors of problems on many of the 5 EQ-5D-3L items.
Claiming compensation from the state’s third-party, no-fault insurers for transport and work-related injury was a predictor of poorer outcome, even after adjusting for potential confounders such as age, socioeconomic status, and the nature of injuries sustained. The finding of poorer long-term outcomes in compensable patient groups is not new [8,32–34]. In our cohort, compensable patients had a higher prevalence of problems on each item of the EQ-5D-3L at each time point after injury than noncompensable patients. However, the compensable group did continue to improve beyond 24-months postinjury in self-care and usual activities, whereas noncompensable patients showed either no change or a decline in outcome from 24- to 36-months postinjury (S1 to S5 Tables). These findings suggest that the recovery process for compensable patients is slower than for other patients. The difference in prevalence of problems in usual activities between compensable and noncompensable cases halved over the study timeframe. The reasons for the differences in recovery rates between compensable and noncompensable patients is not fully known, but a number of factors may contribute. The complexity of navigating compensation agency processes has been suggested as a factor that may contribute to the disparities in health outcomes observed, with previous studies highlighting the association between stressful interactions with compensation agencies and poorer patient-reported outcomes [35,36]. Others have raised the potential for illness behaviour directed towards secondary gain from compensation agencies [37,38].
Increasing age has been identified previously as a predictor of poorer health status and functional outcome [8,10,11,14,33,39], although we found that the relationship between age and outcome differed between the EQ-5D-3L items. There was a dose-response-like increase in adjusted risk of reporting problems with each increase in age group for all items except the anxiety and depression item, for which older patients experienced significantly lower adjusted risk of reporting problems compared to the younger patients. The lower rates of anxiety and depression in older patients may reflect different life circumstances and pressures, and an overall lower risk of mental health issues in older age groups. Notably, recovery was poorer for older patients for most EQ-5D items (S1 to S5 Tables). As the trauma population ages, the burden of injury and the load on rehabilitation and disability services is expected to increase. Identification of interventions to improve outcomes in older trauma patients will be needed to mitigate this.
Our findings, combined with the results of previous studies [8,12,14,40], confirm that socioeconomic status is an important factor in determining outcomes following injury. Less than a university level of education was associated with greater adjusted risk of poorer outcomes on all EQ-5D-3L items, and there was a dose-response-like relationship between socioeconomic disadvantage and most outcomes. Studies of trauma patients have shown low health literacy, particularly in disadvantaged groups, with limited understanding of injuries and postsurgical instructions [41,42]. Others have shown poorer outcomes in surgical patients with lower health literacy [43]. Knowing what could improve one’s outcome and how to access appropriate services in a complex healthcare system is not easy for any patient and is especially challenging for those with low health literacy. These factors, along with service delays, waiting lists, and lack of financial resources in disadvantaged areas, could explain the disparity in outcomes by education level and socioeconomic status. Reducing health literacy demands on patients and improving health literacy amongst trauma populations could lead to better outcomes for patients [44].
Patients with a spinal cord injury demonstrated significantly poorer outcomes. The findings further highlight the life impacts of spinal cord injuries and support the need for continued investment in prevention and treatment options for this devastating injury type. In contrast, patients with isolated chest or abdominal injuries experienced better outcomes, although 30% continued to report problems on 3 of the 5 EQ-5D-3L items at 36-months postinjury.
In our study, the ARR of reporting problems on the mobility, usual activities, pain or discomfort, and anxiety or depression items was significantly higher for women compared to men. This difference has been identified in numerous trauma outcome studies [8,12,14,33,45,46] and remains unexplained. Differences in the psychological impact of the injury, and the influence of different social roles and responsibilities, have been suggested. Notably, intent of injury and the presence of a preexisting mental health, drug, or alcohol problem were not predictors of any outcome other than anxiety or depression. The prevalence of anxiety or depression problems was 15% to 20% higher in victims of interpersonal violence and 8% to 10% higher in patients injured through self-harm when compared to patients injured in unintentional events. These findings highlight the complex psychosocial needs of patients injured in intentional events.
The key strengths of this study were the large sample size, population-based design, prospective data collection, standardised approach to follow-up, use of a validated and recommended measure of health status in trauma, and high rates of follow-up. Additionally, the volume of missing data was low. Values were likely to be missing at random, allowing use of multiple imputation methods, and the complete case analyses were consistent with the results presented in this study from the imputed datasets. Nevertheless, there was responder bias, with patients injured in intentional events, younger, and less seriously injured patients being less likely to participate; therefore, these patient subgroups were underrepresented in the study findings.
In conclusion, this large-population cohort study of hospitalised major trauma patients demonstrates ongoing and dynamic differences in recovery trajectories across injury groups to 3-years postinjury. These findings highlight the high prevalence of ongoing problems following serious injury and have significant implications for trauma system design, injury rehabilitation programs, compensation schemes, and estimates of injury burden. Investment in interventions designed to prevent major trauma overall, and to reduce the impact of injury, is clearly needed.
|
10.1371/journal.ppat.1000418 | Molecular Mechanisms of Recombination Restriction in the Envelope Gene of the Human Immunodeficiency Virus | The ability of pathogens to escape the host's immune response is crucial for the establishment of persistent infections and can influence virulence. Recombination has been observed to contribute to this process by generating novel genetic variants. Although distinctive recombination patterns have been described in many viral pathogens, little is known about the influence of biases in the recombination process itself relative to selective forces acting on newly formed recombinants. Understanding these influences is important for determining how recombination contributes to pathogen genome and proteome evolution. Most previous research on recombination-driven protein evolution has focused on relatively simple proteins, usually in the context of directed evolution experiments. Here, we study recombination in the envelope gene of HIV-1 between primary isolates belonging to subtypes that recombine naturally in the HIV/AIDS pandemic. By characterizing the early steps in the generation of recombinants, we provide novel insights into the evolutionary forces that shape recombination patterns within viral populations. Specifically, we show that the combined effects of mechanistic processes that determine the locations of recombination breakpoints across the HIV-1 envelope gene, and purifying selection acting against dysfunctional recombinants, can explain almost the entire distribution of breakpoints found within this gene in nature. These constraints account for the surprising paucity of recombination breakpoints found in infected individuals within this highly variable gene. Thus, the apparent randomness of HIV evolution via recombination may in fact be relatively more predictable than anticipated. In addition, the dominance of purifying selection in localized areas of the HIV genome defines regions where functional constraints on recombinants appear particularly strong, pointing to vulnerable aspects of HIV biology.
| Recombination allows mixing portions of genomes of different origins, generating chimeric genes and genomes. With respect to the random generation of new mutations, it can lead to the simultaneous insertion of several substitutions, introducing more drastic changes in the genome. Furthermore, recombination is expected to yield a higher proportion of functional products since it combines variants that already exist in the population and that are therefore compatible with the survival of the organism. However, when recombination involves genetically distant strains, it can be constrained by the necessity to retain the functionality of the resulting products. In pathogens, which are subjected to strong selective pressures, recombination is particularly important, and several viruses, such as the human immunodeficiency virus (HIV), readily recombine. Here, we demonstrate the existence of preferential regions for recombination in the HIV-1 envelope gene when crossing sequences representative of strains observed to recombine in vivo. Furthermore, some recombinants give a decreased proportion of functional products. When considering these factors, one can retrace the history of most natural HIV recombinants. Recombination in HIV appears not so unpredictable, therefore, and the existence of recombinants that frequently generate nonfunctional products highlights previously unappreciated limits of the genetic flexibility of HIV.
| Pathogens, and viruses in particular, are subject to strong selective pressures during infection and often have characteristically high degrees of genetic variation [1]. Recombination is an important evolutionary mechanism that contributes to this genetic diversification. By creating novel combinations of pre-existing genetic polymorphisms in a single replication cycle, recombination enables greater movements through sequence space than can be achieved by individual point mutations. As a consequence, recombination provides access to evolutionary “shortcuts”. In addition, since recombination generally involves genes that already encode functional products, the probability of producing viable progeny is higher compared to the insertion of an equivalent number of random point mutations [2]. However, the generation of recombinant forms is not an unconstrained process. Genes and genomes generally evolve through the slow accumulation of point mutations, which often requires the progressive insertion of compensatory mutations at “linked” sites. This coevolution permits the preservation of epistatic interactions. By simultaneously introducing several substitutions, recombination has the potential to substantially perturb such coevolved intra-genome interaction networks [2],[3], impairing the functionality of the genes involved. Thus, the balance between the advantages of taking evolutionary shortcuts and the risk of chimeras being dysfunctional [2] determines the role played by recombination in the evolution of a given gene or organism.
Several studies have focused on the impact of recombination on the evolution of proteins, particularly in relation to directed evolution experiments [4],[5]. Two major factors have a large influence on the functionality of recombinants proteins. The first is the position of recombination breakpoint (the region where the sequence shifts from that of one parental sequence to the other) relative to the location of genetic polymorphisms within the gene. Recombinants involving a large number of non-synonymous substitutions will in fact have a low probability of being functional [2]. The second factor is the position of the breakpoints in relation to the boundaries of discrete protein folds. Breakpoints near the boundaries of these domains will in general have a smaller impact on protein folding, and hence protein function, than breakpoints occurring within them [3],[6],[7]. Recent work on Begomoviruses corroborated these findings by demonstrating that recombination events found in natural viral populations are significantly less disruptive of protein folding than randomly generated recombinants [8].
Adaptation of pathogens, either to on-going immune pressures within individual hosts or following transmission to new hosts of the same or different species, can result in infectious outbreaks that constitute major threats for public health [9]–[12]. The human immunodeficiency virus (HIV) is an extremely recombinogenic pathogen in which recombination has been implicated in key aspects of viral pathogenesis such as immune evasion [13], transmissibility [14], the evolution of antiretroviral resistance [15],[16] and cross-species transmission [9],[12]. Indeed, the remarkable genetic flexibility of HIV is underlined by its large genetic diversity. The HIV-1 population is subdivided into three groups, named M, N and O, with group M (which is responsible for the vast majority of the infections worldwide) being further subdivided into nine subtypes (named A, B, C, D, F, G, H, J and K) [17].
Although recombination in HIV has been shown to occur at all phylogenetic levels (intra- and inter-subtype, as well as inter-group, reviewed in reference [18]), the most widely noted impact of recombination on the genetic diversification of this virus is the frequent natural occurrence of inter-subtype recombinants in parts of the world where multiple subtypes co-circulate [19]–[22]. When the same inter-subtype recombinant is transmitted between multiple individuals, and has therefore the potential to be of epidemiological significance, it is termed a Circulating Recombinant Form (CRF) [17]. As with the HIV-1 subtypes, CRFs form distinct clusters in phylogenetic trees and some of them contribute substantially to the pandemic.
Sufficient inter-subtype recombinant sequences have been sampled to permit the detailed characterisation of variation in the locations of breakpoints both within individual genes [22],[23], and entire genomes [24],[25],[32]. This makes HIV a particularly useful model for studying the forces that shape pathogen populations within the context of global epidemics. Here we focus on recombination within the envelope gene (env). This gene encodes two polypeptides (gp120 and gp41) that form a heterodimer at the surface of the viral particle. Trimers of these heterodimers are the functional units that are responsible for binding to the cellular receptors and co-receptors and ultimately lead to viral entry into target cells [26]. The two protein products of env are also the targets of all the neutralising antibodies identified to date [27]. By using a tissue culture system to characterise inter-subtype recombinants generated within env in the absence of selection, and assaying the functionality of recombinant genes, we produce an empirical model of HIV recombination that accurately describes recombination patterns found in viruses sampled throughout the HIV pandemic.
We used different combinations of env sequences from primary HIV-1 isolates belonging to either different group M subtypes or group O (see Materials and Methods for the list of parental isolates used) to determine the distribution of breakpoints occurring within the HIV env gene in the absence of selection. We chose combinations of isolates belonging to subtypes that are co-circulating in regions of the world from which natural inter-subtype recombinant forms have emerged [28].
In order to quantify variations in recombination rates across env we used a previously described experimental system where human T cells are transduced with HIV-1 replication-defective vectors pseudotyped with the Vesicular Stomatitis Virus (VSV) envelope [29]. As this system mimics a single cycle of viral infection in which reverse transcription products neither influence cellular survival, nor confer a specific phenotype to the transduced cells, recombinants that were produced during reverse transcription were not subjected to any selection. After cloning of the reverse transcription products in E. coli, the system enabled identification of the recombinants based on the presence of a lacZ reporter gene (Figure 1). Given that known input sequences were used, such an approach enables the accurate and unambiguous localization of the breakpoint position to precise regions bounded by nucleotides that differ between the two parental sequences.
The regions of the envelope gene that were studied were chosen so as to obtain 700 to 1,500 nucleotides overlapping windows, spanning the whole of env. For each of seventeen different combinations of parental sequence pairs (Figure 2A), a recombination rate per nucleotide and per reverse transcription run was calculated within a 50 nucleotides sliding window (with 10 nucleotides step size). These were plotted as a function of the location of the window along the gene. To evaluate whether recombination-prone regions exist within the population, data from the 17 different pairs of parental sequences were pooled and an average recombination rate was computed for the different regions, and plotted as function of the position along the env gene (Figure 2B, top panel). Peaks and troughs were apparent all along the gene, with regions refractory to recombination being more common in the gp120 coding portion than in the gp41 region. The probability that breakpoints were more or less clustered across env than could be accounted for by chance (given the null hypothesis that breakpoint positions occur randomly) was determined by a permutation test (Figure 2B, bottom panel). Six major recombination-prone or “hot” regions (shaded light blue areas in Figure 2B) could be defined as env regions where breakpoint clusters were bounded by statistically significant breakpoint “cold spot” (p<0.05). Each of the six identified breakpoint clusters contained at least one breakpoint cluster that constituted a statistically significant recombination “hot-spot” (p<0.01). While these recombination-prone regions covered only slightly more than half of the whole gene (55.3%), they included 81.6% of all the breakpoints (337/413) mapped. These six hot regions are areas where recombination occurs preferentially during HIV replication, irrespective of the parental strains involved.
We next investigated the fate of these recombinants with respect to their establishment in the natural HIV-1 population. The fixation of a recombinant gene within a population is dependent on the interplay of multiple factors. Nevertheless, an obligatory component of evolution is undoubtedly the elimination by purifying selection of viruses that express dysfunctional proteins. To evaluate how profoundly this aspect of natural selection might influence the pattern of breakpoints generated by the mechanism of recombination, we determined the relative functionality of a subset of recombinant env genes.
In addition to encoding the proteins that coat the viral membrane, env also encodes a well-known functional RNA structure, the Rev responsive element (RRE). For the recombinants containing breakpoints in the RRE region the functionality of this RNA module was therefore also tested. Being involved in the regulation of the timing and the balance among the various forms of unspliced and partially or completely spliced RNAs, RRE is essential for viral replication [30]. Failure to properly regulate this process results in either a decrease or complete halt in viral production [30]. The functionality of chimaeric RREs was tested by measuring viral titres obtained upon transfection of cells with a plasmid containing the proviral sequence of the molecular clone NL4.3 of HIV-1, in which we had replaced the native NL4.3 RRE with that of the various chimaeric RREs. To uncouple the effects on RNA-folding caused by the introduced RRE sequences from those altering the amino acid sequence of expressed proteins, we used a variant of NL4.3 that does not express Env (NL4.3-Env−) [31], and a plasmid encoding the wild-type Env was co-transfected to complement the production of gp120 and gp41 proteins. In order to increase the statistical power of the analysis, additional chimaeric RREs were constructed using parental sequences other than those employed in our cell culture recombinant generation experiments (following a PCR procedure described in Materials and Methods) and tested for their functionality. As can be seen in Table 1, the viral titres obtained with every chimaeric RRE sequence we tested were both similar to those obtained with non-recombinant parental RRE sequences and markedly higher than that observed when the RRE was replaced with a non-viral sequence (see Materials and Methods). This result therefore clearly indicated that recombinants generated by breakpoints within the RRE generally retain the functionality of this element.
To determine the functionality of individual recombinant envelopes at the protein level, full-length recombinant envelope genes containing breakpoints of interest were constructed by successive PCR, as described in Material and Methods. Each full-length recombinant gene was then cloned in the pcDNA3.1 expression vector, and used to transfect HEK 293T cells together with the pNL4.3-Env−-Luc plasmid, to generate viral particles pseudotyped with the recombinant envelope of interest. The functionality of the recombinant envelopes was then tested after transduction of HEK 293T-CD4+-CCR5+ cells at a multiplicity of infection of 0.1, by measuring luciferase expression in these cells 48 hours after transduction. Since target cells cannot synthesize new viral envelope proteins, infection was limited to reverse transcription and, potentially, integration. The luciferase values observed therefore reflected the relative success of viral entry into the target cells. For this analysis recombinants derived from parental env sequences that yielded the strongest positive signals in this single cycle test were chosen (parental sequences A-Q461, C-CAP210, G-1033 and O-32, see Table 2 for their relative genetic distance) due to the higher reliability of the luciferase signal. The parental env sequences were used as controls. As for the functional analysis of the RRE, additional recombinants involving combinations of parental sequences – other than those involved in the experiments of recombination in cell culture, but carrying breakpoints in the same regions – were also tested. These additional recombinant env sequences were generated by PCR, as described for the reconstitution of the full-length env gene.
Luciferase values determined for each recombinant were plotted as a function of the corresponding breakpoint position (Figure 3). Recombinants with breakpoints falling within the six hot regions indicated in Figure 2B were preferentially characterized. It was apparent that most of the severely defective recombinants contained breakpoints in hot regions 2 and 3 of the recombination rate distribution (Figure 3). Given this data, we approximated a probability of Env functionality being disrupted by breakpoints falling within each of the six high recombination-rate regions. Since the parental sequences themselves were not uniformly functional (Figure 3), a situation that is probably common in nature, for each recombinant an estimate of loss of functionality was calculated by dividing the luciferase value obtained with that recombinant by the one of the least functional parental sequence involved in its generation. Recombinants displaying values between those of the two parental sequences were considered to retain functionality (and assigned a functionality value of 1). Of note, none of the recombinants yielded functionality values higher than that of the most functional parent from which it was generated. Values from recombinants containing breakpoints within the same region of the six hot regions were pooled, and a functionality loss value for each region was averaged (Figure 3). The most significant loss of functionality was observed in regions 2, 3, and 6.
Having defined a pattern of recombination in the absence of selection and the approximate probabilities of recombination events in various parts of env yielding fully functional products, we were interested in determining whether our experimental data could explain breakpoint patterns observed in circulating recombinants. The distribution along the whole HIV genome of 691 recombination breakpoints within HIV-1 group M full genome sequences from the LANL HIV Sequence Databases (http://hiv-web.lanl.gov/) was inferred. The same approach used in Figure 2B to define the probability that at any region of the genome the breakpoints were more clustered than would be expected by chance was used, with a 200 nucleotides window. A previous analysis of HIV recombinants modelled the distribution of breakpoints and indicated a significant clustering of breakpoints in the 5′ and 3′ ends of the envelope gene and a lack of breakpoints between these regions [32]. Our new analysis (Figure 4) confirmed the propensity for breakpoints to be located at the 5′ and 3′ ends of the env gene and the lack of breakpoints in the majority of its internal regions in recombinants from the database.
In order to compare our experimentally determined breakpoint distribution to that found in recombinants from the HIV Sequence database, a higher-resolution view of the breakpoint distribution within the env gene was determined using the positions of 133 unambiguously unique recombination breakpoints detectable within 230 env sequences. Following the same procedure described above, but using a 50 nucleotides window to enable detection of breakpoint clusters at the same resolution as in our experimental system, we identified a series of recombination hot- and cold-regions within the gene (Figure 5A, purple graph). In a similar way to the breakpoint distribution detected in cell culture, various hot regions could be defined (light-purple boxes at the bottom of Figure 5A), which corresponded remarkably well to recombination hot regions 1, 5 and 6 seen in cell culture (light-blue boxes). Whereas the other hot regions identified in cell culture had no corresponding counterparts in the natural breakpoint distribution, there was close correspondence between the cold-spots detected in both distributions.
Next we used the SCHEMA-based method [8] to investigate whether or not this breakpoint distribution exhibits evidence of purifying selection acting on recombinants with disrupted protein folding. This analysis indicated that breakpoints observable in natural viruses tend to occur in regions within env that were predicted to have a significantly lower impact on protein folding than randomly placed breakpoints (p<1.0×10−4 for gp120 and p = 8.9×10−3 for gp41, see Protocol S1). To investigate whether accounting for variations in the functionality of recombinants might reconcile the natural and experimental breakpoint distributions, we first approximated the combined effects of mechanistic recombination rate variation (Figure 2B) and selection for fully functional recombinants (Figure 3) on the distribution of breakpoints in cell culture. Selection “corrected” recombination rate estimates were then used to determine the distribution of 133 expected breakpoints. The resulting distribution was used to evaluate the probability of clustering of breakpoints (green graph in Figure 5A). Only regions 1, 4, 5 and 6 remained areas of significant clustering (light-green boxes at the bottom of Figure 5A), a pattern very close to that found in HIV recombinants sampled from nature, with the exception of region 4 for which there was substantially less evidence of recombination within natural recombinants than was expected based on our empirical model. Indeed, when compared to the distribution found for the 133 breakpoints encountered in the natural HIV recombinants (Figure 5B), a remarkable overlap was observed, with the discrete statistically significant breakpoint clusters being consistently recaptured by our empirical model of env recombination. The substantial difference of recombination rates in region 4 was also clear.
Through the functional characterization of HIV envelope genes generated by recombination in the absence of selection, we retrace the early steps shaping patterns of inter-subtype env recombination found in the HIV-1 pandemic. We observe that the mechanism of recombination alone defines regions where recombination occurs at significantly higher rates than elsewhere along the gene. The existence of such regions is strongly suggestive of spatially conserved features in HIV genomes that either promote or restrict recombination between different isolates. The distribution of breakpoints within the gp120 encoding region (Figure 2B) is likely due to the distribution of conserved and variable regions, the latter restricting recombination because of the low degree of local sequence identity between the parental sequences [32],[33].
Within genomic regions where sequence identity is high, a trigger for recombination could be the presence of secondary structures [34]. The highest recombination peak within the second region in Figure 2B (corresponding with the C2 portion of gp120) coincides with a recombination hot-spot that is determined by the presence of a stable RNA hairpin structure [29],[35],[36], while the fourth hot region (Figure 2B) corresponds to the RRE RNA structure that is highly conserved amongst all HIV isolates [37]. It is therefore possible that RNA secondary structures also contribute to the high rates of recombination observed at some of the other recombination hot regions. Noteworthy, the functionality of the RRE was retained even when crossing genetically distant isolates as for inter-group M/O recombinants (Figure 2A), supporting the possibility that regions of the genome harbouring functional RNA structures, which are generally more conserved within the population, provide a mechanism for crossing distantly related retroviruses and are possibly important for recombination of RNA viruses in general.
With respect to selection of recombinant genes at the protein level, experiments involving lattice proteins have shown that genes encoding proteins that tolerate mutations also tend to be recombination tolerant [2]. Since the env gene displays a degree of diversity between isolates from different HIV-1 group M subtypes ([38] and references herein) that is two to three times higher than the genome average, we anticipated that the manifest mutation tolerance of env might predispose it to high recombination tolerance. However, we show that this is not the case with certain regions within the gp120 encoding portions of env (particularly region 2 described in the present work in Figure 3) tending not to tolerate recombination well.
Viruses with small genomes (including all RNA viruses) tend to use overlapping genes expressed in different reading frames and to encode proteins that have multiple functions. The HIV envelope encodes for such proteins [26], and the subtle biochemical equilibrium that regulates their functionality is very possibly limiting tolerance to recombination. The low recombination tolerance of the gp120-encoding region could only be imprecisely predicted based on computational estimates of recombination induced protein fold disruption using the SCHEMA algorithm [3]. This may have been due to either our SCHEMA analyses being based on incomplete gp41 and gp120 structures or the fact that the structures used only reflected a single conformation of these two proteins. Therefore this analysis neither takes into account the conformational changes required for Env functionality, nor the quaternary arrangement of the proteins within Env trimers. Despite these issues, the SCHEMA analysis indicated that, amongst the HIV env sequences sampled from nature, selection has been acting against recombinants with disrupted protein folding (Table S1). Unravelling the molecular reasons for the reduced functionality of certain recombinants could provide valuable insights into the nature of the molecular interaction networks required for proper Env function.
The specific determinants of viral fitness (or in vivo replicative capacity) are complex and poorly understood at present. The fixation of a recombinant gene within a population is likely to depend on the interplay of multiple factors. Although combining cell culture functionality data with recombination rate heterogeneity is an oversimplified view of this process, the pattern of recombination predicted by our empirical model matches remarkably well the breakpoint distributions observed in nature (Figure 5B). The only major deviation from this was constituted by the fourth recombination hot region observed in cell culture, which was absent from the natural breakpoint distribution (Figures 2B and 5B). Determining the reasons for this discrepancy will improve our understanding of the mechanisms governing the success of recombinants in nature.
Although the host immune response certainly plays a significant role in the selection of recombinant variants in vivo [13], the similarities between the natural and experimental breakpoint distributions suggest that the forces responsible for the selection of recombinants in vivo only have limited impact on inter-subtype breakpoint patterns in env. This is most likely due to a combination of factors including mainly the complex epistatic interactions within env, the high density of fitness-determining loci within this gene, and the biochemical mechanism of recombination, which collectively constrain the fixation of genetic variability introduced by recombination. Negative fitness effects associated with recombination in env, however, should decrease with decreasing parental genetic distances [3],[6],[39] and therefore, in the context of intra-subtype recombination, the selective constraints on recombinants should be more relaxed than we have found them to be here.
Considering recombination in env in the context of the rest of the HIV genome, it is apparent that env displays the most dramatically variable natural breakpoint distribution of all HIV genes [24],[32], and it constitutes the only gene within which there is an extended region with limited recombination (Figure 4). Nevertheless, although less marked, breakpoint distribution patterns reminiscent of those found in env, with alternate clusters and troughs are also identifiable in several other regions of the genome such as gag and pol [32] (Figure 4). Although little information is presently available either on differential mechanistic predispositions to recombination across these regions, or on the functionality of the resulting products, it is tempting to speculate that underlying rules such as we have defined here for env may also be operational in these other cases.
In conclusion, by experimentally reproducing the generation of HIV-1 recombinants, we demonstrate that the distinctive distribution of breakpoints found in natural viruses is strongly shaped by both the mechanism of recombination, and the relative functionality of the recombinant genes. Thus, HIV evolution might not be the relentlessly unpredictable process it sometimes seems, and exploiting this evidence to pre-empt and counter the most favoured evolutionary tactics of this virus may ultimately be an efficient means by which we can devise effective vaccines and improve drugs against the virus.
HEK 293T, and CD4+CCR5+ 293T cells were grown in Dulbecco's modified Eagle's medium supplemented with 10% foetal calf serum, penicillin, and streptomycin (from Invitrogen, CA, USA), and maintained at 37°C with 10% CO2. MT4 cells were maintained in RPMI 1640 medium supplemented with 10% foetal calf serum and antibiotics at 37°C with 5% CO2.
The parental isolates used in this study were A-115, A-120, A-899 [33], A-859, A-905 (from S. Saragosti) and A-Q461 (Gene Bank: AF407156) for subtype A isolates; B-THRO (Gene Bank: AY835448), for subtype B; D-126, D-89, D-122, [33] and D-21.16 (Gene Bank: U27399), for subtype D; C-CAP210 (Gene Bank: DQ435683), for subtype C; G-858, G-914, (from S. Saragosti) and G-MP1033 (from M. Peeters, Gene Bank: AM279365), for subtype G; O-35 and O-32, for group O (from S. Saragosti).
Single cycle recombination assays were performed using a system previously developed by our laboratory [29]. HIV-1 env fragments from group M subtypes A, C, D and G, and from group O viral DNA were amplified by PCR from infected PBMCs obtained from patients and cloned in plasmids (called genomic plasmids), which differ for the genetic marker present downstream (in the sense of reverse transcription) of the sequence in which recombination is studied (Figure 1). All constructs were verified by sequencing. The trans-complementation plasmids, pCMV R8.2 [40] encoding HIV-1 Gag, Pol, and accessory proteins, and pHCMV-G [41] encoding the Vesicular Stomatitis Virus envelope protein were co-transfected into 293T cells with the two genomic plasmids to produce defective retrovirus particles which were then used to transduce MT4 cells as previously described [29]. The reverse transcription products were purified from the cytoplasmic fraction of transduced cells using the method described by Hirt [42]. The purified double stranded DNA was digested with DpnI for 2 h at 37°C (in order to eliminate possible contaminating DNA of bacterial origin) prior to PCR amplification as previously described [29]. The amplified product was purified after electrophoresis on agarose gel, digested with PstI and BamHI, ligated to an appropriate plasmid vector and used to transform E. coli. Plating on IPTG/X-Gal containing agar plates allowed blue/white screening of recombinant and parental colonies, respectively [29]. The frequency of recombination was determined by computing the number of blue colonies over the total number of colonies as described in reference [29]. Recombination breakpoints were identified by full-length sequencing of the env portion of the recombinant clones.
The recombinant and parental sequences of each pair of isolates tested were aligned using CLUSTAL X [43]. The breakpoint location of each recombinant was determined as being the central position of the interval bounded by the two closest nucleotide sites that were characteristic of each of the parental sequences). Recombination rates were calculated as follows. We define each recombination window studied with each pair of parental sequences as RwXYa–b, for a recombination window involving isolates X and Y, spanning position a to position b of env (reference sequence HXB2); a 50 nucleotides window was then considered (XYa–bSwi, for a sliding window starting at position i of env), beginning from the 5′ border of the sequence studied and the number of breakpoints (indicated as XYa–bni) falling within the window was counted. The resulting recombination rate per nucleotide in the sliding window XYa–bSwi iswhere XYa–bN is the total number of breakpoints characterized for the RWXYa–b pair, and 50 is the size in nucleotides of the sliding window, and F the frequency of recombination observed in the whole region studied, as defined in the previous chapter. The sliding window was then displaced with a 10 nucleotides increment (resulting in XYa–bSwi+10, XYa–bSwi+20, … ) across the recombination window, and XYa–bRi+10, XYa–bRi+20, … were computed. The various R values were reported in the graph as a function of the position of the midpoint of the window along the gene (i.e. the position of the 25th nucleotide of each sliding window). For the pooled dataset reported in Figure 2B, the analysis based on the sliding window was repeated. If Swpi stands for the sliding window at position i for the pooled dataset, Rpi for the corresponding recombination rate, and q is the number of recombination window including position i, recombination rate at position i is calculated asTo statistically test for the presence of recombination hot and cold-spots in the experimentally determined recombination breakpoint distributions we used a modification of a permutation test described previously [44]. Unlike in analyses of natural recombinants, the breakpoint positions approximated in our experimental procedure were not subject to biases introduced by underlying degrees of parental sequence nucleotide variability and patchiness of parental sequence sampling. Rather than explicitly accounting for these biases when placing randomised recombination breakpoints as in the permutation test described by Heath et al. [44], our modification of the test involved the completely randomised placement of recombination breakpoints. The test essentially involved the randomised recreation of 10,000 versions of our real dataset with each version containing exactly the same number of breakpoints between the same 17 parental sequence pairs observed in the real dataset. From breakpoint distributions determined for each of these 10,000 randomised datasets we were able to work out confidence intervals for expected breakpoint density variation given the completely random occurrence of recombination.
For simulating the distribution of 133 breakpoints based on the combined effects of (i) the mechanistic recombination rate and (ii) selection for functional recombinants, local recombination rate data used to generate the graph in Figure 2B were first multiplied by the respective functionality scores given in Figure 3 for each corresponding region, yielding “functionality corrected” rates for each region. Once the expected breakpoint distribution of 133 unique recombinants determined by this method, the number of breakpoints present in a 50 nucleotides rolling window, sliding with a 10 nucleotides increment was calculated and plotted (in Figure 5B) as function of the position along the gene. Deviations from expected degrees of breakpoint clustering given the null hypothesis of random breakpoint locations, was tested using the same modification of the Heath et al., [44] permutation test detailed above.
Full-length sequences of recombinant env genes were reconstituted, using an overlapping PCR procedure. We separately amplified the region from the 5′ end of the acceptor gene (using primer Topo5′ annealing to positions 5966–5990 of the reference strain HXB2) to the breakpoint position (using a specific primer encompassing the region of the breakpoint) and from the 3′ end of the donor gene (primer Donor3′, HXB2 positions 8785–8819) to the breakpoint position (also in this case with a specific primer). These PCR products, overlapping by approximately 30 nucleotides around the breakpoint site, were mixed at equal ratios and used as templates to generate the full-length recombinant env gene using primer Topo5′ and Donor3′. All PCR reactions were run with Phusion DNA polymerase (Finnzymes, Finland) for 30 cycles. PCR products were gel purified and ligated to pCDNA3.1 Topo (Invitrogen, CA, USA). For RRE functionality assays, a portion of the envelope gene containing the RRE of pNL4.3-Env−-Luc (nucleotides 7646 to 8046) was replaced with the corresponding sequence of parental or recombinant envelope genes or, as a negative control, a 400 nt sequence from the Drosophila melanogaster desoxynucleoside kinase gene (ΔdNK). All constructs were verified by sequencing.
HIV particles were produced by co-transfection of HEK 293T cells with an expression vector for a CCR5-tropic (ADA) HIV-1 envelope [45] kind gift of Dr. M. Alizon, together with a pNL4.3-Env−-Luc containing either a parental or recombinant RRE sequence or ΔdNK. Forty-eight hours post transfection, supernatants were filtered trough a 0.45 µM filter and p24 levels were determined using the HIV-1 p24 enzyme-linked immunoabsorbent assay kit (PerkinElmer Life Sciences, MA, USA).
Reporter HIV-1 particles were produced by co-transfection of HEK 293T cells with pNL4.3-Env−-Luc and either an empty expression vector or an expression vector encoding either a parental or a recombinant env. For each individual recombinant variant, prior to their use for transfection, clones were verified by sequencing of the region encoding the recombinant gene as well as the vector-encoded promoter for its expression. Supernatants, containing virus stock, were harvested 48 h post transfection, and filtered trough a 0.45 µM filter. Production of viral particles was tested using an enzyme linked immunoassay for HIV-p24 antigen detection (Perkin Elmer, MA, USA) and 20 ng of p24 were used to infect 105 293T CD4+-CCR5+ cells in 24 wells plates. Forty-eight hours later, cells were washed twice in PBS and lysed in 25 mM Tris phosphate, pH 7.8, 8 mM MgCl2, 1 mM dithiothreitol, 1% triton X-100, and 7% glycerol for 10 min in a shaker at room temperature. The lysates were centrifuged and the supernatant was used to measure luciferase activity using a GloMax 96 Microplate Luminometer (Promega, WI, USA) following the instruction of the luciferase assay kit (Promega, WI, USA). For samples that yielded negative results in the luciferase assay, plasmids from at least three independent bacterial clones were tested.
The HIV-1 group M envelope sequence alignment was retrieved from the Los Alamos National Laboratory (LANL) HIV Sequence Database (http://hiv-web.lanl.gov/). The alignment was reduced to subtype reference sequences (3 strains for each where available), 53 CRF strains (2 strains for each where available) and finally 197 apparently unique recombinants. Recombination was analyzed using the RDP [46], GENECONV [47], BOOTSCAN [48], MAXCHI [49], CHIMAERA [50], SISCAN [51], and 3SEQ [52] methods implemented in the program rdp3beta30 [53]. Default settings were used throughout except that: (1) only potential recombination events detected by four or more of the above methods, coupled with phylogenetic evidence of recombination were considered significant; (2) sequences were treated as linear; and (3) a window size of 30 variable nucleotide positions was used for the RDP method. Using the approach outlined in the rdp3 program manual (http://darwin.uvigo.es/rdp/rdp.html), the approximate breakpoint positions and recombinant sequence(s) inferred for every potential recombination event, were manually checked and adjusted where necessary using the phylogenetic and recombination signal analysis features available in rdp3. Breakpoint positions were classified as unknown if they were (1) detected at the 5′ and 3′ ends of the alignment but could have actually fallen outside the analysed region; or (2) within 20 variable nucleotide positions or 100 total nucleotides of another detected breakpoint within the same sequence (in such cases it could not be discounted that the actual breakpoint might not have simply been lost due to a more recent recombination event). All of the remaining breakpoint positions were manually checked and adjusted when necessary using mainly the MAXCHI and 3SEQ methods (using three sequence scans and the MAXCHI matrix method) but also the LARD matrix method (generated by the LARD two breakpoint scan; [54]), and the CHIMAERA method as tie breakers. The distribution of unambiguously detected breakpoint positions of all unique recombination events were analysed for evidence of recombination hot- and cold-spots with rdp3 as described by Heath et al. ([44]; a window size of either 50 or 200 nucleotides and 10 000 permutations). A normalised version of the breakpoint distribution plot described in that study was used in which the local probability values of breakpoint numbers (determined by a permutation test that takes into account that local degrees of sequence diversity influence the delectability of recombination events) were plotted instead of absolute breakpoint numbers.
PDB files detailing the three dimensional structures of both gp120 (PDB ID: 2B4C, determined by X-ray diffraction, resolution of 3.3 Å, 338 amino acids, [55]), and gp41 (PDB ID 1AIK, determined by X-ray diffraction, resolution of 2 Å, 70 amino acids, [56]) were obtained from http://www.rcsb.org. It is important to point out that these structures are partial and that we therefore only analysed a fraction of the structural interactions involved in Env folding. We performed SCHEMA predictions of recombination induced fold disruptions using the set of natural HIV env recombinants (described above) essentially as described in Lefeuvre et al. ([8]; See Protocol S1, Supplementary Analyses, for a description of the SCHEMA method). This involved: (1) computing protein fold disruption, or E, scores for each natural recombinant with identifiable parents; (2) based on every pair of parental sequences identified for the observed set of recombinants, simulating every possible recombinant that could have been produced with these parental sequence pairs that involved the exchange of the same number of non-synonymous polymorphisms as were exchanged during the actual recombination events; (3) calculating E scores for each of these simulated recombinants; and (4) using a permutation test to determine whether mean E scores calculated for the natural recombinants were significantly lower than mean E-scores for the same set of recombinants randomly drawn from the simulated recombinant datasets (Table S1). If fewer than 5% of simulated datasets had an average E score lower than that of the actual dataset (p<0.05) then this was taken to indicate that predicted fold disruptions incurred by real events were significantly less severe than if the observed distribution of breakpoints was uninfluenced by negative selection acting against recombinants with disrupted protein folding.
|
10.1371/journal.pgen.1003693 | QTL Analysis of High Thermotolerance with Superior and Downgraded Parental Yeast Strains Reveals New Minor QTLs and Converges on Novel Causative Alleles Involved in RNA Processing | Revealing QTLs with a minor effect in complex traits remains difficult. Initial strategies had limited success because of interference by major QTLs and epistasis. New strategies focused on eliminating major QTLs in subsequent mapping experiments. Since genetic analysis of superior segregants from natural diploid strains usually also reveals QTLs linked to the inferior parent, we have extended this strategy for minor QTL identification by eliminating QTLs in both parent strains and repeating the QTL mapping with pooled-segregant whole-genome sequence analysis. We first mapped multiple QTLs responsible for high thermotolerance in a natural yeast strain, MUCL28177, compared to the laboratory strain, BY4742. Using single and bulk reciprocal hemizygosity analysis we identified MKT1 and PRP42 as causative genes in QTLs linked to the superior and inferior parent, respectively. We subsequently downgraded both parents by replacing their superior allele with the inferior allele of the other parent. QTL mapping using pooled-segregant whole-genome sequence analysis with the segregants from the cross of the downgraded parents, revealed several new QTLs. We validated the two most-strongly linked new QTLs by identifying NCS2 and SMD2 as causative genes linked to the superior downgraded parent and we found an allele-specific epistatic interaction between PRP42 and SMD2. Interestingly, the related function of PRP42 and SMD2 suggests an important role for RNA processing in high thermotolerance and underscores the relevance of analyzing minor QTLs. Our results show that identification of minor QTLs involved in complex traits can be successfully accomplished by crossing parent strains that have both been downgraded for a single QTL. This novel approach has the advantage of maintaining all relevant genetic diversity as well as enough phenotypic difference between the parent strains for the trait-of-interest and thus maximizes the chances of successfully identifying additional minor QTLs that are relevant for the phenotypic difference between the original parents.
| Most traits of organisms are determined by an interplay of different genes interacting in a complex way. For instance, nearly all industrially-important traits of the yeast Saccharomyces cerevisiae are complex traits. We have analyzed high thermotolerance, which is important for industrial fermentations, reducing cooling costs and sustaining higher productivity. Whereas genetic analysis of complex traits has been cumbersome for many years, the development of pooled-segregant whole-genome sequence analysis now allows successful identification of underlying genetic loci with a major effect. On the other hand, identification of loci with a minor contribution remains a challenge. We now present a methodology for identifying minor loci, which is based on the finding that the inferior parent usually also harbours superior alleles. This allowed construction for the trait of high thermotolerance of two ‘downgraded parent strains’ by replacing in each parent a superior allele by the inferior allele from the other parent. Subsequent mapping with the downgraded parents revealed new minor loci, which we validated by identifying the causative genes. Hence, our results illustrate the power of this methodology for successfully identifying minor loci determining complex traits and with a high chance of being co-responsible for the phenotypic difference between the original parents.
| Many genetic traits are quantitative and show complex inheritance. Because these traits are so prevalent in nature, understanding the underlying factors is important for various biological fields and for applications like industrial biotechnology and agricultural practice [1]. Recently, baker's yeast Saccharomyces cerevisiae has become an important subject for studies in quantitative genetics [2], [3]. In particular the availability of high-density genetic markers, the ease of performing experimental crosses and the powerful technologies for precise genetic modification [4], [5], do not only allow efficient QTL mapping but also rapid identification of causative genes and their experimental validation and interaction analysis. S. cerevisiae displays many quantitative traits that are also important in other cell types, including industrial microorganisms and cells of higher, multicellular organisms. Such properties include thermotolerance [6] and oxidative stress tolerance [7], the capacity to produce small molecules, such as acetic acid [8] and ethanol tolerance [9], [10]. Other quantitative traits that have been studied in yeast include transcriptional regulation [11], sporulation efficiency [12], telomere length [13], cell morphology traits [14], mitochondrial genome instability [15], global gene expression [16], evolution of biochemical pathways [17] and resistance to chemicals [18].
A major remaining challenge in quantitative trait studies is the efficient mapping of minor quantitative trait loci (QTLs) and identification of their causative genes. Minor QTLs have a subtle influence on the phenotype, which is easily masked by epistasis [19], gene-environment interactions [20], low association to the phenotype because of limited sample size and complex interactions with other QTLs. Minor QTLs are important because together they can produce in an additive or synergistic manner equally dramatic effects on the phenotype as major QTLs. Actually, the work of Bloom et al. [21], in which a large panel of individually genotyped and phenotyped yeast segregants was used, has shown that for 46 quantitative traits, the assembly of all detected loci could explain nearly the entire additive contribution to the heritable variation. The minor QTLs identified should be truly relevant for the trait of interest in the original parent strains and not generated in some unrelated way during the mapping procedure.
Several methods have been reported to identify minor QTLs. Sinha et al. [22] used a targeted backcross strategy to first eliminate a major QTL. Subsequent mapping revealed a novel allele that had an epistatic interaction with the first major QTL. A disadvantage of backcrossing is the reduction of genetic diversity, which likely leads to loss of minor QTLs. In a different approach, Lorenz and Cohen [23] fixed major QTLs either in the superior parent or in the inferior parent and successfully identified minor QTLs by linkage analysis by repeating the QTL mapping with the new parent strains. A potential problem caused by elimination of major QTLs in one parent is that the phenotypic difference between the two parent strains is reduced. This may make it more difficult to evaluate the phenotype of the extreme segregants in comparison with the superior parent. Parts et al. [24] used many millions of segregants and multiple inbreeding steps to facilitate the detection of statistically significant minor QTLs. The use of such a high number of segregants, however, is only feasible for selectable phenotypes. Swinnen et al. [10] made use of more stringent phenotyping, i.e. tolerance to higher ethanol levels, which revealed several additional minor QTLs. The disadvantage of this approach is that higher stringency of phenotyping requires higher numbers of segregants to be phenotyped to obtain enough segregants with the superior phenotype. In the study of Bloom et al. [21], aimed at identifying the source of missing heritability, linkage analysis was performed with a large panel of individually genotyped and phenotyped yeast segregants, which enabled detection of many QTLs with a small effect.
In this work we have extended previous approaches to identify minor QTLs to QTL mapping by pooled-segregant whole-genome sequence analysis and we eliminated the effect of major QTLs in both parents. Our method is based on the observation that superior haploid segregants of natural or industrial diploid strains usually contain mutations that to some extent compromise rather than promote the trait of interest. As a result genetic mapping with such segregants usually reveals QTLs, which are linked to the inferior parent rather than to the superior parent. This allows the construction of two new parent strains, which are both downgraded for the trait of interest by replacement of a superior allele with an inferior allele from the other parent. This maintains a large phenotypic difference between the new parent strains. They also retain all genetic diversity, in particular all remaining minor QTLs. We show the effectiveness of this approach by first mapping QTLs involved in high thermotolerance of a selected yeast strain compared to a control strain, identifying causative genes linked to the superior and inferior parent, constructing two downgraded parent strains and repeating the genetic mapping with the new parents. This revealed several new minor QTLs, which we validated by identifying the causative gene in two QTLs. Interestingly, the two novel causative genes identified in this study are both involved in pre-mRNA splicing, which suggests an important role for RNA processing in conferring high thermotolerance.
We have screened a total of 305 natural and industrial isolates of S. cerevisiae for their ability to grow at high temperature, i.e. 40–41°C, on solid YPD plates. Not a single yeast strain was able to grow with a reasonable rate at 42°C. The strain MUCL28177 showed very good growth at 41°C and was chosen for further analysis. After sporulation, we selected a haploid segregant MUCL28177-21A, further referred to as 21A, which also showed excellent growth at 41°C compared to the control strain BY4742. Strain 21A was crossed with the laboratory strain BY4742, that is unable to grow at 41°C. The hybrid 21A/BY4742 diploid strain grew at least as well as the 21A strain at 41°C, indicating that the high thermotolerance of 21A is a dominant characteristic. Phenotyping of 950 segregants of the 21A/BY4742 diploid strain revealed a range of thermotolerance. It resulted in 58 segregants with similar growth at high temperature as 21A. The growth of the original strain MUCL28177, the parent strains 21A and BY4742, the hybrid diploid strain 21A/BY4742 and ten representative segregants with varying thermotolerance, is shown in Figure 1.
The 58 thermotolerant segregants were pooled based on dry weight and genomic DNA isolated from the pool. Genomic DNA samples from the pooled segregants and from parent strain 21A were sequenced. The sequence reads obtained were aligned with the sequence of the reference S288c genome, which is essentially the same as that of the inferior parent strain BY4742. A set of quality-filtered SNPs to be used as genetic markers, was acquired essentially as described before [10]. For each chromosome, the SNP variant frequency was modeled using an additive logistic regression model [10], [25]. The results are shown in Figure 2. In the top panel, the raw SNP frequencies are plotted against the chromosomal position along with the modeled frequency (smoothed lines). The middle panel shows contrasts between selected pools and an unselected pool along with 95% simultaneous confidence bands. Upward and downward deviations from 0 indicate putative QTLs containing causative alleles from the superior and inferior parent, respectively. Normally, only linkage with the superior parent strain is expected. However, since the original MUCL28177 diploid strain is a natural isolate, it is likely heterozygous. Hence, the 21A segregant may contain recessive mutations that compromise to some extent thermotolerance in spite of the fact that its overall thermotolerance was only slightly lower than that of the MUCL28177 parent strain.
We calculated 2-sided p-value profiles along the chromosome that were adjusted for multiple testing (Text S1 online: Supplementary Methods) and five regions show significant p-values (0.05 significance level, Figure 2). We chose four regions with the smallest p-values for further analysis (Table S1 online). For these loci, selected SNPs were scored in individual thermotolerant segregants (up to 62 after additional segregant isolation and phenotyping) and a binomial exact test with FDR adjusted p-values was used for assessing statistical significance [10], [26]. Three QTLs (QTL1, QTL2 and QTL3) were confirmed to exhibit statistically significant linkage to the high thermotolerance phenotype (0.05 FDR level, Table 1). QTL1 and QTL2 showed linkage with the genome of the superior 21A parent strain, while QTL3 showed linkage with the genome of the inferior BY4742 parent strain. We concentrated our work first on QTL1 and QTL3, because they showed the strongest linkage to the superior and inferior parent, respectively. The subtelomeric regions often show deviations from the 50% value of the SNP variant frequency, but this is also observed in the unselected pool. It may be caused by complications with the mapping of repetitive sequences, which are known to be commonly present in subtelomeric regions. We have analysed for instance the right subtelomeric region of chromosome X, in the mapping with the original parents, using SNP detection in the individual segregants and found a p-value that failed to indicate significant linkage (results not shown).
We first fine-mapped QTL1 by scoring eight selected SNPs in individual thermotolerant segregants, which reduced the size of the locus to about 60,000 bp (Figure 3A). Detailed analysis of the 21A sequence of this region showed that 22 out of the 33 genes and putative ORFs present contained at least one non-synonymous mutation in the ORF compared to the BY4742 sequence (Figure 3A). Next we applied reciprocal hemizygosity analysis (RHA) [6] to identify causative gene(s) in QTL1. RHA is used to test for a possible contribution to the phenotype of each allele of the candidate gene in a hybrid genetic background. For each of the 22 genes with non-synonymous mutations, we constructed two 21A/BY4742 hybrid strains in which either the 21A or the BY4742 allele was deleted, so that each strain only contained one specific allele of the candidate gene. Comparison of the growth at high temperature (41°C) of the two hybrid strains did not show any difference for the 22 candidate genes, except for MKT1 (Figure 3B, Figure S1 online and data not shown). The hybrid strain with the MKT121A allele showed better growth than the strain with the MKT1BY4742 allele. We further confirmed the relevance of MKT1 by demonstrating that MKT1 deletion reduced thermotolerance in the 21A strain background (Figure S2 online). Since 21A with either mkt1Δ or MKT1BY4742 showed the same growth at 40.7°C and since BY4742 showed the same growth at 40.7°C as BY4742 mkt1Δ, the MKT1BY4742 allele behaves as a loss of function allele for thermotolerance when assayed under our conditions and in our haploid strain backgrounds (Figure S2 online).
In a previous QTL mapping study of thermotolerance with a clinical isolate of S. cerevisiae and the lab strain S288c, the MKT1 allele of the clinical isolate was also identified as a causative gene [6] and in a follow-up study, out of two polymorphisms in Mkt1, D30G and the conservative substitution K453R, the D30G mutation was identified as the causative mutation [27]. Sanger sequencing of MKT121A confirmed that Mkt1-21A has the same mutations. END3 and RHO2, which are located close to MKT1 in the same QTL, were also reported to have an allele-specific contribution to thermotolerance [6]. However, in the current experimental setup, the RHO2 alleles from our two genetic backgrounds did not produce a difference in thermotolerance, while for END3 there may be a slight difference (Figure 3B). Sequence alignment using the Illumina sequencing data shows that END321A lacks the causative SNP (C773T) found in END3YJM145 [27]. In the case of RHO2 it is known that SNPs in the 3′UTR of RHO2YJM145 are responsible for the phenotypic effect on thermotolerance. RHO221A contains the same SNPs in its 3′UTR except for insertion of an A six base pairs downstream of the ORF. Hence, this insertion in RHO2YJM145 may cause the growth advantage at high temperature.
QTL3 is linked to the genome of the inferior parent strain, indicating that BY4742 contains a superior genetic element for thermotolerance in this region. We fine-mapped QTL3 by scoring seven selected SNPs in 62 thermotolerant segregants individually. This reduced the locus to 40,000 bp (Figure 4A). Detailed analysis of the 21A sequence in this region revealed 13 genes and putative ORFs with at least one non-synonymous mutation (Figure 4A).
To accelerate identification of the causative genes in this region, we first performed ‘bulk RHA’. Instead of comparing alleles for each single gene, we first made a reciprocal deletion in the hybrid strain of a fragment with multiple genes. We divided the 40,000 bp region of QTL3 into two fragments, with the first containing 11 genes and the second 14 genes (Figure 4A, 4B). For each fragment, we constructed two hybrid strains with one strain containing only the fragment from the 21A background and the other only the fragment from the BY4742 background (Figure 4B). Comparison of growth at high temperature (41°C) showed that FRAGMENT1BY4742 conferred better growth at high temperature than FRAGMENT121A, while there was a much smaller difference between the strains with FRAGMENT2BY4742 or FRAGMENT221A (Figure 4C).
We then applied RHA for the six individual genes of FRAGMENT1 that had at least one non-synonymous mutation (Figure 4C). This identified PRP42BY4742 as a superior allele for thermotolerance compared to PRP4221A, whereas for the other genes there was no allele-specific difference in thermotolerance (Figure S3 online). We also tested growth at high temperature of strains containing a heterozygous deletion of either the complete FRAGMENT1 or only the PRP42 gene together on the same plate. We found the growth at 41°C to be similar whether the complete FRAGMENT1 or only the PRP42 gene from either BY4742 or from 21A was deleted (Figure S4 online). This suggests that PRP42 was likely the only causative gene in FRAGMENT1 and thus also seems to exclude the other genes without non-synonymous mutation in their ORF as possible causative gene. As an additional control, we also performed RHA with the seven genes of Fragment 2 with a non-synonymous mutation in their ORF and we did not find any difference between the alleles from the two parent strains in conferring thermotolerance (data not shown).
PRP4221A has eleven mutations compared to PRP42BY4742, with three of them being non-synonymous and the other eight synonymous (Table S2 online). The three polymorphisms in Prp42, H296Y, F467S, and E526Q, are non-conservative substitutions, but it is difficult to predict a possible effect on the function or structure of the protein. They are located in domains without strong conservation (data not shown). Since no mutation was present in the promoter and terminator region, the difference in thermotolerance conferred by the two PRP42 alleles is likely due to the change in protein sequence and thus in functionality. We have investigated the presence of these mutations in 22 other yeast strains, isolated from various sources, and of which the whole genome has been sequenced (Table S2 online), and found that among the three non-synonymous mutations, C886T is unique to 21A, whereas the other two mutations (C1400T and C1576G) are present in all other strains except in the lab strains S288c, CEN.PK113-7D and W303. If we assume that the inferior PRP42 allele is rare (like the inferior MKT1 allele in S288c), then C886T is the best candidate for the causative mutation. On the other hand, we cannot exclude that C886T is only one of the causative mutations, that it requires interaction with one or more of the other mutations or that a combination of the other SNPs is causative for the phenotype.
We next constructed two downgraded parent strains each with their own superior allele replaced by the inferior allele of the other parent: 21ADG: 21A mkt1Δ:: MKT1BY4742 and BY4742DG: BY4742 prp42Δ::PRP4221A. Growth at 41°C of 21ADG was reduced compared to 21A, confirming the importance of MKT121A for high thermotolerance in 21A (Figure 5A). At 41°C, BY4742 and also BY4742DG are not able to grow (Figure 5A). Hence, we reduced the temperature to 40.7°C, which allowed to demonstrate reduced growth of BY4742DG compared to BY4742 (Figure 5B). Also at 41°C, we could demonstrate the beneficial effect of PRP42BY4742 compared to PRP4221A by comparing growth of the 21ADG/BY4742 and 21ADG/BY4742DG hybrid strains (Figure 5A). The availability of the four hybrid diploid strains also allowed us to demonstrate that in this background the effect of the MKT1 and PRP42 genes on thermotolerance is independent. The hybrid diploids, 21ADG/BY4742 and 21A/BY4742DG, each with replacement of one superior allele, both showed reduced growth at 41°C compared to the original hybrid of the parent strains, 21A/BY4742, while the hybrid of the two downgraded parent strains, 21ADG/BY4742DG, in which both superior alleles are replaced, showed further reduced growth (Figure 5A). (In this figure all strain pairs were put on the same plate.)
Figure 5 shows that both at 41°C and 40.7°C, the two downgraded parent strains, 21ADG and BY4742DG, still show a strong difference in thermotolerance. We sporulated the 21ADG/BY4742DG diploid strain and phenotyped 2464 segregants for thermotolerance. Examples are shown in Figure 5B. The segregants showed a range of thermotolerance and also transgressive segregation [28], since some of the segregants showed poorer thermotolerance than the inferior BY4742DG parent (e.g. segregant 9 in Figure 5B) while others showed better thermotolerance than the superior 21ADG parent (e.g. segregant 8 in Figure 5B). This suggests the presence of additional QTLs and causative genes influencing thermotolerance.
From the 2464 segregants derived from the diploid 21ADG/BY4742DG we selected 58 thermotolerant segregants that grew at 40.7°C at least as well as the 21ADG superior parent strain, and repeated the pooled-segregant whole-genome analysis. We have used the same set of SNPs as generated in the previous sequencing of the 21A parent strain compared to S288c, for the mapping of QTLs linked to thermotolerance. A total of ten regions have a 2-sided p-value low enough for significance (Figure 2). Interestingly, two regions can be discerned with a clear difference between the original and downgraded pool (Figure 2, Table S1 online). The previous peak indicating linkage of one or more causative elements in the region between about 400,000 bp and 600,000 bp on chromosome XIV with the superior parent 21A (QTL1) has shifted to a more upstream position in the mapping with the 21ADG downgraded superior parent (QTL4). In the region between 600,000 bp and 800,000 bp on chromosome XII, there is a new conspicuous peak, indicating linkage with the 21ADG superior parent (QTL5). We confirmed the statistical significance of these two new QTLs by scoring selected SNPs in the individual thermotolerant segregants and performing a binomial exact test (Table 2). For the remaining seven regions, the SNPs showed about 50% variant frequency in the unselected pool (Figure S5 online). This suggests that the putative weak linkage from these regions is not caused by allelic incompatibilities. In addition, the significant association of the causative element(s) in QTL3 with the inferior parent (71% of the segregants had the genotype of the inferior parent, as determined by individual segregant genotyping) observed in the first mapping was completely abolished in the second mapping (52% of the segregants had the genotype of the inferior parent), which reaffirms that PRP42 is the only causative gene in this locus.
In a previous QTL mapping study of thermotolerance [22], the authors identified the NCS2 allele of a clinical isolate as a superior allele compared to the inferior allele from the S288c control strain. Since NCS2 is located in the central region of QTL4 and since the NCS221A allele contains the same mutation (A212T) as identified in the previous study, we have tested whether NCS221A is also a causative allele in our genetic background. For that purpose, we performed RHA for NCS2 using a hybrid diploid strain constructed from the two downgraded parent strains. We found that the NCS221A allele supported higher thermotolerance compared to the NCS2BY4742 allele, indicating that also in our genetic background the NCS2 allele from the superior strain acted as a causative gene (which does not preclude the presence of other causative genes). Deletion of the inferior NCS2BY4742 allele in the hybrid diploid strain also caused a conspicuous drop in thermotolerance (Figure S6 online).
Fine-mapping of QTL5 by scoring six selected SNPs individually in all 58 thermotolerant segregants enabled us to reduce the size of the QTL from 150,000 bp to 40,000 bp (Figure 6A). We then divided this region into three fragments and performed bulk RHA with each fragment in the 21ADG/BY4742DG diploid strain (Figure 6A). (The fragments had an overlap of one gene.) Evaluation of thermotolerance with the pairs of reciprocally deleted hemizygous strains revealed that FRAGMENT121A and FRAGMENT221A conferred higher thermotolerance than the corresponding fragments from the inferior BY4742DG parent. For FRAGMENT3 there was no difference (Figure 6B). We then performed RHA with all individual genes of Fragments 1 and 2 containing non-synonymous mutations in their ORF (as indicated in Figure 6A). However, for none of the genes tested there was a different effect on thermotolerance of the two alleles (data not shown). We then applied RHA to the remaining genes in FRAGMENT2 and found that the SMD221A allele conferred higher thermotolerance compared to the SMD2BY4742 allele (Figure 6B). Hence, it apparently acted as a causative allele in both FRAGMENT1 and FRAGMENT2, since it was the only gene present in the overlap between the two fragments. The observation that replacement of FRAGMENT121A with FRAGMENT1BY4742 caused a similar reduction in thermotolerance compared to the replacement of FRAGMENT221A with FRAGMENT2BY4742 is consistent with SMD2 being the only causative gene in QTL5.
We confirmed by Sanger sequencing that SMD221A only displayed SNPs in the promoter and terminator region as compared to SMD2BY4742 (data not shown). Hence, a difference in expression level may be responsible for the difference in thermotolerance. We have compared SMD2 transcription levels in different strains and with incubation at different temperatures. We found a higher level of SMD2 expression for 21A compared to BY4742 in cells growing exponentially in liquid cultures (YPD at 30°C) and also 21ADG showed a higher level of SMD2 expression under these conditions than BY4742DG (Figure 6C). The difference in SMD2 expression level is also clear for the 21A/BY4742 RHA pairs, but there is no significant difference for the 21ADG/BY4742DG RHA pairs (Table S3 online). This indicates that the mechanism of SMD2 in influencing thermotolerance cannot be solely due to differences in its transcript level, and other mechanisms such as post-transcriptional regulation may play a role.
In the cross with the original parents, the QTL5 region did not show any indication of linkage to the genome of the superior parent strain 21A, with 37 out of 58 thermotolerant segregants of 21A/BY4742 having the SMD221A allele (confirmed by genotyping the individual segregants, data not shown). We have also applied RHA for SMD2 in the original 21A/BY4742 hybrid. Interestingly, we could not detect any difference in thermotolerance at the two temperatures tested (40.7°C and 41°C) (Figure 7A). Knowing that 21ADG/BY4742DG lack only two superior alleles as compared to 21A/BY4742 and both PRP42 and SMD2 encode proteins forming the same spliceosomal complex, we constructed double hetero-allelic mutations for PRP42 and SMD2 in the 21A/BY4742 background, and evaluated thermotolerance of the strains. In the hybrid with the inferior PRP42 allele, the superior SMD2 allele caused higher thermotolerance compared to the inferior SMD2 allele, whereas in the hybrids containing the superior PRP42 allele, the two SMD2 alleles did not influence thermotolerance differently (Figure 7B). The identification of SMD2 as a causative gene for thermotolerance indicates that our new approach of mapping with the downgraded parent strains is able to reveal minor loci and causative genes that escape detection in QTL mapping with the original parents, in this specific case because of epistatic interaction.
We expressed the two PRP42 alleles from a centromeric plasmid in the parent 21A strain (Figure S7A online) and in the 21A prp42Δ strain (Figure S7B online). In both cases, there was no difference in thermotolerance between the strains. On the other hand, comparison of the thermotolerance of strain 21A and that of the two heterozygous RHA strains showed that the RHA strain expressing the 21A allele had clearly lower thermotolerance than the other two strains (Figure S7C online). The thermotolerance of the heterozygous RHA strain expressing the superior PRP42 allele from BY4742 was not higher than that of the 21A strain. These results show that the BY4742 allele of PRP42 is not able to enhance the thermotolerance level of the 21A strain further, apparently indicating that other factors become limiting for thermotolerance. One such other factor may be SMD2. In the 21A strain it is present for 100% in the superior form, while in the heterozygous RHA strain, it is only present for 50% in the superior form. Hence, a dosage effect of SMD2 may possibly be limiting for thermotolerance in the heterozygous RHA strain expressing the superior PRP42 allele from BY4742. The difference in ploidy or in the genetic constitution between the haploid 21A strain and the diploid RHA hybrid strains may also play a role, although this seems to be contradicted by the fact that we mapped the superior PRP42 allele using haploid segregants of the superior and inferior parents. Also in the study of Sinha et al. [27], replacement of the inferior allele of MKT1 with the superior allele in the S288c strain did not cause the expected improvement in thermotolerance.
Identification of QTLs with minor effects on complex traits remains a difficult issue in quantitative genetics [29]. Major approaches used up to now have been fixing of major QTLs in a single parent and repeating the QTL mapping procedure either with backcrosses or regular crosses between the parents [22], [23], [30], the use of very high numbers of segregants [24], more stringent phenotyping to enhance the detectability of the minor QTLs [10] or genotyping and phenotyping single segregants [21]
In this study, we have extended the approach of fixing major QTLs to mapping by pooled-segregant whole-genome sequence analysis. In addition, we fixed a major QTL in each parent strain to create a downgraded superior and a downgraded inferior parent strain. The benefit of downgrading both parents, especially in pooled-segregant mapping, is that it keeps a large phenotypic difference between the parental strains. This makes the isolation of a sufficient number of segregants with extreme phenotype easier or at least makes the evaluation of their phenotype in comparison with that of the superior parent easier. In addition, it may enhance the chances that the minor QTLs identified are truly relevant for the phenotypic difference between the original parents and not generated in some unrelated manner in the mapping procedure. The advantages of fixing major QTLs in one parent for linkage mapping have recently been demonstrated [23]. Fixing major QTLs in both parents may have a similar advantage for linkage mapping as in pooled-segregant analysis, since it increases the potential of a wider range of phenotypic variation and thus also for more reliable selection of segregants with an extreme phenotype.
If enough phenotypic variability is obtained in the segregants, one could in principle map with parents that do not differ at all in the trait-of-interest. However, in this case one is mapping mutations in the background of the strains that affect the trait-of-interest. This is not the general purpose of our work. In our case, the goal was to identify the mutant alleles that are responsible for the elevated thermotolerance in the superior parent strain, so that these alleles after their identification could be transferred to other industrial yeast strains to enhance their thermotolerance. In principle, one could also do a second round of mapping with upgraded parents. However, we believe that mapping with downgraded parents has a higher chance of revealing additional minor QTLs because it eliminates epistatic interactions with the major QTLs and also because elimination of the major QTLs enhances requirement for the presence of minor QTLs if the screening of the phenotype is performed at a similar stringency.
Our approach is based on the observation that the causative genetic element(s) in some QTLs is(are) linked to the inferior rather than to the superior parent. This is likely due to the fact that genetic mapping in yeast is performed with haploid strains derived from natural or industrial diploid strains that generally harbor a single copy of many recessive alleles. As a result of the presence of negative, recessive mutations, positively acting QTLs and causative genes will be identified that are linked to the inferior rather than the superior parent. This has also been observed in several previous mapping studies [10], [18], [24]. It indicates that linkage of QTLs to the inferior parent is not an uncommon phenomenon and, moreover, may significantly increase when the influence of major QTLs is weakened or when genetic linkage in the genome is reduced.
Identification of the causative gene in QTL1, linked to the superior parent, and in QTL3, linked to the inferior parent, allowed us to construct both a downgraded, superior and a downgraded, inferior parent strain using targeted allele replacement. Repeating the genetic mapping with the downgraded parent strains successfully revealed new minor QTLs and thus established the effectiveness of this approach. Moreover, we validated the new QTLs 4 and 5 by identifying the causative genes. QTL4 contained a causative gene previously identified for high thermotolerance in another yeast background [22], further underscoring the effectiveness of this approach. Interestingly, our identification in the cross with the downgraded parent strains of new QTLs linked to both superior and inferior parent, allows in principle to construct further downgraded parent strains and repeat the mapping to identify additional minor QTLs with significant linkage.
An advantage of our approach is that it keeps all genetic information from both superior and inferior parents, whereas in backcrossing approaches, 50% of the genetic information of the superior parent is lost. As a result, minor QTLs may be missed. Furthermore, backcrossing creates regions that are identical between the new parents, i.e. the F1 segregant and the inferior parent, which makes it impossible to identify in the next cross QTLs linked to the inferior parent in these regions. Although the phenotype of the downgraded parent may not be as extreme as that of the F1 segregant normally selected for backcrossing, it has all the genetic diversity to generate segregants with a phenotype as extreme as obtained in the backcross.
Another advantage compared to backcrossing of repeating the QTL mapping after fixing causative genes in the parents is that it can reveal new minor QTLs and causative genes located closely to or even within the previously identified QTL. If the superior alleles that have been replaced in the downgraded parent strains were the only causative gene in their QTL, this QTL should disappear completely in the second cross. In our case, this happened with QTL3, for which there was no linkage anymore with the segregants of the downgraded parents. On the other hand, if other causative genes exist within the QTL in addition to the fixed gene, the QTL will likely remain present in the second mapping, allowing identification of the remaining causative gene(s). This happened in our case with QTL1, which shifted to a slightly more upstream position. In the new QTL, which was called QTL4, we could subsequently confirm NCS2 as the causative gene. The presence of multiple causative genes located close to each other within a single QTL has been found before [6], [10], [22]. To resolve closely located QTLs in the first cross an impractical number of F1 segregants is easily required [31]. Recently, multiple, random inbreeding with all F1 segregants was used to enhance recombination between the genomes of the parents and thus reduce linkage in the genome. This resulted in a higher resolution of genetic mapping, facilitating detection of closely located minor QTLs and also strongly reduced the number of candidate genes in the centre of the QTL [24].
The appearance of new minor QTLs in the second mapping, with QTL5 and its causative gene SMD2 as a striking example, raises the question why these QTLs were not detected in the first mapping. One plausible explanation is interaction between causative genes from different QTLs, which has been identified by Lorenz et al. [23]. In our study we identified a negative interaction between the SMD2 and PRP42 alleles, which can explain the absence of QTL5 in the first mapping. In the latter, the presence of the superior PRP42 allele in the selected thermotolerant segregants could compensate for the presence of an inferior SMD2 allele. In the second mapping, after removal of the superior PRP42 allele, the effect of the superior SMD2 allele now apparently became more significant, causing a higher chance for this allele to be present in the thermotolerant segregants.
Thermotolerance of growth, which is the ability to grow at elevated temperatures, has been a favourite trait in quantitative genetics with yeast [6], [22], [24], [27], [32], [33]. It is easily scored on solid nutrient plates, it is highly relevant for several industrial applications with yeast and is a typical characteristic of clinical isolates of S. cerevisiae. To date, several genes have been identified in natural yeast strains with an allele-specific contribution to thermotolerance. The QTLs identified in our study did not overlap with the regions in which these genes are located, except for QTL1 (MKT1) and QTL4 (NCS2). The diverse biological functions of these genes underscores our limited understanding of this phenotype, since apparently none of these genes has a function that can be directly linked in a known mechanistic manner to sustaining high thermotolerance.
In this study, we have identified PRP42BY4742 and SMD221A as two novel and naturally-occurring superior alleles for high thermotolerance. Yeast Prp42 was identified as an essential protein for U1 small nuclear ribonucleoprotein (snRNP) biogenesis, which has a high similarity to Prp39 [34]. SMD2 encodes a core protein Sm D2 that is part of the spliceosomal U1, U2, U4, and U5 snRNPs [35]. These snRNPs function in pre-mRNA splicing by recognizing short conserved sequences from 5′ to 3′ at the exon-intron junctions and assemble into active spliceosomes [36]. Interestingly, the related function of these two genes suggests an important role for RNA processing in growth at high temperature. Further analysis revealed an allele-specific interaction between PRP42 and SMD2. This is consistent with the previous evidence for direct interaction between the human homologues of these gene products as revealed by crystal structure determination of human spliceosomal U1 snRNP [37].
The MKT1 gene has been found as a causative gene in several QTL mapping studies with various phenotypes and using diverse genetic backgrounds, but always with the S288c/BY background for the control parent [10], [15], [30], [38]. Mkt1 appears to control gene expression at a post-transcriptional step [39], which may explain why its deficiency produces effects on such a diversity of phenotypes.
To allow faster identification of causative genes in the mapped QTLs, we have applied bulk RHA, which evaluates multiple adjacent genes simultaneously. The successful identification of causative genes (PRP42 and SMD2) using this approach confirms the effectiveness of this method. A possible advantage of this strategy over RHA with single alleles is that it can take into account genetic interactions [19] between the genes in the deleted region. If two closely located genes can compensate for each other, bulk RHA may detect their effect as opposed to single gene RHA. Another advantage of bulk RHA is its high efficiency, especially in cases where QTLs cannot be reduced to a small size with only few genes in the centre because of a limited number of segregants available for fine-mapping. In general, this will be the case with phenotypes that require a high workload for scoring. In our experience, with bulk RHA one can easily evaluate a region with a size of 20 kb, which encompasses on average between 6 and 12 genes in yeast. On the other hand, bulk RHA carries possible pitfalls. When a region used for bulk deletion carries both positively acting and negatively acting genes, as was found in previous studies [6], [10], simultaneous deletion of both can result in the absence of any phenotypic effect. Hence, a negative result with bulk RHA does not necessarily imply the absence of causative genes.
In this paper we have shown that identification of new minor QTLs involved in complex traits can be successfully accomplished by crossing parent strains that have both been downgraded for a single QTL. Using this approach we have identified new QTLs and new causative genes, revealing an important role for RNA processing in high thermotolerance. This method has the advantage of maintaining all relevant genetic diversity and enough phenotypic difference between the two parent strains and thus significantly increases the chances of identifying minor QTLs. In principle, successive rounds of minor QTL mapping could be performed in this way by sequentially downgrading the two parent strains further, making use each time of a causative gene identified in a QTL linked to the superior parent and in a QTL linked to the inferior parent.
The following yeast strains were used: prototrophic and heterothallic diploid strain MUCL28177, which was isolated from orange juice in the region of Strombeek-Bever, Belgium, its haploid segregant MUCL28177-21A, referred to as 21A, and BY4742 (Matα his3Δ1 leu2Δ0 ura3Δ0 lys2Δ0) [40]. Yeast cells were grown in YPD medium containing 1% (w/v) yeast extract, 2% (w/v) bacteriological peptone, and 2% (w/v) glucose. 1.5% (w/v) Bacto agar was used to make solid nutrient plates. Transformants were grown on YPD agar plates containing 200 µg/ml geneticin. Mating, sporulation and isolation of haploid segregants were done using standard protocols [41].
Strains were inoculated in liquid YPD and grown in a shaking incubator at 30°C overnight. The next day the cells were transferred to fresh liquid YPD at an OD600 of 1 and grown for 2 to 4 h to enter exponential phase. The cell cultures were then diluted to an OD600 of 0.5 and 5 µl of a fourfold dilution range was spotted on YPD agar plates, which were incubated at different temperatures. Growth was scored after two days incubation for all conditions. All spot tests were repeated at least once, starting with freshly inoculated cultures. Repetitions of the thermotolerance assays may show slight differences in growth intensity. Hence, the strains to be tested were always spotted together with the relevant controls on the same plate.
SNPs were scored in individual segregants by PCR. At a given chromosomal location, two SNPs spacing between 500 and 1,500 bp were chosen for the design of specific primers. For a given SNP, two primers either in the forward or reverse direction, were designed with one mismatch at their 3′ ends. First, a gradient PCR was applied using genomic samples of 21A and BY4742 as templates, with each template tested with two primer combinations (primer pair based on the sequence of BY4742 and primer pair based on the sequence of 21A). The annealing temperature at which the best distinguishing power was obtained with the two parents was used for scoring of the SNPs in the individual segregants.
All the ORFs of non-essential genes in the centre of the QTL were deleted separately in both 21A and BY4742. PCR-mediated gene disruption was used [46]. Plasmid pFA6a was used as a template to amplify a linear DNA fragment containing the kanMX4 cassette [47], with 50 bp homologous sequences for the target regions at both ends. Transformants growing on YPD geneticin plates were verified by PCR with several combinations of internal and external primers. The verified haploid deletion strains were subsequently crossed with the matching wild type haploid to generate the hybrid diploids. For RHA with essential genes and fragments containing multiple genes, transformation was performed directly in the hybrid diploid. External SNPs primer pairs together with internal primers within the kanMX4 cassette were used in different combinations to determine in which parent the allele or the fragment had been deleted. For each heterozygous deletion hybrid, at least two isogenic strains were made and evaluated for thermotolerance. The growth of strains in the RHA test should always be compared within the strain pairs and not between the strain pairs, since the loss of one copy of a gene can cause an effect on the growth of the strains under non-restrictive conditions or even under restrictive conditions if the gene is important for the phenotype and because of the variability between different thermotolerance assays..
The replacement of MKT121A with MKT1BY4742 in 21A was performed by a two step transformation. For the first transformation, a linear DNA fragment with the AMD1 gene from Zygosaccharomyces rouxii flanked by 50 bp sequences that are homologous to the two sides of the MKT1 ORF was amplified from plasmid pFA6a-AMD1-MX6 [48] by PCR, and transformed into 21A. Transformants were grown on YCB (Yeast Carbon Base 1.17%, phosphate buffer 3%, Bacto agar 2%) plates containing 10 mM acetamide. Single colonies were checked for the correct replacement with the use of external and internal primers. For the second transformation, colonies were transformed with a linear DNA fragment containing the MKT1BY4742 ORF, together with ∼100 bp downstream and upstream. Transformants were grown on YNB galactose (0.17 Yeast Nitrogen Base w/o amino acids and ammonium sulfate, 1.5% Difco agar, 0.01% galactose, pH 6.5) containing 100 mM fluoroacetamide. Colonies were first checked for the presence of MKT1 by PCR, and then confirmed by DNA sequencing.
The replacement of PRP42BY4742 with PRP4221A in BY4742 was performed in a two step transformation. For the first transformation, a URA3 gene was inserted ∼50 bp downstream of the PRP42 ORF in BY4742. Colonies growing on –URA plates were confirmed to have a correct insertion by PCR. For the second transformation, a linear DNA fragment containing the ORF of PRP4221A together with ∼400 bp downstream and upstream was transformed into the previous colonies, and the transformants were grown on 5-FOA plates. Colonies were first checked for the right DNA polymorphism by SNP primer pairs, and then confirmed by DNA sequencing.
All sequence data have been deposited in the Sequence Read Archive (SRA) at the National Center for Biotechnology Information (NCBI) and can be accessed with account number SRA058979.
|
10.1371/journal.pntd.0002847 | In Vitro and In Vivo Trypanocidal Activity of H2bdtc-Loaded Solid Lipid Nanoparticles | The parasite Trypanosoma cruzi causes Chagas disease, which remains a serious public health concern and continues to victimize thousands of people, primarily in the poorest regions of Latin America. In the search for new therapeutic drugs against T. cruzi, here we have evaluated both the in vitro and the in vivo activity of 5-hydroxy-3-methyl-5-phenyl-pyrazoline-1-(S-benzyl dithiocarbazate) (H2bdtc) as a free compound or encapsulated into solid lipid nanoparticles (SLN); we compared the results with those achieved by using the currently employed drug, benznidazole. H2bdtc encapsulated into solid lipid nanoparticles (a) effectively reduced parasitemia in mice at concentrations 100 times lower than that normally employed for benznidazole (clinically applied at a concentration of 400 µmol kg−1 day−1); (b) diminished inflammation and lesions of the liver and heart; and (c) resulted in 100% survival of mice infected with T. cruzi. Therefore, H2bdtc is a potent trypanocidal agent.
| The protozoan parasite Trypanosoma cruzi causes Chagas disease, a condition that affects the poorest regions of Latin America mainly. The chronic phase of this disease disables thousands of patients, constituting an important public health issue. The pharmacotherapy that is currently applied to treat the disease emerged many decades ago, is ineffective in most patients, mainly during the chronic phase, and has serious side effects. In a recent study, we showed that the compound 5-hydroxy-3-methyl-5-phenyl-pyrazoline-1-(S-benzyldithiocarbazate) (H2bdtc) is a potential drug candidate against the in vitro trypomastigote form of Tulahuen strains of T. cruzi. Here we report that H2bdtc loaded into solid lipid nanoparticles (H2bdtc-SLNs) displays good trypanocidal activity against the trypomastigote form of the Y strain of T. cruzi both in vitro and in vivo. Our in vivo experiments revealed that H2bdtc-SLN is 100 times more active than benznidazole (BZN), the drug that is commercially available to treat Chagas disease. Surprisingly, this compound has no side effects on the T. cruzi acute phase. Hence, we propose that H2bdtc-SLNs possesses interesting anti-Trypanosoma properties.
| T. cruzi parasites are transmitted by insect vectors (triatomine bugs). T.cruzi is the causative agent of Chagas disease, which is silent and can remain asymptomatic for years [1], [2], [3]. A century after its discovery, this disease remains a serious public health issue—it is closely associated with human poverty and political instability as well as with little investment in drug development. According to the World Health Organization (WHO), between seven and eight million people are infected with T. cruzi worldwide, primarily in Latin America [4], [5], [6]. One in every four Chagas patients develops a fatal symptom of the disease due to lack of adequate diagnosis and treatment.
Nifurtimox and benznidazole (BZN) are currently available to treat the disease [7], [8], [9], [10]. However, neurological side effects have led commercial nifurtimox production to be discontinued [11]. As for BZN, although it is mainly effective during the acute phase of the infection, it presents undesirable side effects such as rash and gastrointestinal symptoms [12], so patients often fail to comply with the treatment [8]. Long treatment periods (30, 60, or 90 days) and appropriate pediatric formulations not available (administration of the medication to children often requires tablet fractionation) also limit BZN use [9], [11], [13]. A further concern is that no effective treatment for the symptomatic chronic phase of Chagas disease exists, so the patients usually receive palliative drugs at this stage [14], [15]. Therefore, a number of researchers are making considerable efforts to find new drugs to combat this disease.
Dithiocarbazates display notable biological and pharmacological properties, including anticancer [16], [17], antimicrobial [17], [18], [19], and insecticidal activities [20]. A recent study has shown that cyclic compounds derived from S-dithiocarbazate and 1,3-diketones exhibit significant trypanocidal activity [21]: in particular, 5-hydroxy-3-methyl-5-phenyl-pyrazoline-1-(S-benzyldithiocarbazate) (previously referred to as H2L2a [21]) which was renamed H2bdtc in this reference in this work (Figure 1) constitutes a potential drug lead to develop a new agent against the trypomastigote form of Tulahuen strains of T. cruzi [21]. Nevertheless, the lipophilic character of H2bdtc may limit its administration and result in low oral bioavailability [22].
Drug delivery systems can help to circumvent this problem. Because lipids have excellent physiological acceptability and can promote drug absorption as well as selective lymphatic uptake, researchers have focused on lipid-based drug release systems [23]. In particular, solid lipids constitute solid lipid nanoparticles (SLNs) at room and body temperature. Since SLNs consist of biocompatible and biodegradable lipids with low or no human toxicity, they can function as drug delivery systems [24], [25]. SLNs offer many advantages: they protect the drug against degradation, enable controlled drug release, and dismiss the use of organic solvents. Moreover, SLNs can be produced on a large scale, meeting industrial requirements [24], [26].
Our group has used H2bdtc in vitro experiments involving the Tulahuen strain (group I) [21]. The resistances of this group and of the Y strain (group II) have been reported to be different, based on phosphatase activities in T. cruzi homogenates [27]. Tulahuen had an optimum phosphatase activity at pH 4.0 and the Y strain at pH 7.0 [27]. Also in chronic phase has been associated with T. cruzi II-restricted infections [28]. In this sense, evaluating the trypanocidal activity of H2bdtc against the Y strain could support the use of this compound as a new drug against T. cruzi. In addition, so far little attention has been paid to the use of SLNs to treat Chagas disease [29], [30]. Therefore, this work investigates the in vitro and in vivo trypanocidal activity of free H2bdtc and H2bdtc encapsulated into SLNs (H2bdtc-SLNs) against the Y strain of T. cruzi and compares results with data obtained for the currently available drug BZN.
BZN, a product manufactured by Lafepe, Brazil, was used as a reference drug. The synthesis of H2bdtc has been described previously [21]. RPMI Medium 1640 supplemented with 5% bovine fetal serum (GIBCO, Grand Island, NY, USA), 100 IU mL−1 penicillin G, and 100 mg mL−1 streptomycin (Gibco-BRL, Grand Island, NY, USA) was employed. Dimethyl sulfoxide (DMSO) and propidium iodide (PI) were obtained from Sigma-Aldrich Chemicals Co. (St. Louis, MO, USA). Stearic acid and sodium taurodeoxycholate were purchased from Sigma-Aldrich (St. Louis, MO, USA), Lipoid S 100 (soya lecithin) was acquired from Lipoid (Ludwigshafen, KOLN, Germany), and Amicon Ultra 15, MWCO 100 K, was provided by Millipore (Billerica, MA, USA).
SLNs were prepared using a microemulsion method [31]. Briefly, the desired amount of sodium taurodeoxycholate (0.12% w/v) was dissolved in hot aqueous phase, which was added to melted stearic acid (0.95% w/v) containing soya lecithin (0.48% w/v) and H2bdtc (0.02% w/v). The mixture was emulsified via magnetic stirring at 90.0±2.0°C, until a thermodynamically stable microemulsion formed. The SLNs dispersion was obtained by cooling the hot microemulsion in cold water (2–5°C) under vigorous stirring at 20,000 rpm for 10 min (IKA-T25 Ultra-turrax, Germany) at 1∶20 ratio (microemulsion/cold water). Next, the SLNs aqueous dispersion was subjected to high-pressure homogenization (EmulsiFlex – C3, Germany) at 500 bars for 10 min.
The particle size and dispersity of the H2bdtc-SLNs dispersion were measured via photon correlation spectroscopy (PCS) [32]; the zeta potential was determined on the basis of the electrophoresis mobility of the nanoparticles using the Zetasizer ZS Nano 90 (Malvern Instruments, UK.). The samples were diluted (1∶10) with distilled water at 25.0°C.
The morphology of H2bdtc-SLNs was assessed using an atomic force microscope (ICON Bruker, USA). The samples were prepared by immersing freshly cleaved mica (Muscovite Mica Substrates Sheets, SPI Supplies, China) in SLNs aqueous dispersion and stored overnight, at room temperature, to complete the drying process. The samples were evaluated by AFM in the intermittent contact mode (tapping mode) by scanning the surface of mica (2 µm×2 µm in area) using a rectangular silicon cantilever with a spring constant of 40 N m−1 vibrating at a frequency of 320 kHz. Imaging was performed at room temperature, and the topology image was used to determine the morphology of H2bdtc-SLNs [33].
The total H2bdtc content in the H2bdtc-SLNs was determined by UV-vis spectroscopy at 400 nm (UV Spectrophotometer UV 1800, Shimadzu, Japan). First, a defined amount of H2bdtc-SLNs was dissolved in dimethyl sulfoxide. The amount of encapsulated drug was indirectly measured after centrifuging the H2bdtc-loaded SLNs for 40 min at 6000 rpm (1605 G), at 25°C in a centrifuge (Heraeus Megafuge 16 R Thermo Scientific, USA) equipped with a membrane concentrator (Amicon Ultra 15, MWCO 100 K, Millipore Corporation, USA). The filtrate was diluted with dimethyl sulfoxide (1∶1), and the concentration of free H2bdtc in the diluted filtrate was determined using the same conditions employed to measure the total H2bdtc content used during the loading procedure (section 2.2). The amount of H2bdtc loaded into SLNs was calculated by subtracting the amount of free H2bdtc in the filtrate from the total amount of H2bdtc used during loading (26). EE (%) was determined using the following equation [26], [34].
Partition coefficients for the H2bdtc were determined in triplicate in an n- octanol/water system following a published procedure [35]. Measurements of H2bdtc n-octanol/water partition coefficients were carried out using the shake-flask method. H2bdtc was dissolved in aqueous solution previously saturated with n-octanol at a concentration of 1 mg/mL and mixed with the same volume of octanol also previously saturated with water. Samples were stirred for 30 min, separate in two phases, and centrifugated for 10 min at 2000 rpm. The amount of H2bdtc in the aqueous phase was quantified by UV-visible spectroscopy.
Female Swiss mice (6 to 8 weeks old) were bred and maintained at the Department of Biochemistry and Immunology, School of Medicine of Ribeirao Preto, University of São Paulo, Ribeirão Preto, Brazil. The mice were maintained in microisolator cages under standard conditions; they were fed with food and water ad libitum.
All the in vivo procedures were performed in accordance with the guidelines issued by the Brazilian College of Animal experimentation (COBEA) and received prior approval by the Ethics Committee on Animal Experimentation – CETEA (n° 006/2011) of the School of Medicine of Ribeirão Preto.”
All the experiments were conducted using the trypomastigote form of the Y strain of T. cruzi (Lineage type II). For the in vitro experiments, parasites were grown in a fibroblast cell line (LLC-MK2). For the in vivo experiments, mice were intraperitoneally inoculated with 2.0×103 bloodstream trypomastigote forms, which had been derived from previously infected Swiss mice.
The trypanocidal activities of free H2bdtc, H2bdtc-SLNs, and BZN against the trypomastigote form of the T. cruzi Y strains were evaluated as described previously [36]. To this end, the trypomastigote culture at a concentration of 6.5×106 parasites mL−1 was re-suspended in RPMI 1640 medium with 5% FBS. Triplicate cultures were treated with one of the investigated drugs and maintained at 37.0±0.1°C in a humidified atmosphere of 5% CO2. To test parasite viability, the number of motile forms was determined using a previously described method [37]. The concentration of compound corresponding to 50% trypanocidal activity after 24 h of incubation was expressed as the IC50try (inhibitory concentration for the trypomastigote form).
Spleen cells isolated from C57BL/6 mice, macerated in RPMI 1640 medium (Gibco), and filtered using a 100-µm pore filter were used to evaluate the cytotoxicity in vitro. The isolated cells were centrifuged at 1500 rpm for 10 min, and erythrocytes were lysed in lysis buffer for 5 min, at room temperature. Cells were washed, counted, and resuspended at 6.5×106 mL−1 in RPMI medium containing 5% fetal bovine serum. The spleen cells were seeded to a 96-well microplate (n = 2) and incubated for 24 h with H2bdtc diluted in dimethyl sulfoxide (DMSO, final H2bdtc concentration not exceeding 0.5%) or H2bdtc-SLNs (concentrations ranging from 125 µM to 0.24 µM in serial dilutions). BZN (Roche) was used as the reference drug; Tween 20 was employed as positive control for cell death. After the incubation period, the cells were washed and incubated with propidium iodide at a final concentration of 10 µg mL−1. Cell cytotoxicity was measured on a flow cytometer (FACSCantoII - BD), and the data were analyzed using the FlowJo program (Tree Star).
Female Swiss mice aged between 6 and 8 weeks, weighing between 20 and 25 g, were infected with 2.0×103 blood trypomastigotes per animal. A total of four experimental groups consisting of seven Swiss mice each were included in the study. Treatment started at day 5 post-inoculation (p.i.). BZN, free H2bdtc and H2bdtc-SLNs were orally administered at 4 µmol kg−1 (BZN 1.0 mg kg−1 day-1/free H2bdtc and H2bdtc-SLNs 1.4 mg kg−1 day−1) per day for 10 consecutive days. The following treatments were applied: Group 1 = PBS control group; infected and not treated, Group 2 = infected and treated with BZN, Group 3 = infected and treated with free H2bdtc, and Group 4 = infected and treated with H2bdtc-SLNs. To evaluate parasitaemia and mortality, seven animals from each group were used. Seven animals were killed at day 22 p.i. (early mortality), to quantify inflammation of the heart and liver and to measure creatine kinase-MB (CK-MB) and glutamic-pyruvic transaminase (GPT) production.
Parasitemia was analyzed on alternate days from day 7 p.i.; to this end, 5 µL of fresh blood was collected from the animal tail. The count of 100 fields was performed via direct observation under a light microscope [38]. Mortality was inspected on a daily basis until day 60.
Groups of seven mice were euthanized at day 20 p.i., and portions of the heart and liver were fixed in paraffin for histological analysis. To assess inflammatory infiltration via light microscopy DP71 (Olympus Optical Co, Japan), tissues were sectioned at a 5-µm thickness and stained with hematoxylin-eosin (H&E). Each tissue section was imaged 25 times, and the percentage of the area occupied by cellular infiltrates was determined using the Image J program.
Quantitative PCR was used to determine the amount of parasitic DNA in heart tissues. Briefly, DNA was purified from 25 mg of heart tissue using a QIAamp DNA Mini Kit (Qiagen), according to the manufacturer's instructions. Each PCR reaction comprised 40 ng of genomic DNA; 0.3 µM of the T. cruzi-specific primers TCZ-F 5′-GCTCTTGCCCACAMGGGTGC-3′ (M = A or C) TCZ-R 5′-CCAAGCAGCGGATAGTTCAGG-3′ [39], which amplify a 182-bp product; 7.65 µL of GoTaq qPCR Master Mix; and H2O (final total volume of 15 µL).
The reactions were performed using the Real-Time PCR System. The cycling program involved a denaturation cycle of 95.0°C for 10 min, followed by 40 cycles of the three steps of the amplification phase: 95.0°C for 15 s, 55.0°C for 30 s, and 72.0°C for 15 s. The melting phase was performed at 95.0°C for 15 s and at 60.0°C for 1 min, followed by a 0.3°C ramp and then 95.0°C for 15 s. During the melting phase, the acquisition setting was set at step. The data were analyzed with StepOne Software version 2.2.2.
The cardiac and hepatic lesions of mice infected with T. cruzi, treated or not, were assessed by measuring the creatine kinase-MB (CK-MB) and glutamic-pyruvic transaminase (GPT) levels, respectively, in the serum at day 22 p.i. The CK-MB levels were measured using a CK-MB kit (Liquiform, Brazil), as previously described [40]. Absorbance was measured on a microplate spectrophotometer (EMAX Molecular Devices Corporation, California, EUA). The color produced from this reaction was measured at a wavelength of 340 nm; the results are expressed in U/I. GPT was analyzed using an ALT/GPT kit (Liquiform, Brazil), according to the manufacturer's instructions. The colorimetric assay determines the amount of pyruvate produced according to the Reitman and Frankel method, from the formation of 2,4-dinitrophenylhydrazine [41]. The color produced by this reaction was measured at a wavelength of 505 nm.
Data are expressed as the mean SEM. Student's t-test was used to analyze the statistical significance of the variation between the infected and control assays. Differences were considered statistically significant when P<0.05. The differences in droplet size, dispersity, zeta potential, and entrapment efficiency values achieved during the stability test were evaluated via a one-way ANOVA analysis of variance followed by Tukey post-test analysis. The differences were considered statistically significant when P<0.05. All the analyses were performed using PRISM 5.0 software (Graph Pad, San Diego, CA, US).
The H2bdtc showed a lipophilic character (Log P (o/w) = 2.69±0.03) and were efficiently encapsulated in this manner in SLNs. On the basis of Photon Correlation Spectroscopy (PCS), H2bdtc-SLNs had diameter of 127.4±10.2 nm and dispersity lower than 0.3; the zeta potential revealed a negative surface charge (−56.1±4.4 mV) (Table 1). The entrapment efficiency was 98.16±1.12, showing that the drug dispersed well within the lipid matrix. Atomic force microscopy images revealed that H2bdtc-SLNs particles were spherical, with an average diameter of approximately 180 nm (Figure 2), agreeing with the PCS results.
We assessed the in vitro trypanocidal activity of free H2bdtc, H2bdtc-SLNs and BZN after 24 h of incubation with T. cruzi trypomastigotes forms. Free H2bdtc presents IC50try (inhibitory concentrations against bloodstream trypomastigote) as 0.50±0.12, H2bdtc-SLNs as 1.83±0.18 and BZN 0.50±0.39 µM (Figure 3A). We also measured the cytotoxicity of free H2bdtc and H2bdtc-SLNs in spleen cells of Swiss mice; none of the tested drugs was significantly cytotoxic (Figure 3B). Hence, both free H2bdtc and H2bdtc-SLNs displayed similar in vitro trypanocidal activity to BZN; this activity was not associated with general cytotoxicity but rather with specific activity against the parasite.
We performed in vivo experiments to investigate the controlled release behavior of H2bdtc from H2bdtc-SLNs; we also compared the activities of free H2bdtc and H2bdtc-SLNs against T. cruzi. We decided to use an H2bdtc-SLNs concentration of 4 µmol kg−1 day−1. During the in vivo treatments on the basis of preliminary in vivo results obtained for the Y strain of T. cruzi, which revealed that H2bdtc had low level of parasitemia (Supporting Information: Figure S1). In all the infected groups, parasitemia peaked at day 9 p.i., with gradual parasite elimination from the bloodstream after day 11 p.i. It is worth noting that we employed BZN concentrations 100 times lower than that used for Chagas patients. H2bdtc-SLNs eliminated 70% of the circulating parasites at the peak of infection, whereas free H2bdtc and the positive control BZN eliminated 48 and 15% of the parasites, respectively, as compared with the control group treated with PBS (Figure 4A). In agreement with the data revealing reduced parasitemia, mice treated with H2bdtc-SLNs presented 100% survival rate (Figure 4B), similar to the result achieved with BZN administered at a clinical dose of 400 µmol kg−1 day−1 (100 times more concentrated than the concentration used herein). Compared with the control group (PBS), groups treated with free H2bdtc and BZN exhibited a survival rate of 57%.
Encouraged by the in vivo results, we evaluated how free H2bdtc, H2bdtc-SLNs, and BZN affected the cardiac and hepatic tissues of the surviving animals. Infected mice treated with H2bdtc-SLNs presented reduced cardiac inflammation (Figure 5A) and heart lesions were absent (Figure 5B), as established by the absence of CK-MB, the enzyme released into plasma during cardiac lesion. Treatment with free H2bdtc diminished cardiac damage by 50% as compared with therapies with BZN or PBS. Concerning the ability of the tested compounds to reduce the liver damage caused by the parasite, H2bdtc-SLNs decreased inflammatory infiltration in the liver and hepatic toxicity more effectively, as assessed by measuring the glutamic-pyruvic transaminase (GTP) levels in the serum (Figure 5C, 5D). Considering all these results, it is possible to infer that treatment with H2bdtc per se reduced exacerbation of the inflammatory response on T. cruzi target organs and, consequently, tissue damage. Loading of H2bdtc into nanoparticles afforded even better results, producing no lesion in the heart tissue.
Because it is well established that parasites play an important role in cardiac damage during T. cruzi infection [42], [43], we quantified T. cruzi DNA derived from the heart tissues of mice treated with the tested compounds via real-time PCR. Treatment with H2bdtc-SLNs reduced the parasite burden significantly more effectively as compared with the other tested drugs (Figure 6), indicating that killing the parasites is most likely the mechanism through which H2bdtc-SLNs acts to diminish tissue lesions and enhance mice survival.
Chagas disease has often been pointed out as being a major neglected disease; the drugs that are currently available to treat this disease are little effective [5], [12]. Efforts have been made to provide the affected populations with new compounds to treat the disease. In the past few years, researchers have tested many substances against T. cruzi. In particular, H2bdtc, which belongs to the class of S-dithiocarbazates, is efficient against the parasite [21]. A 24-h UV-vis study into the stability of H2bdtc in aqueous solution did not evidence any changes in the spectrum of this compound. Nevertheless, this drug is poorly soluble in water (1.50×10−6 M), which has limited its use to treat Chagas disease. Because H2bdtc is lipophilic (Log P(o/w) = 2,69±0,03) and SLNs constitute effective oral drug delivery systems, we loaded H2bdtc into this type of lipid.
We prepared the SLNs by the microemulsion method [31], [44], to avoid the use of organic solvents. The resulting SLNs had diameter of approximately 120 nm, dispersity lower than 0.3, and spherical shape, which made these lipids suitable for oral administration [45], [46], [47], [48]. The zeta potential measurement allowed us to predict the stability of the colloidal dispersion. Charged particles have high zeta potential–negative or positive–and usually do not aggregate [49]. The zeta potential results revealed that the SNPs prepared here had a negative surface charge (−56.1±4.4 mV), indicating that the system was physically stable. Loaded and unloaded SNPs had similar zeta potentials, attesting that the tested drug was completely and uniformly dispersed inside the lipid matrix [50].
Encapsulation did not change the in vitro trypanocidal activity of H2bdtc, which was higher than the activity of BZN at the same concentration used here. The IC50 values obtained for H2bdtc against the trypomastigote form of the Y strain of T. cruzi were comparable with or superior to those of previously reported active compounds [51], [52]. Other papers have also described the use of colloidal drug carriers such as liposomes and nanoparticles to treat Chagas diseases [53], [54]. Treatment of T. cruzi infection with a BZN-loaded liposome increased BZN levels in the liver and blood. Intravenous administration of free BZN and BZN-encapsulated liposome at 0.2 mg of BZN per kilogram of body weight revealed three-fold higher BZN accumulation in the liver in the second case. Nevertheless, encapsulation failed to improve the in vivo BZN efficacy [55].
Liposome instability prevents their use as drug delivery systems [56]. Fortunately, we verified that free H2bdtc and H2bdtc-SLNs were not toxic to the spleen cells of Swiss mice, which encouraged us to directly test the effect of H2bdtc formulations in vivo using a murine model of acute Chagas disease.
Treatment started with a relatively low oral dose of free H2bdtc and H2bdtc-SLNs (4 µmol kg−1 day−1) as compared with currently employed doses of the commercially available BZN and compounds tested in the literature [57], [58]. H2bdtc-SLNs, Free H2bdtc and BZN reduced the presence of parasites in the blood of infected mice in 70, 48 and 15% respectively. H2bdtc-SLNs maintained 100% survival rate of infected mice, whereas 43% of the mice treated with free H2bdtc or BZN at the same concentration succumbed to the disease. It is noteworthy that free SLN and PBS elicited similar levels of parasitemia (Supporting Information: Figure S2). Therefore, the use of SLNs as a drug delivery system increased the oral bioavailability of the target drug, as previously described [59], [60], [61], [62], [63]. H2bdtc loading into SLNs overcame the problems inherent to the poor water solubility of the compound and may be could make it more accessible to the parasite (however, detailed pharmacokinetic data will be presented in a separate forthcoming paper). Additionally, some authors have proposed that drugs loaded into SLNs measuring 20–500 nm are absorbed by lymphatic transport, which reduces the first-pass metabolism [62], [63].
Analysis of histological sections of liver and heart tissues (Supporting Information: Figure: Figure S3 and Figure S4) revealed that the inflammatory infiltrate decreased in all the treated groups as compared with the control. The reduction was more pronounced in mice treated with H2bdtc-SLNs, possibly because parasitemia was lower in this case. This corroborated with findings from previous studies [64], [65] and confirmed that the parasite elicited intense inflammation especially in the cardiac tissues. The fact that the heart tissues of mice treated with H2bdtc-SLNs were perfectly preserved agreed with the notion that the presence of inflammatory infiltrates is associated with cardiac tissue damage [66], [67] and also with parasitic load [42], [43]. Indeed, mice treated with H2bdtc-SLNs exhibited significantly lower parasite burden as compared with the other groups. Hence, the reduced parasitism elicited by H2bdtc-SLNs helps to preserve the heart tissues of mice infected with T. cruzi, allowing us to conclude that H2bdtc is a potent trypanocidal agent.
Investigation into how H2bdtc interacts with possible targets represents a theme for future studies. For the time being, we must bear in mind that triazoles and thiosemicarbazones are well known for inhibiting cruzain, a protein belonging to the family of cysteine proteases and which is the most abundant protein in T. cruzi. Cruzain is essential for parasite development and survival within host cells [68]. H2bdtc bears pyrazole and dithiocarbazate parts, which are similar to triazoles and thiosemicarbazones, respectively, and could account for its trypanocidal action.
A mechanism of action similar to that of BZN probably does not occur. The BZN mode of action involves intracellular reduction of the nitro group, to produce highly reactive free radicals and/or electrophilic metabolites that could affect other systems, especially host systems, contributing to the cytotoxic effects observed in BZN-treated patients [69].
It is worth noting that cysteine proteases are very important for parasites; however, the lack of redundancy with respect to their mammalian hosts makes these proteases interesting targets for the development of new therapeutic agents [70]. Altogether, our findings show that H2bdtc-SLNs are a possible drug candidate to treat Chagas disease: it is more efficient against T. cruzi than the drugs used in current therapies.
|
10.1371/journal.pgen.1004650 | Topoisomerase II Is Required for the Proper Separation of Heterochromatic Regions during Drosophila melanogaster Female Meiosis | Heterochromatic homology ensures the segregation of achiasmate chromosomes during meiosis I in Drosophila melanogaster females, perhaps as a consequence of the heterochromatic threads that connect achiasmate homologs during prometaphase I. Here, we ask how these threads, and other possible heterochromatic entanglements, are resolved prior to anaphase I. We show that the knockdown of Topoisomerase II (Top2) by RNAi in the later stages of meiosis results in a specific defect in the separation of heterochromatic regions after spindle assembly. In Top2 RNAi-expressing oocytes, heterochromatic regions of both achiasmate and chiasmate chromosomes often failed to separate during prometaphase I and metaphase I. Heterochromatic regions were stretched into long, abnormal projections with centromeres localizing near the tips of the projections in some oocytes. Despite these anomalies, we observed bipolar spindles in most Top2 RNAi-expressing oocytes, although the obligately achiasmate 4th chromosomes exhibited a near complete failure to move toward the spindle poles during prometaphase I. Both achiasmate and chiasmate chromosomes displayed defects in biorientation. Given that euchromatic regions separate much earlier in prophase, no defects were expected or observed in the ability of euchromatic regions to separate during late prophase upon knockdown of Top2 at mid-prophase. Finally, embryos from Top2 RNAi-expressing females frequently failed to initiate mitotic divisions. These data suggest both that Topoisomerase II is involved in the resolution of heterochromatic DNA entanglements during meiosis I and that these entanglements must be resolved in order to complete meiosis.
| Proper chromosome segregation during egg and sperm development is crucial to prevent birth defects and miscarriage. During chromosome replication, DNA entanglements are created that must be resolved before chromosomes can fully separate. In the oocytes of the fruit fly Drosophila melanogaster, DNA entanglements persist between heterochromatic regions of the chromosomes until after spindle assembly and may facilitate the proper segregation of chromosomes during meiosis. Topoisomerase II enzymes can resolve DNA entanglements by cutting and untwisting tangled DNA. Decreasing Topoisomerase II (Top2) levels in the ovaries of fruit flies led to sterility. RNAi knockdown of the Top2 gene in oocytes resulted in chromosomes that failed to fully separate their heterochromatic regions during meiosis I and caused oocytes to arrest in meiosis I. These studies demonstrate that the Top2 enzyme is required for releasing DNA entanglements between homologous chromosomes before the onset of chromosome segregation during Drosophila female meiosis.
| In most organisms, crossing over between homologs during meiosis ensures their faithful segregation at the first meiotic division. However, in Drosophila melanogaster females, the 4th chromosomes are always achiasmate, and the X chromosomes normally fail to crossover in 6–10% of oocytes [1]. Nonetheless, Drosophila females can segregate these achiasmate chromosomes with high efficiency, demonstrating the existence of a system (termed the distributive system) to segregate homologous chromosomes that fail to recombine [2].
Heterochromatic regions on the achiasmate chromosomes are both necessary and sufficient for the proper segregation of achiasmate homologs, and homologous heterochromatic regions remain tightly paired throughout prophase of Drosophila female meiosis [3]–[5]. However, during prometaphase I, achiasmate chromosomes move dynamically on the spindle before properly biorienting and then congress into a mass with the chiasmate chromosomes at metaphase I [6], [7]. During these movements, achiasmate 4th and X chromosomes are connected by heterochromatic threads, which may play a role in the mechanism by which heterochromatin mediates chromosome segregation [6], [8]. How these threads are formed is unknown, but they could potentially arise from stalled replication intermediates [9].
Evidence for such connections between segregating meiotic chromosomes was first observed in crane fly spermatocytes [10]. Cutting the centromere from the arm of a segregating anaphase I chromosome led to re-association of the severed arm with its homolog on the opposite half-spindle, supporting the idea that chromosomes are able to maintain physical connections during meiosis I and that these connections can generate the force necessary to bring chromosomal regions together [10]. Additionally, chromosomal associations were observed during anaphase I in D. melanogaster sperm mutant for components of the condensin complex [11]. These studies indicate that the thread-like structures connecting chromosomes may be a conserved mechanism for segregating meiotic chromosomes.
To prevent loss of genetic material, such connections between homologs need to be resolved before anaphase I. Topoisomerase II enzymes are capable of creating double-strand breaks in DNA to resolve DNA entanglements during replication and transcription [12]. This function makes topoisomerase II or topoisomerase II-like enzymes possible candidates for resolving heterochromatic DNA threads between homologs during meiosis. Unfortunately, the study of topoisomerase II enzymes has been limited in meiosis due to the requirement of these enzymes to resolve DNA concatenations caused by replication in mitosis, as well as other potential roles in recombination, transcription and chromosome condensation [12]. Thus, most strong loss-of-function mutations of topoisomerase II enzymes are lethal. Examining the function of topoisomerase II during meiosis is further complicated by the presence of two topoisomerase II enzymes in many organisms.
Various studies have tried to address these issues either by chemically inhibiting topoisomerase II enzymes, such as in mice [13]–[16], or by induction of conditional mutations of topoisomerase 2 (top2), as in yeast [17]–[19]. Shifting a temperature-sensitive top2 mutant of Saccharomyces cerevisiae to the restrictive temperature during meiosis led to arrest just prior to spindle assembly [19]. Shifting a temperature-sensitive top2 mutant of Schizosaccharomyces pombe to the restrictive temperature during meiosis led to arrest during the first meiotic division [17]. In both yeasts, earlier stages of meiosis appeared normal at restrictive temperatures [17]–[19]. Additionally, blocking recombination in the top2 mutants of both types of yeast resulted in yeast that could progress past their initial arrest but were still unable to successfully complete meiosis II [17], [18]. Based on these studies, it was concluded for both types of yeast that Top2 was required to resolve DNA entanglements that form between recombinant homologs at meiosis I [17], [18].
D. melanogaster contains only a single gene encoding a topoisomerase II enzyme, Top2, and null mutations in Top2 in Drosophila are lethal [20]. Heterozygous combinations of weak loss-of-function mutations are viable in some cases, but ovarian development is either so severely disrupted to prevent analysis or the mutations only mildly decrease Top2 function and/or protein levels [21]. These factors have made it difficult to fully assess the role that Top2 plays in the resolution of heterochromatic DNA threads at later stages of Drosophila female meiosis [21].
A solution to these difficulties was created when the Transgenic RNAi Project (TRIP) at Harvard Medical School made available a Top2 RNAi-expressing line that is inducible in the female germline using the maternal alpha-tubulin GAL (matαGAL) driver [22]. The matαGAL driver appears to start expressing robustly at approximately stage 3 of the Drosophila ovary [23]. By expressing Top2 RNAi with the matαGAL driver, Top2 levels can be decreased after the completion of the ovarian mitotic divisions where Top2 is essential and after the initiation of recombination where Top2 could potentially play a role. We will show that Top2 RNAi expression during prophase of Drosophila female meiosis leads to specific defects in the ability of heterochromatic regions to fully separate and for the achiasmate 4th chromosomes to move precociously towards the spindle poles in prometaphase I, despite the completion of spindle assembly. More importantly, we will demonstrate that the defect in chromosome separation leads to the inability of oocytes to successfully complete meiosis I, illustrating that the separation of heterochromatic regions is essential for the proper completion of meiosis.
Because strong loss-of-function alleles of Top2 are lethal [20], an RNAi construct targeting the Drosophila Top2 gene was expressed using the maternal α-tubulin GAL driver ({matalpha4-GAL-VP16}V37 or matαGAL for short) [22]. This allowed us to examine the effect of decreased Top2 levels starting at approximately stage 3 of the ovary during mid-prophase [23]. By Western blot, the level of Top2 protein present in Top2 RNAi/matαGAL oocytes was reduced compared to control lines, including flies heterozygous for only the matαGAL driver or only the Top2 RNAi construct (Figure S1). The knockdown was not complete, as a weak band of Top2 was typically present.
We analyzed Top2 RNAi-expressing oocytes using immunofluorescence with an antibody recognizing α-tubulin to mark the meiotic spindle and one recognizing histone 3 phosphorylated at serine 10 (H3S10p), which marks nuclei that have entered prometaphase I of meiosis and fortuitously highlights the DNA threads connecting achiasmate chromosomes during D. melanogaster female meiosis [8]. During prometaphase I in wild-type oocytes, achiasmate chromosomes biorient, and the obligately achiasmate 4th chromosomes move toward opposite poles of the bipolar spindle. H3S10p-positive threads are frequently observed emanating from, and often connecting, these chromosomes (Figure 1A) [8]. The achiasmate chromosomes then congress back to the chiasmate chromosomes by metaphase I and the chromosome mass forms a lemon-shaped structure [7].
Preparations of oocytes from two-day-post-eclosion mated females are enriched for prometaphase I oocytes [7]. Achiasmate homologs (usually the 4th chromosomes) were observed fully separated from the main chromosome mass in 40.0% (10/25) of such prometaphase I-enriched oocytes from mothers bearing only the Top2 RNAi construct (Top2 RNAi/+) and in 40.7% (11/27) of oocytes from mothers bearing only the driver (matαGAL/+). However, in similar preparations of Top2 RNAi/matαGAL oocytes, achiasmate chromosomes were rarely observed completely separated from the chiasmate chromosomes (3.7% of oocytes, 1/27). In addition, in 44.4% (12/27) of Top2 RNAi/matαGAL oocytes, one or more DNA projections were present that extended toward the spindle poles, with three of these oocytes containing two projections (Figure 1B–D). These projections stained positive for H3S10p and did not appear to connect chromosomes (Figure 1B–D). Similar projections were not observed in control oocytes heterozygous for only the Top2 RNAi construct or the matαGAL driver (N = 25 and 27, respectively). Thus, in Top2 RNAi/matαGAL oocytes, achiasmate homologs fail to separate from the main chromosome mass, as is observed in control oocytes. Rather, we often observe abnormal DNA projections emanating from that mass toward the pole. The nature of these projections is discussed below.
The location of centromeres in relationship to the DNA projections was assessed using an antibody recognizing Centromere Identifier (CID), the Drosophila CENP-A homolog [24]. In wild-type prometaphase I oocytes, a CID focus is typically observed leading each achiasmate 4th chromosome that has moved toward the spindle poles, while the six CID foci of the chiasmate chromosomes are located within the chromosomal mass such that homologous centromeres are oriented toward opposite spindle poles (Figure 2A). In Top2 RNAi-expressing oocytes, CID foci could be observed at or near the tip of the DNA projections in 97.5% (39/40) of projections analyzed, indicating that centromere-led movements of the chromosomes towards the spindles poles are the cause of the DNA projections (Figure 2B, C).
More than one CID focus could be seen within some projections (Figure 2B, C), with five foci being the maximum number observed within a single projection within one oocyte. In all but four oocytes examined, the CID foci displayed some degree of clustering, making acquiring an accurate average number of CID foci within the projections impossible. However, Figure 2B shows an example where all the CID foci appear to be present within the two projections. This finding suggests that the projections are not simply the 4th chromosomes becoming entangled with the autosomes and stretching as they attempt to move toward the poles, but that centromeres of chiasmate chromosomes are also moving toward the poles in some Top2 RNAi/matαGAL oocytes. These results demonstrate that decreased Top2 levels generally affect the movement and organization of centromeres within the chromosome mass. The data also suggest that the DNA protrusions are caused by the centromeres of both achiasmate and chiasmate chromosomes being pulled or pushed towards the spindle poles while other parts of the chromosomes are still anchored at the center of the spindle. Such movements would result in portions of the chromosomes being stretched out behind the centromeres.
To determine whether defects in spindle assembly were the cause of the abnormal centromere–led DNA projections, we examined spindle morphology in Top2 RNAi/matαGAL and control oocytes. In Top2 RNAi/+ oocytes, 87.5% (21/24) of spindles were bipolar and 75.0% (18/24) were tapered at both ends of the spindle. In matαGAL/+ oocytes, 100% (27/27) of spindles were bipolar and 74.0% (20/27) were fully tapered at both ends. Despite the abnormal DNA projections described above, chromosomes in Top2 RNAi/matαGAL oocytes were also able to organize a bipolar spindle in 88.5% (23/26) of oocytes, though in some cases one or both sides of the spindle were elongated to accommodate the chromosomal projections (Figure 1B–D). However, tapering of both ends of the spindle was only seen in 57.7% (15/26) of oocytes. This decrease in spindle tapering at both ends of the spindle may be due to the abnormal DNA projections rather than direct defects in spindle assembly caused by decreased Top2 levels.
The ability of Top2 RNAi/matαGAL oocytes to organize a bipolar spindle is further illustrated by live imaging. Video S1 shows a bipolar spindle quickly forming after germinal vesicle breakdown in a Top2 RNAi/matαGAL oocyte. After the completion of spindle assembly, the spindle remained bipolar and the achiasmate chromosomes remained associated with the autosomes for the duration of imaging. In 11 time-lapses of Top2 RNAi/matαGAL oocytes undergoing germinal vesicle breakdown, eight successfully formed a bipolar spindle. Additionally, all nine oocytes that had already completed spindle assembly by the start of live imaging maintained bipolar spindles for the duration of imaging. These results argue strongly that defects in spindle assembly are not the primary cause of the centromere-led abnormal projections in Top2 RNAi/matαGAL oocytes.
The regions of DNA that stain brightest with DAPI are typically the heterochromatic regions, such as those near the centromeres and a large portion of the 4th chromosomes. At metaphase I, the DAPI-bright DNA regions are oriented towards opposite spindle poles and are located at the ends of the chromosome mass [7]. In Top2 RNAi/matαGAL oocytes, DAPI-bright regions could be observed in aberrant configurations. For example, in Figure 2B, a single DAPI-bright region is present in the middle of the chromosome mass, and in Figure 2C a single DAPI-bright region is present at one end of the chromosome mass. To investigate this further, we used an antibody recognizing histone 3 trimethylated on lysine 9 (H3K9me3), a chromatin modification associated with heterochromatin [25], [26]. While in Top2 RNAi/+ oocytes the H3K9me3 signals were observed oriented towards both spindle poles, in Top2 RNAi/matαGAL the H3K9me3 signal was present on the DNA within the projections (Figure 3). H3K9me3-positive regions could also be seen splayed across the chromosome mass (Figure 3). These results suggest that the projections are composed of heterochromatic sequences, and we will show below that knockdown of Top2 causes a failure of heterochromatic regions to properly orient.
As the DNA projections appeared to be composed of heterochromatin, as based on the H3K9me3 antibody, we next wanted to determine whether these projections were simply caused by the failure of the achiasmate 4th chromosomes to move fully toward the spindle poles. Because heterochromatic threads connect achiasmate chromosomes during prometaphase I, we also wanted to know how heterochromatic regions would be affected by decreased Top2 levels. Finally, we wanted to determine whether homologous chromosome orientation would be affected by Top2 RNAi expression.
We utilized fluorescent in situ hybridization (FISH) probes recognizing heterochromatic regions of each chromosome to look at the separation of specific heterochromatic regions and to assess biorientation of homologs in females bearing structurally normal X, 2nd, 3rd, and 4th chromosomes. We first examined a probe recognizing the 359-bp satellite, which is primarily localized to the X chromosome, with a minor region on the 3rd chromosome [27]. In prometaphase I and metaphase I control Top2 RNAi/+ and matαGAL/+ oocytes, two large, well-separated and bioriented fluorescent 359-bp signals were observed in 96.4% (80/83) and 96.3% (79/82) of oocytes, respectively (Figure 4A, B). Furthermore, in immunoFISH studies using the 359-bp probe and an α-tubulin antibody to mark the spindle, the X chromosomes were properly oriented toward opposite poles of the bipolar spindle in 100% (16/16) of Top2 RNAi/+ oocytes (Figure S2A).
However, when we followed the X chromosomal heterochromatin using the 359-bp probe in Top2 RNAi/matαGAL oocytes, none of the oocytes (0/81) displayed two bioriented 359-bp foci. Instead, the 359-bp heterochromatic region failed to fully separate in 96.3% (78/81) of Top2 RNAi/matαGAL oocytes that should have been in prometaphase I or metaphase I, based on egg chamber stage (Figure 4A, C–E). In an immunoFISH experiment with α-tubulin, the lack of separation of the 359-bp heterochromatic region could be observed even on fully formed bipolar spindles in Top2 RNAi/matαGAL oocytes. Two bioriented 359-bp foci were not observed in any of the 16 oocytes examined, which is to say that the 359-bp regions failed to separate for 100% (16/16) of the spindles. This result indicates that the failure to separate heterochromatic regions is not likely to be due to a defect in meiotic progression or spindle assembly after germinal vesicle breakdown (Figure S2B, C).
We then asked whether the heterochromatin of the achiasmate 4th chromosomes would fail to separate in Top2 RNAi/matαGAL oocytes. A FISH probe recognizing the AATAT repeat present throughout the 4th chromosomes was examined. Two large, well-separated and bioriented AATAT probe foci were observed in 92.8% (77/83) of Top2 RNAi/+ control oocytes (Figure 4A, B) and in 97.6% (80/82) of matαGAL/+ oocytes (Figure 4A). However, separation of the AATAT 4th chromosome probe was strongly impaired in Top2 RNAi/matαGAL oocytes. The 4th probe failed to separate in 44.4% (36/81) of Top2 RNAi/matαGAL oocytes (Figure 4A, C). In 33.3% (27/81) of Top2 RNAi/matαGAL oocytes, the 4th probe signal was highly elongated, often extending into one or more projections (Figure 4A, D, E). This phenotype suggests that the AATAT repeat of the 4th chromosomes has become stretched out and indicates that at least some of the projections are composed of 4th chromosome heterochromatin. In 8.6% (7/81) of oocytes, two foci could be distinguished, but the foci were still oriented toward the same side of the DNA mass, indicating a failure in 4th chromosome biorientation. However, 13.6% (11/81) of oocytes did display two separated and bioriented foci, indicating that at least in some cases this region of the 4th chromosome can successfully separate and biorient in Top2 RNAi/matαGAL oocytes. These data demonstrate that decreased Top2 levels result in defects in 4th chromosome biorientation and the full separation of the AATAT heterochromatin repeat.
Heterochromatin threads have been observed connecting the 4th and X homologs during prometaphase I in Drosophila oocytes [6]. The presence of these threads may make heterochromatic regions of these chromosomes more sensitive to decreased Top2 levels. To determine whether a heterochromatic region of the chiasmate 2nd chromosomes would also be affected by decreased Top2 levels, we examined a heterochromatic probe recognizing the AACAC repeat on the right arm of the 2nd chromosome in prometaphase I and metaphase I oocytes [27]. Two bioriented AACAC foci were present in 96.3% (78/81) of Top2 RNAi/+ oocytes and 97.6% (82/84) of matαGAL/+ oocytes (Figure 5A, B). In contrast, Top2 RNAi/matαGAL oocytes displayed only a single focus for the AACAC probe in 71.4% (60/84) of oocytes that had likely completed spindle assembly based on egg chamber stage, indicating that, like heterochromatic regions on the X and 4th chromosomes, this heterochromatic region of 2R is also defective in its ability to separate (Figure 5A, C, D, E). In 22.6% (19/84) of oocytes, two maloriented AACAC probe foci were observed. The probe was stretched in 2.4% (2/84) of Top2 RNAi/matαGAL oocytes and two bioriented foci were observed in only 3.6% (3/84) of oocytes. These results indicate that a decreased Top2 level affects the separation of the AACAC heterochromatic region and the proper biorientation of the 2nd chromosomes during meiosis I.
Finally, we asked whether a heterochromatic region of the chiasmate 3rd chromosomes would show similar defects. We utilized a FISH probe to the heterochromatic Dodeca satellite on the 3rd chromosomes. Two bioriented Dodeca foci were present in 98.8% (80/81) of Top2 RNAi/+ control oocytes and 98.8% (83/84) of matαGAL/+ control oocytes during prometaphase I and metaphase I (Figure 5A, B). Upon Top2 RNAi expression, the Dodeca heterochromatic repeat failed to separate in 60.7% (51/84) of oocytes (Figure 5A, C). In 13.1% (11/84) of Top2 RNAi/matαGAL oocytes, the Dodeca FISH signals were highly stretched, often extending into a projection (Figure 5A, D, E). This observation demonstrates that the projections can be composed of chiasmate chromosome heterochromatin as well as that of the achiasmate 4th chromosomes. The Dodeca probe was bioriented in 16.7% (14/84) of oocytes, while two foci were maloriented in 8.3% (7/84) of oocytes. This illustrates that 3rd chromosome heterochromatin is also affected by decreased Top2 expression. We also observed that in 36.4% (30/83) of Top2 RNAi/matαGAL oocytes the 2nd and 3rd chromosomes segregated away from each other (Figure 5E) and that in 24.1% (20/83) of oocytes, both sets of 2nd and 3rd chromosomes were oriented in the same direction. These results once again illustrate that Top2 is involved in the proper biorientation of chiasmate chromosomes during meiosis I.
In conclusion, heterochromatic regions on all four chromosomes showed defects in their ability to fully separate at prometaphase I and metaphase I in Top2 RNAi/matαGAL oocytes. Additionally, these chromosomes displayed a failure to properly biorient. These data support the idea that Top2 is involved in releasing the bonds that hold heterochromatic regions together during prophase [5]. Even if heterochromatic regions could become fully separated at anaphase I in Top2 RNAi/matαGAL oocytes, the orientation of homologous chromosomes toward the same pole would lead to high levels of chiasmate and achiasmate chromosome missegregation [7], [28]. Although the primary conclusion to be drawn here is that Top2 is required to separate heterochromatic regions on all four chromosomes, perhaps the more interesting inference is that the lack of heterochromatic separation argues strongly for heterochromatic entanglements that affect all four heterochromatic regions tested.
To ensure that the defects in heterochromatic separation observed with the Top2 RNAi construct were caused specifically by the 21-nucleotide sequence targeting Top2, we constructed a mutated RNAi construct and tested its effects on spindle assembly and heterochromatin separation (see Materials and Methods). The Top2 RNAi construct shares homology with a second locus (CG33296) at 18 of the 21 nucleotides. Rather than randomly mutagenizing or scrambling the original construct, the Top2 RNAi construct was mutated to match the CG33296 gene, even though a meiotic function has not been speculated for this gene product nor has it been reported to show ovarian expression (FlyBase). We hoped the targeted mutagenesis of the Top2 RNAi construct would provide a better control for the specificity of the construct compared to a randomly scrambled construct that would only address the potential general effects of RNAi induction in the ovary.
Bipolar spindles could be observed in oocytes from CG33296 RNAi/matαGAL mothers (Figure S3A) and at least one 4th chromosome was separated from the autosomes in 27.6% (8/29) of oocytes. FISH experiments of CG33296 RNAi/matαGAL oocytes showed that all four heterochromatic FISH probes were well separated and properly bioriented in the majority of oocytes (97.8% [44/45] 359 bp, 96.7% [29/30] AACAC, 100% [30/30] Dodeca, and 97.8% [44/45] AATAT) (Figure S3). These data support the conclusion that the effects of the Top2 RNAi construct are due to decreased Top2 levels rather than off-target RNAi effects.
Dernburg et al. [5] noted that Drosophila oocytes undergo a modified diplotene phase, in which euchromatic regions appear to separate as early as stage 3–4, while heterochromatic regions remain tightly paired until prometaphase I. Thus, we did not expect that the RNAi knockdown generated in Top2 RNAi/matαGAL oocytes performed in these studies would impair separation of euchromatic regions. To verify this hypothesis, we examined mid-prophase oocyte nuclei to determine whether euchromatic regions could separate in Top2 RNAi/matαGAL oocytes. Dernburg et al. [5] observed separation of the euchromatic histone locus in 54.5% (36/66) of mid-prophase oocytes. Similarly, we observed that Top2 RNAi/matαGAL oocytes exhibited two foci of fluorescence for a BAC probe to polytene band 3C of the X chromosome in 51.1% (23/43) of prophase oocytes compared to 48.9% (22/45) of Top2 RNAi/+ control oocytes (Figure S4). Using a BAC probe to polytene bands 7DE of the same chromosome, 57.6% (19/33) of Top2 RNAi/matαGAL oocytes were observed to have two foci compared to 47.9% (23/48) of Top2 RNAi/+ control oocytes (Figure S4). These results suggest that while Top2 is required for the separation of heterochromatic regions following nuclear envelope breakdown, Top2 is either not required for the separation of euchromatic regions during mid-prophase, or the matαGAL driver does not knock down the level of Top2 enough to affect separation of euchromatic regions. However, these results do not rule out the possibility that there are earlier euchromatic entanglements that are formed during replication and then resolved prior to Top2 RNAi induction. Our results are consistent with our view that DNA entanglements that persist into mid-to-late prophase, as revealed by Top2 knockdown, are specific to the heterochromatin.
A defect in the ability to properly separate heterochromatic regions during prometaphase and metaphase of meiosis I would likely result in a failure to properly complete meiosis and enter into the first mitotic divisions after fertilization. In embryos from Top2 RNAi/+ control mothers, we observed normal embryonic development in 30/30 (100%) embryos, indicating that meiosis was successfully completed (Figure 6A). In contrast, 0.0% (0/20) of embryos from Top2 RNAi/matαGAL mothers initiated proper embryonic development, with 85.0% (17/20) containing only two nuclei: a small nucleus with a centriolar spindle that was presumed to be the paternal pronucleus and a larger, round nucleus (Figure 6B, C). This round nucleus was surrounded by α-tubulin that was not organized into a bipolar shape. Because this nucleus did not resemble the typical rosette structure of embryos, it was presumed to be the oocyte nucleus that had failed to exit meiosis I. The three exceptions are described in the legend of Figure 6. This absence of mitotic entry led to a complete failure of embryos from Top2 RNAi-expressing mothers to hatch (0.0% [0/447]).
We have shown that knockdown of Top2 during prophase of meiosis I in Drosophila oocytes results in defects in homolog segregation and sterility. Heterochromatic regions of all four chromosomes failed to properly separate, leading to a failure of chromosomes to properly biorient during meiosis I. Additionally, achiasmate chromosomes showed defects in their ability to move away from the autosomes towards the spindle poles in prometaphase I. Instead, abnormal chromosomal projections were present. These DNA projections displayed several differences compared to the heterochromatic threads observed in wild-type prometaphase I oocytes. Specifically, the projections did not appear to directly connect two chromosomes and, more importantly, contained centromeres, which are not present in wild-type DNA threads. These attributes suggest that chromosomes initiate centromere-led movement but are anchored by DNA entanglements at the center of the spindle, resulting in the stretching out of chromosomal regions.
One might imagine that one component of these defects reflects a role of Top2 in resolving chiasmata. Indeed, in yeast, conditional mutants of top2 caused a meiotic cell cycle arrest that was partially alleviated by simultaneously eliminating recombination, indicating that in yeast, some of the targets of Top2 during meiosis are recombination dependent [17], [18]. Several lines of evidence suggest that this is not the case in Drosophila. First, the matαGAL driver used in the Drosophila oocytes does not appear to be strongly expressed until after recombination is thought to be finished [23]. Second, recombination is suppressed in heterochromatic regions near the centromeres [29]. Therefore, it seems more likely that the defects in heterochromatic separation during Top2 knockdown are due to the failure to resolve the heterochromatic threads observed during prometaphase I rather than a failure of Top2 to resolve DNA entanglements during the repair of double-strand breaks initiated in early prophase. Third, the AATAT heterochromatic repeat of the 4th chromosomes also shows defects in separation. Since the 4th chromosomes, which never undergo crossing over and are thus obligately achiasmate, also fail to properly separate heterochromatic regions [1], the defects in 4th chromosome heterochromatin separation cannot be due to homologous connections formed as a result of recombination. Finally, the heterochromatic threads observed during prometaphase I in Drosophila oocytes are not dependent on recombination, since they are observed in mutants of the Drosophila spo11 homolog, mei-W68 (Figure S5). While Top2 appears to affect the heterochromatic regions by resolving DNA entanglements that are recombination independent, it is unknown whether or not Top2 plays a role in the resolution of chiasmata at anaphase I. Since knockdown of Top2 results in metaphase I arrest, a role at anaphase I in chiasmata resolution cannot be assessed under these conditions.
The observation that 4th chromosome sequences are present within the DNA projections in Drosophila oocytes is not surprising given that the 4th chromosomes often move precociously towards the spindle poles during prometaphase I. However, it is more difficult to explain the stretching of 3rd chromosome heterochromatic sequences into some projections and the multiple CID foci within the projections. These results may indicate that the centromeres of the chiasmate chromosomes are also attempting to move towards the poles in Top2 RNAi-expressing oocytes. One possibility is that these projections are the consequence of a failed attempt by the oocyte to separate the heterochromatic regions. The four heterochromatic regions examined varied in the extent that they failed to separate. These differences may be due to a difference in the number of DNA entanglements between homologous heterochromatic regions in the oocytes.
The 359-bp region displayed the highest failure to separate upon Top2 knockdown. Several lines of evidence have suggested that the 359-bp heterochromatic region of the X may be handled differently by the cell than other heterochromatic regions. First, Ferree and Barbash [30] demonstrated that the hybrid lethality between D. simulans females and D. melanogaster males is due to the formation of anaphase bridges containing the 359-bp repeat region during mitosis in hybrid embryos. The authors speculate that D. simulans may lack factors for proper condensation of the 359-bp region. More recently, Ferree et al. [31] demonstrated that the lethality caused by some circularized X-Y ring chromosomes is also due to anaphase bridging of the 359-bp repeat in embryos. In both instances, Top2 localized to the anaphase bridges. Additionally, during the mitotic divisions of the Drosophila germarium, the 359-bp region is highly paired while the AACAC and Dodeca regions of the 2nd and 3rd chromosomes are mostly unpaired [32]. These results, as well as the complete failure of the 359-bp region of the X to separate in over 90% of oocytes when Top2 is knocked down, suggest that the 359-bp region may be especially prone to form DNA entanglements during replication (both between homologs and sisters) or that these entanglements may be processed differently than those in other heterochromatic regions. Additionally, in vivo studies of Top2 cleavage sites during mitosis in Drosophila showed a major cleavage site in the 359-bp repeat [33]. A failure to cleave this site when Top2 levels are decreased in meiosis I may contribute to the high failure of the 359-bp repeat to separate in Top2 RNAi oocytes.
In a parallel study examining decreased Top2 levels during Drosophila male meiosis, in which recombination does not occur [34], Mengoli et al. (cosubmitted) observed phenotypes similar to those seen by us in oocytes. Homologs, as well as sister chromatids, frequently failed to separate during meiosis I and meiosis II in Drosophila males, despite the formation of bipolar spindles similar to Drosophila oocytes. Homologs were stretched out into anaphase bridges at meiosis I, a phenotype which has similarities to the stretched out chromosomal projections in oocytes. These results indicate that Topoisomerase II may resolve similar DNA connections in both sperm and oocytes. Thus, the cells are responding in a similar fashion to deal with the failure of the resolution of these homologous connections. It is worth noting, however, that Mengoli et al. (cosubmitted) observed defects in euchromatic regions as well as heterochromatic regions of the chromosomes during male meiosis, while in oocytes, only heterochromatic regions were affected. This difference likely reflects the fact that Mengoli et al. (cosubmitted) examined mutations affecting Top2 levels at the start of meiosis while in Drosophila females the Top2 RNAi construct is not expressed until mid-late prophase.
Decreasing the level of Top2 in mitotic cell types causes phenotypes that are both similar to and divergent from those phenotypes observed during Drosophila female and male meiosis (Mengoli et al. cosubmitted). Top2 RNAi in mitotic Drosophila S2 cells led to the formation of DNA projections [35]. Although these chromosomal projections appeared similar to the ones in oocytes, CID foci were not observed within the DNA projections in S2 cells, while one or more CID foci were present in the projections in oocytes. This suggests that while the projections in oocytes are at least partially centromere led, the S2 projections are composed of the arms of the chromosomes. Top2 RNAi expression in S2 cells results in extensive chromosome lagging, chromosome bridging during anaphase, and chromosome missegregation [36]. Oocytes expressing Top2 RNAi seem to arrest in metaphase I before chromosome bridging can manifest, but chromosome missegregation is evident in oocytes as well.
We should also note that knockdown of Top2 in Kc cells leads to a decrease in euchromatic pairing without affecting the pairing of heterochromatic regions [37], [38]. However, evidence suggests Top2 is acting in different ways in each system. In meiosis, heterochromatic regions remain associated at higher levels than euchromatic regions during mid to late prophase [5], while euchromatic pairings persist longer than heterochromatic pairings during mitosis in cultured cells [37], [38]. Our data suggest that the lack of heterochromatin dissociation is due to the failure of Top2 to resolve DNA entanglements within the heterochromatin, while there is no evidence for the persistence of similar entanglements between euchromatic regions upon knockdown of Top2 in cell culture.
The study by Mengoli et al. (cosubmitted) suggests that different phenotypes manifest in Drosophila larval neuroblasts depending on the residual level of Top2. Additionally, different cell types, for example sperm, were more sensitive to decreases in Top2 levels. This study, as well as others in Drosophila [21], [39], illustrates the complexity of understanding the function of Top2 in resolving various types of DNA entanglements. For example, expression of the Top2 RNAi construct using the nanos-Gal4:VP16 driver, which is expressed beginning in germline stem cells [40], led to minimal ovarian development (Figure S6), suggesting that high levels of Top2 expression are likely necessary to resolve DNA entanglements caused by replication in the germline stem cell divisions and/or the cystoblast divisions. This hinders the examination of the role of Top2 in such processes as replication, recombination, and chromosome condensation early in oogenesis.
Knocking down topoisomerase II enzymes using RNAi or chemically inhibiting its two isoforms in mitotic human cell lines leads to a number of defects, including entangled chromosomes, chromosome segregation defects, cell cycle delays, and in some cases cell cycle arrest [41]–[43]. Most interesting is that chemically inhibiting Topoisomerase IIα in HeLa cells has been reported to increase the number and duration of PICH (Plk1-interacting checkpoint helicase)-positive ultrafine DNA bridges that connect centromeres during anaphase of mitosis, including to non-centromeric regions of the chromosomes. These results indicate that Topoisomerase II enzymes resolve DNA entanglements prior to anaphase in addition to those observed at the centromeres in mitotic cells [44], [45]. These PICH-positive ultrafine bridges have several similarities to the DNA threads observed during prometaphase I of Drosophila oocytes, in that they are composed of heterochromatin and connect segregating chromosomes. While mitotic DNA entanglements are between sister chromatids and some of the meiotic entanglements are likely between homologs, the results suggest that Topoisomerase II enzymes may play a conserved role in resolving chromosomal entanglements in mitosis and meiosis.
Determining the mechanism by which Topoisomerase II functions to resolve mitotic entanglements may provide insight into its potential role in resolving the meiotic threads observed by Hughes et al. [6]. In HeLa cells, centromeric cohesion appears to protect centromeric DNA threads from resolution until anaphase I when this cohesion is lost, and a similar mechanism is believed to protect centromeric concatenations at centromeres until anaphase II during the mouse male meiotic divisions [46], [47].
Based on these studies, it is at least possible that in Drosophila oocytes entanglements may form during replication in both heterochromatic and euchromatic regions, but euchromatic entanglements may be more accessible to resolution by Top2 immediately after replication. These entanglements would be resolved before the Top2 RNAi construct is induced. Heterochromatic entanglements may be protected from early resolution due their conformation after replication or the presence of heterochromatin binding proteins. Top2 would be unable to resolve these entanglements until the karyosome reorganizes for prometaphase I or until tension is provided on the DNA by microtubules, as has been proposed for the resolution of entanglements by Top2 in yeast [48]. At this stage Top2 levels would be reduced in Top2 RNAi-expressing oocytes, leading to a failure to resolve these entanglements. Alternatively, heterochromatic regions may be particularly prone to forming entanglements due to the repetitive nature of heterochromatic DNA, and thus more sensitive to decreased Top2 levels.
Topoisomerase II enzymes have also been implicated in regulating chromosome condensation in a number of cell types and organisms [12]. For example, Mengoli et al. (cosubmitted) reported that the centric heterochromatin appeared undercondensed in some neuroblasts from Top2 RNAi-expressing larvae. Additionally, strong knockdown of Top2 levels in Drosophila S2 cells led to a large and quantitative change in chromosome condensation [35].
These observations led us to ask whether some of the phenotypes observed in this study might be the consequence of the effect of Top2 depletion on chromosome condensation, especially in the pericentric heterochromatin. Global chromosome condensation in Top2 RNAi-expressing oocytes looked similar to wild-type oocytes, but small changes in chromosome condensation would be obscured by the close proximity of the chromosomes at prometaphase I and the high level of condensation of the chromosomes. It is thus possible that chromosomes are undergoing mild decreases in condensation, particularly in heterochromatic regions, when Top2 levels are decreased. Decreased condensation could contribute to the stretched out phenotype observed with the 3rd and 4th chromosome FISH probes and to the centromere-led chromosomal projections. However, the effects on condensation that we observe appear to be too weak to account for the entanglements and stretching that we observe.
These observations lead to a speculative model for the cause of the defects in Top2 RNAi-expressing oocytes. In wild-type oocytes, DNA entanglements form between the highly repetitive DNA sequences of heterochromatic regions. These entanglements could form during replication by stalled replications forks that can occur in the repetitive heterochromatic regions or by intertwinings that could form as chromosomes are replicated in close proximity. Entanglements within the heterochromatic regions would not be immediately resolved. Therefore, these entanglements could help hold heterochromatic regions of homologous chromosomes tightly together during prophase, including those chromosomes that, like the 4th chromosomes, fail to undergo recombination [5]. As germinal vesicle breakdown approaches, many of these entanglements would have to be resolved by Top2 in order for chromosomes to separate and biorient properly during prometaphase I and metaphase I, possibly assisted by karyosome reorganization and/or microtubule attachments to the chromosomes. The chromatin threads observed during prometaphase I could be the entanglements that failed to be resolved during late prophase or those protected from resolution to facilitate the biorientation of achiasmate chromosomes. Top2 and/or other enzymes would then resolve these final DNA threads by anaphase I. In oocytes with decreased levels of Top2, many of the DNA entanglements would not be resolved during meiosis I, leading to a failure of homologous heterochromatic regions to separate. In some cases, heterochromatic regions appear to attempt separation, but the DNA entanglements hold the chromosomes together at one or more places and the rest of the heterochromatic regions of the chromosomes become highly stretched out. The centromere-led DNA projections apparently occur when chromosomes attempt separation despite the existence of heterochromatic entanglements. Since the heterochromatic regions would still be locked together at egg activation when meiosis resumes, chromosomes would be unable to segregate to opposite spindle poles at anaphase I and ultimately, the oocytes would fail to exit meiosis I. Our results indicate that Top2 plays an important role in resolving homologous DNA entanglements in Drosophila oocytes. These results also suggest that the formation of such entanglements (by whatever mechanism) may be a characteristic of the meiotic process.
Flies were maintained on standard food at 25°C. Fly stocks used for RNAi experiments were y w; spapol, w; {matα4-GAL-VP16}V37 (Bloomington 7063), y1 sc1 v1; P{y[+t7.7] v[+t1.8] = TRiP. GL00338}attP2 (Bloomington 35416) an RNAi construct targeting the Top2 (CG10223) gene, and y v; CG33296 RNAi (described below). To obtain control flies containing only one copy of the driver or RNAi construct, the designated stocks were crossed to y w; spapol flies. Transheterozygotes mutant for mei-W68 (CG7753) were made from the following stocks: y/BS Y; mei-W68Z1049 cn bw/SM6a and y/y+ Y; mei-W68Z4572 cn bw/Cyo [49]. The genotype of the nanos-Gal4:VP16 driver flies was y w/y+ Y; nanos-Gal4:VP16; spapol.
Immunostaining of late stage oocytes was carried out as described [8] under conditions to limit activation. For experiments enriching for prometaphase I oocytes, females were yeasted for 2–3 days with males. For preparations enriching for metaphase I oocytes, virgin females were yeasted for 4–5 days post-eclosion [7]. Ovaries were treated with the primary and secondary antibodies described below and with 1.0 µg/mL 4′6-diamididino-2-phenylindole (DAPI) to label the DNA and mounted in ProLong Gold (Invitrogen). Immunostaining of embryos was carried out as described [50]. The DNA was labeled with 2.5 µg/mL Hoechst 34580 (Invitrogen) and mounted in ProLong Gold (Invitrogen).
Primary antibodies were used at the following concentrations: rat anti-α-tubulin (AbD Serotec, NC 1∶250), mouse anti-α-tubulin DM1a (Sigma-Aldrich 1∶100), rat anti-CID [51] 1∶1000), rabbit anti-trimethylated-histone-3 at lysine 9 (AbCam 1∶250) and rabbit anti-phosphorylated-histone 3 at serine 10 (Millipore 3∶1000). Secondary Alexa-488, Alexa-555, or Alexa-647 conjugated antibodies (Molecular Probes) were used at a dilution of 1∶400.
FISH was carried out as described [52], with the following modifications. Alexa Fluor 488 was conjugated to a region of the 359-bp repeat on the X chromosome, and the AATAT repeat on the 4th chromosome was conjugated to Cy3 [5], [30]. Both probes were denatured at 91°C and hybridization was carried out at 31°C. For the Alexa Fluor 488-labeled AACAC probe on the right arm of chromosome 2 and the Alexa Fluor 555-labeled Dodeca probe on the right arm of chromosome 3, the samples were denatured at 92°C and hybridization was at 37°C. Heterochromatic probes were made by Integrated DNA Technologies. Samples with BAC probes to euchromatic regions of the X chromosome were denatured at 92°C and hybridization was carried out at 37°C. BAC probes for 3C (BACR03D13) and 7DE (BACR39F18) were labeled with ARES Alexa Fluor 488 or 647 DNA labeling kits (Invitrogen). BAC DNA was digested with AluI, HaeIII, MseI, MspI, RsaI and MboI. Fragments were precipitated and then resuspended in water. The DNA was then denatured at 100°C for 1 minute and chilled immediately on ice. 20 µl of 5X terminal deoynucleotidyl transferase buffer, 20 µl of 25 mM Cobalt(II) chloride, 2.5 µl of 2 mM amionally dUTP from the ARES DNA labeling kit, 5 µl of 1 mM unlabeled 2′-deoxythymidine 5′-triphosphate, and 1 µl (400 units) Terminal deoxynucleotidyl Transferase (Roche) were added to 51.5 µl of the BAC DNA at room temperature. The reaction was allowed to proceed for 2 hr at 37°C. DNA was precipitated and the dried pellet resuspended in water. DNA was mixed with Ares labeling kit buffer and Ares dye that had been dissolved in dimethylsulfoxide. Sample was incubated in the dark for 1–2 hr at room temperature. The reaction was quenched with 1 M hydroxylamine. Probe was purified with a Qia-quick column (Qiagen). Labeled DNA was pelleted, allowed to dry and resuspended in water. Immuno-FISH was carried out as described [28].
For analysis of fixed ovaries, the DeltaVision microscopy system was used (Applied Precision, Issaquah, WA). The system is equipped with an Olympus 1X70 inverted microscope and high-resolution CCD camera. The images were deconvolved using the SoftWoRx v.25 software (Applied Precision). Embryos were imaged using an LSM-510 META confocal microscope (Zeiss) with a Plan-APO 40X objective (1.3 NA) with a zoom of 0.7 or a 63X Plan-apochromat (1.4 NA) with a zoom of 2. Images were acquired using the AIM software v4.2 by taking a Z stack and transformed into 2D projections using AIM software v4.2.
The sequence targeting the Top2 gene is CACGAAGATATCCAACTACAA (top) and TTGTAGTTGGATATCTTCGTG (bottom). This 21-bp oligo matched a sequence of CG33296 at 18/21 nucleotides. Oligos were created by Operon Biotechnologies Inc. to change the three differing nucleotides to precisely match CG33296. The 21-bp sense and anti-sense targeting sequences were CTCGAAGATATCCAACTTTAA and TTAAAGTTGGATATCTTCGAG. Oligos were annealed and ligated into a Valium22 vector that was digested with NheI and EcoRI according to the instructions provided by TRiP. Ligated product was transformed into competent cells. Genetic Services, Inc injected purified vector into y sc v; attP2 flies and y+ v+ flies were selected and sequence-verified.
Numerous females and males were allowed to acclimate to grape plates with yeast paste. Flies were transferred to fresh grape plates with yeast paste and females were allowed to lay eggs for 1–3 hours. Eggs within a grid on the plates were scored for hatching approximately 48 hr later.
Live imaging was performed on prometaphase I oocytes as described [6]. Oocytes were injected using standard microinjection procedures with an approximately 1∶1 ratio of porcine rhodamine-conjugated α-tubulin minus glycerol (Cytoskeleton) and Quant-iT OliGreen ssDNA Reagent (Invitrogen) diluted 0.7 fold with water. Oocytes were imaged using an LSM-510 META confocal microscope (Zeiss) with an alpha plan-fluar 100X (1.4 NA) objective and a 1.5 zoom. Images were acquired using the AIM software v4.2 by taking a 10-series Z stack at 1 micron intervals with 20 seconds between acquisitions, which resulted in a set of images approximately every 45 seconds. Images were transformed into 2D projections and concatenated into videos using the AIM software v4.2.
For each genotype, ovaries from virgin, yeast-fed females were dissected in cold 1X PBS, the ovaries were teased apart, and stage 14 oocytes were selected over the course of 2 hr. Stage 14 oocytes were homogenized in 50 µL of cold lysis buffer containing 150 mM NaCl, 50 mM Tris (pH 6.8), 2.5 mM EDTA, 2.5 mM EGTA, 0.1% Triton-X, and protease inhibitor cocktail (Sigma-Aldrich). Ovary lysates were cleared by centrifugation twice at 14,000 rpm for 15 min at 4°C. Lysates were assayed by Bradford and concentrations adjusted before samples were combined with 2X SDS sample buffer and boiled for 5 min, and the solubilized proteins were analyzed by Western blot using standard techniques. The primary antibody used for Western blot was rabbit anti-Top2 [21] at a dilution of 1∶5000 and α-tubulin (Serotec) at a dilution of 1∶5000. Immunoreactivity was detected using an alkaline phosphatase-conjugated rabbit secondary antibody (Jackson ImmunoResearch) and the nitroblue tetrazolium and 5-bromo-4-chloro-3-indolyl phosphatase (NBT/BCIP, Invitrogen) reagents.
|
10.1371/journal.pntd.0005272 | The Antiviral RNAi Response in Vector and Non-vector Cells against Orthobunyaviruses | Vector arthropods control arbovirus replication and spread through antiviral innate immune responses including RNA interference (RNAi) pathways. Arbovirus infections have been shown to induce the exogenous small interfering RNA (siRNA) and Piwi-interacting RNA (piRNA) pathways, but direct antiviral activity by these host responses in mosquito cells has only been demonstrated against a limited number of positive-strand RNA arboviruses. For bunyaviruses in general, the relative contribution of small RNA pathways in antiviral defences is unknown.
The genus Orthobunyavirus in the Bunyaviridae family harbours a diverse range of mosquito-, midge- and tick-borne arboviruses. We hypothesized that differences in the antiviral RNAi response in vector versus non-vector cells may exist and that could influence viral host range. Using Aedes aegypti-derived mosquito cells, mosquito-borne orthobunyaviruses and midge-borne orthobunyaviruses we showed that bunyavirus infection commonly induced the production of small RNAs and the effects of the small RNA pathways on individual viruses differ in specific vector-arbovirus interactions.
These findings have important implications for our understanding of antiviral RNAi pathways and orthobunyavirus-vector interactions and tropism.
| A number of orthobunyaviruses such as Oropouche virus, La Crosse virus and Schmallenberg virus are important global human or animal pathogens transmitted by arthropod vectors. Further understanding of the antiviral control mechanisms in arthropod vectors is key to developing novel prevention strategies based on preventing transmission. Antiviral small RNA pathways such as the exogenous siRNA and piRNA pathways have been shown to mediate antiviral activity against positive-strand RNA arboviruses, but information about their activities against negative-strand RNA arboviruses is critically lacking. Here we show that in Aedes aegypti-derived mosquito cells, the antiviral responses to mosquito-borne orthobunyaviruses is largely mediated by both siRNA and piRNA pathways, whereas the piRNA pathway plays only a minor role in controlling midge-borne orthobunyaviruses. This suggests that vector specificity is in part controlled by antiviral responses that depend on the host species. These findings contribute significantly to our understanding of arbovirus-vector interactions.
| Orthobunyaviruses are endemic in tropical and subtropical regions worldwide and are transmitted by mosquitoes, midges, ticks or other arthropods. The Orthobunyavirus genus within the Bunyaviridae family comprises at least 30 viruses that can cause disease in humans, including Oropouche virus (OROV; febrile illness), La Crosse virus (LACV; encephalitis) and Ngari virus (haemorrhagic fever) [1]. In addition, infection by orthobunyaviruses such as Cache Valley virus (CVV) and Schmallenberg virus (SBV) can lead to disease in animals [2].
Bunyamwera virus (BUNV) is the prototype virus of both the Orthobunyavirus genus and the family. Like most viruses in the genus, the BUNV genome possesses a tripartite, single-stranded negative sense RNA genome, in which the small (S) segment encodes the nucleocapsid (N) protein and the nonstructural protein NSs in overlapping reading frames, the medium (M) segment encodes a viral glycoprotein precursor (in the order Gn-NSm-Gc) for two envelope glycoproteins Gn, Gc and a nonstructural protein NSm, and the large (L) segment encodes the RNA-dependent RNA polymerase. This genome structure is generally reflected by most orthobunyaviruses with some differences for example in the presence or length of NSs [1].
BUNV was originally isolated from Aedes spec. mosquitoes in the Semliki Forest in Uganda in 1943 [3] but has since also been found in Culex spec. and Mansonia spec. (see [4] and references on BUNV therein) and Ochlerotatus spec. [5]. BUNV infections cause febrile illness and (rarely) encephalitis in humans in Sub-Saharan Africa, in particular in Nigeria and the Central African Republic, with wild rodents likely to be serving as amplifying reservoir [6]. Cache Valley virus (CVV) belongs to the Bunyamwera serogroup and is enzootic throughout North and South America [7, 8]. It was first isolated from Culiseta inornata mosquitoes in the Cache Valley in Utah, United States of America in 1956 [8], and has since been shown to be transmitted by mosquitoes of the Culiseta, Anopheles, Aedes, Culex and Ochleratatus genera [9, 10]. A small number of CVV infections in humans have been reported [11–13], where infection rarely leads to serious disease. In ruminants, including sheep and cattle, CVV causes spontaneous abortions and multiple congenital malformations [14–16]. Large mammals including deer, horses and sheep are known to serve as amplifying hosts [6].
In addition to mosquito-borne orthobunyaviruses some members of the genus are exclusively transmitted by biting midges (e.g. SBV, OROV and Sathuperi virus [SATV]) [17–22], ticks (Tete group viruses Bahig and Matruh) [23] or are mosquito/insect-specific [24, 25].
SBV and SATV are closely related orthobunyaviruses of the Simbu serogroup. SBV was first discovered in 2011 in Germany and the Netherlands [26] and infections have since been reported in many European countries [27]. Infections can lead to reduced milk yield, fever, fetal malformations and abortions in ruminants (primarily sheep, goats and cattle) [26, 28]; human infections have not been reported. SATV was isolated from mosquitoes in India in 1957 [29], and was later detected in cattle and biting midges in Nigeria [30, 31]. More recently, SATV was detected in Japan in 1999 [32]. To date little information is available on its pathogenicity in ruminants [22]. Both SBV and SATV are transmitted by Culicoides sp. biting midges [22, 31, 33–35].
The exogenous small interfering (exo-si) RNA and Piwi-interacting (pi)RNA pathways have previously been described as important mosquito antiviral responses limiting the replication of positive-strand RNA flaviviruses and togaviruses [36, 37]. The activity of the exo-siRNA pathway is mediated by two key proteins, the endoribonuclease Dicer-2 (Dcr2) and Argonaute-2 (Ago2). Dcr2 cleaves long virus-derived double-stranded RNAs (dsRNAs) into 21 nucleotides (nt) long small interfering RNA (viRNA) duplexes. These viRNAs are then incorporated into the multiprotein RNA-induced silencing complex (RISC), where presumably one strand is retained (guide strand) by Ago2 to detect, bind and catalyze the degradation of complementary single-stranded (ss)RNA such as viral mRNA. Indeed, 21 nt viRNAs were produced in mosquitoes or mosquito-derived cell lines upon infection with several arboviruses, for example flaviviruses (dengue virus, DENV; West Nile virus, WNV), alphaviruses of the Togaviridae family (chikungunya virus, CHIKV; Semliki Forest virus, SFV; Sindbis virus, SINV; o’nyong’nyong virus, ONNV), bunyaviruses (Rift Valley fever virus, RVFV; and SBV) as well as reoviruses (bluetongue virus, BTV) [38–42]. Silencing experiments in mosquitoes have confirmed the antiviral activity of the exo-siRNA pathway against DENV, SINV, CHIKV and ONNV in vivo [42–46].
In addition to the exo-siRNA pathway, the piRNA pathway has been shown to be a contributor to antiviral immunity in mosquitoes or derived cells [38, 47]. In Drosophila, the piRNA pathway is involved in the epigenetic regulation of the expression of transposable elements (TE) in the germline, and thus preserves the genome integrity [48–51]. However, PIWI proteins have also been detected in somatic cells [49, 52–54]. The production of transposon-specific piRNAs is complex and relies on proteins of PIWI clade, in particular Piwi, Aubergine (Aub) and Argonaute-3 (Ago3). piRNAs are believed to be generated via a primary processing pathway and a secondary ping-pong amplification loop [51, 52]. Primary piRNAs have a 5’ uridine (U1) bias. Secondary piRNAs have a 10 nt overlap with primary piRNAs and contain an adenine at position 10 (A10 bias). Mature piRNAs are generally 26–32 nt in length. The piRNA pathway is functional in mosquito germline and somatic cells [38, 39, 47, 55–57]. Interestingly, in mosquitoes a loss of Aub and a diversification of Piwi proteins has occurred [58]. This gene diversification has been linked with a gain of function of the piRNA pathway and a new role in antiviral immunity since the pathway has also been found to target a number of mosquito-borne viruses including DENV, CHIKV, SFV and RVFV [38, 39, 47]. Further, in the Ae. aegypti-derived Aag2 cell line an antiviral effect of Piwi4 against SFV has been directly demonstrated [47]. Recently it was shown that Ago3 and Piwi5 (and Piwi6 to a lesser extent) are needed for the generation of SINV and DENV specific piRNAs in Aag2 cells [59, 60]; however, the viral RNA substrate that induces this pathway is unknown. Importantly, the respective contribution of the two small RNA pathways to immune defenses against negative-sense RNA arboviruses has not been studied. In short, little is known about the interactions of viruses and vectors, and antiviral responses of vectors, which may govern viral infection, dissemination and transmission.
In this study we investigated the antiviral activities of the mosquito exo-siRNA and piRNA pathway against two mosquito- and three midge-borne orthobunyaviruses. Using reporter BUNV and SBV that express Nano luciferase we compared these responses in vector-virus and non-vector-virus interactions. We performed small RNA sequencing and showed that mosquito as well as midge-borne viruses produce virus-specific siRNAs and piRNAs in mosquito cells. Interestingly we found that silencing of Ago2 and Piwi4 in Aag2 cells led to increased viral replication of mosquito-borne orthobunyaviruses, in contrast to midge-borne orthobunyaviruses where only Ago2 silencing increased virus replication. Additionally, silencing of other piRNA pathway members (Piwi5, 6 and Ago3) affected virus replication differently depending on whether the virus was mosquito- or midge borne. These findings indicate that RNAi pathways play a crucial role in the control of orthobunyavirus replication; however, the piRNA pathway in Aag2 cells seems to be adapted to specific virus-vector combinations and may have important consequences for arbovirus tropism and vector specificity.
Our previous work has shown that small RNAs with viRNA and piRNA characteristics are produced in non-vector mosquito cells infected with the midge-borne SBV [40] (S1A Fig). To determine if this is similar for a mosquito-borne virus, Ae. aegypti-derived Aag2 cells were infected with BUNV (Fig 1A and 1B) and small RNAs were isolated, sequenced and mapped to the virus genome and antigenome. As shown in Figs 1C and 2A, 21 nt long small RNAs were produced from all three segments which mapped along the genome and antigenome in a cold and hot spot pattern. For the L segment, these 21 nts viRNAs were the major small RNA species produced. Moreover, small RNAs of 24–30 nts with piRNA-specific features (A10, U1 bias), which mapped across the genome and antigenome in a hot and cold spot pattern (Fig 2B), were produced for all three segments (Fig 2C); however, they were the major small RNA species only for M and S segments with a bias for small RNAs mapping to the antigenome. In contrast, 24–30 nt small RNAs mapping to the L segment had a bias for the genome (Fig 1C). Similar results were obtained in BUNV-infected Ae. albopictus-derived U4.4 cells (S2 Fig).
To determine if similar small RNAs are produced for BUNV and SBV in midge cells; experiments were repeated with infected C. sonorensis KC cells. As previously reported no piRNA-like molecules could be detected for SBV in KC cells, in contrast, 21 nt small RNAs were mapped to the genome and antigenome of all three segments [40] (S1B Fig). Similar results were observed for BUNV small RNAs in infected KC cells (Fig 1D). Although small RNAs of 24–28 nts could be detected in KC cells, they lacked the piRNA-specific features (A10, U1 bias and 10 nts overlap of sense and antisense small RNAs) (Fig 1D).
Overall, our data showed that BUNV infection induced small RNA patterns comparable with SBV [40] in mosquito-derived cells as well as Culicoides-derived cells. In addition, differences in small RNA patterns were found between mosquito and midge cells.
Previously, luciferase expressing alphaviruses were employed to investigate the antiviral RNAi response in arthropod cells [47, 61, 62]. To obtain similar tools for bunyaviruses, Nano luciferase (NL) expressing BUNV (BUNV-NL) and SBV (SBV-NL) were constructed in the course of this study. BUNV NSm protein, which is dispensable in viral replication in tissue culture, consists of the ectodomain, transmembrane, cytoplasmic domain and type-II transmembrane domain that also serves as internal Gc signal peptide [63, 64]. For BUNV-NL, the 62 residues of the NSm cytoplasmic domain (residues 395 to 455) was replaced by the NL coding region between residues 394 and 456, resulting in a chimeric NSm-NL fusion protein (S3A Fig). NSm-NL has a similar molecular weight to that of N protein (28.67 versus 26.67 kDa) and was indistinguishable from the N protein band in the protein profile of the reporter virus (S3B Fig). BUNV-NL exhibited smaller plaque size than the wildtype virus (S3C Fig). The reporter virus could readily infect Aag2 cells, similar to wildtype virus (S3D Fig). The SBV-NL reporter virus was constructed in the same way as BUNV-NL and successful infection of Aag2 cells, comparable to wildtype SBV, was verified by immunostaining (S3D Fig).
To assess the antiviral role of the small RNAi pathways in BUNV and SBV infected Aag2 mosquito cells, knockdown experiments were performed. Transcripts of the different RNAi pathway key effectors (Ago2, exo-siRNA pathway; Ago1, miRNA pathway; Piwi4-6 and Ago3, piRNA pathway) were silenced by transfection of sequence-specific dsRNA. The effect of the silencing was evaluated by BUNV-NL or SBV-NL infection at 24 hours post-transfection (p.t.) at low MOI (0.01) and subsequent luciferase detection at 48 hours p.i. (Figs 3A, 3B, 4A and 4B); dsRNA specific to eGFP was used as negative control.
Silencing of Ago2 led to an increase in luciferase expression for both BUNV-NL and SBV-NL (Fig 3A and 3B). In contrast, silencing of Piwi4 had no effect on SBV-NL, but resulted in an increase of BUNV-NL replication. Interestingly, silencing expression of other piRNA pathway related genes resulted in a significant decrease of BUNV-NL (Piwi6 and Ago3 knockdowns) and SBV-NL (Piwi5 knockdown) replication (Fig 4A and 4B). A decrease of luciferase expression was also observed for BUNV-NL following Ago1 silencing, compared to an increase of luciferase expression for SBV-NL (Fig 3A and 3B). These results suggested differences in the ability of the piRNA and miRNA pathway to interact with BUNV and SBV.
To determine if this was virus-specific or due to vector versus non-vector arbovirus interaction, silencing experiments were repeated with the mosquito-borne Cache Valley orthobunyavirus (CVV) as well as the midge-borne Sathuperi (SATV) orthobunyaviruses after their ability to grow in Aag2 cells was verified (S4A Fig).
Similar to BUNV, silencing of Ago2 or Piwi4 promoted CVV replication, whereas Ago1 knockdown reduced CVV replication (Fig 3C). Moreover, CVV-infected, Piwi4-silenced Aag2 cells showed cytopathic effects not observed for any of the other knockdown experiments with this virus (S4B Fig). For midge-borne SATV, an increase in replication was observed in cells silenced for Ago2 and Ago1. However, no significant effect on SATV replication was observed in cells treated with Piwi4 dsRNA (Fig 3C). Successful silencing of transcripts by sequence specific dsRNAs was verified by quantitative RT-PCR (Figs 3D and 4C).
In short, Ago2 silencing led to a consistent increase in virus replication for mosquito and midge-borne orthobunyaviruses, supporting an antiviral activity of the exo-siRNA pathway. Similar effects were observed for Piwi4 silencing in the case of mosquito-borne viruses, but not midge-borne viruses.
Silencing of Ago1 resulted in a decrease of replication of the tested mosquito-borne viruses, suggesting an importance of the miRNA pathway for a successful infection in Aag2 cells for these viruses. In contrast, silencing of Ago1 resulted in an increase of SBV and SATV, albeit only slightly significant.
The RNAi response is a major antiviral response in arthropods against arbovirus infection. The activated exo-siRNA pathway plays a role in a variety of organisms, including mosquitoes and the model insect D. melanogaster. In contrast, the antiviral activity of the piRNA pathway and the production of viral specific piRNA molecules have been restricted to mosquitoes, especially Aedes spp. Besides, interactions between the miRNA pathway and viruses have been reported in several organisms [65, 66], acting either pro- or antiviral. This can be by expression changes of vector/host miRNAs or viral encoded miRNAs which can either directly target the virus or have host/vector targets, resulting in changes of the cell environment. Previous research has often used D. melanogaster to investigate the interaction between the insect RNAi response and different viruses including arboviruses; however little is known about the specificity of the antiviral response in vector- or non-vector arbovirus interactions. Knockdown experiments of key proteins of the exo-siRNA, miRNA and piRNA pathway in mosquito cells and orthobunyavirus infection, either mosquito- or midge-borne, supports the broad antiviral activity of the exo-siRNA pathway, based on an observed increase in virus replication upon Ago2 silencing for all viruses used. In contrast, the miRNA pathway seemed to be only important for virus replication in the case of a virus-vector match. miRNA expression is often species or even tissue specific and some miRNA-arbovirus interactions have been reported [67]. Whether a similar interaction is important for BUNV and CVV infection in mosquitoes has to be investigated in the future.
Infections with both mosquito-borne and midge-borne viruses were able to induce viral specific piRNAs in the used mosquito cell line. However, the antiviral activity of the piRNA pathway was only confirmed for the mosquito-borne viruses in the Piwi4 knockdown cells. Interestingly knockdown of other piRNA pathway members indicated that they may have pro-viral activities, knockdown of Piwi6 and Ago3 had replication suppressive effects on the mosquito-borne BUNV while knockdown of these proteins had no effect on the midge-borne SBV, whereas knockdown of Piwi5 reduced SBV replication but not BUNV replication. The meaning and biological relevance of these observations is not yet clear. Co-silencing/infection experiments and sequencing of virus-derived small RNAs from such cells may give clues to how different Piwi proteins interact, their role in piRNA-like small RNA production and potentially a hierarchy of protein effector and regulatory functions within this pathway. Perhaps, similar to other RNA binding proteins some of these proteins are important for promoting virus replication though this would require in depth analysis of RNA protein interactions and/or protein-protein interaction studies. Little is known about the piRNA pathway in mosquitoes, the production of virus specific piRNAs and how/if they target the virus. Recently, it has been shown that Piwi5 (and to a lesser extent Piwi6) and Ago3 are needed for virus-derived piRNA production [59, 60]; however, no antiviral activity has previously been linked to these transcripts. In contrast, Piwi4 seems not to be involved in virus-derived piRNA production [59]; however knockdown of Piwi4 can result in antiviral activity for mosquito-borne viruses, like SFV [47]. Similar antiviral activity of Piwi4 is observed for the mosquito-borne orthobunyaviruses: CVV and BUNV, but not for midge-borne orthobunyaviruses (SBV and SATV). This is especially striking as both the mosquito-borne BUNV as well as midge-borne SBV produce similar amounts of virus-specific piRNA molecules in infected Aag2 cells. This would suggest that the antiviral activity of Piwi4 is species specific and either acts as an effector protein downstream of the virus specific-piRNA production or through a different as yet unidentified pathway. Interestingly, no viral specific piRNAs have been reported in midges so far; although this has only been investigated in the C. sonorensis-derived KC cell line and not whole midges [40].
Interestingly, it has been shown that infection of mosquitoes and mosquito-derived cultured cells with arboviruses can lead to generation of virus derived cDNA forms, which are important for mosquito tolerance to virus infection and survival [68]. These DNA forms have the potential to become the template for small RNA production and could therefore determine the action of RNAi pathways on acute arbovirus infection. If the generation of cDNA forms in mosquitoes and derived cells is mosquito-borne virus specific (not occurring with e.g. midge borne viruses) is not known, but could explain the observed species specific antiviral activity of Piwi4.
Overall, these results show the importance to investigate the antiviral RNAi response in vector cells to understand the complex interaction between virus-vector interplay and its effect on tropism.
BSR-T7/5 cells [derived from the BSR clone of baby hamster kidney cells-21 [BHK-21] and stably expressing T7 RNA polymerase [69]; a kind gift of Dr. K.K. Conzelmann, Max-von-Pettenkofer Institut, Munich, Germany] were maintained in Glasgow minimal essential medium (GMEM) supplemented with 10% tryptose phosphate broth (TPB), 10% fetal calf serum (FCS), and 1 mg/mL geneticin. BHK-21 cells were maintained in GMEM supplemented with 10% TPB, 10% newborn calf serum (NCS) and 1% penicillin/streptomycin (P/S). Sheep choroid plexus cells (CPT-Tert) [70] were grown in Iscove's modified Dulbecco's media (IMDM) supplemented with 10% FCS and 1% P/S. Vero E6 cells were grown in Dulbecco’s modified Eagle’s medium (DMEM) and 10% FCS. BSR-T7/5, BHK-21, CPT-Tert and Vero E6 cells were grown at 37°C and 5% CO2. Ae. aegypti-derived Aag2 cells were maintained in L-15 medium supplemented with 10% TBP, 10% FCS and 1% P/S at 28°C.
Plasmids to generate full-length BUNV antigenome RNA transcripts [pT7riboBUNL(+), pTVT7RBUNM(+), and pT7riboBUNS(+)] have been described previously [71, 72]. pTVT7RBUNM-NL was generated by PCR-directed internal deletion to replace the coding region of the NSm cytoplasmic tail (residues 395 to 456) with that of Nano luciferase (NL). A five amino acid linker, GASGA, was inserted between the NSm transmembrane domain (TMD) and the N-terminus of NL. To facilitate the cloning of NL, a unique KpnI restriction enzyme site was introduced at nt 1419 in the BUNV M segment cDNA (S3A Fig). The plasmids, pUCSBVST7, pUCSBVMT7 and pUCSBVLT7, used to rescue SBV have been described previously [73]. pUCSBVT7-NL was generated through the introduction of two unique restriction sites, MIuI and XhoI, in the SBV M segment (provided by M. Varela; University of Glasgow). The restrictions sites were used to delete 90 nt of NSm (corresponding to nt 1235–1325 in JX853180.1) and NL was subsequently cloned in the deletion site. NL is a small luciferase subunit (19 kDa) from the deep sea shrimp Oplophorus gracilirostris with significantly increased luminescence expression and signal half-life as well as specific activity in mammalian cells compared to both Firefly and Renilla luciferases. NL uses a novel imidazopyrazinone substrate (furimazine) [74].
Rescue experiments were performed as essentially described previously, with a modification [75]. Briefly, BSR-T7/5 cells were transfected with a mixture of plasmids comprising 0.5 μg each of pT7riboBUNL(+), pT7riboBUNS(+) and either TVT7RBUNM(+) or TVT7RBUNM-NLuc cDNA. At 4 hours p.t., 2 ml growth medium was added and incubation continued for 5–11 days at 33°C until cytopathic effect (CPE) was evident. Virus titre was determined by plaque assay on BSR-T7/5 cells. Cells were fixed in 4% formaldehyde and plaques stained in 0.01% toluidine blue. The rescue of SBV-NL was performed using pUCSBVST7, pUCSBVMT7-NL and pUCSBVLT7 as described for BUNV but with the exception that 1 μg of each plasmid was used and that the plaque assay was performed using BHK-21 cells.
BUNV, BUNV-NL, SBV, SBV-NL, CVV and SATV stocks were grown in BHK-21 cells. Cells were infected with viruses at a multiplicity of infection (MOI) of 0.01 PFU/cell and incubated at 33°C. Virus-containing cell supernatant was harvested when CPE was evident (usually 2–4 days p.i.), cleared by centrifugation and stored at -80°C. Virus titres of BUNV and CVV were determined by plaque assay on BHK-21 cells and those of SBV and SATV on CPT-Tert cells.
Procedures for metabolic radiolabelling and immunoprecipitation of BUNV proteins were described previously [76]. Briefly, at 24 hours p.i., BSR-T7/5 cells were labelled with [35S]methionine (50 Ci) for 2 hours and then lysed, on ice, in 300 μl of non-denaturing RIPA buffer (50 mM Tris-HCl [pH7.4], 1% Triton X-100, 300 mM NaCl, 5 mM EDTA) containing a cocktail of protease inhibitors (Roche). For immunoprecipitation assays, BUNV viral proteins were immunoprecipitated with anti-BUNV antibody, a rabbit antisera raised against purified BUNV [77] that had been conjugated to magnetic Protein A-Dynabeads (Life Technologies). Viral proteins were analysed by SDS-PAGE under reducing conditions.
The plaque phenotype of BUNV and BUNV-NL in BSR cells (routinely grown in GMEM supplemented with 10% foetal calf serum at 37°C in a 5% CO2 incubator) was investigated by plaque assay. Cells were seeded into 12-well plates at a density of 1.2 x 105 cells/well and left to adhere overnight. Cells were infected with BUNV or BUNV-NL at a MOI of 0.05 and cells were fixed with 4% formaldehyde-PBS and stained with 0.1% crystal violet blue solution at 3 days post infection.
Growth of BUNV, SBV, CVV and SATV in Aag2 cells was assessed. Briefly, Aag2 cells were seeded into 24-well plates at a density of 1.7 x 105 cells/well and left to adhere overnight. Cells were then infected with viruses at MOI 1 (BUNV, CVV) or MOI 0.01 (SBV, SATV) and culture supernatant was harvested at different time points p.i. Viral titres were determined by plaque assays on BHK-21 cells (BUNV, CVV) or CPT-Tert cells (SBV, SATV).
Aag2 cells were seeded into 24-well glass bottom plates at a density of 1.7 x 105 cells/well and infected with BUNV, BUNV-NL, SBV or SBV-NL at a MOI of 0.01 for 48 hours. Cells were fixed and stained with anti-BUNV N or–SBV N antibody [78, 79]. Goat anti-rabbit Alexa Fluor 488 antibody (Molecular Probes) was used to detect primary antibodies and cells were mounted using Vectashield mounting media containing DAPI (Vectorlabs). Images were taken using the EVOS FL Cell Imaging System.
Small RNA sequencing of BUNV-infected Aag2 and U4.4 cells was carried out by Edinburgh Genomics (University of Edinburgh) using the Illumina HighSeq 2000 platform, as previously described [40]. In short, 2.6 x 106 Aag2 cells/well were seeded into a 6-well plate and left to adhere overnight. Cells were infected with BUNV at a MOI of 10. At 24 hours p.i., RNA was isolated from individual wells using 1 ml TRIzol (Life Technologies) followed by purification, sequencing, analysis and mapping of virus specific small RNAs using viRome [80]. Small RNA sequences were submitted to the European Nucleotide Archive (accession number PRJEB15203). Data for SBV are referenced under [40]. The actual number of reads for each experiment is shown S1 Table.
Silencing of Ae. aegypti Ago2, Piwi4, and Ago1 was performed in Aag2 cells using dsRNAs as previously described [81]. dsRNA targeting eGFP was used as control. Silencing of transcripts was confirmed by qRT-PCR using Fast SYBR Green Master Mix (Applied Biosystems) and the corresponding primers [S2 Table; [81]]; on an ABI 7500 Fast real-time PCR instrument. Ae. aegypti S7 ribosomal transcript was used as housekeeping gene for relative quantification as previously described [81]. At 24 hours p.t., cells were infected with BUNV-NL, SBV-NL, CVV or SATV at a MOI of 0.01. At 48 hours p.i. cells were either lysed in passive lysis buffer and NL luminescence was measured (BUNV-NL and SBV-NL) or RNA was isolated using TRIzol (CVV and SATV), followed by cDNA production and qRT-PCR analysis. cDNA synthesis was performed using random hexamer primers (Promega) and SuperScript III reverse transcriptase (Life Technologies). qRT-PCR was performed using Fast SYBR Green Master Mix using corresponding primers (S2 Table). Ae. aegypti S7 ribosomal transcript was used as housekeeping gene.
|
10.1371/journal.ppat.1002747 | Proteomic Profiling of the TRAF3 Interactome Network Reveals a New Role for the ER-to-Golgi Transport Compartments in Innate Immunity | Tumor Necrosis Factor receptor-associated factor-3 (TRAF3) is a central mediator important for inducing type I interferon (IFN) production in response to intracellular double-stranded RNA (dsRNA). Here, we report the identification of Sec16A and p115, two proteins of the ER-to-Golgi vesicular transport system, as novel components of the TRAF3 interactome network. Notably, in non-infected cells, TRAF3 was found associated with markers of the ER-Exit-Sites (ERES), ER-to-Golgi intermediate compartment (ERGIC) and the cis-Golgi apparatus. Upon dsRNA and dsDNA sensing however, the Golgi apparatus fragmented into cytoplasmic punctated structures containing TRAF3 allowing its colocalization and interaction with Mitochondrial AntiViral Signaling (MAVS), the essential mitochondria-bound RIG-I-like Helicase (RLH) adaptor. In contrast, retention of TRAF3 at the ER-to-Golgi vesicular transport system blunted the ability of TRAF3 to interact with MAVS upon viral infection and consequently decreased type I IFN response. Moreover, depletion of Sec16A and p115 led to a drastic disorganization of the Golgi paralleled by the relocalization of TRAF3, which under these conditions was unable to associate with MAVS. Consequently, upon dsRNA and dsDNA sensing, ablation of Sec16A and p115 was found to inhibit IRF3 activation and anti-viral gene expression. Reciprocally, mild overexpression of Sec16A or p115 in Hec1B cells increased the activation of IFNβ, ISG56 and NF-κB -dependent promoters following viral infection and ectopic expression of MAVS and Tank-binding kinase-1 (TBK1). In line with these results, TRAF3 was found enriched in immunocomplexes composed of p115, Sec16A and TBK1 upon infection. Hence, we propose a model where dsDNA and dsRNA sensing induces the formation of membrane-bound compartments originating from the Golgi, which mediate the dynamic association of TRAF3 with MAVS leading to an optimal induction of innate immune responses.
| In response to pathogens, such as viruses and bacteria, infected cells defend themselves by generating a set of cytokines called type I interferon (IFN). Since Type I IFN (namely IFN alpha and beta) are potent antiviral agents, understanding the cellular mechanisms by which infected cells produce type I IFN is required to identify novel cellular targets for future antiviral therapies. Notably, a protein called Tumor Necrosis Factor receptor-associated factor-3 (TRAF3) was demonstrated to be an essential mediator of this antiviral response. However, how TRAF3 reacts in response to a viral infection is still not totally understood. We now demonstrate that, through its capacity to interact with other proteins (namely Sec16A and p115) that normally control protein secretion, TRAF3 resides close to the nucleus in uninfected cells, in a region called the ER-to-Golgi Intermediate Compartment (ERGIC). Following viral infection, the ERGIC reorganizes into small punctate structures allowing TRAF3 to associate with Mitochondrial AntiViral Signaling (MAVS), an essential adaptor of the anti-viral type I IFN response. Thus, our study reveals an unpredicted role of the protein secretion system for the proper localization of TRAF3 and the antiviral response.
| Following exposure to pathogen-associated molecular patterns (PAMPs), the innate immune response and the subsequent inflammatory reaction rely on evolutionarily conserved receptors termed pattern-recognition receptors (PRRs) [1]. These signalling receptors can be expressed at the cellular membrane (Toll-like receptors (TLRs) 1, 2, 4, 5, and 6), in acidic endosomes (TLRs 3, 7, 8, and 9), or in the cytoplasmic compartment (the double-stranded RNA (dsRNA)-activated kinase (PKR); the RIG-I-like helicases (RLH): retinoic-acid-inducible gene I (RIG-I), melanoma differentiation antigen 5 (MDA5), and LGP2; the HIN-200 family members: Absent In Melanoma 2 (AIM2) and interferon (IFN)-inducible IFI16 protein [2]; the DNA-dependent activator of interferon regulatory factors (IRFs) (DAI) and the nucleotide-binding oligomerization domain (NOD) receptors). RIG-I and MDA5 have been characterized as important cytoplasmic sensors for viral RNA [3]–[6]. Once activated by dsRNA molecules, RIG-I and MDA5 are recruited to the mitochondrial adaptor protein know as Mitochondrial AntiViral Signaling (MAVS) (also called IPS-1, Cardif and VISA) in order to trigger signalling cascades leading to IRF-3 and NF-κB activation, two essential players involved in the establishment of a cellular antiviral state [7]–[10].
Tumor Necrosis Factor (TNF) receptor-associated factors (TRAFs) are part of a family of adaptor proteins that bridge the intracellular domains of multiple receptors, such as TNFR, IL1R, and TLRs, to downstream effectors involved in the inflammatory and innate immune signalling pathways. The TRAF family is composed of seven members, TRAF1 through TRAF7. They all share a C-terminal TRAF domain, which is composed of a coiled-coil domain followed by a conserved receptor-interacting domain. This domain mediates self-association and interaction with receptors or signalling proteins. Their N-terminal regions are composed of one or more zinc-finger motifs and, with the exception of TRAF1, a RING-finger domain that mediates E3 ubiquitin ligase activity and signalling [11]. All mammalian TRAFs localize to the cytoplasm except TRAF4, which is found in the nucleus. Importantly, gene deletion studies have identified TRAF3 as a critical mediator involved in the induction of the type I interferons (IFNs) by the RLH pathway [12], [13].
TRAF3 has originally been shown to associate with TNF receptors (e.g. BAFFR, CD40, LTβR, RANK, CD30, and Fn14), which are activators of the non-canonical NF-κB pathway [14]–[17]. TRAF3 acts as a negative regulator in this pathway by promoting the recruitment of the TRAF2-cIAP1-cIAP2 E3 ligase complex to NF-κB-inducing kinase (NIK) in order to control its rapid turnover in resting cells [18], [19]. However, in the RLH pathway, the adaptor protein TRAF3 acts as a positive regulator. Its interaction with MAVS and TRADD is important to trigger IRF-3 phosphorylation through the adaptor molecule TANK and the IKK-related kinases TBK1 and IKKi [20], [21]. The TRADD-mediated recruitment of FADD and RIP1 to MAVS also enhances the interaction between TANK and TRAF3. A model was then proposed in which TRADD simultaneously organizes FADD- and RIP1-mediated NF-κB signalling on one hand and TRAF3- and TANK-mediated IRF-3 signalling on the other [21], [22]. However, this possible mechanism of action requires further investigation to determine how TRAF3 is recruited to the mitochondrial adaptor protein MAVS upon viral infection.
Here, we have used a proteomics-based strategy to identify novel TRAF3 interacting proteins that are implicated in the induction of type I IFN. Using this approach, we have identified two novel TRAF3 interactors, Sec16A (also known as KIAA0310) and p115 (also known as USO1), which have characterized roles in the Endoplasmic Reticulum (ER)-to-Golgi vesicular transport system. Both proteins were shown to play a primary role in the anterograde trafficking at the ER-Golgi interface by influencing the assembly and transport of coat protein complex II (COPII) vesicles. Sec16A assembles on the ER membrane and forms organized scaffold defining ER exit sites (ERES) where COPII assembly occurs [23]–[25]. The coiled-coil myosin-shaped molecule p115 was demonstrated to be an important tethering adaptor, which mediates vesicle tethering at the ER [26], Endoplasmatic Reticulum-Golgi Intermediate Compartment (ERGIC) [27], and in conjunction with tether proteins giantin and GM130 at the cis-Golgi [28], [29].
Since novel essential mediators of the type I IFN response were recently found to be associated with the ER or the exocyst pathway, such as STING (also called MITA, ERIS, and MPYS), Sec61β and Sec5 [30]–[33], we postulated that Sec16A and p115 may exert a similar function through the ER-to-Golgi transport compartments. Co-immunoprecipitation experiments and confocal microscopy confirmed the association and co-localization of TRAF3 with p115 and Sec16A. Importantly, overexpression of p115 or Sec16A increased the type I IFN response, whereas their knockdown impaired the induction of antiviral genes. Interestingly, the Golgi apparatus fragmented into cytoplasmic punctate structures following both RLH and cytoplasmic DNA sensor pathway activation, allowing TRAF3 to colocalize and associate with MAVS. Our study identifies p115 and Sec16A as new scaffold proteins involved in the establishment of the antiviral state.
In order to find novel players involved in the type I IFN pathway, we have used a functional proteomics approach based on FLAG affinity purification and mass spectrometry analysis (AP/MS). HEK293 cells stably expressing FLAG-TRAF3 were harvested, subjected to IP with an anti-FLAG antibody under native conditions and FLAG-TRAF3 complexes were analyzed by liquid chromatography coupled to tandem mass spectrometry (LC-MS/MS). In parallel, multiple AP/MS analyses were performed from cells expressing the FLAG alone. Following standard database searches, stringent statistical filtering was performed using SAINT (see Methods). Proteins detected with AvgP≥0.7 were manually inspected for frequency of detection across a database of ∼1000 AP/MS analyses, and proteins frequently detected in AP-MS experiments were removed. This resulted in the identification of 12 interaction partners for TRAF3, including TBK1, a well-known TRAF3 interactor [12], [13]. Surprisingly, Sec16A and p115, two proteins involved in ER-to-Golgi vesicular trafficking, were found to associate with TRAF3 immunocomplexes with a high confidence (Figure 1A and Figure S1).
To confirm these interactions, we performed conventional co-immunoprecipitation experiments by overexpressing the candidate tagged-proteins with FLAG-TRAF3 in 293T cells. The interactions between FLAG-TRAF3 and Myc-p115 or EGFP-Sec16A were clearly detected (Figure 1B–C). To further substantiate the interaction network between TRAF3, Sec16A and p115, we additionally established a pool of HEK293 cells stably expressing FLAG-p115 and analyzed its physiological interactors by FLAG affinity purification and LC-MS/MS, followed by analysis with SAINT. FLAG-p115 was found to be associated with nine proteins (after filtering), including Sec16A and GM130 (also known as GOLGA2), an established physical partner of p115 [34] (Figure 1D and Figure S1). The interaction of EGFP-Sec16A with Myc-p115 was further confirmed by co-immunoprecipitation experiments (Figure 1E). However, endogenous TRAF3 was not recovered in our FLAG-p115 analysis. This result may be explained by the fact that p115 has many higher-abundance interactors and/or is part of alternative complexes independent from TRAF3. However, overexpressed Myc-TRAF3 was recovered from FLAG-p115 complexes when the latter were immunoprecipitated from 293T cells co-expressing both constructs (Figure 1F). The TRAF3 interactome network identified by functional proteomics (Figure S1B) suggests the presence of at least a fraction of TRAF3 in close proximity to the Golgi network.
Members of the TRAF family often share common interacting partners. For example, TRADD and RIP1 strongly bind to TRAF1, TRAF2 and TRAF3 [21], [35], whereas the mitochondrial anti-viral signaling protein MAVS interacts with TRAF2, TRAF3 and TRAF6 [10], [20]. To verify the binding selectivity of the newly identified TRAF3 interactors, we next performed co-immunoprecipitation experiments in 293T cells overexpressing FLAG-tagged TRAF2, TRAF3 or TRAF6, TRAF molecules involved in type I IFN and inflammatory responses, along with Myc-p115 or EGFP-Sec16A. Only Myc-p115 was found to be enriched in FLAG-TRAF3 immunocomplexes (Figure S1C). A similar result was obtained with EGFP-Sec16A, except that a weak enrichment was observed with TRAF2 when compared to TRAF3 (Figure S1D).
To further validate the interaction between TRAF3 and these new interactors, we next analyzed their subcellular localization by confocal microscopy. The vesicle-tethering protein p115 is known to colocalize and interact specifically with the NH2 terminus of the cis-Golgi protein GM130 ([36] and see Figure 2A, panel 2). Upon ectopic expression of Myc-p115 and FLAG-TRAF3, we observed a co-localization of these two proteins (Figure 2A, panel 1). FLAG-TRAF3 was also observed to localize to the Golgi apparatus where it exhibits a high degree of overlap with the cis-Golgi marker GM130 (Figure 2A, panel 3). p115 was previously reported to be present in the ERGIC, through an interaction involving activated Rab1 [37], [38]. This cellular localization of FLAG-p115 can be visualized with the conventional ERGIC marker, ERGIC53 (Figure 2A panel 4). Notably, FLAG-TRAF3 (or Myc-TRAF3 (unpublished data)) was also present in the ERGIC (Figure 2A, panel 5). No significant colocalization was detected between TRAF3 and the ER marker calnexin (Figure 2A, panel 6), the lysosomal compartments (Figure 2A, panel 7) and the mitochondrial network (Figure 2A, panel 8).
In HeLa cells, Sec16A was demonstrated to define ERES [24], [25], localizing to punctate structures on the ER membrane. This pattern was reproduced in this study (Figure 2B, panel 1). Since Rab1 recruitment of p115 to ERES [26] represents an essential step for the subsequent docking of ER-derived vesicles to the ERGIC [39], we next examined the colocalization of EGFP-Sec16A and FLAG-p115. The two proteins clearly colocalized at the perinuclear region (Figure 2B, panel 2). FLAG-TRAF3 and EGFP-Sec16A also mainly colocalized at the perinuclear region in HeLa cells (Figure 2B, panel 3). Moreover, a colocalization of FLAG-TRAF3 with endogenous Sec16A at ERES distributed in the cytoplasm could be observed. However, some FLAG-TRAF3 punctae also appeared in close proximity to those containing Sec16A (Figure 2B, panel 4, compare arrows). Ectopic expression of FLAG-TRAF2 and FLAG-TRAF6 revealed that only TRAF3 exhibits a cellular Golgi-like distribution (Figure S2A) and colocalizes with endogenous Sec16A (Figure S2B) or Myc-p115 (Figure S2C). Importantly, endogenous staining of TRAF3 revealed that the majority of TRAF3 proteins localized to the juxtanuclear region containing both the cis-Golgi marker GM130 and the ERGIC marker ERGIC53 (Figure 2C).
To further confirm the localization of TRAF3 to the Golgi apparatus, we next treated the cells with nocodazole. Microtubule depolymerization is known to result in the reorganization of the Golgi complex into characteristic mini-stacks, which appear as punctate structures throughout the cell [40]. Nevertheless, FLAG-TRAF3 was detected to colocalize with GM130 and Myc-p115 in cells treated with nocodazole (Figure S3, panels 1, 2 and 3). Treatment with brefeldin A (BFA), leads to relocalization of the components of the cis-Golgi matrix to cytoplasmic punctate structures (also called remnants) that appear close to ERES [41], [42]. GM130 and p115 are cis-Golgi proteins, which are known to be relocalized to these remnants [41]. FLAG-TRAF3 was also relocated to cytoplasmic remnants upon BFA treatment, where it co-localized with GM130 and Myc-p115 (Figure S3, panels 4 and 5). Altogether, results from our pharmacological experiments and confocal microscopy strongly suggest that TRAF3 localized to ER-to-Golgi transport compartments, where it tightly associates [43], [44].
It has been proposed that a structurally intact TRAF3 molecule is required for its biological function. Indeed, TRAF3 lacking its N-terminal RING or the C-terminal TRAF domain lacks antiviral activity [20]. We therefore examined the subcellular localization of TRAF3 deletion mutants in reconstituted TRAF3 knockout MEF cells. Removal of the N-terminal Ring Finger domain (Figure 3A, panel 1), the N-terminal Ring and Zinc finger domains (Figure 3A, panel 2) or the C-terminal TRAF domain (Figure 3A, panel 3) resulted in TRAF3 molecules that no longer colocalize with the Golgi marker GM130. Furthermore, coimmunoprecipitation experiments in 293T cells revealed that immunocomplexes containing p115 are detected only with full length TRAF3 and that Sec16A-containing immunocomplexes required at least the isoleucine zipper and the TRAF domain (Figure 3B). Moreover, TRAF3 is known to interact with several substrates containing a particular motif (PxQxS/T) called the TRAF interaction motif (TIM) [45]. The mutation of two amino acids located in the TIM-binding pocket of TRAF3, Y440 and Q442, abrogates these interactions [20]. Interestingly, a strong interaction was detected between FLAG-TRAF3 Y440/Q442A and Myc-p115 or EGFP-Sec16A (Figures 3C and 3D), implying that this interaction is independent of the TIM motif. Thus, it is not clear yet whether TRAF3 interacts directly with Sec16A or p115 or requires other components such as TFG ([46] and see Figure S1). Collectively, these data suggest that an intact TRAF3 molecule is required for its proper localization and interaction with components of the ER-to-Golgi vesicular pathway.
Our data demonstrate that TRAF3 does not associate with the mitochondrial network in resting cells (Figure 2A, panel 8). However, TRAF3 was demonstrated to link the mitochondrial membrane-bound protein MAVS to the activation of TBK1, which is required for IRF3/7 phosphorylation and type I IFN induction in response to viral infection [20], [47]. Therefore, we next addressed the subcellular localization of endogenous TRAF3 upon viral infection and RNA/DNA sensor pathway activation. Intracellular delivery of the double-stranded RNA mimicry molecule, poly I:C, or the dsDNA mimicry agent poly dA:dT resulted in disorganization of the ribbon-like structure of the Golgi apparatus, giving rise to the formation of Golgi ministacks containing GM130 (Figure 4A, arrows in panel 2 and 3). Importantly, the localization of endogenous TRAF3 followed these Golgi fragments. Similar observations were made in cells infected with RIG-I inducers, Sendai virus (SeV), Respiratory Syncytial Virus (RSV) and Influenza virus (Figure 4B).
Additionally, we addressed the association of TRAF3 with p115- and Sec16A-containing complexes upon PAMP exposure. In unstimulated cells, a weak but constitutive association of endogenous TRAF3 with endogenous Sec16A and p115 was detected (Figure 5 A and B). However, upon viral infection or transfection with poly I:C or poly dA:dT, immunocomplexes containing endogenous TRAF3 were enriched with p115 and Sec16A. Importantly, the induced association of TBK1 with TRAF3 closely mirrored the presence of p115 and Sec16A. (Figure 5B). From these results we hypothesized that the localization of TRAF3 to the ER-to-Golgi compartment and the Golgi fragmentation of the latter into punctate structures might be required for the proper positioning of TRAF3 with MAVS.
To verify this hypothesis, loss-of-function experiments were conducted using HeLa cells exposed to siRNA duplexes targeting Sec16A and p115. As previously observed for p115 and Sec16A [24], [25], [48], [49], reducing the expression level of Sec16A or p115 led to a drastic disorganisation of the Golgi paralleled by a relocalization of TRAF3 as observed by the formation of small punctate structures (Figure S4A, panels 2 and 4; Figure S4B, panel 2). However, the majority of these GM130 positive punctae do not colocalize with TRAF3 and thus appear to be different from those observed following dsRNA and dsDNA sensing (compare Figure S4A, panel 4 with Figure 4).
Next we examined the effect of reducing the expression level of Sec16A or p115 on the ability of TRAF3 to colocalize with MAVS upon SeV infection. As expected, TRAF3 localization reorganized into punctate structures following SeV infection, allowing a significant proportion of TRAF3 to colocalize with MAVS (Figure 6A, panel 2). This effect was severely compromised by reducing the expression of Sec16A or p115 (Figure 6A, compare panels 4 and 6 with panel 2). Additionally, co-immunoprecipitation experiments revealed that TRAF3 formed an immunocomplex with MAVS upon SeV infection. Interestingly, silencing the expression level of p115 or Sec16A clearly blunted the ability of TRAF3 to bind to MAVS upon SeV infection (Figures 6B and 6C). Thus loss-of-function experiments targeting p115 and Sec16A led to a mislocalization of TRAF3 and its subsequent incapacity to associate with MAVS upon RLH pathway activation.
This prompted us to ask whether enforced retention of TRAF3 at the ER-to-Golgi compartment could negatively influence the type I IFN response. In order to verify this, a TRAF3 mutant containing a COPI and COPII sorting signal peptide [50], namely “AKKFF” [51], at its C-terminal end was generated and used in confocal microscopy and reporter gene assays. Confocal microscopy experiments revealed that addition of dilysine and dihydrophobic residues to the C-terminal end of TRAF3 resulted in the formation of large TRAF3 aggregates which failed to colocalize with the Golgi marker GM130 upon infection with SeV (Figure 7A). Consequently, the ability of the TRAF3-AKKFF mutant to mediate TRAF3-dependent synergistic activation of the IFNβ promoter was drastically reduced (Figure 7B), which is likely due to less binding to MAVS (Figure S5).
Altogether, these results indicate that the localization of TRAF3 to the ER-to-Golgi compartment is involved in the proper positioning of TRAF3 within the mitochondrial network and the induction of type I IFN innate immune response.
The results presented above suggest a role for the ER-to-Golgi compartment in TRAF3-dependent innate immune response. To investigate whether Sec16A or p115 play a role in the type I IFN response, we overexpressed both proteins in Hec1B cells and assessed NF-κB and IRF-3 transcription factor activation using reporter gene assays. Without any stimulation, overexpression of either protein did not significantly activate the IFNβ promoter. However, following viral infection, the response was increased in cells overexpressing Sec16A or p115 (Figure 8A). Overexpression of Sec16A and p115 also increased the activation of the ISG56 promoter (IRF3-dependent promoter) (Figure 8B) and the NF-κB-dependent promoter (Figure 8C) following SeV infection. Moreover, we observed a synergistic effect on IFNβ promoter activity when Sec16A or p115 were co-expressed with MAVS (Figure 8D), TBK1 (Figure 8E) and, interestingly, the TLR3 essential effector TRIF (Figure 8F). Similar results were also obtained for the ISG56 promoter and the NF-κB-dependent promoter (Figure S6). To further substantiate that the positive transcriptional effect of Sec16A and p115 is dependent on TRAF3, TRAF3-knockout MEF cells were transfected with p115 and Sec16A in the presence or absence of TRAF3 and used in the IFNβ promoter reporter assay. As suspected, the enhanced promoter activation, induced by ectopically expressed p115 and Sec16A, was entirely dependent on the presence of TRAF3 (Figure 8G). Thus, when expressed in relatively low amounts in Hec1B and MEF cells (not shown), p115 and Sec16A positively participate in a TRAF3-dependent type I IFN response, probably reflecting the ability of a subpopulation of cytoplasmic TRAF3 to further associate with the ER-to-Golgi components under these conditions of mild ectopic expression.
Interestingly, several recent studies have demonstrated that overexpression of p115 or Sec16A in highly transfectable cell lines and depletion of Sec16 or p115 resulted in identical cellular outcomes (i.e. Golgi fragmentation (see Figure S4 and Figure 6) and delayed ER-to-Golgi transport), thereby suggesting that they are required in stoichiometric amounts [24], [25], [52]. Thus, when ectopically expressed in high amounts in 293T cells, p115 and Sec16A were expected to blunt TRAF3-dependent transcriptional activation. Indeed, transfection of increasing amounts of p115 or Sec16A efficiently blunted TRIF-, RIG-I-, and MAVS-induced IFNβ promoter activation (Figure S7A–C) as well as NF-κB promoter activation (data not shown). Importantly, adding increasing amounts of TRAF3 in this specific reporter gene assay dose-dependently reversed the inhibitory effect of p115 and Sec16A, once more substantiating the relationship that exist between Sec16A, p115 and TRAF3 (Figure S7D). Moreover, transfection of these plasmids also blunted TBK1-induced ISRE promoter activation (Figure S7E), but did not affect the transactivation response induced by the use of a constitutively active form of IRF-3 (IRF3-5D) (Figure S7F), suggesting that the ER-to-Golgi compartment plays upstream of IRF-3 in type I IFN signalling.
To further confirm the implication of p115 and Sec16A in the type I IFN response, loss-of-function experiments were conducted next. As suspected, an RNAi approach targeting Sec16A and p115, which leads to Golgi fragmentation (see Figure S4 and Figure 6) significantly diminished Ifnb, ifit1 (ISG56), and oas1 mRNA induction following poly I:C and poly dA:dT transfection and SeV infection (Figure 9). To verify whether this approach affected IRF-3 activation and the induction of an IRF-3-dependent antiviral protein [53], we next verified the phosphorylation state of IRF-3 and the induction of ISG54 in HeLa cells expressing either shRNA duplexes targeting p115 and Sec16A or cells expressing a non-targeting (NT) shRNA duplex. The phosphorylation of IRF-3 and the expression of ISG54 were readily observed upon SeV infection, poly I:C and poly dA:dT transfection in HeLa cells expressing the NT shRNA duplex but was clearly reduced in cells expressing different shRNA duplexes targeting p115 (Figure 10A) and Sec16A (Figure 10B). Altogether, these data indicate that TRAF3 localization to the ER-to-Golgi vesicular pathway is necessary for a proper type I IFN response.
Gene disruption strategies have revealed that TRAF3 plays a major role in the type I IFN response [12], [13]. However, how TRAF3 assembles into functional signalling complexes is still not fully understood. In general, TRAF3 is thought to reside in the cytosol and translocate to surface membrane receptors upon engagement of CD40 or other TNFR family members [54]. Akin to its role in MyD88-dependent cytokine production and TRIF-dependent type I IFN production [55], TRAF3 conceivably also has the capacity to associate with endosomal compartments enriched in TLR3, TLR4, TLR7, TLR8 and TLR9 receptors [56]. Additionally, upon RLH activation TRAF3 interacts with MAVS and TRADD to trigger IRF-3 phosphorylation through the adaptor molecule TANK and the IKK-related kinases TBK1 and IKKi [20], [21]. However, how TRAF3 associates with MAVS upon RLH activation remains unanswered.
Herein, we report that TRAF3 localizes to the ER-to-Golgi compartments through its ability to interact with p115- and Sec16A-containing complexes. A pharmacological approach using the microtubule depolarizing agent nocodazole led to the redistribution of TRAF3 into small punctate cytoplasmic structures discrete from the ER along with both Golgi matrix proteins p115 and GM130. Both the structure and positioning of the Golgi apparatus have been shown to be highly dependent on the microtubule cytoskeleton [57]. Interestingly, a link between TRAF3 and the microtubule network has been already established in a previous study through its interaction with Microtubule-Interacting Protein that associates with TRAF3 (MIP-T3) [58]. TRAF3 was dissociated from this complex upon CD40L stimulation and, consequently, it was suggested that microtubule association of TRAF3 could be responsible for directing TRAF3 to defined membrane microdomains in the cell. A similar scenario is proposed here where, in response to viral infection, the association of TRAF3 with complexes containing p115 and Sec16A at the ER-to-Golgi vesicular pathway may play an important role in positioning TRAF3 with MAVS (see Figure 11). Indeed, the following findings suggests a role for Sec16A and p115 in the TRAF3-mediated RLH type I IFN response: (1) Sec16A and p115 are found in immunocomplexes containing TRAF3, but not TRAF2 or TRAF6; (2) inactivation of TRAF3 by deletion of its N-terminal RING finger domain and the C-terminal TRAF domain displaces TRAF3 from the ER-Golgi transport compartments; (3) in non-treated cells, TRAF3 colocalizes and tightly associates with p115, Sec16A, ERGIC53 and GM130, markers of the ER-to-Golgi vesicular compartment; (4) activation of the RLH pathway leads to reorganization of the Golgi apparatus into punctate structures containing TRAF3 and GM130; (5) an increased association between TRAF3, Sec16A, p115 and TBK1 is observed in virally-infected, dsRNA- and dsDNA-transfected cells; (6) mild overexpression of both proteins enhances SeV-, TBK1- and MAVS-stimulated IFNβ, ISG56 and NF-κB promoter induction; (7) knocking down the expression level of p115 or Sec16A affects the cellular distribution of TRAF3, impairs its capacity to associate with MAVS and diminishes the type I IFN response following poly I:C or polydA:dT transfection and SeV infection; and (8) enforced retention of TRAF3 at the ER-to-Golgi compartment by the addition of a COPI and COP II sorting signal peptide impairs TRAF3 recruitment to the cis-Golgi and diminishes the type I IFN response. Thus, we propose that these two trafficking proteins, Sec16A and p115, form a complex with TRAF3 at ER-to-Golgi transport compartments in order to ensure its proper recruitment to the mitochondrial network during a viral infection. Interestingly, enforced expression of Sec16A or p115 also increases TRIF-mediated IFNβ promoter activation, reinforcing the role for the ER-to-Golgi vesicular compartment in TLR3 and TLR4 signalling, as recently reviewed [59].
In support of our findings, the ER-to-Golgi transport compartment seems to also host several proteins involved in type I IFN signalling such as TRADD [21], the translocon [31] and potentially the exocyst [32] (Clement and Servant, unpublished observations). How these proteins cooperate with TRAF3 at the ER-to-Golgi transport compartments is currently unclear and will be the objective of future studies. Nevertheless, all these data suggest a model where vesicles and/or membranes originating from reorganized ER-to-Golgi compartments come in close proximity with the mitochondrial network in order to facilitate the assembly of a functional MAVS signalling complex.
In addition to its role in the RNA sensing pathways, STING is now considered an important effector of innate immune signalling in response to DNA pathogens [60]. Interestingly, STING is an ER-resident protein, which in response to dsDNA treatment, was recently demonstrated to traffic from the ER to the Golgi [61], [62] giving rise to punctate structure formation [62]. It is likely that the use of dsDNA (polydA:dT) used in our study might activate both the RNA-dependent pathway (through RNA polymerase III [63]) and the recently described DNA-dependent pathway (through IFI16 [2]), allowing TRAF3-loaded punctae to interact with both MAVS and STING respectively for proper innate immune signalling (Figure 11). Even though this needs to be investigated further, we speculate that the membranous network composed of the ER, Golgi and mitochondria provides a convenient platform on which antiviral cell-signalling complexes are arranged and optimally activated.
It is noteworthy that, as a common feature, plus-stranded RNA viruses have the ability to induce cytoplasmic membrane rearrangements that facilitate their replication. Consequently, the formation of these RNA replication complexes results in dramatic reorganization of the secretory pathway of host cells [64]. For example, poliovirus-infected cells accumulate membranous vesicles derived from COPII vesicles [65] whereas Kunjin virus induces “convoluted membranes” that contain markers from the ERGIC [66]. The precise role for this internal membrane rearrangement in the virus propagation and virus-host interaction requires further investigation. Nevertheless, localization of TRAF3 and TRADD to these vesicular transport compartments could represent a cellular strategy to increase the rate of RNA detection and the formation of an effective signalling complex at the mitochondrial membrane. The observation that TRADD translocates from the cytoplasm to the mitochondria during Influenza A virus infection supports this model [67]. Additionally, recent observations highlight the fact that viruses have evolved a variety of mechanisms involving the Golgi apparatus to specifically block TRAF3 recruitment into a functional signalling complex. Notably, the SARS Coronavirus M protein, a Golgi localized protein, was recently found to impede the formation of a TRAF3-TANK-TBK1/IKKi complex at the Golgi apparatus [68]. The NY-1 strain Hantavirus glycoprotein (Gn) was also shown to disrupt TRAF3-TBK1 interaction by interacting with TRAF3 through its cytoplasmic tail [69].
The notion of cellular proximity to favor exchanges and signalling events between organelles has been an intense field of interest for many years. Recently, mitofusin 2 present on the ER was shown to tether the ER to mitochondria in order to promote efficient Ca2+ uptake into the mitochondria for oxidative phosphorylation purposes. Interestingly, mitofusin 2 was also shown to inhibit RLH pathway signalling by interacting with the C-terminal of MAVS [70]. Furthermore, the Golgi localization of the glycolipid GD3 is important for its transport to the mitochondria after TNF-α stimulation [71], [72]. Membrane scrambling between Golgi and mitochondria following Fas stimulation is another example pointing to the connection between different cellular organelles [73]. Moreover, signalling at the Golgi apparatus and endosomes has been observed for different types of membrane-bound receptors [56], [74] and protein kinase cascades [75].
Although Bouwmeester and colleagues reported an NF-κB-inducing kinase-dependent interaction between Sec16A and NF-κB 2/p100 in an exhaustive study mapping the human TNF-α/NF-κB signal transduction network [76], a role for the ER-to-Golgi vesicular pathway in RLH-induced innate immune response was still unknown until now. Future characterization of the TRAF3 interactome will undoubtedly help to understand the molecular relevance of the specific subcellular localization of TRAF3 for an optimal type I IFN response.
Commercial anti-GM130 antibody was purchased from BD Transduction (San Jose, CA). The monoclonal anti-FLAG epitope (M2), the polyclonal anti-FLAG and the anti-β-actin (clone AC-74) were obtained from Sigma (Oakville, Ontario, Canada). The c-Myc (9E10) monoclonal antibodies, as well as the polyclonal p115 (H-300) and TRAF3 (C-20, H-20, and G-6) antibodies were purchased from Santa Cruz (Santa Cruz, CA). The anti-GFP (monoclonal 1218) antibody and the polyclonal goat anti-GFP antibody were obtained from ABCAM (Cambridge, MA) and US Biological (Swampscott, MA) respectively. The anti-ERGIC53 and anti-calnexin antibodies were from Enzo Life Sciences, anti-p-IRF3 Ser398 was from Millipore (Billerica, MA) and anti-ISG54 was from Novus Biologicals (Littleton, CO). The polyclonal anti-Sec16A and p115 antibodies were obtained from Bethyl Laboratories and Santa Cruz. The plasmid encoding for EGFP-Sec16A was a kind gift of Dr. David Stephens (University of Bristol, UK). Human TRAF3 and p115 cDNAs were amplified from the MGC bank collection and respectively subcloned in pcDNA3 and pTag2B (FLAG) or pTag3B (Myc) vectors (Invitrogen, Burlington, ON, Canada). Human TRAF6 cDNAs were purchased from Origene (Rockville, MD) and subcloned in the pTag2B/3B vectors. The pFLAG-CMV2-TBK1 and pFLAG-TRAF3 Y440/Q442A were gifts from Drs. John Hiscott (McGill University). pcDNA3.1-FLAG-MAVS construct was from Rongtuan Lin (McGill University). The pcDNA3-His-TRIF construct was from Dr. Daniel Lamarre (Université de Montréal). The pRK5-TRAF2-FLAG was obtained from Dr. Nathalie Grandvaux (Université de Montréal). pFLAG-CMV2 TRAF3 deletion mutants (1–117, 1–381, 114–568, 259–568 and 389–568) were from Dr. Carl Ware (La Jolla Institute for Allergy and Immunology). The IFNß reporter plasmid, pGL3-IFN-ß-LUC was described previously [77] as well as the ISG56-luciferase [78] and the NF-κB p2(2)TK reporter plasmids [77]. The pFLAG-TRAF3 mutant with C-terminal retention motif AKKFF was generated by PCR and subcloned in pCDNA3.1 (+) and pMRX-ires-puro (a kind gift from Dr. Shoji Yamaoka, Tokyo Medical and Dental University, Japan). Poly I:C was purchased from GE HealthCare (Waukesha, WI) and transfected with Lipofectamine2000 (Invitrogen) at final concentrations of 1.0 to 2.5 µg/ml. Poly dA:dT was from InvivoGen and used at 1 µg/ml. BFA and nocodazole were obtained from Calbiochem and used at a final concentration of 5 µg/ml.
HeLa, Hec1B, HEK 293, HEK 293T, HEK 293 QBI cell lines and TRAF3 knockout MEF cells (a kind gift from Dr. John Hiscott, McGill University) were maintained in Dulbecco's modified Eagle Medium supplemented with 10% fetal bovine serum. All DNA transfections in human cell lines were performed with Lipofectamine 2000 (Invitrogen) according to the manufacturer's protocol. Transient transfection of immortalized MEF cells was performed by microporation with the Microporator Apparatus (Montreal Biotech) according to the manufacturer's instructions. Sendai virus (SeV) was obtained from Specific Pathogen-Free Avian Supply (North Franklin, CT) and used at 200 HAU/ml. Respiratory Syncytial Virus (RSV.A2) (a kind gift from Nathalie Grandvaux, Université de Montréal) was used at a MOI of 3. Influenza A (PR8) virus was a kind gift from Dr. Rongtuan Lin (McGill University).
Preparation of whole cell extracts, co-immunoprecipitation studies, Native-PAGE and immunoblot analysis were performed as described previously [79]. A RIPA buffer (50 mM Tris-HCl, pH 7.4, 100 mM NaCl, 5 mM EDTA, 50 mM sodium fluoride, 40 mM β-glycerophosphate, 1 mM sodium orthovanadate, 1% Triton X-100, 0.1% SDS, 0.5% sodium deoxycholate, and protease inhibitors mixture (Sigma)) was used for the extraction of the TRAF3 AKKFF mutant. Antibodies were used as recommended by the manufacturers.
For immunofluorescence, cells were fixed with 4% paraformaldehyde (PFA) in PBS for 20 min followed by permeabilization with 0.1% Triton X-100 for 5 min. Cells were washed with PBS (pH 7.2) and blocked with 0.5% BSA in PBS. Anti-FLAG antibody (M2, Sigma) was used at 1∶1000, anti-GM130; 1∶100, anti-ERGIC53; 1∶100, anti-FLAG polyclonal antibody; 1∶400, anti-GFP (ABCAM); 1∶100, anti-Myc 9E10; 1∶100, anti-TRAF3; 1∶200, anti-P115; 1∶100, anti-Sec16A; 1∶200, and anti-MAVS; 1∶100. Secondary fluorophore-conjugated antiserum (Alexa Fluor 488 and 564) was obtained from Molecular Probes (Eugene, OR) and used at 1∶500 in PBS 0.5% BSA. The nucleus was revealed by 4′,6-diamidino-2-phenylindole (DAPI) staining. The confocal micrographs represent a single optical section through the plane of the cell. Images were acquired with LSM v3.2 software (Zeiss) on a LSM 510 inverted microscope (Zeiss, Germany) with a plan-apochromat 63×/1.4 oil disc lens using 405 nm in conjunction with a LP 505 for DAPI, 488 nm in conjunction with a BP 505–530 for Alexa 488, and 543 nm in conjunction with BP 560–615 for Alexa 568. Images were assembled in Adobe Photoshop CS 3.0.
FLAG-affinity purification was performed as described previously [80] with the following modifications. Detergent concentration in the lysis buffer was 0.5% NP-40; the lysis buffer was added at 4 ml/g wet cell pellet, and cells were subjected to passive lysis (30 minutes) followed by one freeze-thaw cycle and centrifugation. Immunoprecipitation was performed on the cleared lysate by adding 25 µl packed FLAG M2 beads (Sigma) and incubating for two hours. Beads were washed three times in lysis buffer, and three times in 50 mM ammonium bicarbonate. Samples were eluted with ammonium hydroxide, lyophilized in a speed-vac, resuspended in 50 mM ammonium bicarbonate (pH 8–8.3), and incubated at 37°C with trypsin overnight. The ammonium bicarbonate was evaporated, and the samples were resuspended in HPLC buffer A2 (2% acetonitrile, 0.1% formic acid), then directly loaded onto capillary columns packed in-house with Magic 5 µm, 100A, C18AQ. MS/MS data was acquired in data-dependent mode (over a 2 hr acetonitrile 2–40% gradient) on a ThermoFinnigan LTQ equipped with a Proxeon NanoSource and an Agilent 1100 capillary pump. Acquired RAW files were converted to mgf format using ProteoWizard. The searched database was human RefSeq (version 45). *.mgf files were searched with the Mascot search engine (version 2.3) using the following variable parameters: semi trypsin digestion, one missed cleavage allowed, asparagine deamidation and methionine oxidation. The fragment mass tolerance was 0.6 Da (monoisotopic mass), and the mass window for the precursor was +/−3 Da (only +2 and +3 charge ions were processed). Mascot results were parsed for further analysis into a LIMS system developed at the Samuel Lunenfeld Research Institute [81]. Scoring of specific interactors for FLAG-TRAF3 and FLAG-p115 was performed using the statistical tool SAINT (Significance Analysis of INTeracome). SAINT converts label free quantification, such as spectral counts, for each prey protein identified in a purification of a bait into the probability of true interaction between the two proteins [82], [83]. SAINT can calculate a probability of interaction even for proteins proteins frequently detected in AP-MS experiments, providing that a quantitative enrichment is detected in the purification of the sample [84]. For each bait, two biological replicates were used. Twelve negative control runs (consisting of cells expressing the FLAG tag alone) were processed in parallel and combined into 5 virtual controls for SAINT modeling. SAINT calculates scores differently depending on the availability of negative control purifications, and thus the implementation for spectral count data incorporating control purification data was used (details are described in [82]). The probability score was first computed for each prey in independent biological replicates separately (iProb). Then the final probability score for a pair of bait and prey proteins was calculated by taking the average of the probabilities in individual replicates (AvgP); final results with AvgP≥0.5 were further inspected. A manual cross-reference against a database containing >1000 independent FLAG AP-MS runs was finally performed to identify potential proteins frequently detected in AP-MS experiments and were removed from the final dataset.
HeLa cells were transfected with 40 nM siRNA using Lipofectamine2000 (Invitrogen). siRNA p115, Sec16A and the non-targeting pool siRNA duplexes were purchased from Dharmacon (Lafayette, CO). Sequences are as follow: Sec16A (#3: 5′-ggagagcuuucgcgcugua-3′; #4: 5′-ccucaguccucuagcgugu-3′) and p115 (#3: 5′-guuauuauguggagguuug-3′; #4: 5′-ugauggagguauaguaguu-3′). shRNA vectors targeting p115 (TRCN0000065070, TRCN0000065071, TRCN0000065072) and Sec16A (TRCN0000246015, TRCN0000246016) and non-Targeting control shRNA were purchased form Sigma (St. Louis, MO). Lentiviral vector production and transduction was conducted as described previously [85].
After stimulation, total RNA was extracted from HeLa cells using Trizol reagent (Invitrogen). 2 µg of RNA was reverse transcribed using the High Capacity cDNA Reverse Transcription Kit with random primers (Applied Biosystems) as described by the manufacturer. SYBR green PCR reactions were performed using 2 µl of cDNA samples (25–50 ng), 5 µl of the Fast SYBR qPCR Master Mix (Applied Biosystems) and 10 pmol of each primer in a total volume of 10 µl. The IFN qRT-Primer set for real-time quantification of the IFN response (IFNβ, ISG56 (ifit1) and OAS1) was purchased from InvivoGen (San Diego, CA). The ABI PRISM 7900HT Sequence Detection System (Applied Biosystems) was used to measure the amplification level. All reactions were run in triplicate and the average Cts were used for quantification. TBP (TATA binding protein) was used as endogenous control.
Subconfluent Hec1B, HEK 293 QBI and 293T cells in 24 well-plates were transfected with 25 ng of pRL-TK reporter (renilla luciferase for internal control) and 125 ng of pGL3-IFN-β-LUC, pGL3-ISG56-LUC, pGL3-ISRE or pGL3-NF-κB-LUC using the conventional CaPO4 transfection protocol (for Hec1B, HEK 293 QBI cells and 293T cells) or Lipofectamine 2000 (for TRAF3 knockout MEF cells). Cells were harvested 24 h post-transfection, lysed in passive lysis buffer (Promega, Madison, WI), and assayed for dual-luciferase activity using 10 µl of lysate according to the manufacturer's instructions. All firefly luciferase values were normalized to renilla luciferase to control for transfection efficiency.
Statistical analyses were performed using GraphPad Prism version 5.0 for Mac (GraphPad Software, San Diego, CA). Comparison of two groups was carried out using a two-tailed unpaired t-test, and comparison of more than two groups was carried out with one-way ANOVA and a Bonferroni posttest. Statistical significance was accepted at a P-value below 0.05.
TRAF2; 7186, TRAF3; 7187, TRAF6; 7189, Sec16A; 9919, p115; 8615, MAVS; 57506, STING; 340061, ISG56; 3434, ISG54; 3433, TRIF; 148022, TBK1; 29110, RIG-I; 23586, MDA5; 64135, LGP2; 79132, AIM2; 9447, IFI16; 3428, IRF3; 3661, NIK; 9020, TRADD; 8717, TANK; 10010, IKKi; 9641, GM130; 2801, RIP1; 8737, IFNβ: 3456, NFκB: 4790, OAS1; 4938, ERGIC53; 3998, calnexin; 821, Sec61β; 10952; Sec5; 55770.
|
10.1371/journal.pbio.1000494 | Chromosomal Redistribution of Male-Biased Genes in Mammalian Evolution with Two Bursts of Gene Gain on the X Chromosome | Mammalian X chromosomes evolved under various mechanisms including sexual antagonism, the faster-X process, and meiotic sex chromosome inactivation (MSCI). These forces may contribute to nonrandom chromosomal distribution of sex-biased genes. In order to understand the evolution of gene content on the X chromosome and autosome under these forces, we dated human and mouse protein-coding genes and miRNA genes on the vertebrate phylogenetic tree. We found that the X chromosome recently acquired a burst of young male-biased genes, which is consistent with fixation of recessive male-beneficial alleles by sexual antagonism. For genes originating earlier, however, this pattern diminishes and finally reverses with an overrepresentation of the oldest male-biased genes on autosomes. MSCI contributes to this dynamic since it silences X-linked old genes but not X-linked young genes. This demasculinization process seems to be associated with feminization of the X chromosome with more X-linked old genes expressed in ovaries. Moreover, we detected another burst of gene originations after the split of eutherian mammals and opossum, and these genes were quickly incorporated into transcriptional networks of multiple tissues. Preexisting X-linked genes also show significantly higher protein-level evolution during this period compared to autosomal genes, suggesting positive selection accompanied the early evolution of mammalian X chromosomes. These two findings cast new light on the evolutionary history of the mammalian X chromosome in terms of gene gain, sequence, and expressional evolution.
| Some evolutionary theories predict that the X chromosome will be enriched for genes with male functions. However, recent studies showed there had been gene traffic in which autosomal male-biased genes were retroposed from X-linked parental genes. A question remains about whether this pattern also holds for all types of new genes. Herein, using comparative genomic analysis, we dated all human and mouse genes to the vertebrate phylogenetic tree. We found that the X chromosome evolved with two bursts of gene origination events. The recent burst includes mainly male-biased genes in contrast to older X-linked genes that are often female-biased in expression. Meiotic sex chromosome inactivation contributes to this dynamic since it silences the older but not the younger X-linked genes. The older burst was after the split of eutherian mammals and the marsupial opossum, and the genes from this burst were quickly incorporated into transcriptional networks of multiple tissues, especially in the brain. The transcriptional expansion, together with the rapid protein evolution of the preexisting old X-linked genes, suggests that positive selection was acting in the early evolution of the mammalian X chromosome. These two lines of findings revealed extensive gene evolution in the mammalian X chromosome.
| In mammals and Drosophila, the X chromosome usually differs dramatically from autosomes since it is hemizygous in males [1]. Sexual antagonism (beneficial for one sex, but deleterious for the other) enriches male-biased genes on the X chromosome, if alleles are generally recessive, and on the autosome if they are generally dominant [2]–[3]. On the other hand, inactivation of the X chromosome during spermatogenesis [4]–[5] drives the accumulation of male-biased genes on the autosomes where they can be expressed in the meiotic or post-meiotic phase [6]–[7]. These two processes can explain the gene traffic between the X and autosomes in Drosophila [8] and mammals [9]–[10] as well as the excess of male-biased genes on the autosomes [11]–[12].
However, recent analyses of male-biased genes identified several X-linked genes that originated in the last 1–3 million years (myr) in Drosophila [13]–[15]. Whether or not these data implicate an effect of evolutionary time on the chromosomal location of male-biased genes remains unknown. In our investigation of how the various evolutionary forces impact the chromosomal distribution of sex-biased genes, we focused particularly on how the age of genes affects their chromosomal locations. By dating when genes arose in humans and mouse, we found male-biased genes were distributed at different locations in different phases of mammalian evolution: young male-biased genes are enriched in the X chromosome, but older male-biased genes favor autosomal locations. Interestingly, this redistribution seems to be associated with feminization of the X chromosome with more X-linked old genes expressed in ovaries.
Besides the recent gene gain contributed by emergence of male-biased genes on the X chromosome, we found another burst of gene gain on X chromosome immediately after the divergence of opossum and eutherian mammals. Accelerated protein evolution and transcriptional evolution of X-linked genes reveal positive selection occurring in this period. These data support the recent notion [10],[16] that our X chromosome originated in the therian ancestor instead of the common ancestor of all mammals.
These two lines of findings significantly extend our knowledge of the origination and evolution of X chromosomes in mammals.
We inferred the origination times of genes based on the presence and absence of orthologs in the vertebrate phylogeny and assigned 19,935 human and 21,122 mouse protein-coding genes into different evolutionary branches (Figure 1; Table S1, S2; Materials and Methods). We found that 1,828 human genes are primate-specific (branches 8 to 12 of Figure 1A) and 3,111 mouse genes are rodent-specific (branches 8 to 11 in Figure 1B) [17]–. In subsequent analyses, except if specified elsewhere, we define them as young genes and the remaining as old genes.
Compared to previous reports [10],[25], our method identified young genes more conservatively. For example, Church et al. identified up to 2,941 primate specific genes, considerably more than we found [25]. Also, for the 67 human genes that intersect between our dataset and [10], we assigned 44 (66%) genes onto the same phylogenetic branch as they did. For the remaining 23 cases in conflict, we assigned 22 to older branches (Table S3) since we used a larger number of outgroup species.
We tracked the relative gene abundance of individual chromosomes across 450 myr and identified two bursts of genes occurring on the X chromosome (Figure 2). One burst (branches 5–7) postdated the divergence of eutherian mammals (human or mouse) and marsupials (opossum) and the other occurred recently after the split of human and chimp and after the split of mouse and rat, respectively. For both peaks, the X chromosome contributes to 8%∼14% of genes, while it only accounts for 3% of genes in the first 300 myr of vertebrate evolution. In contrast, autosomes tend to vary less in their relative contribution to the whole genome (Figure S1). As the major contributor generating new genes, DNA-level duplication accounts for 73∼95% of genes of these two peaks. If we only use DNA-level duplicates, the pattern remains the same.
Considering that many more genes arose in branch 5 compared to branch 6 or 7 (1,200∼1,400 versus 400∼500, Figure 1), the old peak seems to be best explained by the hypothesis that the X chromosome emerged in the therian ancestor and subsequently recruited many genes in an accelerated evolution of sex-related functions, as found with retrogene-based chromosomal movement studies [26]. In contrast, the recent burst reveals a rapid addition of new genes into the mammalian X chromosome, which may be independent of major chromosomal changes.
Based on human body index data (GSE7307, Materials and Methods) and mouse tissue profiling data [27] at the NCBI GEO database [28], we identified genes with sex-biased expression (Materials and Methods). As shown in Figure 3, both human and mouse demonstrate a similar pattern regarding the proportion of male-biased genes and the age of the branch in which they arose.
For younger branches (less than 50 myr), male-biased genes are enriched in the X chromosome compared to autosomes (∼50% versus ∼30%, Chi-square test p<0.05), which might be driven by fixation of recessive male-beneficial alleles under sexual antagonism. This pattern decreases for genes originating in earlier branches. Male-biased genes older than 300 myr are overrepresented on the autosomes (∼30% versus ∼15%, p = 1×10−9). This pattern was independently supported by an Affymetrix exon array panel with larger coverage of new genes (Figure S2). Thus, the recent peak observed in Figure 2 could be attributed to a burst of male-biased genes on X chromosome younger than 50 myr. Figure 3 also demonstrates that the X chromosome consists of a similar or even higher proportion of male-biased genes compared to autosomes from 90 myr ago (branch 7) to 130 myr ago (branch 5). Thus, many of the genes gained in the first, older peak may also have male-biased expression.
Notably, the proportion of female-biased genes on branch 5 was greater on the X chromosome compared to autosome (39% versus 20% in Table 1). In contrast, for branches 6 and 7, the proportion of female-biased genes is around 15% for both the X chromosome and autosomes (Table S4). Again, this suggests that the newly originated X chromosome was subjected to enhanced positive selection and recruited an excess of both male- and female-biased genes.
The earlier peak in Figure 2 indicates the mammalian X chromosome emerged before the divergence of eutherian and marsupial [10]. Thus, the nascent X chromosome changed remarkably, gaining an excessive number of genes. If this scenario is true, those preexisting genes on the ancestral X chromosome might have accumulated many evolutionary changes during this period (branch 5), as did genes linked to the neo-X chromosome in Drosophila [29]. That means we would expect these ancient genes on the X chromosome to show signatures of positive selection. To test this scenario, we investigated the evolutionary path of ancient genes shared by vertebrates by comparing the ratio between non-synonymous substitution rate and synonymous substitution rate (Ka/Ks) (Materials and Methods). In other words, we compared the Ka/Ks of X- and autosomal-linked old genes in separate evolutionary periods. Across evolution of 450 myr, the X chromosome did not show significantly higher Ka/Ks except in branch 5 (Table 2), which strongly corroborates the hypothesis that the X chromosome did not acquire sex-chromosome status until this period.
We extended this analysis to genes gained since branch 5. We directly estimated the proportion of replacement substitutions (α) based on polymorphism and divergence data in [30] and a maximum-likelihood method implemented in the DoEF package [31]. As shown in Table S5, young genes generally show higher α compared to old genes, and X-linked male-biased genes show the highest α, 0.501. This pattern shows that positive selection instead of neutrality drives the evolution of X-linked genes arising since branch 5, especially those with male-biased expression.
However, positive selection of nucleotide substitutions can only suggest that initial fixation may also be driven by positive selection. More direct evidence comes from copy number polymorphism (CNP) data in Drosophila, which showed that the X chromosome is subject to stronger purifying selection than autosomes [32]. In human, it was also noted that the X chromosome shows a paucity of CNPs [33]. Together with bursts of adaptive fixations occurred on the neo-X of Drosophila [29], it is likely that positive selection instead of drift accounts for two bursts of genes on the X chromosome.
As we noted before, enrichment of young male-biased genes on the X declines for those originating in earlier evolutionary branches. Using expression data from mouse spermatogenesis, we compared different age groups to investigate which force underlies such a demasculinization process (Table 3). As previous studies such as [7] found, old genes are expressed more in the pre-meiosis stage (spermatogonia) but are silent from meiosis (pachytene spermatocyte) to post-meiosis (round spermatid). In terms of whole testes, however, old X-linked genes are underrepresented (Table 3). New genes show a distinct pattern: while often expressed in spermatogonia, they are not silent in meiosis. Moreover, a much greater proportion of new genes on the X are expressed in the post-meiosis stage compared to genes on the autosome (70% versus 27%, Chi-square test p = 5×10−10). This is consistent with a previous observation of X-linked postmeiotic multicopy genes [34], the vast majority of which we found were very young (Materials and Methods). Such a pattern suggests that the young X-linked genes are not affected by MSCI. An independent microarray dataset of mouse spermatogenesis [35] confirms high expression of X-linked young genes in spermatid (Figure S3). In addition, we note that the customized array by Khil et al. was comprised mainly of old and conserved genes, with only 1.7% of the set being young genes. In contrast, the Affymetrix array data [36] we used covered 14,923 Ensembl genes, 3.9% of which are young genes.
This striking contrast between young and old genes suggests that MSCI plays an important role in determining the age-dependent chromosomal distribution of male-biased genes. In order to investigate how this contrast occurred in such a short time, we analyzed four major cell types including sertoli cells, spermatogonium, spermatocyte, and spermatid between mouse [35] and rat [37]. We used the Euclidean distance of relative abundance (RA) to measure how orthologous genes have diverged in their expression (Materials and Methods). Consistent with a previous comparison of human and chimpanzee [38], the testis expression of genes on the X chromosome diverge more between rat and mouse than genes on autosomes (Wilcoxon rank sum test p = 4×10−6, Figure 4). Furthermore, X-linked young genes show significantly higher divergences, compared to all other three groups (p<0.05).
While we found that expression in various spermatogenesis stages is generally conserved [35] with only about 3% divergence (Figure S4), X-linked young genes show the largest expression divergence in spermatid. Specifically, after the split of mouse and rat 37 myr ago [39], young X-linked genes show 6.9% divergence in spermatid, which is much higher than the genomic average for spermatid, 3.3% (Wilcoxon rank sum test p = 0.002). This increased divergence suggests that, although these genes seem to escape MSCI and preferentially transcribe in post-meiosis, the expression profile is not conserved. It remains unknown whether these genes get up-regulated or down-regulated in one species. But if the latter case were true, it indicates that the high post-meiotic expression would be silenced by MSCI in later evolution. This could also explain how the different pattern between young and old genes in Table 3 is achieved.
We investigated the distribution of female-biased genes on chromosomes and its correlation with gene ages. Interestingly, female-biased genes are distributed in a pattern symmetrical to male-biased genes (Figure S5 versus Figure 3): the old X-linked genes are more often female-biased, while young genes are not.
We characterized ovary expression of genes using the Affymetrix mouse exon array panel data. Consistent with Figure S5, ovary expression also depends on the age of the gene's origination. Specifically, young autosomal genes show significantly higher expression in ovaries than young X-linked genes (Wilcoxon rank sum test p = 5×10−12, Figure 5). However, old X-linked genes generally show higher expression in ovaries (p = 5×10−7). Thus, as gene age increases, this expressional excess of autosomal genes reverses and older X-linked genes show significantly higher expression in ovaries.
It can be argued that such an age-dependent pattern of expression is not a specific property of ovary evolution and other organs might also show a similar pattern. To test this possibility, we investigated gene expression in the major organs: brain, heart, kidney, liver, lung, muscle, spleen, and thymus. All these tissues, except for brain, showed a significant excess of expression for new genes (branch≥5) on autosomes compared to that of X-linked genes (Wilcoxon rank sum test p<0.01, Figure S5). However, for old genes (branch≤4), they are evenly distributed (p>0.05).
The brain shows a unique pattern. Young genes (branch>7) are relatively abundant on autosomes (p = 0.001, Figure S5), but old genes (branch≤7) are overrepresented on the X chromosome (p≤0.01). This is consistent with previous findings that X chromosome is enriched with genes expressed in brain [1],[40]. Notably, different from ovaries, enrichment in the brain did not show clear age dependence, since genes originating from branches 5 to 7 presented the most significant excess (Figure S6).
The coincidence that the X chromosome is enriched with both ovary-expressed and brain-expressed genes occurring in branch 5 (Table 1; Figure S5) motivated us to perform more thorough transcriptional profiling to get a more complete picture of how genes from this evolutionary period are transcribed.
We investigated mouse exon atlas data (GSE15998) to ask whether X-linked genes are more frequently expressed in the tissue of interest across different age groups. We clustered tissues by the proportion of X-linked genes expressed versus the proportion of autosomal genes expressed and identified three major groups: nervous system, testes, and all other tissues (Figure 6). Remarkably, the X-linked genes originating in branch 5 are transcriptionally permissive with a larger proportion of them expressed in many tissues compared to autosomal genes. This excess is most pronounced for brain samples.
Consistently, human data revealed that a greater proportion of X-linked genes emerging on branch 5 are expressed more widely than autosomal genes originating in this period, which is strongest for the brain (Figure S7). Since human and mouse share a similar pattern, parsimony suggests this striking transcriptional pattern of branch 5 derived genes is ancestral. Notably, none of these genes show sex bias in human brain profiling data [41], which suggests they might be important for both sexes.
We have described evolutionary patterns of protein-coding genes, which could be driven by natural selection in various forms like sexual antagonism or MSCI. If, however, such a pattern is a product of some mutational bias of gene origination, we would not detect similar evolutionary patterns in non-coding RNA genes, such as X-linked miRNAs. Therefore, we investigated the chromosomal distribution of miRNA genes annotated in miRBase [42] and found that miRNA duplicates are distributed in a pattern similar to that observed for protein-coding genes (Table S6). Specifically, both human and mouse show significant miRNA gene gain in branches 5 to 7 compared to the proportion of all miRNA genes (18∼22% versus 10∼13%, Fisher's Exact Test p<0.05). Moreover, they also show an excess for the youngest branch. Although it is not significant for the human data due to small sample size, it is for mouse (p = 0.02).
Like protein-coding genes, a larger proportion of X-linked miRNAs originating in branch 5 are transcribed in nine tissues (statistically significant for six of them) surveyed on Agilent chip [43] compared to autosomal genes (Table S7; Materials and Methods). Moreover, semi-quantitative PCR data of X-linked miRNAs in 12 tissues [44] show 9 out of 13 (69%) young genes are expressed higher in testes than at least six non-testis tissues. However, this percentage drops to 23% for old X-linked genes (9 out of 39, Fisher's Exact Test p = 0.005). Consistent with protein-coding genes, these data also show that old genes have moderate or high expression in ovaries and the young genes show only trace levels of expression (Wilcoxon rank sum test p = 0.01).
The age-dependent locations and expression profiles of miRNAs support that it is evolutionary forces, rather than some mutation bias intrinsic to a certain type of gene, which account for the dynamics of X-linked gene evolution.
It is known that the X chromosome can be divided into five evolutionary strata because of step-wise repression of recombination [45]–[47]. The X-conserved region (XCR) consists of the oldest strata 1 and 2, while the X-added region (XAR) includes younger strata 3 that is shared by primates and rodents, and much younger 4 and 5 that were derived within primates [46],[48]. Since sexual antagonism or other sex related forces like the faster-X process (see Discussion) depends on hemizygosity of the X chromosome in male, we expect the accordance between bursts of gene gain with the formation of corresponding strata. If these forces shape the evolution of gene content on the X chromosome, we should find that X-linked genes originating at a given time period should accumulate only in the strata already formed at that time. In other words, we should find a correlation between the ages of genes and the strata in which they are located. Consistent with these predictions, Figure 7 shows that the older strata 1 to 3 are associated with relatively older genes, while strata 4 or 5 are enriched with younger genes (one sided Fisher's Exact Test p = 0.03). This finding parallels the temporal correspondence between the occurrence of strata and the out-of-X retrogene traffic [49].
Our analyses demonstrated that the X chromosome evolved dramatically on both the sequence and expression levels after the split of eutherian mammal and marsupials. Specifically, the X chromosome showed a burst of gene gain during this time, and many of these genes quickly invaded the transcriptional network of various tissues, especially the brain. Furthermore, genes predating the birth of the X showed rapid protein-level evolution. A straightforward interpretation is that the newborn mammalian X was subjected to strong positive selection similar to the neo-X chromosome in Drosophila [29]. Moreover, the X-linked genes arising in branch 5 seem to have played important roles, as shown by their broad expression. Their transcription pattern suggests that the early evolution of placental mammals was associated with rapid changes in the brain. Furthermore, analysis of gene ontology showed that many of these genes mainly played regulatory roles in transcription and metabolism (Table S8). Thus, regulatory change contributed by gene gain on the X chromosome was extensively involved in the initial evolution of eutherian mammals. The fact that this peak ranges between branches 5 and 7 suggests remodeling of incipient X chromosome might take about 90 myr (−160∼−70 myr, Figure 2), which is consistent with one report based on retrogene movement [26]. However, the selective pressures driving this dramatic change in branch 5 appear to be smaller in subsequent branches (Table 2).
Our analyses reveal chromosomal redistribution of X-linked male-biased genes. Sexual antagonism may contribute to the initial fixation of X-linked recessive alleles as described previously [2],[7]. The faster-X hypothesis was initially proposed to fix more mutations on the X chromosome only if they are recessive and beneficial [1]. Recently, it was observed that this force was most pronounced for male-biased genes [50]. This suggests that the faster-X process could also be involved in the emergence of young X-linked male-biased genes, as the hypothesized sexual antagonism might.
These young X-linked male-biased genes could be later silenced by MSCI as suggested by Table 3, Figure 4, and Figure S4. At least two processes could be involved in this switch. First, we found a statistically significant excess of male-biased retrogenes generated in the X→A movement process and X-enrichment of the female-biased parental genes for both human and mouse (Table S9). Thus, the demasculinization and feminization of the X chromosome could be coupled in retrogene traffic. Moreover, our RA analysis (Figure S6) extends the out-of-the-testes hypothesis [51] to non-retroposed new genes. We found that new genes generally acquire transcription in more tissues during evolution although they are initially enriched in testes. With increasing MSCI and expanding expression breadth, X-linked male-biased genes might become unbiased or even female-biased as Figure S6 shows.
If new strata on the X chromosome represent regions that did not develop recombination repression until recently, the genes encoded in these regions will often escape MSCI [45]. Thus, it is expected that the X-linked male-biased genes more likely escape MSCI when located on young strata or pseudoautosomal regions (PARs). However, out of 13 young male-biased genes in humans, the relatively young strata 4 and 5 encode only one (Table S10), which does not significantly differ much from the expected number based on its genomic size. How then did the remaining 12 genes, those situated on older strata, escape from MSCI?
It was proposed that the excess of inverted repeats (IRs) encoded by human and mouse X chromosome could protect genes contained by these IRs from MSCI [52]. IRs suppress MSCI through formation of cruciforms or other unusual chromatin structures. Moreover, cancer/testis (CT) genes that are often expressed in normal testes and in cancerous tissues frequently overlap with IRs [52]. Given that X-linked CT genes underwent recent expansion [53], it is not surprising that some of them could form highly homologous IRs. In fact, 8 out of 13 young X-linked male-biased genes are CTs (Table S10). Thus, the high IR abundance on the mammalian X chromosome might be one reason that these genes can be transcribed in meiosis or postmeiosis.
Furthermore, out of 12 genes encoded by PARs and covered by unique probes (Table S11), there is only one (8%) male-biased gene, PPP2R3B, which is shared by human and mouse. Thus, different from our intuition, PARs do not harbor an excess of male-biased genes compared to the remaining strata (18%) and to autosomes (24%). Albeit of small sample size, this observation suggests that sex-related forces like sexual antagonism or faster-X process account for the observed excess of young X-linked male-biased genes.
There are only limited number of genes with unique probes on strata 4 (five) and 5 (eight). For the remaining strata, stratum 3 is enriched with male-biased genes, which is much higher than stratum 1 (27% versus 17%, one-sided Fisher's Exact Test p = 0.02) and stratum 2 (27% versus 15%, p = 0.03). This pattern suggests that stratum 3 recruits more young male-biased genes and there was not enough evolutionary time to be feminized as occurred in the oldest strata 1 and 2.
As shown in Figure 2, the emergence of young male-biased genes peaks in recent evolution of human and mouse. However, this peak started 30 myr ago (before the divergence of mouse and rat) in the rodent lineage, while the peak appeared in the last 5 myr in human lineage.
This difference is consistent with the fact that the mouse X encodes more young male-biased genes than the human X. Specifically, male-biased genes account for 52% and 74% of X-linked young genes in human and mouse, respectively (Figure 3; one sided Fisher's Exact Test, p = 0.07). Exon array data are similar (Figure S2; 45% versus 76%, one sided Fisher's Exact Test, p = 2×10−8). Origination of significantly more male-biased young genes suggests that stronger positive selection acts on rodents and could explain why the recent peak of gene gain (Figure 2) began earlier in the mouse lineage than in the human.
We downloaded Ensembl [54] release 51 (November, 2008) as the basic gene dataset for our analyses. We used MySQL V5.0.45 to organize the data, BioPerl [55] and BioEnsembl [56] to fold the pipeline, and R V2.8.0 [57] to perform all statistical analyses.
We developed a genome-alignment based pipeline to infer the origination time of a given genomic region by modifying a previous gene-alignment based method [58]. We analyzed UCSC [17] netted chained file for human (hg18) and mouse (mm9) to verify whether a given human/mouse locus has a reciprocal syntenic alignment in the outgroup genome such as chimpanzee, rat, chicken, and so on. In other words, we investigated whether a best-to-best match could be found between human/mouse loci and outgroup loci regardless of chromosomal linkage. In this way, we can identify orthologous genes; even those with different chromosomal location due to fusions or translocations such as those found in XAR region will be identified as well. Then, in order to handle occasional sequencing gaps, we scanned multiple outgroups and assigned this locus to a specific branch by following a parsimony rule. Compared to the previous method [58], our strategy is independent of gene annotation of outgroups and robust with gene translocation. Thus, we generated a more stringent young gene dataset (as described in the Result section). And, as Figure S8 shows, we have not assigned most genes encoded by XAR as young genes simply because this region changed the linkage by fusing to X chromosome. Conversely, several genes originated in branch 5 are located in strata 1 and 2 that are not XAR (Figure S8), also supporting that our pipeline is robust with gene translocations.
Notably, for regions without reliable synteny, our method might not work. This situation would be most pronounced for telomeres, which tend to be repetitive and prone to recombine [59] and thus have very limited synteny. For example, we dated 17 genes situated on PARs of the X chromosome (Table S11). For three genes encoded by PAR2, repeats contribute less than 16% of the gene loci based on UCSC annotation [17]. Accordingly, our age assignments for these three genes are always consistent with those inferred by tree reconstruction provided by Ensembl [54]. In contrast, for 14 genes linked with PAR1, repeats are prevalent with a median contribution of 55% to the gene loci. In this case, our results are consistent for only three out of nine cases with Ensembl age information.
We slightly modified the previous pipeline [58],[60]–[61] and classified young genes as DNA-level duplicates, RNA-level duplicates (retrogenes), and de novo genes. Briefly, we performed all-against-all BLASTP search for human and mouse proteins. It was reported previously that retrogenes can recruit other neighboring genome regions with introns after being retroposed [51]. Thus, in order to define a new gene as retrogene, we requested that in the aligned region between the most similar paralog (candidate parental gene) and child genes, the former contain at least one intron and the latter to be intronless. Otherwise, it will be classified as DNA-level duplicates. Notably, if there is no hit with BLAST evalue cutoff 10−6 found [58] and no annotated paralog by Ensembl [54], the gene will be defined as de novo.
In order to avoid non-specific probes and to cover more recently annotated genes, we used the customized array annotation files (released on November, 2008) downloaded from University of Michigan [62], HGU133Plus2_Hs_ENSG (Affymetrix Human 133 plus 2) and Mouse4302_Mm_ENSG (Affymetrix Mouse Genome 430 2.0 Array) for human and mouse, respectively. For exon array analysis, we used HuEx-1_0-st-v2,U-Ensembl49,G-Affy.cdf and MoEx-1_0-st-v1,U-Ensembl50,G-Affy,EP.cdf generated by Aroma.affymetrix team [63]. Thus, we excluded some candidate young genes that were too similar to their paralogs and did not have specific probes.
Based on R [57] and Bioconductor platform [64], we used RMA [65] to normalize and generate gene-level intensity for 3′ gene array and Aroma.affymetrix to normalize and summarize gene-level signal for exon arrays. We used MAS5 to call expressional presence and absence for 3′ gene array. In case of exon array, we used Affymetrix dabg (detection above background) algorithm to generate chip specific background signal and then compared gene-level signal to this background with Wilcoxon rank sum one-tail test. Considering multiple-testing issues, we converted all p values to q values using the qvalue package [66]. The q value of 0.01 was used as the cutoff. For Agilent miRNA array, we used “gIsGeneDetected” column generated by Agilent Feature Extraction software to define presence or absence calls [67]. We required a gene to be present in all replicates to be considered a presence and a gene to be absent in all replicates to be considered an absence. We removed all ambiguous cases from the final statistics.
We used the LIMMA package [68] to call expressional difference, with a false discovery rate corrected p of 0.05 used as the cutoff. Although we compared testis and ovary, we used the term “male-bias” or “female-bias” rather than “testis-bias” or “ovary-bias.” The reason is that these two datasets are nearly equivalent. A previous study showed that the proportion of germline male-biased genes is much higher than that of somatic male-biased genes (20% versus 2%) [12].
For meta-analyses of mouse and rat spermatogenic data, we followed the concept of RA and euclidean distance (d) to measure the between-species expression divergence [69]. Specifically, we defined RA as the proportion of expression intensity of one tissue out of all tissues and d as the sum of the square of RA difference for all tissues between mouse and rat, i.e., .
We mapped 20 out of 33 representative genes in [34] to our gene age data using unique NCBI gene names. Remarkably, 16 (80%) are rodent-specific, with 11 of them originating after the mouse and rat split. We note here that this dataset does not overlap with what we described in Table 3, since Table 3 only presents genes with unique probes, which 19 of these 20 genes do not have.
We downloaded the vertebrate-wide 44-way coding sequence alignment from UCSC. UCSC known genes mapping to multiple Ensembl genes were discarded. For Ensembl genes mapping to multiple UCSC known genes, we retained only one UCSC gene with the longest coding region. Then, considering that low quality assembly often causes unreliable estimation of Ka/Ks [70], we extracted 17 species with relatively better quality (Figure 1) and then removed all in-frame stop codons or gaps in the alignment. According to our age dating information, taxa conflicting with the age were removed. Based on the species tree (Figure 1), we estimated Ka/Ks for each branch using free ratio model in PAML [71].
We downloaded Gene Ontology (GO) annotations for Ensembl V51. We used the program analyze.pl V1.9 of TermFinder package [72] to identify those significant terms for new genes, with multiple test corrected p of 0.05 as the cutoff and the whole genome as the background. Herein, TermFinder was updated to V0.83, which corrected a mistake in calculating false discovery rate [73].
|
10.1371/journal.pcbi.0030141 | A Balanced Memory Network | A fundamental problem in neuroscience is understanding how working memory—the ability to store information at intermediate timescales, like tens of seconds—is implemented in realistic neuronal networks. The most likely candidate mechanism is the attractor network, and a great deal of effort has gone toward investigating it theoretically. Yet, despite almost a quarter century of intense work, attractor networks are not fully understood. In particular, there are still two unanswered questions. First, how is it that attractor networks exhibit irregular firing, as is observed experimentally during working memory tasks? And second, how many memories can be stored under biologically realistic conditions? Here we answer both questions by studying an attractor neural network in which inhibition and excitation balance each other. Using mean-field analysis, we derive a three-variable description of attractor networks. From this description it follows that irregular firing can exist only if the number of neurons involved in a memory is large. The same mean-field analysis also shows that the number of memories that can be stored in a network scales with the number of excitatory connections, a result that has been suggested for simple models but never shown for realistic ones. Both of these predictions are verified using simulations with large networks of spiking neurons.
| A critical component of cognition is memory—the ability to store information, and to readily retrieve it on cue. Existing models postulate that recalled items are represented by self-sustained activity; that is, they are represented by activity that can exist in the absence of input. These models, however, are incomplete, in the sense that they do not explain two salient experimentally observed features of persistent activity: low firing rates and high neuronal variability. Here we propose a model that can explain both. The model makes two predictions: changes in synaptic weights during learning should be much smaller than the background weights, and the fraction of neurons selective for a memory should be above some threshold. Experimental confirmation of these predictions would provide strong support for the model, and constitute an important step toward a complete theory of memory storage and retrieval.
| A critical component of any cognitive system is working memory—a mechanism for storing information about past events, and for accessing that information at later times. Without such a mechanism, even simple tasks, such as deciding whether to wear a heavy jacket or a light sweater after hearing the weather report, would be impossible. Although it is not known exactly how storage and retrieval of information is implemented in neural systems, a very natural way is through attractor networks. In such networks, transient events in the world trigger stable patterns of activity in the brain, so by looking at the pattern of activity at the current time, other areas in the brain can know something about what happened in the past.
There is now considerable experimental evidence for attractor networks in areas such as inferior temporal cortex [1–3], prefrontal cortex [4–9], and hippocampus [10,11]. And from a theoretical standpoint, it is well understood how attractor networks could be implemented in neuronal networks, at least in principle. Essentially, all that is needed is an increase in the connection strength among subpopulations of neurons. If the increase is sufficiently large, then each subpopulation can be active without input, and thus “remember” events that happened in the past.
While the basic theory of attractor networks has been known for some time [12–14], moving past the “in principle” qualifier, and understanding how attractors could be implemented in realistic, spiking networks, has been difficult. This is because the original Hopfield model violated several important principles: neurons did not obey Dale's law; when a memory was activated, neurons fired near saturation, much higher than is observed experimentally in working memory tasks [1,15]; and there was no null background state—no state in which all neurons fired at low rates.
Most of these problems have been solved. The first, that Dale's law was violated, was solved by “clipping” synaptic weights; that is, by using the Hopfield prescription [12], assigning neurons to be either excitatory or inhibitory, and then setting any weights of the wrong sign to zero [16,17]. The second, building a Hopfield-type network with low firing rate, was solved by adding appropriate inhibition [18–23] (importantly, this was a nontrivial fix; for discussion, see [23]). The third problem, no null background, was solved either by making the units sufficiently stochastic [18–21] or adding external input [14,20–23].
In spite of these advancements, there are still two fundamental open questions. One is: how can we understand the highly irregular firing that is observed experimentally in working memory tasks [24]? Answering this question is important because irregular firing is thought to play a critical role both in how fast computations are carried out [25] and in the ability of networks to perform statistical inference [26]. Answering it is hard, though, because, as pointed out in [27], with naive scaling the net synaptic drive to the foreground neurons (the neurons that fire at elevated rate during memory) is proportional to the number of connections per neuron. Consequently, because of the high connectivity observed in cortex, the mean synaptic drive is much larger than the fluctuations, which implies that the foreground neurons should fire regularly. Moreover, as pointed out by Renart et al. [28], even for models that move beyond the naive scaling and produce irregularly firing neurons, the foreground neurons still tend to fire more regularly than the background neurons, something that is inconsistent with experiments [24].
Several studies have attempted to get around this problem, either directly or indirectly [22,27–29]. Most of them, however, did not investigate the scaling of the network parameters with its size (i.e., with the number of neurons and connections). So, although parameters were found which led to irregular activity, it was not clear how those parameters should scale as the size of the network increased to realistic values. In the two that did investigate scaling [27,28], irregular firing was possible only if a small fraction of neurons was involved in each memory; i.e., only if the coding level was very small. Although there have been no direct measurements of the coding level during persistent activity, at least to our knowledge, experiments in superior temporal sulcus [30] suggest that it is much larger than the one used in these models. We should point out, though, that the model of Renart et al. [28] is the only one in which the foreground neurons are at least as regular as the background neurons.
The second open question is: what is the storage capacity of realistic attractor networks? That is, how many different memories can be stored in a single network? Answering this is critical for understanding the highly flexible and seemingly unbounded memory capacity observed in animals. For simple, albeit unrealistic, models the answer is known: as shown in the seminal work of Amit, Gutfreund, and Sompolinsky [31], the number of memories that can be stored in a classical Hopfield network [12] is about 0.14 times the number of neurons. For slightly more realistic networks the answer is also known [16,19,21,27,32–38]. However, even these more realistic studies lacked biological plausibility in at least one way: connectivity was all–all rather than sparse [19,21,33,38], the neurons were binary (either on or off, with nothing in between) [16,19,21,32,33,37], there was no null background [16,32,33,35,37,38], the firing rate in the foreground state was higher than is observed experimentally [16,27,32,33,36,37], or the coding level was very small [27,36].
Here we answer both questions: we show, for realistic networks of spiking neurons, how irregular firing can be achieved, and we compute the storage capacity. Our analysis uses relatively standard mean-field techniques, and requires only one assumption: neurons in the network fire asynchronously. Given this assumption, we first show that neurons fire irregularly only if the coding level is above some threshold, although a feature of our model is that the foreground neurons are slightly more regular than the background neurons. We then show that the maximum number of memories in our network—the capacity—is proportional to the number of connections per neuron, a result that is consistent with the simplified models discussed above. These predictions are verified with simulations of biologically plausible networks of spiking neurons.
To address analytically the issues of irregularity and storage capacity in attractor networks, we consider a model in which neurons are described by their firing rates. Although firing rate models typically provide a fairly accurate description of network behaviour when the neurons are firing asynchronously [39,40], they do not capture all features of realistic networks. Therefore, we verify all of our predictions with large-scale simulations of spiking neurons.
Our network consists of two populations, one excitatory and one inhibitory, with NE neurons in the former and NI in the latter. (In general we use E for excitation and I for inhibition.) We represent the firing rate of the ith neuron in pool Q(=E,I) by vQi. As we show in the section “Fast fluctuations,” and discuss below, the time evolution equations for the firing rates are given by
where τE and τI are the excitatory and inhibitory time constants, hQi is the synaptic input to the ith neuron in pool Q, and FQ(h) is a function that tells us the steady state firing rate of a neuron receiving synaptic input h. This function, which has a relatively stereotyped quasi-sigmoidal shape, can be determined analytically (or semi-analytically) for specific noise models [41–43], and numerically for more realistic models [40]. The synaptic drive, hQi, is related to the activity of the presynaptic neurons via
where
is the synaptic weight from the jth neuron in pool R to the ith neuron in pool Q, and
is the external, purely excitatory, input to neurons in pool Q. Finally, the steady-state firing rate of each neuron is determined by setting dvEi/dt and dvIi/dt to zero, yielding the equation
The bulk of our analysis focuses on solving Equation 3; we use the dynamics, Equation 1, only when investigating stability. Our goal is to determine the conditions that support retrieval states—states such that subpopulations of neurons have elevated firing rates.
Since the gain functions, FQ(h), that we use in Equation 1 play such a central role in our analysis, we briefly justify them here; for additional details, see the section “Fast fluctuations.” These gain functions come from an average over the fast temporal fluctuations of the synaptic input—basically, filtered spikes. Calculating the temporal fluctuations self-consistently is a hard problem [44], but, fortunately, it's not a problem we have to solve. As we show in the section “Fast fluctuations,” in the limit that each neuron receives a large number of connections, the temporal fluctuations experienced by all the excitatory neurons have the same statistics, as do the temporal fluctuations experienced by all the inhibitory neurons. Thus, we can use a single function, FE(h), for the excitatory neurons, and another function, FI(h), for the inhibitory ones. Of course, we won't be able to calculate the shape of FQ without knowing the structure of the temporal fluctuations. However, as we show below, the precise shapes of the gain functions don't play a strong role in our analysis.
As discussed above, much of our focus in this paper is on solving Equation 3. For even moderate size networks, this corresponds to solving thousands of coupled, highly nonlinear equations, and for large networks that can number into the millions. We do not, therefore, try to find a particular solution to this equation, but instead look for a statistical description—a description in terms of probability distributions over excitatory and inhibitory firing rates. The main tool we use is self-consistent signal-to-noise analysis [48,49]. The idea behind this analysis is to treat the synaptic input (hEi and hIi in Equation 3) as Gaussian random variables. Solving Equation 3 then reduces to finding, self-consistently, their means and variances.
Because hEi and hIi consist of 2K (very weakly) correlated terms, where
naive central limit arguments tell us that the standard deviations of these quantities should be smaller than their means by a factor of K1/2. It would seem, then, that in the kinds of high connectivity networks found in the brain, where K is on the order of 5,000–10,000, neuron-to-neuron fluctuations in firing rate would be small, on the order of K−1/2. By the same reasoning, temporal fluctuations in the firing rates would also be small, again on the order of K−1/2. Neither of these, however, are observed in biological networks: there are large fluctuations in firing rate both across neurons and over time [24,50–53].
To resolve this apparent contradiction, one need only notice that hEi and hIi consist of both positive and negative terms (the first and third terms in Equation 2 are positive; the second is negative). If these terms approximately cancel—to within O(K−1/2)—then both the mean and standard deviation of the synaptic drive will be on the same order, and network irregularity will be restored. As showed by van Vreeswijk and Sompolinsky in a groundbreaking set of papers [25,54], under fairly mild conditions this cancellation occurs automatically, thus placing networks very naturally in what they called the balanced regime. In this regime, fluctuations across both neurons and time are large. Whether networks in the brain really operate in the balanced regime is not completely clear, although recent experimental evidence has come down strongly in favour of this hypothesis [55,56].
While the work of van Vreeswijk and Sompolinsky was extremely important in shaping our understanding of realistic recurrent networks, their focus was primarily on random connectivity. The situation, however, is more complicated in attractor networks. That's because these networks consist of three classes of neurons rather than two: background excitatory neurons and background inhibitory neurons, as found in randomly connected networks, but also foreground excitatory neurons. Our goal in the next several sections is to understand how all three classes can be balanced, and thus fire irregularly.
Our mean-field analysis gave us two predictions. The first is that if the background synaptic weights, the
, scale as K−1/2, the foreground weights, A, scale as K−1, and the coding level, a, is sufficiently high, then both the background and foreground neurons should operate in the balanced regime and the neurons should fire irregularly. The second prediction is that the number of memories that can be stored is proportional to the number of excitatory connections per neuron, KE.
To test these predictions, we perform simulations with large networks of spiking neurons. We start by finding, for a particular network size, parameters such that both foreground and background neurons exhibit irregular activity. We then increase the size of the network while scaling the synaptic weights according to the above prescriptions. If the larger networks continue to exhibit irregular activity, then our predicted scalings are correct. To test the relation between storage capacity and number of connections per neuron, we calculate the storage capacity for networks with different sizes. A linear relation would indicate a scaling consistent with our predictions.
In this paper we addressed two questions. First, can all the neurons in an attractor network—both background and foreground—exhibit irregular firing? And second, what is the storage capacity in networks of realistic spiking neurons? To answer these questions, we applied self-consistent signal-to-noise analysis to large networks of excitatory and inhibitory neurons, and we performed simulations with spiking neurons to test the predictions of that analysis.
Our primary finding is that two conditions must be met to guarantee irregular firing of both foreground and background neurons. The first is proper scaling with the number of connections per neuron, K: the strength of the background weight matrix must scale as K−1/2 and the strength of the structured part of the weight matrix (the part responsible for the memories) as K−1. What this scaling does is guarantee “balance,” meaning the network dynamically adjusts its firing rates so that the mean input to a neuron is on the same order as the fluctuations, independent of K. This in turn guarantees that the degree of irregular firing is independent of K.
While balance is a necessary condition for irregular firing, it is not sufficient. That's because balance ensures only that the mean and fluctuations are independent of K, but does not rule out the possibility that the mean is much larger than the fluctuations, which would result in regular firing. To ensure that this does not happen, a second condition must be satisfied: the coding level, a, must be above some (K-independent) threshold. This condition is needed to ensure that the coupling between background and foreground neurons is sufficiently strong to stabilize a low firing rate foreground state on the unstable branch of the m-nullcline (see Figure 1).
The analysis that led to predictions of irregular firing also quite naturally provided us with information about the capacity of attractor networks—the maximum number of patterns that could be stored and successfully retrieved. What we found, under very general conditions, was that this maximum, denoted pmax, is linear in the number of excitatory connections per neuron, KE. This scaling relation has been observed in studies of simplified attractor networks [16,32,34], but, as discussed in the Introduction, those models did not include all the features that are necessary for a realistic recurrent networks. Thus, the analysis performed here is the first to show that the number of memories is linear in KE in biophysically plausible networks.
Note that there are other types of scaling, different from what we proposed, which can result in irregular firing of both foreground and background neurons. What is critical is that the net input a foreground neuron receives from the other foreground neurons should be O(1). We achieved this by letting the structured part of the connection matrix (the second term in Equation 11a) be O(1/K), and using a coding level, a, that was O(1). However, this is not the only possible combination of connection strengths and coding levels, and in the two other studies that address both scaling and irregularity in memory networks [27,28], different combinations were used. In the model proposed by van Vreesjwik and Sompolinsky [27], the structured part of their connection matrix was a factor of K1/2 larger than ours; to balance that, the coding level was a factor of K1/2 smaller. In the model proposed by Renart et al. [28], the structured part of the synaptic weights was K times larger than ours, so their coding level had to scale as O(1/K). Whether such low coding levels are consistent with reality needs further investigation; however, data from studies conducted on selectivity of neurons to visual stimuli suggests that it is too low [30]. In addition to the very low coding level that these two models require, they also exhibit non-biologically high foreground firing rate. Nevertheless, the model of Renart et al. [28] does have one advantage over others: the foreground neurons are as irregular as, or even more irregular than, the background neurons, something our model does not achieve (see next section).
Although our simulations showed irregular activity, we found that the mean CV was only about 0.8. This is smaller than the values measured in vivo, which are normally close to, or slightly above, one [24,50–53]. In addition, in our simulations the CV showed a small, but consistent, decrease with firing rate (see the left column in Figure 7). This is due to the fact that with the scaling that we chose, the fluctuations in the input current to foreground and background neurons are the same but the mean current to the foreground neurons is higher (see the section “Fast fluctuations”). This decrease in the CV disagrees slightly with a study by Compte et al. [24], who found that the CV in prefrontal cortex does not depend on the mean firing rate, at least in a spatial memory task. While there are many possible reasons for this discrepancy, a likely one arises from the fact that the neurons in our network contained only two time scales, the membrane and synaptic time constants, and both were short: 10 ms for the former and 3 ms for the latter. Real neurons, however, have a host of long time scales that could contribute to irregularity [60]. In addition, in vivo optical imaging [61–63] and multi-electrode [64] studies indicate that the background activity varies coherently and over long time scales, on the order of seconds, something we did not model. Both of these would increase the CV, although how much remains to be seen.
Although multiple time scales could certainly increase irregularity, it is not the only possible way to do this. As discussed in the Introduction and in the previous section, the model proposed by Renart et al. [28] also increases irregularity, and is consistent with the experimental results of Compte et al. [24]. However, it requires a very small coding level (a ∝ 1/K), and fine-tuning of the parameters.
In conventional models of persistent activity [14,22,29], the foreground activity necessarily lies on the concave part of the excitatory gain function, FE(hE), whereas the background activity lies on the convex part. Since the inflection point of realistic gain functions is typically near the firing threshold [42,43], this type of bistability is called suprathreshold bistability [22,28]. Because the concave part of the gain function is typically at high firing rate, with suprathreshold bistability it is hard to have either low foreground firing rate or high CV. Consequently, there has been interest in understanding whether it is possible to have subthreshold bistability; that is, whether it is possible for both foreground and background solutions to lie on the subthreshold part of the gain function [28].
The model presented here can in fact show subthreshold bistability: as discussed in the section “Reduced mean-field equations in the infinite K limit,” increasing the coding level, a, brings the foreground firing rate very close to the background rate. Therefore, for sufficiently large a, the foreground state would be on the convex part of the transfer function. Our model, and the recently proposed model by Renart et al. [28], are the only ones that can show subthreshold bistability.
One rather striking feature of our networks is that they all produce a highly bimodal distribution of firing rates: as can be seen in the first and third columns of Figure 7, the background neurons fire at a much lower rate than the foreground neurons—so much lower, in fact, that they form a distinct, and easily recognizable, population. This occurs because the patterns we store—the
—are binary, which makes the average input current to every neuron in the foreground exactly the same. This feature is potentially problematic, as the distinction between foreground and background rates observed in experiments is not nearly as striking as the one in Figure 7 [65]. However, this feature is not essential to our analysis, for two reasons. First, as discussed in the section “Building a balanced network” (see especially Figure 6), we deliberately made the background firing rate low to increase the capacity. Second, it is easy to extend our analysis to real valued patterns in which the elements of the
are drawn from a continuous distribution [34]. Under this, more realistic, scenario, it should be possible to match the statistics of the response seen in the cortex. This will be the subject of future work.
In our model, every time a new pattern is learned, the weights change by an amount proportional to K−1. This is a factor of K−1/2 smaller than the background weights. Since weight changes are unlikely to be under such fine control, it is natural to ask whether errors during learning will lead to a major reduction in storage capacity. The answer, of course, depends on the size of the errors. In the section “Fine-tuning in the learning rule,” we show that errors can be larger than the weight changes by a factor of (K/p)1/2, with only a small change in storage capacity. More specifically, every time a pattern is learned, noise of O((Kp)−1/2) can be added to the synaptic strength, and the network will retain its ability to store and recall patterns.
Although this result tells us that the noise in the weight changes can be large compared with the structured part, the fine-tuning problem is not entirely eliminated: the noise must still be a factor of p1/2 smaller than the background weights. Because of the low storage capacity found in these networks (at most 2.5% [23]), even when K is as large as 10,000, 1/p1/2 is on the order of 6%. It seems plausible that biological machinery has evolved to achieve this kind of precision. However, for networks with larger capacity, the requirements on the precision of the weight would be more stringent.
It is also possible to have a probabilistic learning rule for which the changes in the weights are on the same order as the background weight, but this decreases the capacity significantly, by a factor of
(see the section “Fine-tuning in the learning rule,” Equation 78; we thank Carl van Vreeswijk for pointing this out). Although this probabilistic learning rule guarantees a balanced state with irregular background and foreground firing, it has the drawback that the storage capacity scales as
rather than K.
Although we showed that pmax ∝ KE, we did not compute analytically the constant of proportionality. In our simulations, this constant was small: from Figure 9, pmax is about 0.01 KE, which means that for KE = 10,000 we can store only about 100 patterns. It is important, though, to note that we made no attempt to optimize our network with respect to other parameters, so the constant of proportionality 0.01 is unlikely to be a fundamental limit. In fact, Latham and Nirenberg [23] were able to store about 50 patterns in a network with 2,000 excitatory connections, 2.5 times larger than our capacity. Interestingly, the only substantial difference between their network and ours was that in theirs the background activity was generated by endogenously active neurons rather than by external input.
Can we further increase the scaling factor? One potential mechanism is to decrease the coding level, a, since, at least in simple models [33,34,37], the maximum number of patterns that could be stored and retrieved is inversely proportional to the coding level. But, as we showed in the section “Storage capacity,” realistic networks do not exhibit this 1/a scaling. Consequently, sparse coding cannot be used as a way to improve the storage capacity in our network. Simplified models also suggest that one can increase the storage capacity by a factor of 3–4 by using other schemes, such as non-binary patterns [34], or spatially correlated patterns [66]. Whether these techniques can be extended to the kind of network we have studied here is not clear, and requires further investigation. However, an increase beyond a factor of 3–4, to a capacity above about 0.1, seems unlikely within this class of networks.
In any case, there is a limit to the number of memories that can be stored in a single attractor network with a fixed number of connections per neuron, no matter how many neurons in the network. This suggests that, in order to make the best use of the existing connections, realistic working memory systems must be composed of interconnected modules. In this paradigm, each module would consist of an attractor network [67–69]. Such modular structure naively suggests a combinatorial increase in storage capacity; however, understanding how to achieve such an increase has proved difficult. For simple models whose storage capacity could be calculated analytically, either no increase in the storage capacity [67] or a modest increase [69] was found. It is yet to be determined how modular networks could be implemented in realistic networks of spiking neurons, and what their storage capacity would be.
The starting point for essentially all of our analysis is Equation 1, which, when combined with Equation 2, tells us that the time evolution of the firing rate of each neuron is purely a function of the firing rates of the other neurons. At a microscopic level, though, each neuron sees as input a set of spikes, not rates. However, for our model, rate-based equations do apply, as we show now.
In a spiking, current-based network, the input, hQi(t), to the ith neuron in population Q has the form
where
is the time of the kth spike on the jth neuron, fR(t), which mimics the PSP, is a non-negative function that integrates to 1 and vanishes for t < 0 and t large (greater than a few tens of ms). In a slight departure from our usual convention, R can refer to external input (R = ex) as well as excitatory and inhibitory input (R = E,I).
Our first step is to divide the input, hQi, into a mean and a temporally fluctuating piece. The mean, which is found by time-averaging the right-hand side of Equation 34a and using the fact that fR(t) integrates to 1, is simply
where 〈···〉t represents a temporal average. The temporally fluctuating piece of the input can then be written
The fluctuations, δhQi, have zero mean by construction, and their correlation function, CQi(τ), is defined to be
Assuming that hQi is Gaussian (which is reasonable if there are a large number of neurons and they are not too correlated), then the firing rate depends only on the mean, 〈hQi(t)〉t, and the correlation function, CQi(τ). If the correlation function is independent of i, then the only i-dependence in the firing rate is through the mean input, and we recover Equation 1. What we now show is that, for our model, CQi does not depend on i.
To understand the behaviour of CQi, we express it in terms of
; using Equation 36a, we have
Under the assumption that the neurons are very weakly correlated, only the terms with j = j′ survive, and this expression simplifies to
Let us focus on the sum on j on the right-hand side of this expression. For Q ≠ E or R ≠ E, this sum is given by (see Equations 6b–6d)
For sparsely connected networks,
is independent of
. Consequently, we can replace
on the right-hand side of Equation 38 by its average, c, and the right hand side becomes independent of i.
For Q = R = E, the situation is more complicated, as
has an additional dependence on Aij, the structured part of the connectivity. Specifically using Equation 6a and again replacing
by its average, c, we have
As discussed in the section “Mean-field equations,” Aij receives contributions from two sources: the p − 1 patterns that are not activated, and the one pattern that is. The non-activated patterns are not correlated with δSj, so they can be averaged separately in Equation 39, and thus do not produce any i-dependence. The activated pattern, on the other hand is correlated with δSj. However, the connection strength for the one activated pattern is smaller than
by a factor of K−1/2 (see the section “Strong synapses and the balanced condition”). Consequently, in the high connectivity limit, we can ignore this contribution, and the right-hand side of Equation 39 is independent of i. This in turn implies that CQi depends only on Q.
The upshot of this analysis is that the only i-dependence in the firing rate comes from 〈hQi(t)〉t. Moreover, comparing Equations 2 and 35, we see that 〈hQi(t)〉t is exactly equal to hQi, the input current to the firing rate function, FQ, that appears in Equation 1. Thus, for the model used here, the rate-based formulation is indeed correct. What we do not do is compute FQ, as that would require that we compute the correlation function, CQ(τ), self-consistently, which is nontrivial [44]. However, our results depend very weakly on the precise form of FQ, so it is not necessary to have an explicit expression for it.
In this section, we derive the mean-field equations for the model described in the section “Model.” As discussed in the main text, the derivation of these equations revolves around finding the distributions of
and δhIi, the fluctuations around the mean excitatory and inhibitory synaptic input (both quantities are defined implicitly in Equations 11–13). The main assumption we make is that
and δhIi are zero mean Gaussian random variables, so all we need to do is find their variances self-consistently. In addition, primarily for simplicity (and because it is reasonable in large networks in the brain), we assume that the number of connections is small compared with the number of neurons, so c ≪ 1.
Our first step is to simplify the expressions for our main order parameters, νE, m, and νI. In the context of the self-consistent signal-to-noise analysis, “simplify” means “replace sums by Gaussian integrals.” To see how to do this, note that, for any function g,
where Var[·] indicates variance, exact equality holds in the NE → ∞ limit (but approximate equality typically holds when NE is only a few hundred), and
A similar expression applies, of course, to δhIi.
Applying the sum-goes-to-integral rule to Equation 16, we have
where the average over ξ is with respect to the probability distribution given in Equation 5.
To complete Equation 40, we need the variance of
and δhIi. It is convenient to break the former into two pieces,
, where the first, δhEi, is associated with the background neurons, and the second, δhmi, is associated with the foreground neurons (both will be defined shortly). Then, examining Equations 11–15, and performing a small amount of algebra, we find that
and
Here δμ,v is the Kronecker delta; it is 1 if μ = ν and zero otherwise. In addition, for notational convenience, we have returned the superscript “1” to ξi. For the rest of the section, we will use
and ξi interchangeably.
Let us focus first on the contribution from the background, Equation 41. Since
is equal to c on average, the mean of both terms on the right hand side of Equation 41 is zero. Moreover, these terms are uncorrelated, so their variances add. The variance of the QRth term is then
where the angle brackets represent an average over the distribution of
. Because
and
are independent when j ≠ j′, only terms with j ≠ j′ produce a nonzero average. Thus, all we need is the variance of
, which is given by
(the last approximation is valid because, as mentioned above, we are assuming c ≪ 1). Performing the sums over j and j′ and collecting terms, we have
The term on the right-hand side,
, is the second moment of the firing rate of the neurons in pool R. Inserting Equation 43 into 41, we find that
The last quantity we need is the variance of δhm. A naive approach to computing it proceeds along lines similar to those described above: assume all the terms in the sum over j and μ in Equation 42 are independent, so that the variance of δhm is just pNE (the number of terms in the sum) times the variance of each term. This yields, with rather loose notation for averages and ignoring the
correction associated with μ = 1,
All the averages in this expression are straightforward:
, 〈ξ2〉 = a, 〈(ξ − a)2〉 = a(1 − a), and
was defined in Equation 43. Putting all this together and defining ρ2 to be the variance of δhm, we have
While Equation 45 turns out to be correct, our derivation left out a potentially important effect: correlations between the patterns,
, and the firing rates, νEj in Equation 42. These correlations, which arise from the recurrent feedback, turn out to scale as c, and so can be neglected [32,35,70,71]. Rather than show this here, we delay it until the end of the section (see the section “Loop corrections vanish in the small c limit”).
To write our mean-field equations in a compact form, it is convenient to define the total excitatory variance,
Then, combining Equations 3, 40, 44, and 45, the mean-field equations become
where the subscript z indicates a Gaussian average,
and, recall, α = p/KE (Equation 28).
Finally, it is convenient to explicitly perform the averages over ξ that appear in Equation 47. Defining
the relevant averages become
The functions
and
that we used in Equation 17 are equivalent to the ones defined in Equation 48, although we had suppressed the dependence on the standard deviation and dropped the superscript.
Equation 47 constitutes our full set of mean-field equations. A key component of these equations is that the number of memories, p, enters only through the variable α, which is p/KE. Thus, the number of memories that can be embedded in a network of this type is linear in the number of connections.
To correctly treat the loop corrections in our derivation of the variance of δhm, we need to be explicit about the correlations between the patterns,
, and the firing rates, νEj, in Equation 42. We start by defining the i-dependent overlap,
, as
Inserting this into Equation 42 leads to
Each of the terms
is a Gaussian random variable whose variance must be determined self-consistently. This can be done by inserting Equation 3 into Equation 50 to derive a set of nonlinear equations for the
. There are two types of terms to consider: the activated memory, for which μ = 1, and the non-activated memories, for which μ ≠ 1. However, in the large p limit we can safely ignore the one term corresponding to μ = 1. Thus, considering the contributions from memories with μ ≠ 1, we have
Taylor expanding around
and defining
where a prime denotes a derivative, we have
We can write Equation 52 in matrix form as
where I is the identity matrix, the ith component of mμ is equal to
, and the matrices Λμ and Ξμ are given by
To solve Equation 53, we need to invert I − Λ, in general a hard problem. However, what we show now is that Λ has only one O(1) eigenvalue, with the rest
. This allows us to write the inverse in terms of a single eigenvector and adjoint eigenvector, a simplification that allows us to perform the inversion explicitly.
The spectrum of the random matrix, Λμ, is determined primarily by the mean and variance of its components [72]. In the large NE limit, these are given by
where 〈···〉ij indicates an average over i and j, and we used the fact that
and
are independent.
Given that Λμ is an NE × NE matrix, the fact that the mean and variance of its elements are
and O((KENE )−1), respectively, implies that it has one eigenvalue that is O(1) and NE − 1 eigenvalues that are
[72]. Letting vk and
be the eigenvector and adjoint eigenvector of Λμ whose eigenvalue is λk, we can solve Equation 53 for mμ,
where “·” represents dot product. Letting k = 0 correspond to the O(1) eigenvalue and explicitly separating out this component, the expression for mμ becomes
and
Since v0 and
are vectors whose components are all the same, without loss of generality we can choose v0 = (1,1,…1)/NE and
. Combining this choice with Equation 55 and using Equation 54b for Ξμ, we have
We are now in a position to return to Equation 51 and compute the variance of δhm (which, recall, is denoted ρ2). Treating, as usual, all the terms in Equation 51 as independent, we have
To compute
we use Equation 57 and the fact that the off-diagonal elements average to zero, and we find that
To derive this expression, we again used 〈(ξ − a)2〉 = a(1 − a).
Our final step is to insert Equation 59 into Equation 58. Ignoring the two terms in brackets in Equation 59 that are a factor of c smaller than the first, and using the fact that
, this leads to the expression for ρ2 given in Equation 45. Consequently, loop corrections vanish, and we can use our naive estimate for the variance of δhm.
Ignoring the two terms in brackets in Equation 59 is strictly correct for infinitely diluted networks; i.e., networks with c → 0. When c is nonzero but small, the terms in the brackets can be ignored safely unless λ0 → 1. However, as we now show, λ0 → 1 is precisely the point where the background becomes unstable. Thus, it is not a regime in which we can operate.
The significance of the limit λ0 → 1 can be seen by replacing Equation 53 by its dynamical counterpart (see Equation 1),
When the largest eigenvalue of Λμ exceeds 1, the unactivated memories become unstable, and retrieval of just one memory is impossible. As discussed above, the largest eigenvalue of Λμ is λ0. Consequently, loop corrections are necessarily important (no matter how dilute the network is) at precisely the point where the unactivated memories, and thus the background, become unstable.
To determine stability, we need to write down time-evolution equations for the order parameters, and then linearize those around their fixed points. For νE, νI, and m, which are linear combinations of the firing rates, this is straightforward—we simply insert their definitions, Equation 16, into the time-evolution equations for the individual firing rates, Equation 1. For the variances,
and
, the situation is much more difficult, as these quantities do not admit simple time-evolution equations [73]. Fortunately, we expect the effects of the variances to be small—as discussed in the main text, their primary effect is to smooth slightly the gain functions, something that typically (although presumably not always) stabilizes the dynamics. Alternatively, if we assume that the variances are functions of νE, νI, and m, (meaning we give them instantaneous dynamics), we can rigorously neglect them. This is because derivatives of the gain functions with respect to νE and νI are large, on the order of K1/2, while derivatives with respect to the variances are O(1). Thus, as a first approximation, we will ignore these variables, and consider only the dynamics of νE, νI, and m. Because of this approximation, we expect our stability boundaries to be off by a small amount.
Combining Equation 1 and Equation 47, the time-evolution equations for νE, νI, and m may be written
To simplify notation, it is convenient to define
Then, linearizing Equation 61 by letting νE → νE + δνE, νI → νI + δνI, and m → m + δm, we have
where the notation ϕa,b indicates a derivative of ϕa with respect to the argument specified by b (for example, ϕE,I = ∂ϕE/∂νI and ϕI,M = ∂ϕI/∂m). Since ϕI is independent of m (which means ϕI,m = 0), the equation for the eigenvalues, denoted λ, becomes
Equation 63 is a cubic equation in λ, and thus not straightforward to solve. However, in the large K limit it simplifies considerably. That's because derivatives with respect to νE and νI are O(K1/2), which follows because the ϕ's depend on νE and νI through hE and hI, and the latter are proportional to K1/2 (see Equation 13). Defining the O(1) quantities
R = νE,νI,m and Q = νE,νI, Equation 63 becomes (ignoring O(K−1/2) corrections)
Examining Equation 64, it follows that if the eigenvalue, λ, is O(K1/2), then the term ϕm,m − 1 and the last term in brackets can be neglected. There are two such eigenvalues, and they are given by
Both eigenvalues are negative if
Since ϕI,I < 0, the first condition is satisfied if τI is sufficiently small. For the second condition, from Equations 13, 20, and 62, we see that
where the constant of proportionality is positive. Since the condition for the stability of the background is D > 0 [54], we see that 65b is satisfied whenever the background is stable. Thus, for τI sufficiently small and the background stable, the two O(K1/2) eigenvalues are negative.
The third eigenvalue is O(1), so when computing it we can drop all the K−1/2 λ terms. Denoting this eigenvalue λm, we thus have
Using a prime to denote a derivative with respect to hE and noting that (see Equation 62)
Equation 66 reduces to
where prime denotes a derivative.
Comparing Equations 62 and 49, we see that ϕE,m = aϕm,m, which leads to
This expression strongly emphasizes the role of the coding level, a: if it were zero, the only stable equilibria would be those with ϕm,m < 1, which would imply high firing rates for foreground neurons (see Figure 1B).
Although Equation 67 tells us the stability of an equilibrium, it is not in an especially convenient form, as it does not allow us to look at a set of nullclines and determine instantly which equilibria are stable and which are not. However, it turns out that it is rather easy to determine the sign of λm for a given set of nullclines simply by looking at them. To see how, we make use of the expressions for ϕE and ϕm (Equations 62a and 62c) to reduce the right-hand side of Equation 67 to an expression with a single derivative. Our starting point is the definition
where hE(m) is given by Equation 26; the solutions of the equation Ψ(m) = m correspond to network equilibria. The advantage of this one-dimensional formulation is that, as we show below, the condition λm < 0 is equivalent to dΨ/dm < 1. Thus, by plotting the function Ψ(m) versus m and looking at its intersections with the 45° line, we can find the equilibrium values of m, and, more importantly, we can easily determine which of them is stable and which is unstable.
To show that dΨ/dm < 1 is equivalent to the condition λm < 0, we note first of all that
where, recall, a prime denotes a derivative. By combining these expressions with Equation 67, and performing a small amount of algebra, the condition λm < 0 can be written
To see how this compares to dΨ/dm, we use Equation 68 to write
Then, using Equation 25, which tells us that
this expression becomes
Comparing Equations 69 and 70, we see that the condition dΨ/dm < 1 is equivalent to λm < 0. Thus, it is only when Ψ(m) intersects the 45° line from above that, the equilibrium is stable. Since Ψ(m) is bounded, if there are three equilibria, the smallest one must be stable, the middle one unstable, and the largest one again stable. Thus, we can look at the nullcline plots and immediately determine stability (see below and Figure 10).
As an example, we revisit Figure 2. In terms of our specific form for the gain functions, Equation 23, and with hE(m) given by Equation 26, the equation for m becomes
This equation is solved graphically in Figure 10A where we plot Ψ(m) versus m for the same values of β used in Figure 2 and with a = 0.005. Intersections with the 45° line correspond to solutions of Equation 71, and thus to network equilibria.
As we saw in the sections “Reduced mean-field equations in the infinite K limit” and “An example: Nullclines for a simple gain function,” the main factor that determines the number and location of the intersections, and thus the ability of the network to exhibit retrieval states, is β. For β = 0.1 and 0.25, there is just one intersection at m = 0, while for intermediate values of β, β = 0.5 and 1.2, two additional intersections appear. Increasing β even further moves one of the solutions to negative m and destabilizes the background, but this is not shown. We can now easily see that the curves in Figure 10A with β = 0.1 and 0.25 have a single stable intersection at m = 0 (meaning that the solutions with m = 0 in Figures 2A and 2B are stable); the curves with β = 0.5 and β = 1.2 have two stable intersections, one at m = 0 and one at large m (and thus the solutions at m = 0 in Figure 2C are stable, those at intermediate m are unstable, and those with large m are again stable).
Although we see bistability, the firing rate for the retrieval state is unrealistically high—on the order of 100 Hz, near saturation. As discussed in the main text, we can reduce the firing rate by increasing a. This is done in Figure 10B, where we plot Ψ(m) versus m but this time for a = 0.05 and β = 1.2. Again there are three intersections (corresponding to the three intersections between the m-nullcline with β = 1.2 and the hE-nullcline with a = 0.05 in Figure 2C). With this higher value of a, the upper intersection is now in a biologically realistic range.
When we performed network simulations, we found that the memory strength,
, did not exhibit exactly the predicted 1/K scaling. Here we ask whether the departure from predictions that we observed can be explained by finite K corrections. These corrections, as we will see shortly, are on the order of K−1/2. Since in our simulations K is as small as 1,500, these corrections are potentially large.
Our starting point is the exact set of reduced mean-field equations, which is found by combining Equations 18 and 19,
When K is large we can solve these equations by perturbing around the K → ∞ solutions, which we denote hE0, m0, and hI0 (these are the solutions to Equation 21). The zeroth step in this perturbation analysis is to replace hE and hI by hE0 and hI0 where they appear in brackets (and thus multiply K−1/2). This gives us a new set of equations,
where
For the inhibitory firing rate, it is easy to see the effect of finite K: hI is shifted relative to hI0 by an amount proportional to δνI. Only slightly more difficult are hE and m, for which we have to consider how δνE affects the nullclines. Fortunately, only the hE-nullcline is affected, and we see that it shifts in a direction given by the sign of δνE. In particular,
(We consider −hE since, by convention, we plot our nullclines in a space with −hE on the y-axis.) Thus, if δνE is positive then the hE-nullcline shifts down relative to hE0, while if it is negative the nullcline shifts up.
In our simulations we set β to βmin, the minimum value of β that allows retrieval of one memory. To determine how K affects βmin, then, we need to know how to adjust β so that we keep the grazing intersection as K changes. Fortunately, the hE-nullcline depends on K but not on β, and the m-nullcline depends on β but not on K. Thus, all we need to know is how the m-nullcline changes with β. Using Equation 72b, it is easy to show that at fixed m,
The numerator in this expression is clearly positive, and, for equilibria to the left of the peak of the m-nullcline, the denominator is also positive (see the section “Stability analysis”). Thus, increasing β causes the m-nullcline to move up.
Combining Equation 74 and 75, we have the following picture,
where “up” corresponds to movement in the m −(−hE) plane. To complete the picture, we need to know how δνE depends on K. From 73, we see that δνE ∝ K−1/2 [JIIhE0 − JEIhI0] = K−1/2[−|JII| hE0 + |JEI|hI0]. Thus, whether δνE is an increasing or decreasing function of K depends on whether |JII|hE0 is larger or smaller than |JEI|hI0. However, as we have seen, typically hE is negative. Thus, we expect δνE to be proportional to K−1/2 with a positive constant of proportionality, which means that δνE is a decreasing function of K. Combining that with the above picture, we conclude that when K increases, βmin also increases. This is shown explicitly in Figure 11. Moreover, it was exactly what we saw in our simulations: βmin (
in Table 1) was larger than predicted when we increased K (compare
with
in Table 1).
In the model described here, the structured part of the synaptic weights scale as K−1, whereas the background scales as K−1/2. This appears to require fine-tuning, since adjustments to the weights during learning of the attractors have to be a factor of K1/2 times smaller than the background weights; a factor that can be as high as 100.
The first question to ask, then, is: exactly how big is the fine-tuning problem? In other words, how much noise can we add to the learning rule without having a huge effect on the storage capacity? This can be answered by considering a learning rule in which the weight changes during learning a pattern are not quite perfect. Specifically, let us consider the following modification of Equation 4,
where the
are zero-mean, uncorrelated random variables with variance
. The additional noise in this learning rule increases the variance of the quenched noise by an amount
. As a result, if
the effect on storage capacity is an O(1) increase in the quenched noise, and thus the storage capacity still scales as KE.
With the scaling in Equation 77, weight changes during learning of each pattern is a factor of p1/2 smaller than the background weights, and therefore the amount of fine-tuning depends on how many patterns are stored. Because of the low storage capacity found in these networks (at most 2.5% [23]), even when K is as large as 10,000, p−1/2 is on the order of 6%.
We should also point out that it is possible for the weight changes associated with the structured part of the connectivity to be on the same order as the background, although at the expense of storage capacity. Let us consider a third learning rule in which each synapse has a probability q of changing its value during learning,
where the
are Bernoulli variables;
with probability q and 0 with probability 1 − q. Let us define the coupling strength slightly differently than in Equation 10,
where, as usual, β ∼ O(1). With this definition, the mean memory strength,
, is again β/KEa(1 − a), as in Equation 10. But by setting
, the synaptic weight change—if there is one—is
, just as it is for the background weights. However, there is a major drawback: as is easy to show, the variance associated with the structured part of the connectivity increases by a factor of KE, so the maximum number of patterns scales as
rather than KE. We thus use Equation 4 for Aij in all of our analysis.
|
10.1371/journal.pmed.1002203 | Zika Virus Infection as a Cause of Congenital Brain Abnormalities and Guillain–Barré Syndrome: Systematic Review | The World Health Organization (WHO) stated in March 2016 that there was scientific consensus that the mosquito-borne Zika virus was a cause of the neurological disorder Guillain–Barré syndrome (GBS) and of microcephaly and other congenital brain abnormalities based on rapid evidence assessments. Decisions about causality require systematic assessment to guide public health actions. The objectives of this study were to update and reassess the evidence for causality through a rapid and systematic review about links between Zika virus infection and (a) congenital brain abnormalities, including microcephaly, in the foetuses and offspring of pregnant women and (b) GBS in any population, and to describe the process and outcomes of an expert assessment of the evidence about causality.
The study had three linked components. First, in February 2016, we developed a causality framework that defined questions about the relationship between Zika virus infection and each of the two clinical outcomes in ten dimensions: temporality, biological plausibility, strength of association, alternative explanations, cessation, dose–response relationship, animal experiments, analogy, specificity, and consistency. Second, we did a systematic review (protocol number CRD42016036693). We searched multiple online sources up to May 30, 2016 to find studies that directly addressed either outcome and any causality dimension, used methods to expedite study selection, data extraction, and quality assessment, and summarised evidence descriptively. Third, WHO convened a multidisciplinary panel of experts who assessed the review findings and reached consensus statements to update the WHO position on causality. We found 1,091 unique items up to May 30, 2016. For congenital brain abnormalities, including microcephaly, we included 72 items; for eight of ten causality dimensions (all except dose–response relationship and specificity), we found that more than half the relevant studies supported a causal association with Zika virus infection. For GBS, we included 36 items, of which more than half the relevant studies supported a causal association in seven of ten dimensions (all except dose–response relationship, specificity, and animal experimental evidence). Articles identified nonsystematically from May 30 to July 29, 2016 strengthened the review findings. The expert panel concluded that (a) the most likely explanation of available evidence from outbreaks of Zika virus infection and clusters of microcephaly is that Zika virus infection during pregnancy is a cause of congenital brain abnormalities including microcephaly, and (b) the most likely explanation of available evidence from outbreaks of Zika virus infection and GBS is that Zika virus infection is a trigger of GBS. The expert panel recognised that Zika virus alone may not be sufficient to cause either congenital brain abnormalities or GBS but agreed that the evidence was sufficient to recommend increased public health measures. Weaknesses are the limited assessment of the role of dengue virus and other possible cofactors, the small number of comparative epidemiological studies, and the difficulty in keeping the review up to date with the pace of publication of new research.
Rapid and systematic reviews with frequent updating and open dissemination are now needed both for appraisal of the evidence about Zika virus infection and for the next public health threats that will emerge. This systematic review found sufficient evidence to say that Zika virus is a cause of congenital abnormalities and is a trigger of GBS.
| In 2015, the mosquito-borne Zika virus caused epidemics of a mild viral illness for the first time in Brazil and then other countries in Latin America and the Caribbean.
In mid to late 2015, clinicians in northeastern Brazil reported unexpected increases in the numbers of babies born with abnormally small heads (microcephaly) and of adults with Guillain–Barré syndrome (GBS), a paralytic condition triggered by certain infections.
In February 2016, the World Health Organization (WHO) declared a Public Health Emergency of International Concern and called for research about the causal relationship between Zika virus and congenital brain abnormalities, including microcephaly, and GBS.
We developed a causality framework for Zika virus and (a) congenital brain abnormalities, and (b) GBS. For each outcome, we developed specific questions in ten different dimensions of causality: temporality; biological plausibility; strength of association; exclusion of alternative explanations; cessation; dose–response relationship; animal experimental evidence; analogy; specificity; and consistency of findings.
We did a systematic review of published and unpublished evidence up to May 30, 2016. We summarised the evidence descriptively. A panel of experts assessed the findings and reached a consensus about causality.
For congenital brain abnormalities, we assessed 72 studies that addressed questions in one or more causality dimensions. Reports of pregnancies affected by Zika virus have come from countries with circulating Zika virus in the Americas, the Pacific region, and West Africa. Clinical reports have documented Zika virus infection in pregnant women followed by foetal abnormalities, particularly with infection in the first trimester. These women did not have any other congenital infection or dengue virus infection. The risk of congenital brain abnormalities could be around 50 times higher in mothers who had Zika virus infection in pregnancy compared with those who did not. In laboratory studies, Zika virus has been shown to cross the placenta and replicate in human brain cells.
For GBS, we assessed 36 studies that addressed questions about one or more causality dimensions. In several countries in the Americas and the Pacific region, a temporal association has been found, with symptoms of Zika virus infection preceding the onset of GBS. In these countries, surveillance reports of cases of GBS followed the pattern of reports of Zika-like illness. During a Zika virus outbreak in French Polynesia in 2013–14, scientists estimated that around one in 4000 people with Zika virus infection developed GBS. The odds of having had a recent Zika virus infection were more than 30 times higher in patients with GBS than those without in a hospital-based study in French Polynesia. Several other infections that can trigger GBS were excluded.
This systematic review found sufficient evidence to conclude that Zika virus is a cause of congenital abnormalities and is a trigger of GBS.
Systematic reviews of evidence about emerging public health threats need to be updated frequently.
| An “explosive pandemic of Zika virus infection” [1] in 2015 caught the world by surprise. The Pan American Health Organization (PAHO) and World Health Organization (WHO) published an alert about a possible association with increases in reports of congenital abnormalities and Guillain–Barré syndrome (GBS) on December 1, 2015 [2]. On February 1, 2016, WHO declared a Public Health Emergency of International Concern [3]. Microcephaly at birth is a clinical finding that can include a range of brain malformations resulting from a failure of neurogenesis [4]. Infections acquired in pregnancy, including cytomegalovirus and rubella, are established causes [4]. GBS is an immune-mediated ascending flaccid paralysis, which typically occurs within a month of an infection, such as Campylobacter jejuni or cytomegalovirus [5]. As of October 20, 2016, 67 countries have reported autochthonous transmission of the mosquito-borne flavivirus Zika since 2015, and 27 of these countries have reported cases of congenital brain abnormalities, GBS, or both [6]. The emergency committee recommended increased research [3] to provide more rigorous scientific evidence of a causal relationship as a basis for the global health response.
Unexplained clusters of rare but serious conditions require urgent assessment of causality, balancing speed with systematic appraisal. Bradford Hill is widely credited for his proposed framework for thinking about causality in epidemiology, which listed nine “viewpoints” from which to study associations between exposure and disease (S1 Text, p2) [7]. Since then, the list has been modified (S1 Text, p2; S1 Table) [8]. Bradford Hill emphasised that his viewpoints were not rules but, taken together, the body of evidence should be used to decide whether there is any other more likely explanation than cause and effect.
The level of certainty required before judging that Zika virus is a cause of microcephaly and GBS is contentious [9]. Most assessments have been based on rapid but nonsystematic appraisals [10–12]. Based on rapid reviews, WHO has stated that there is “scientific consensus that Zika virus is a cause of microcephaly and GBS” since March 31, 2016 [13]. On April 13, a narrative review stated that there was “a causal relationship between prenatal Zika virus infection and microcephaly” [11]. Evidence about the causal relationship between Zika virus and GBS has not yet been assessed in detail. We previously described a causality framework for Zika virus and plans for a systematic review (S1 Text, p3; S2 Table), with a preliminary overview of 21 studies, published up to March 4, 2016 [14]. The objectives of this study are to reassess the evidence for causality and update the WHO position about links between Zika virus and (a) congenital brain abnormalities, including microcephaly, in the foetuses and offspring of pregnant women and (b) GBS in any population, and to describe the process and outcomes of an expert assessment of the evidence.
We describe three linked components: the causality framework, the systematic reviews, and the expert panel assessment of the review findings. The WHO Zika Causality Working Group convened the expert panel of 18 members with specialist knowledge in the fields of epidemiology and public health, virology, infectious diseases, obstetrics, neonatology, and neurology (membership of the expert panel is provided in the Acknowledgments).
In February 2016, we developed a causality framework for Zika virus by defining specific questions for each of ten dimensions, modified from Bradford Hill’s list (S2 Table): temporality; biological plausibility; strength of association; exclusion of alternative explanations; cessation; dose–response relationship; animal experimental evidence; analogy; specificity; and consistency of findings. This review covered 35 questions about congenital brain abnormalities, including microcephaly, and 26 questions about GBS. We also listed seven groups of cofactors that might increase the risk of an outcome in the presence of Zika virus [15].
Our protocol was registered on March 21, 2016 in the database PROSPERO (CRD42016036693) [16]. We report the methods in full according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) [17] in S1 Text (p3-4).
To report our findings, we use the term item for an individual record, e.g., a case report. Occasionally, the same individuals or data were reported in more than one publication (item). To avoid double counting, we organised these items into groups. We chose a primary publication (the item with the most complete information) to represent the group, to which other items were linked (S4 Table, S6 Table).
In a series of web and telephone conferences between April 18 and May 23, 2016, we presented our approach to the assessment of causality, the causality framework, and our synthesis of evidence to the expert panel. We discussed these topics with the experts during the conferences and through email discussions. We then drafted summary conclusions about the most likely explanation for the reported clusters of cases of microcephaly and GBS. The expert panel members reached consensus statements to update the WHO position (Fig 1).
We found 1,091 unique items published from 1952 to May 30, 2016 (S1 Fig, S3 Table). Most excluded items were reviews or editorials and commentaries (44%, n = 479). We included 106 items from 87 groups (Table 1), of which 83% were published in 2016. For both outcomes, the majority of items were clinical, individual-level case reports, case series, or population-level surveillance data.
A total of 72 items belonging to 58 groups addressed questions related to congenital brain abnormalities up to May 30, 2016 [13, 21–93]. Table 2 summarises the characteristics of 278 mother–infant pairs described in included studies.
Table 3 summarises the assessment for each causality dimension, S4 Table provides an extended description of study findings, and S5 Table summarises the quality of the body of evidence.
We found 36 items belonging to 31 groups that addressed questions related to GBS [54–57, 67, 78, 100–122]. We summarise the findings according to clinical characteristics of 118 individuals diagnosed with GBS in Table 4.
Table 5 summarises the reviewers’ assessments by causality dimension, S6 Table provides an extended description of study findings, and S7 Table summarises the quality of the body of evidence.
Based on the evidence identified up to July 29, 2016, the expert panel concluded that
The expert panel recognises that Zika virus alone may not be sufficient to cause either congenital brain abnormalities or GBS. The panel does not know whether these effects depend on as yet uncharacterised cofactors being present, nor does it know whether dengue virus plays a part, as this is carried by the same species of mosquito and has circulated in many countries during the same period.
Up to May 30, 2016, we found evidence that supported a causal association between Zika virus infection and congenital brain abnormalities, including microcephaly, with at least one study addressing one or more specific questions for eight of ten causality dimensions and between Zika virus infection and GBS, with at least one study about one or more specific questions in seven of ten dimensions. There are methodological weaknesses, inconsistencies, and gaps in the body of evidence for both sets of conditions. Studies found after the cut-off for our first searches did not change our conclusions but strengthened the evidence about biological plausibility, strength of association, and exclusion of alternative explanations.
The expert panel’s conclusions support causal links between Zika virus and congenital brain abnormalities and GBS and address Bradford Hill’s question, “…is there any other answer equally, or more, likely than cause and effect?” [7]. The conclusions consider both the epidemiological context of unexpected clusters of different types of neurological conditions in countries that have experienced their first outbreaks of Zika virus infection and the strengths and weaknesses of a systematically reviewed body of evidence about ten dimensions of causality (S4 Table, S5 Table, S6 Table and S7 Table). Empirical observations cannot “prove” causality, however [7, 142], and discussions have been intense [9]. A cause can be identified without understanding all the necessary component causes or the complete causal mechanisms involved [142, 143]. In the case of GBS, the infections that precede it are often referred to as “triggers” of the immune-mediated causal pathways involved in pathogenesis.
The body of evidence about Zika virus and congenital abnormalities (72 items included) has grown more quickly than that for GBS (36 items). Research efforts might have concentrated on congenital abnormalities because clusters of affected infants were so unusual, especially in Brazil, where rubella has been eradicated. In contrast, GBS is an established postinfectious neurological disorder, and some commentators have already assumed Zika virus as another cause [12]. Whilst only one case-control study from French Polynesia has been published so far [112], clusters of GBS during outbreaks of Zika virus infection have been reported from several other countries [144], and case-control studies are ongoing in Brazil, Colombia, Mexico, and Argentina.
Comparative studies from French Polynesia suggest that the risk of both microcephaly or of GBS is at least 30 times higher in people who had Zika virus infection compared to those who did not [47, 94, 112], although confidence intervals are wide. The true effect size might be weaker because the earliest studies investigating causality are often overestimates [145]. Even if the methods of forthcoming studies in Brazil [42] and elsewhere reduce confounding and bias, the increase in the risk of disease amongst those with Zika virus infection is likely to remain substantially raised. Inconsistencies in the evidence still need investigation, however. Disease clusters were not seen in Africa [146], but congenital abnormalities and GBS are rare complications that might not be detected in countries with small populations or poor surveillance systems. In the case of microcephaly, terminations of potentially affected pregnancies might have resulted in underascertainment [147].
Current evidence does not show which specific environmental and host factors interact with Zika virus. A cofactor that increases the risk of neurological damage could help to explain why surveillance reports show clusters of microcephaly or GBS in some geographical areas but not others. Dengue virus has been suggested as a possible cofactor or another component cause [143]. One major limitation to interpretation of data about causality and cofactors is the lack of accurate and accessible diagnostic tools, owing to the short duration of viraemia, cross-reactivity with other flaviviruses, and lack of standardisation [148].
The strengths of our study are that we appraised evidence of causality systematically but rapidly and transparently within a structured framework. We searched both published and unpublished sources. The systematic review process could not eliminate publication bias but reduced the risk that only positive reports in favour of causation would be evaluated. There were limitations to the process, too. Our search strategy did not cover the literature about analogous conditions or cofactors systematically. We did not select studies or extract data in duplicate, but additional reviewers checked the extracted data independently. The included studies used a variety of case definitions for microcephaly and GBS, and we could not standardise these, so misclassification is possible, but this limitation did not change the overall conclusions. Our rapid assessment of quality was not quantitative; we did not find a tool that covered all review questions and study designs appropriately and were not able to standardise the GRADE tool across study designs in the time available [20].
The conclusions of the expert panel facilitate the promotion of stronger public health measures and research to tackle Zika virus and its effects. The evidence gaps that we identified provide researchers with research questions, and WHO has published a Zika virus research agenda [149]. Better diagnostic tests will allow more accurate assessment of Zika virus in tissues and of population-level immunity. Research about Zika virus and acute flaccid paralysis is needed to define the clinical and electrophysiological pattern, mechanisms of causality, and to distinguish between the roles of autoimmunity and direct effects on anterior horn cells or neurons. Basic research will also further the development of vaccines, treatments, and vector control methods. For the populations currently at risk, cohort studies are needed to determine both absolute and relative risks of pregnancies affected by asymptomatic and symptomatic Zika virus infection and the role of cofactors and effect modifiers, and to define the full range of physical and developmental abnormalities that comprise the congenital Zika virus syndrome.
Our systematic review deals with multiple neurological disorders and more detailed questions about causality than other reviews. We reached the same conclusion as Rasmussen et al. [11], but the larger number of studies allowed a more comprehensive and balanced summary of evidence and of evidence gaps. In addition, our review addresses the association between Zika virus and GBS. We also plan to examine other acute neurological disorders (S1 Text).
Our review will quickly become outdated because the pace of new publications is outstripping the time taken for the review process. The concept of a “living systematic review” has been proposed as a way to combine rigour with timeliness for intervention research [150] through the development of methods to incorporate new evidence as soon as it is available and make evidence summaries available immediately. We are working on methods to produce a living systematic review of the Zika causality framework that will incorporate new studies, provide frequent open access updates, and allow cumulative meta-analyses of both aggregate and individual patient data from rigorous prospective studies as these become available. The declaration by journal editors to improve access to data during public health emergencies [151, 152] could be combined with the living systematic review approach to improve timeliness and open access to research about causality [153].
In summary, rapid and systematic reviews with frequent updating and open dissemination are now needed, both for appraisal of the evidence about Zika virus infection and for the next public health threats that will emerge. This rapid systematic review found sufficient evidence to conclude that Zika virus is a cause of congenital abnormalities and is a trigger of GBS.
|
10.1371/journal.pntd.0006272 | Management of severe strongyloidiasis attended at reference centers in Spain | Strongyloides stercoralis is a globally distributed nematode that causes diverse clinical symptoms in humans. Spain, once considered an endemic country, has experienced a recent increase in imported cases. The introduction of serology helps diagnosis and is currently replacing microbiological techniques in some settings, but its sensitivity is variable and can be low in immunocompromised patients. Diagnosis can only be confirmed by identification of larvae. Often, this “gold standard” can only be achieved in severe cases, such as disseminated S.stercoralis infection, or S.stercoralis hyperinfection syndrome, where parasite load is high. In addition, these clinical presentations are not well-defined. Our aim is to describe severe cases of S.stercoralis, their epidemiological profile, and their clinical details.
An observational retrospective study of disseminated S.stercoralis infection, or hyperinfection syndrome. Inclusion criteria: aged over 18, with a diagnosis of disseminated S.stercoralis infection, or hyperinfection syndrome, confirmed by visualization of larvae. Patients were identified through revision of clinical records for the period 2000–2015, in collaboration with eight reference centers throughout Spain.
From the period 2000–2015, eighteen cases were identified, 66.7% of which were male, with a median age of 40 (range 21–70). Most of them were foreigners (94.4%), mainly from Latin America (82.3%) or Western Africa (17.6%). Only one autochthonous case was identified, from 2006. Immunosuppressive conditions were present in fourteen (77%) patients, mainly due steroids use and to retroviral coinfections (four HIV, two HTLV). Transplant preceded the clinical presentation in four of them. Other comorbidities were coinfection with HBV, Trypanosoma cruzi, Mycobacterium leprae or Aspergillus spp. All presented with digestive disorders, with 55.6% also presenting malaise. 44.4% of cases had fever, 27.8% skin complaints, and 16.7% respiratory or neurological disorders. One patient presented anemia, and one other nephrotic syndrome. Diagnosis was confirmed by identification of larvae in fresh stool samples (n = 16; 88.9%), concentration techniques (n = 6; 33.3%), larval culture (n = 5; 29.4%), or digestive biopsies (n = 8; 44%). S.stercoralis forms were identified during necropsy in one case. In addition, ten (55%) had a positive serology. All the cases were treated with ivermectin, six (33%) also received albendazole and one case received thiabendazole followed by ivermectin. All needed inpatient management, involving a mean hospitalization stay of 25 days (range 1–164). Two cases received intensive care and eventually died.
Only eighteen cases of disseminated S.stercoralis infection/hyperinfection syndrome were identified from the 15-year period, most of which were considered to have been imported cases. Among those, immunosuppression was frequent, and mortality due to S.stercoralis was lower than previously described.
| Strongyloides stercoralis is a globally distributed worm. It has a free living cycle in wet moist soils, and an autoinfecting cycle affecting humans in their lungs, bowels and skin. Strongyloidiasis is the name of the infection caused by S.stercoralis and it can vary from an indolent state, with no symptoms at all, to a severe clinical condition if the adult worms reproduce and disseminate into the body tissues. This second clinical picture usually presents when the immune system of the host is altered for any reason, most likely other immunosuppressive infections or treatments. Severe clinical conditions are usually confirmed when an increased number of S.stercoralis larvae are found in body tissues, which leads to a broad spectrum of clinical symptoms that are usually fatal without treatment. These conditions are both uncommon and not well defined. We present the experience of eight referral centers in Spain regarding the management of severe strongyloidiasis. We found 18 cases diagnosed over 15 years, the second largest series outside endemic areas, and compiled their epidemiological, clinical and outcome data.
| Strongyloidiasis is an infection caused by Strongyloides stercoralis. Other Strongyloides spp. include S. fuelleborni, which infects some non-human primates and may cause self-limited disease in humans. S.stercoralis is widely distributed, affecting up to 370 million people worldwide according to recent estimates [1, 2], and some experts claim that its prevalence is increasing around the globe [1, 3]. Spain had an estimated prevalence of 0.2% in some rural areas during the first half of the last century [4], especially where wet crops, such as rice, and animal tracks were common.
The epidemiology of the disease in Spain changed notably in the early 80s owing to economic growth, the abandonment of traditional farming techniques and the mechanization of agriculture, along with an improvement in sanitation networks in rural areas. Although, some reports suggest that such transmission may still be occurring, [5, 6], all these advances have led to autochthonous cases becoming rare [5].
Economic growth brings both population ageing and an increase in the availability of diagnosis and treatment for chronic conditions, leading to a higher rate of comorbidities and also immunosuppressions in which severe cases might appear [6]. Recently published data from our group have shown a tenfold increase in hospital admissions where S.stercoralis is involved over the last decade [7].
This parasite undergoes a fascinating life-cycle that alternates between a free-living cycle in moist soil and a parasitic cycle in the host. The latter may last for decades, since the worm can replicate in the host and cause repeated autoinfections [8]. Parasite transmission to humans occurs during its free-living cycle through direct skin contact with previously contaminated soil containing rhabditiform larvae, although transmission through organ transplant has also been described [9, 10]. Therefore, poor sanitary conditions facilitate parasite transmission.
Strongyloidiasis is an indolent infection in most cases, although it can lead to mild digestive, skin or pulmonary symptoms. In some cases, usually due to comorbidities such as immunosuppression, replication of larvae can increase and they can disseminate into other tissues causing acute severe conditions [9]. These life-threatening forms of the disease comprise hyperinfection syndrome and disseminated strongyloidiasis. Both clinical pictures display a variety of symptoms, often with significant digestive involvement, that can produce paralytic ileus. In addition, larvae migration to the lungs can lead to hemoptysis and respiratory distress, and invasion of the brain can present as meningoencephalitis. Translocation of bowel bacteria causing sepsis and meningitis is often seen along with these findings. In both presentations, eosinophilia is an unfrequent finding and mortality can be as high as 62.7% [9]. Unfortunately, confirmation of diagnosis relies on visualization of parasite forms, which is unlikely with low parasite loads. Nevertheless, hyperinfection and disseminated forms are, by definition, situations where parasite load is high, so diagnosis in these cases is often confirmed.
As a Neglected Tropical Disease, strongyloidiasis is often forgotten about even where it is most common, and more so where it was once prevalent, but is no longer. Some authors state that S.stercoralis hyperinfection syndrome is in fact an emerging disease, and point to a lack of awareness among health-care professionals in non-endemic areas [11]. In particular, the clinical presentation of disseminated strongyloidiasis, or hyperinfection syndrome, is poorly defined in the literature, often based solely on case reports [6, 9]. Our aim is to define these severe forms of the disease as seen in our context.
This is an observational retrospective study of disseminated S.stercoralis infection, or hyperinfection syndrome, identified through revision of clinical records. Cases were tracked in collaboration with the Geohelminths Study Group of the Spanish Society of Tropical Medicine and International Health (SEMTSI). The period studied was defined as from 1 January 2000 through 31 December 2015.
Inclusion criteria: aged over 18, with a diagnosis of disseminated S.stercoralis infection, or hyperinfection syndrome, confirmed by biopsy, necropsy, or microbiological evidence of S.stercoralis adult forms, eggs or larvae. Hyperinfection was defined as the presence of signs or symptoms suggesting increased larvae migration and accelerated autoinfection, with microbiological evidence for it. Migration of larvae out of the skin, lungs or digestive tract was regarded as dissemination.
Exclusion criteria: a lack of a parasitological confirmation test, and/or simple classical forms of the disease.
Study variables included age, gender, country of birth, occupational history, whether the case was considered to be imported or autochthonous, duration of residence in Spain, clinical manifestations, diagnostic evidence, treatment received, and main outcomes observed.
Eosinophilia was defined as an eosinophil count in peripheral blood exceeding 4.5×108/L (450/μL) or more than 5% of the circulating leukocytes, in accordance with the consensus of the Spanish Society of Tropical Medicine and International Health expert group [12]. Data were collected through an anonymized on-line questionnaire, completed by study collaborators. Qualitative variables are expressed in absolute and relative frequencies, while continuous variables are described in medians and absolute ranges.
This work was performed in accordance with the ethical standards laid down in the Declaration of Helsinki as revised in 2013. The study protocol was approved by the Ethical Review Board of Vall de Hebron University Hospital (Barcelona, Spain) with the assigned code PR_AG_03–2016. Since this was a retrospective observational study, our Institutional Review Board accepted to proceed to data compilation and analysis with no previous informed consent obtained from the participants. All clinical and epidemiological data were anonymized.
Of the ten reference centers which participated in the study, two found no cases of disseminated strongyloidiasis nor hyperinfection syndrome, while the eight remaining centers were able to recover cases from their clinical records for the period defined. From the years 2000–2015, eighteen cases were identified. The majority were male (n = 12, 66.7%), which implies a sex ratio of 2:1, and the median age was 40 (range 21–69). Most of them were foreigners (n = 17, 94.4%), mainly from Latin America (n = 14, 82.3%) or Western Africa (n = 3, 17.6%). Of these patients, fourteen had information about the length of their period of residence in Spain, the median period being 7.21 years (range 2 months-31 years), and in two cases, it was possible to identify a recent travel history involving a visit to friends and relatives in their native country, respectively six and eleven months before the onset of symptoms. Further epidemiological details are provided in Table 1. Only one autochthonous case was identified, that of a 40-year-old construction worker from the Canary Islands with no travel history outside of Spain. He reported having walked barefoot in mud in a known endemic area in mainland Spain, and later developed hyperinfection syndrome after induced immunosuppression for a renal transplant.
Most of the cases fulfilled the criteria for hyperinfection syndrome (n = 17, 94.4%), and only one case of dissemination was identified and confirmed by necropsy (case 3). This allowed for identification of S.stercoralis in skin, lungs, bowels, brain, kidneys and lymph nodes. The clinical details of each patient are specified in Table 2 and summarized below.
Immunosuppressive conditions were identified in fourteen (77%) patients and were mainly due to prolonged steroids use (n = 9, 50%), followed by retroviral coinfections (four HIV, two HTLV). All the HIV cases were severely immunosuppressed, with CD4 counts of 8, 179, 10 and 22 at diagnosis (cases 5, 8, 9 and 16, respectively), and with detectable viral loads ranging from 390,000 to 4.300,000 copies/mL. In case 16, S.stercoralis severe infection was diagnosed along with HIV infection, while the others developed S.stercoralis hyperinfection syndrome within three months of starting antiretroviral therapy (ART). Of note, cases 5 and 9 had also received high dose steroids for cerebral toxoplasmosis before the onset of severe strongyloidiasis symptoms. Transplant preceded the clinical presentation in three patients, case 3 having received an allogenic blood stem cell transfer, and cases 7 and 13 having received a renal transplant. Case 2 developed hyperinfection syndrome after a renal transplant. Every transplanted case was also receiving steroids. A further two patients were under treatment with steroids; one for Sjögren syndrome, and the other for a type-2 lepromatous reaction (cases 11 and 17 respectively).
Other comorbidities were coinfection with Trypanosoma cruzi in three cases, hepatitis B virus in two cases, and Mycobacterium leprae and Aspergillus spp. in one case each.
All presented with digestive disorders. Ten (55%) patients presented with malaise. Eight patients (44%) had eosinophilia, as defined above, when strongyloidiasis was diagnosed; out of these, five were considered immunosuppressed. Of note, two out of the four HIV patients presented with eosinophilia, and case 16 (who was ART naïve at diagnosis) had an increase in his absolute and relative eosinophilia after starting ART, as happened with case 17 after steroid therapy was discontinued; Eosinophil count ranged from 191 eosinophils/μL (5.3%) to 11.200 eosinophils/μL (65.5%). Eight (44.4%) cases had fever, five (27.8%) had skin complaints, four (22.2%) presented with respiratory complaints and three (16.7%) with neurological disorders, one patient presented anemia, and one presented nephrotic syndrome.
Diagnosis was confirmed by identification of larvae in every case. Nine out of ten participating centers used Ritchie’s concentration technique and one obtained larvae after concentration with ethyl acetate. Larvae were cultured with charcoal culture, Harada-Mori filter paper or blood agar plate. Every case had a large number of filariform larvae in fresh stool samples (n = 16; 88.9%) concentration techniques (n = 6; 33.3%) or larval culture (n = 5; 27%). In some patients, S.stercoralis adult worms, eggs and larvae were identified in digestive and extra-digestive biopsies (n = 8; 44%), often described with an eosinophilic infiltrate around them. S.stercoralis forms were identified during necropsy in one case. In addition, ten (55%) had a positive serology, not performed in the remaining eight patients as the parasite had already been identified otherwise.
All the cases were treated with ivermectin adjusted to 200 mcg per kg a day, with the duration of treatment ranging from two to ten days. Three patients receiving ivermectin for two days had this regimen repeated one or two weeks later. Furthermore, six (33.3%) also received albendazole 400 mg before or concomitantly with ivermectin. One patient was treated with thiabendazole 1750 mg for three days in addition to a single dose of ivermectin. All needed inpatient management with a mean hospitalization stay of 25 days (range 1–164).Two cases (11.1%) needed intensive care and eventually died.
We found eighteen cases of severe S.stercoralis infection during the 15-year period studied in four different regions of Spain, most of which were imported, mainly from Latin America. The male to female ratio was 2:1, and the majority of patients were young adults. Immunosuppression frequently preceded the onset of symptoms, and mortality was lower than hitherto described, with a rate of 11.1%.
Robson et al. sought cases of S.stercoralis hyperinfection syndrome diagnosed in the United Kingdom after the Second World War to our date and described the largest series outside endemic areas [13]. Other recent studies on this topic are case reports, or hospital case-series describing any type of strongyloidiasis regardless of its severity [6, 14–16]. So ours is the biggest case series compilation outside endemic areas after that from Robson et al., who described severe cases over six decades.
One of the strengths of this study is the circumstances of the patients; since they were attended in reference centers, we have been able to fully describe the clinical picture and outcome of severe cases when all health-care resources are available in one place. However, this is also a disadvantage, as our data are not translatable to other clinical settings, nor to endemic areas. A further limitation of our study stems from the use of clinical records, as at times information was lacking due to the fact that it relied on the health worker’s thoroughness in registering data. For this reason, detailed information on exposure to risk factors was not available.
The age, sex and origin of the participants match the epidemiology described in previous hospital series in our country [6, 14–16], probably reflecting the epidemiology among our immigrant population. Clinical presentations are in accordance with those seen in previous case reports of severe strongyloidiasis, with a predominance of digestive tract disorders and general physical malaise. Nevertheless, these clinical symptoms might be undistinguishable from a complicated chronic S.stercoralis infection without microbiological evidence of a high number of larvae. Eosinophilia was a relatively frequent finding (n = 8, 44%) for such a population, since it is less frequently observed when cellular immunity is compromised. Of note, two of our patients with HIV developed eosinophilia once they started ART, and another presented eosinophilic infiltrate in a duodenal biopsy, findings which are in accordance with those described in other series [15, 17].
Regarding comorbidities, a recent systematic revision of severe strongyloidiasis cases described a large number as having received steroids before the onset of symptoms (67%), a 15% HIV coinfection rate, and a further 11.5% of cases being transplant recipients [9]. In our series, immunosuppression could be identified in two-thirds of the cases, predominantly among those receiving steroids and oncological chemotherapy, but four patients seemed to have no factor which would have triggered a massive increase in S.stercoralis larvae.
Serology is a sensitive tool although not highly specific. However, sensitivity can decrease to 42.9% in immnunocompromissed patients [18]. This technique was available in every center of our group. Since all our cases were confirmed with microbiological evidence, serology was only requested in ten of them. Out of these, six presented an immunosuppressive condition, raising the highest sensitivity described in such cases (100%), but we believe these numbers are not significant. Serology has also been recently proposed as a tool to monitor treatment success [16]. Unfortunately, it was not systematically requested during our patients’ follow-up.
Recent research into screening strategies for imported diseases found a significant rate of intestinal parasites among those awaiting treatment for oncohematological malignancies [19]. Steroid treatment and immunosuppressive drugs provided prior to and after transplants were both attributable causes of severe strongyloidiasis in half of the patients in our series, since S.stercoralis infection was thought to have been present before these therapies were started. Nevertheless, other recent reports have also described S.stercoralis infection in donors for both solid organ and stem cell transplants [20, 21]. Since the origin of the donors is unknown for the four patients who had received a transplant, they cannot be ruled out as a possible source of infection.
Some expert groups in tropical diseases have found rates of up to 18.4% of S.stercoralis infection in HIV patients coming from endemic areas, suggesting that this coinfection may be common [22]. Among those infected with HIV in our series, the development of severe strongyloidiasis could be attributed to HIV infection alone in only one patient (case 16), while a further three might have developed S.stercoralis severe infection as a manifestation of an immune reconstitution phenomenon, as described in the literature [23–25]. However, cases 5 and 9 were also receiving high-dose steroids as part of the treatment for a cerebral toxoplasmosis, which probably played a role in triggering a higher larval load.
An evidence-based guideline has been published during the development of this work. It recommends combination therapy with ivermectin and albendazole as the treatment of choice for severe cases [26], which was the empirical treatment given to 33% of our cases. The different treatment strategies used during our study period were chosen according to the published evidence at the time and drug availability, meaning ivermectin was not always present at the center when the patient arrived and albendazole or thiabendazole had to be launched initially, while awaiting ivermectin. Nevertheless, some drug regimens had to be repeated weeks apart, in some cases lasting up to months, since no parasitological and/or clinical improvement was observed with single dosages, probably reflecting prepatent autoinfections arising due to the increased larvae replication.
Buonfrate et al. found that all the patients suffering from severe strongyloidiasis who received ivermectin survived [9]. This treatment is the current drug of choice for the treatment of S.stercoralis infection and was given to every patient in our series [27, 28], what might explain the relatively low mortality described here (11.1%). In addition, they were managed in reference centers where the disease was suspected and diagnosed early, and where the coverage of concomitant sepsis with broad-spectrum antibiotics or intensive care were assured.
Raising awareness about the disease among populations-at-risk and healthcare professionals is strongly recommended [26]. During a S. stercoralis and Trypanosoma cruzi screening campaign, performed in 2016 among Latin-American immigrants in Alicante, Spain, it was found that 92.2% of participants (119/129) had never heard of strongyloidiasis, including none of the ten participants who had a positive S. stercoralis serology (personal communication). Moreover, a questionnaire about five Neglected Tropical Diseases completed by students from Madrid in their final year of Medicine, revealed that less than 18% of the students (18/103) ‘passed the exam’ on strongyloidiasis, this being one of the most worrying results [29].
We conclude that screening for strongyloidiasis should be mandatory for HIV patients, as well as for both transplant recipients and donors coming from endemic areas. Infection should also be ruled out in those diagnosed with HTLV-1 infection, and ideally before the onset of steroid treatment. At the time this paper was under revision, a panel of experts published an evidence-based guideline supporting this recommendation with a Ia grade [30]. It is clear that disease outcomes improve when clinicians are aware of the infection and ivermectin supply is available for patients who require it.
|
10.1371/journal.pntd.0004135 | Seasonal and Spatial Dynamics of the Primary Vector of Plasmodium knowlesi within a Major Transmission Focus in Sabah, Malaysia | The simian malaria parasite Plasmodium knowlesi is emerging as a public health problem in Southeast Asia, particularly in Malaysian Borneo where it now accounts for the greatest burden of malaria cases and deaths. Control is hindered by limited understanding of the ecology of potential vector species.
We conducted a one year longitudinal study of P. knowlesi vectors in three sites within an endemic area of Sabah, Malaysia. All mosquitoes were captured using human landing catch. Anopheles mosquitoes were dissected to determine, oocyst, sporozoites and parous rate. Anopheles balabacensis is confirmed as the primary vector of. P. knowlesi (using nested PCR) in Sabah for the first time. Vector densities were significantly higher and more seasonally variable in the village than forest or small scale farming site. However An. balabacensis survival and P. knowlesi infection rates were highest in forest and small scale farm sites. Anopheles balabacensis mostly bites humans outdoors in the early evening between 1800 to 2000hrs.
This study indicates transmission is unlikely to be prevented by bednets. This combined with its high vectorial capacity poses a threat to malaria elimination programmes within the region.
| The first natural infection of Plasmodium knowlesi was reported 40 years ago. At that time it was perceived that the infection would not affect humans. However, now P. knowlesi is the predominant malaria species (38% of the cases) infecting people in Malaysia and is a notable obstacle to malaria elimination in the country. Plasmodium knowlesi has also been reported from all countries in Southeast Asia with the exception of Lao PDR and Timor Leste. In Sabah, Malaysian Borneo cases of human P. knowlesi are increasing. Thus, a comprehensive understanding of the bionomics of the vectors is required so as to enable proper control strategies. Here, we conducted a longitudinal study in Kudat district, Sabah, to determine and characterize the vectors of P. knowlesi within this transmission foci. Anopheles balabacensis was the predominant mosquito in all study sites and is confirmed as vector for P. knowlesi and other simian malaria parasites. The peak biting time was in the early part of the evening between1800 to 2000. Thus, breaking the chain of transmission is an extremely challenging task for the malaria elimination programme.
| Significant progress has been made fighting malaria in the last decade, decreasing the incidence of cases and mortality by 30% and 47% respectively on a global scale [1] and reducing cases by 76% in Asia Pacific countries [2]. The development and use of better tools for diagnostics and treatment [3] coupled with substantial increases in the coverage of vector control methods such as Long Lasting Insecticide Treated bednets (LLINs) and Indoor Residual Spraying [4] have contributed to these successes.
An additional challenge to malaria elimination is the existence of a zoonotic reservoir of malaria. The primate malaria Plasmodium knowlesi has recently been documented as causing human infections in multiple countries in Southeast Asia [5–11], and is a serious public health problem within Malaysia [12–19]. In the Malaysian state of Sabah, this parasite is now responsible for the greatest number of malaria cases with 815 and 996 cases reported respectively in 2012 and 2013 [20].
This growing burden of P. knowlesi presents a notable obstacle to malaria elimination in Malaysia where historically, most transmission has been due to human-specific parasite species [17]. Since 2011, Malaysia has made great progress towards the elimination of these human malaria species, leading to a target for complete elimination by 2020 [21]. Whether existing elimination targets can be met in the face of increasing P. knowlesi cases with this Plasmodium now causing 38% of human malaria cases in Malaysia in 2012 remains to be seen.
Two features of P. knowlesi make it particularly difficult to control by conventional methods: (1) it has a sizeable zoonotic reservoir in macaques, which means that even if infections are eliminated from humans there remains a risk of future spillover, and (2) current evidence indicates that previously incriminated mosquito vectors of P. knowlesi in Malaysia bite and rest outdoors where control methods such as LLINs and IRS will not be effective [19, 22, 23].
Incrimination of vector species responsible of P. knowlesi transmission is a crucial first step for planning control but limited data is available on vectors of simian malaria in this region. Mosquitoes belonging to the Anopheles leucosphyrus group are thought to be responsible for P. knowlesi transmission. Anopheles hackeri was the first species to be incriminated as a vector, in the coastal area of Selangor [24], followed by An. latens in Kapit, Sarawak [22, 25], An. cracens in Kuala Lipis [14, 23] and An. introlatus in Hulu Selangor [19]. In Vietnam An. dirus, was incriminated as the P. knowlesi vector [26, 27]. The considerable spatial variation in P. knowlesi vector species both within and beyond Malaysia reinforces the need for detailed studies of vector ecology in a localized context to guide appropriate control strategy.
Anopheles balabacensis is hypothesized to be the primary vector of P. knowlesi within the current, extensive transmission foci of P. knowlesi in Sabah. Based on extensive studies carried out in the region in the 1980s [28–31] An. balabacensis was incriminated as a vector of human malaria, and laboratory studies that showed that An. balabacensis can be experimentally infected with P. knowlesi [32]. Since this early work, there has been significant ecological change occurring throughout Sabah due to conversion of forest to palm oil plantations [33–35]. How these changes have impacted the abundance, diversity and transmission potential of P. knowlesi vectors needs to be investigated. Control of P. knowlesi in Sabah requires confirmation that An. balabacensis remains the most likely vector, and characterization of its dynamics within a range of habitats that reflect current land use patterns. For that purpose, we conducted a 12 month longitudinal study within the large, ongoing focus of P. knowlesi transmission in Kudat and Banggi Island (Kudat District) in Sabah, aiming to characterize the abundance and biting behavior of potential vector species and incriminate vector species. These findings will be of importance to guide the development of local vector control programmes to eliminate malaria transmission.
Studies were conducted in three sites: Timbang Dayang (TD) (117°102’92”E, 7°155’85”N) and Limbuak Laut (LL) 117°065’75”E, 7°215’84”N) on Banggi island, and Kampung Paradason (KP) (116°786’35”E, 6°768’37”N on mainland Kudat (Fig 1). These sites were selected to reflect the range of ecotypes broadly representative of the study area in Northern Sabah: small scale farming (TD), secondary forest (LL) and a village settlement (KP). Sightings of macaques and recent human cases of P. knowlesi were reported near all sites. Timbang Dayang is a village with a population of 180 people. It is situated in a hilly landscape where houses are surrounded by small farming areas ~200 meters from the edge of secondary forest. These small farms (>1 hectare) contain mixed agriculture primarily for household consumption, including maize, banana and fruit trees. The mosquito collection site was near the edge of farm approximately 150 meters from the group of houses.
Limbuak Laut (LL) is a village consisting of 144 people, with houses situated on a road bordering closed canopy secondary forest. Mosquito sampling was conducted at a point situated within the secondary forest, at a distance of approximately 500 meters from the edge of the forest.
Paradason in Kudat is a village of 160 people situated in a heavily cultivated area, characterized by swidden farming and small plantations of rubber and palm oil. The area is undergoing a high rate of environmental change, including frequent burning and clearing of land. Little intact secondary forest remains in this area. The local community lives in both individual houses and a traditional communal longhouse shared between six households. Here mosquitoes were sampled at a point near the longhouse (100m), and in an associated garden area 75m away from the first collecting point.
Mosquitoes were collected by human landing collections (HLC) which were carried out monthly at all sampling sites from August 2013 to July 2014 (three nights per month at TD and LL and two nights in KP). Two men per team carried out collections at each site from 1800 to 0600 hrs. Mosquitoes landing on the legs of catchers were captured individually in vials which were then plugged with cotton wool and labelled by hour and collection sites. A supervisor visited the team hourly to ensure collections were being carried out. In TD and LL, collections were conducted by one team each night, whereas two teams (situated ~75 m apart) worked each night in Kudat. Thus a total of six individual human landing catches were performed each month in TD and LL, and eight per month at KP.
In the laboratory Anopheles mosquitoes were identified using the keys of Reid (1968) and Sallum (2005). Specimens morphologically identified as An. balabacensis were further confirmed by PCR and sequencing analysis of ITS2 and CO1 genes [19]. Anopheles mosquitoes were dissected to extract their ovaries, midguts and salivary glands to determine parity, oocyst and sporozoites respectively. All positive midguts and salivary glands, and the corresponding head and thorax of these positive specimens were transferred into individual microcentrifuge tubes containing 95% ethanol for subsequent molecular analysis.
Ethanol was allowed to evaporate completely from specimen tubes by placing them in a Thermomixer (Eppendorf, Germany) at 70°C. Genomic DNA was extracted from the guts and glands using the DNeasy tissue kit (Qiagen, Germany) according to the manufacturer’s protocol. The eluted DNA was kept at -20°C until required. A nested PCR was performed to detect and identify human specific malaria parasites (Plasmodium falciparum, P. vivax, P. malariae and P. ovale) and P. knowlesi found in mosquitoes using primers based on the Plasmodium small subunit ribosomal RNA (ssurRNA) [12, 36]. Primers and protocol used for human malaria and P. knowlesi detection were as developed by Singh et al [12] and Lee et al [37] for other simian malaria. Positive and negative controls were also included for each batch of assays.
Statistical analysis was conducted using PASW Statistics 18 and R programming language for statistical analysis (version 3.2.0). Generalised linear mixed models (GLMM) were constructed to analyse the following parameters of interest: the abundance of An. balabacensis, their time of biting, and the proportion of vectors that were (i) infected with oocysts, (ii) infected with sporozoites, and (iii) that were parous. In all analyses, locality (TD, LL or KP) was fit as a fixed effect. Month was fit alternatively as a fixed (to predict monthly values) or random effect (to test for differences between localities while controlling for seasonal variation).
Poisson and negative binomial distributions were used separately in the analysis of mosquito abundance, while a binomial distribution was assumed in all analysis of proportion data (parity and infection rates). Zero inflation in count data (mosquito abundance) was assessed. Models testing associations between response variables (vector abundance, parity and infection rates) explanatory variables (locality and month) and random effects of sampling night were assessed through comparison on the basis of having higher log-likelihood and lower Akaike information criterion (AIC) values, as well as the result of analysis of variance (ANOVA) of nested models). Tukey post hoc contrasts were used to differentiate the nature of statistical differences between localities. Graphs were produced using GraphPad Prism 6.0.
This project was approved by the NMRR Ministry of Health Malaysia (NMRR-12-786-13048). All volunteers who carried out mosquito collections signed informed consent forms and were provided with antimalarial prophylaxis during participation.
A total of 1884 Anopheles belonging to ten different species was obtained of which An. balabacensis predominated (95.1% of total, Table 1) in all sites. Other species of Anopheles were found in very low numbers and present in one or two localities only. Anopheles balabacensis was the only species from the Leucosphyrus group caught. A total of 379 Culicines were obtained but were not identified to species.
The number of An. balabacensis collected in HLC ranged from ~2–28 per man night, but did not show any clear, consistent trend in seasonality (Fig 2). The pattern of seasonal fluctuation differed between sites (Fig 2). In the forest site (Fig 2A), An. balabacensis abundance was relatively low (<15 per night) and constant across months. In the small farming site, An. balabaensis varied more than 10-fold over the course of a year, with a high in August and November, and low from February-to May and July 2014. Anopheles balabacensis abundance was more variable in the village settlement (Fig 2C). Here the highest monthly density was observed in January (27 per night) with values <1 per night in October and November.
Analysis using GLMM models indicated that the Poisson distribution was generally a better representation of An. balabacensis abundance data than the negative binomial. On the basis of statistical models assuming a Poisson distribution, the Tukey post hoc test indicated that An. balabancensis abundance was significantly higher in the village settlement (KP) than in the two other localities, (KP and LL: (p = 0.04; KP and TD: p = 0.02;Table 2). Controlling for variation across months, An. balabancensis abundance in the village site was estimated to be ~15–20% higher than in the other localities.
As shown in Fig 3 An. balabacensis started to bite as early as 1800 hours and continued to bite throughout the night until early hours of the morning. The peak biting time occurred between 1800 to 2000hrs in both LL and KP (Fig 3A and 3C), accounting for 38% of the total night catch. In TD, biting rates were relatively similar between 1800-2400hrs, then began to fall with a second small peak in the early part of the morning (0300-0400hrs, Fig 3B).
The parous rate of An. balabacensis was more than 50% on most collections, in all sites (Fig 4). The mean parous rate varied between 58 to 65%, with little fluctuation (Fig 4, Table 3). Statistical analysis indicated no evidence of significant variation in parity rates between all 3 sites (p>0.05, Table 2). On the basis of the parous rate, a daily survival rate [38], life expectancy [39] and vectorial capacity values were calculated [40] for An. balabacensis at each site. Estimates of An. balabacensis survival and vectorial capacity were predicted to be higher in LL and TD compared to KP (Table 3). In LL and TD respectively, 24% and 22% of An. balabacensis would be expected to live the 10 days necessary for P. knowlesi to develop into transmission-stage sporozoites, contrasting with only 16% in KP. Those surviving the 10 days would have a further life expectancy of 7 and 6.7 days in LL and TD respectively, compared to 5.4 days in KP. Vectorial capacity was predicted to be highest in LL with an estimated value of 3.85.
Forty five (3%) An. balabacensis out of the 1482 dissected were found to be positive for Plasmodium infection in terms of either sporozoites (14), oocysts (18) or both (13) by microscopy. Of these only 10 salivary glands and three midguts were positive for P. knowlesi by PCR. Besides P. knowlesi other simian malaria parasites were also present as shown in Table 4. This shows that in addition to P. knowlesi, An. balabacensis is also a vector to other simian Plasmodium species as well.
Due complexity of infection the subsequent discussion refers to all Plasmodia. There was no consistent seasonal pattern of mosquito infection rates across sites (Fig 5). In LL sporozoite rates were highest from December to February (4–16.67%). In the TD, sporozoite rates were high in December (5.00%) and in June to July (7.69–12.50%). In March, only three mosquitoes at TD were dissected of which two were found to be positive; one for sporozoites and one for oocyst. Thus, sporozoite rates appear to be extremely high at this time, but it is likely an artifact of low sample size. In KP the highest sporozoite rate was obtained in May 2014 (2.86%). The highest entomological inoculation rate (EIR) was 0.6 in TD in June. Tukey post hoc tests performed on the results of statistical models of An. balabancensis infection rates indicated there was variation between sites. Specificially, sporozoites rates were lower in KP compared to LL (p = 0.04), and oocyst rates were lower in KP than in TD (p = 0.035) (Table 2). Sporozoite rates were estimated to be approximately 2 and 3 times higher in the LL and TD respectively than in KP (Table 2).
Our study provides the first evidence to confirm that An. balabacensis is the vector of the zoonotic malaria P. knowlesi within the substantial foci of human infection in Sabah. It was the predominant species found in all sites with mean biting rates ranging from 6.8 to 8.8. A substantial proportion of An. balabacensis (32.8%) were captured biting outdoors in the early part of the evening (1800–2000), a time when humans would not be expected to be using LLINs, which is the current front line malaria control strategy in Malaysia. In this study all collections were made outdoors, as previous studies have found that this is where the majority of An. balabacensis (~76%) host seek [28, 41]. However, we note that total amount of human exposure to infectious bites from An. balabacensis may be even higher than indicated here if the additional contribution of limited indoor exposure were to be incorporated. In comparing the density and bionomics of An. balabancensis populations between three sites, we found evidence of geographical variation in both their abundance and sporozoite infection rate. Vector abundance was highest in the village site, whereas sporozoite rates were higher in the forest and small farming site than in the village site. However, it is unknown whether these differences are truly the result of habitat-dependent transmission efficiencies, as only one site from each ecotype was sampled. However these findings reinforces the hypothesis that spatial heterogeneity in P. knowlesi exposure risk may be driven by variation in mosquito vector demography in addition to the presence of the reservoir macaque host.
Although it has been postulated that P. knowlesi was present in macaques before the arrival of humans in Southeast Asia [42], and a large number of P. knowlesi malaria cases has been reported from Sabah [20], the identity of the vector remained elusive. Whilst it has been demonstrated by Chin et al [43] that An. balabacensis can transmit P. knowlesi from monkey to man, man to monkey and man to man under experimental conditions, this study is the first to confirm that it acts as a vector under natural conditions. Anopheles balabacensis was also incriminated as the primary vector of human malaria in Sabah in the 1950s [44, 45]; a role that was further supported with extensive studies in the 1980s which confirmed its role as the main vector for human malaria infections [28, 46]. Given An. balabacensis is the likely vector of other primate malaria species in this area, it could also be the conduit for other zoonotic malaria spillovers to humans. This indicates that these Plasmodia species are not partitioned amongst different vector species, and emphasizes that An. balabacensis should be the primary target for all malaria control efforts in the area.
We observed a significant difference in the Anopheles species composition found here relative to previous studies in Sabah. Currently An. balabacensis and An. donaldi constituted > 95% and 1.3% of all Anopheles recorded on Bangii island respectively, while studies in this area in the 1980s estimated the relative proportion of these species to 13.6%, and 39% of Anopheles respectively [29]. In the central region of Sabah An. donaldi was incriminated as the dominant vector for human malaria parasites in studies carried out in 2001–2002 [41]. We did not document infection in An. donaldi within this study, but this may because too few were collected (n = 25) for reliable detection. Thus, we cannot dismiss the possibility that An. donaldi remains in other areas of Sabah where it is most abundant. The cause of this apparent shift in malaria vector species composition over the past 40 years in Banggi Island is uncertain although it coincides with a period of extensive deforestation in Sabah [33, 34]. One possibility is that this is just an artefact of sampling, as here we did not conduct sampling in the exact same locations as historical studies, but instead targeted sites of known human P. knowlesi infection. These sites may have inherently higher densities of An. balabancensis (thus triggering P. knowlesi infection) than other locations within the area. However, there is grounds to hypothesize this could be evidence of long-term shift in species composition in response to the rapid deforestation or prolonged use of interventions such as LLINs or IRS as has been documented elsewhere [47]. In previous work within the Kinabatangan area of Sabah, we have also documented a shift from a high proportion of An. balabacensis to dominance of An. donaldi within the same sites over the period 1980s to 2000 [29, 41]. Regardless of the explanation for the dominance of An. balabacensis within this study the relatively high survival and sporozoite rates in this vector coupled with the potentially increased contact of human-vector-macaques have likely made major contributions to the increase in P. knowlesi cases in the area.
Although Sabah has reported a large number of P. knowlesi cases in the past few years especially in Kudat district, it is hypothesized that people are only getting infected when they visit forested areas. Within our current study sites, the number of malaria cases occurring over the sampling period ranged from 1.9 to 2.5 cases per 100 people [48]. As positive An. balabacensis were present in most months of the year and most of the infective mosquitoes (40%) were captured biting in the early part of the evening between1800 to 2000, people could be exposed when they return from work in or around forested areas. Our preliminary studies now and previously have demonstrated that the Anopheles mosquitoes start biting only after 1800 hrs. The average biting rate reported for An. balabacensis here is much higher than in previous studies conducted in the 1980s (eg. 6.8 to 8.8/night compared to 0.75 to 4.44) [28]. These biting rates are also considerably higher than has been reported for An. latens (0.95 to 4.71 bites per night) in Sarawak [22]. The high density of An. balabacensis in this area combined with its relatively high sporozoite rates with all simian malaria (1.82%) and P. knowlesi in particular (0.67%) indicate it is most likely responsible for the majority of transmission in this area.
In this study all mosquitoes were collected using human bait, thus results are only directly informative for estimating potential human exposure and not transmission between macaques. Ideally parallel collections of mosquitoes attracted towards macaques would have been conducted but this was not possible due to logistical constraints and ethics regulations for working with macaques. Previous work [22, 23] showed that the P. knowlesi vectors in other areas namely An. latens and An. cracens were attracted to both humans and macaques. Furthermore in Palawan Island, Philippines, An. balabacensis was more attracted to a monkey baited trap than traps baited with water buffalo or humans and individuals host seeking on macaques had oocyst and sporozoites (although malaria species unconfirmed) [49]. Thus, although data for mosquitoes biting macaques are not available here, we could expect, that transmission between macaques to be at least as high or much greater than predicted for humans here.
To further resolve the transmission dynamics of P. knowlesi in primates, these studies should be expanded to incorporate assessment of the host preference and choice of An. balabacensis and other potential vectors most directly through analysis of the blood meals in randomly sampled resting females [50]. However, collection of recently blood fed mosquitoes resting outdoors has proved challenging. To overcome this limitation ongoing work is also investigating the use of new sampling methods to increase feasibility of such data collection in the future.
The high rate of parity, survival and sporozoite infections in this mosquito indicates that An. balabcensis is a highly competent vector. With a very high vectorial capacity and life expectancy, An. balabacensis will continue to pose a risk of human infection. As Malaysia moves towards malaria elimination, breaking transmission under these conditions will be extremely challenging, further complicated by the presence of a sizeable macaque reservoir.
Current frontline malaria control measures in this area are insecticide treated bednets and indoor residual spraying but more innovative control methods that specifically target outdoor biting mosquitoes such as the use of repellents or attractive toxic sugar baits will be essential.
|
10.1371/journal.pbio.1001666 | Hypoxic Regulation of Hand1 Controls the Fetal-Neonatal Switch in Cardiac Metabolism | Cardiomyocytes are vulnerable to hypoxia in the adult, but adapted to hypoxia in utero. Current understanding of endogenous cardiac oxygen sensing pathways is limited. Myocardial oxygen consumption is determined by regulation of energy metabolism, which shifts from glycolysis to lipid oxidation soon after birth, and is reversed in failing adult hearts, accompanying re-expression of several “fetal” genes whose role in disease phenotypes remains unknown. Here we show that hypoxia-controlled expression of the transcription factor Hand1 determines oxygen consumption by inhibition of lipid metabolism in the fetal and adult cardiomyocyte, leading to downregulation of mitochondrial energy generation. Hand1 is under direct transcriptional control by HIF1α. Transgenic mice prolonging cardiac Hand1 expression die immediately following birth, failing to activate the neonatal lipid metabolising gene expression programme. Deletion of Hand1 in embryonic cardiomyocytes results in premature expression of these genes. Using metabolic flux analysis, we show that Hand1 expression controls cardiomyocyte oxygen consumption by direct transcriptional repression of lipid metabolising genes. This leads, in turn, to increased production of lactate from glucose, decreased lipid oxidation, reduced inner mitochondrial membrane potential, and mitochondrial ATP generation. We found that this pathway is active in adult cardiomyocytes. Up-regulation of Hand1 is protective in a mouse model of myocardial ischaemia. We propose that Hand1 is part of a novel regulatory pathway linking cardiac oxygen levels with oxygen consumption. Understanding hypoxia adaptation in the fetal heart may allow development of strategies to protect cardiomyocytes vulnerable to ischaemia, for example during cardiac ischaemia or surgery.
| Regulation of oxygen usage in cardiomyocytes is of great medical interest, because adult cardiac tissue is extremely vulnerable to hypoxia during myocardial infarction and cardiac surgery. While some progress has been made toward protecting cardiomyocytes from hypoxia in these circumstances, it has been limited by a lack of understanding of endogenous oxygen-sensing pathways. In contrast to adult cardiac tissue, embryonic cardiomyocytes are highly resistant to hypoxia, although the mechanisms underlying this have hitherto been unclear. Using mice we show that the transcription factor Hand1 is expressed at high levels in the fetal heart, under direct control of HIF1α signaling, a pathway well known to respond to hypoxia. We show that Hand1 expression decreases at birth as the neonate is exposed to higher levels of oxygen. By experimentally increasing Hand1 expression in the neonatal heart, we see lower oxygen consumption in cardiomyocytes and this is caused by Hand1 repressing key regulatory genes involved in cardiomyocyte lipid metabolism. This has the effect of decreasing mitochondrial ATP generation via the tricarboxylic acid cycle. Furthermore, we show that increasing Hand1 expression in adult transgenic hearts is protective against myocardial infarction, suggesting that a hypoxia–Hand1 pathway may also be of importance in the adult heart.
| Adult cardiomyocytes are particularly vulnerable to hypoxia, which can result in cellular dysfunction and death. While some progress has been made towards protecting cardiomyocytes from the deleterious effects of low oxygen levels, cardiac ischaemia continues to result in high mortality and remains a clinical challenge [1]. Fetal cardiomyocytes, in contrast, are adapted to function at extremely low oxygen levels. In the relative hypoxia of the womb, the fetal heart generates ATP predominantly from glucose via glycolysis and energy generation shifts to mitochondrial β-oxidation of lipids and the tricarboxylic acid cycle as the primary source of energy for the heart only after birth, when oxygen levels are abundant [2],[3]. It is striking that to varying degrees this is reversed in the adult failing heart.
Little is known about the mechanisms regulating these changes in cardiac energy metabolism. Specifically, the molecular link between oxygen availability and molecular control of energy metabolism is currently ill-defined. In the mouse, cellular and metabolic changes in the heart occur rapidly after birth, the switch from glycolysis to β-oxidation of lipids and the tricarboxylic cycle occurring within a week of birth [2],[4]–[7]. Under conditions of “stress” such as those leading to cardiac hypertrophy and heart failure, fetal-type glycolytic energy metabolism is known to recommence [8]. This is accompanied in the failing heart by re-expression of several isoforms of enzymes, of transcription factors, and of structural and other proteins normally expressed in the fetal heart. Metabolic remodeling therefore appears to be part of a broader phenotypic switch between adult and fetal states [8]. Analyzing the mechanisms regulating changes in the heart during the transition from fetus to neonate, in particular the changing sensitivity to oxygen levels, may therefore inform efforts towards more effective therapeutic intervention for heart failure. Furthermore, understanding how the fetal heart is adapted to hypoxia may allow development of strategies to protect cardiomyocytes vulnerable to ischaemia, for example during myocardial ischaemic events or cardiac surgery [1].
We have previously reported studies suggesting a role for elevated levels of the cardiac transcription factor Hand1 in the derangement of electrical conductivity and arrhythmia of the failing heart [9]. Here we show that cardiac levels of Hand1 mRNA and protein fall immediately following birth and investigate the control and consequences of this postnatal down-regulation. Expression of Hand1 mRNA is up-regulated by HIF signaling, and Hand1 itself inhibits the expression of a number of genes involved in cardiomyocyte lipid metabolism. Through inhibition of lipid metabolism, Hand1 reduces cellular oxygen consumption and cardiomyocyte vulnerability to ischaemia. We hypothesise that in the cardiomyocyte, Hand1 is part of a pathway linking ambient oxygen levels to oxygen consumption through regulation of cellular lipid metabolism.
Hand1 mRNA is expressed at high levels in the developing embryonic heart [10],[11]. In contrast, only low levels have been reported in adult cardiac tissue [12]. We therefore investigated whether Hand1 expression changed around birth or during subsequent growth and maturation of the heart. Using quantitative RTPCR analysis of cDNA from mouse hearts, we found that Hand1 mRNA levels decrease rapidly in the immediate postnatal period (Figure 1A), whereas those of the related transcription factor Hand2 remain largely unaltered (Figure 1B). We found that protein levels of Hand1 and 2 broadly follow those of mRNA (Figure 1C). In the early embryo, Hand1 is expressed at high levels in the developing trophoblast and subsequently in the developing heart [11],[13]. Both tissues are known to be dependent on hypoxia signaling for normal development [14]–[17], and we therefore investigated whether the steep fall in perinatal cardiac Hand1 levels is due to the rapid change in ambient oxygen levels encountered by the neonate around birth.
We first tested whether ambient levels of oxygen affect Hand1 levels in the adult mouse, incubating wild-type adult mice in 12% oxygen for 2 wk and comparing Hand1 expression levels with those of controls maintained under normoxic conditions. Such exposure to hypoxia resulted in 18-fold elevation of Hand1 mRNA levels compared with controls (n = 4 each group, p = 0.001) (Figure 1D), indicating that cardiac Hand1 levels are directly or indirectly hypoxia-inducible.
We next tested whether neonatal Hand1 expression is directly dependent on HIF1α signaling by examining Hand1 mRNA levels in transgenic mice showing constitutive hypoxic signaling. Cardiac-specific deletion of VHL in αMHC-cre::VHL(fl/fl) neonates results in constitutive cardiac expression of HIF1α [18]. At p0.5, Hand1 mRNA and protein were significantly elevated in αMHC-cre::VHL(fl/fl) hearts compared with nontransgenic controls (Figure 1E,F). Furthermore, chromatin immunoprecipitation (ChIP) showed binding of HIF1α to two of the five canonical HIF binding sites located in 5 kb of sequence upstream of the murine Hand1 transcriptional start site (Figure 1G).
We then used the inducible XMLC2-rtta::tet-Hand1 system (XMLC2-Hand1) [9] to determine the effect of prolonging Hand1 expression in the neonatal heart. Inducing persistent Hand1 transgene expression from the day before birth results in approximately 2.5-fold increase in Hand1 mRNA, corresponding to 60% of fetal levels at p0.5 (Figure 2A). In subsequent studies, we examined the phenotype of male pups at p0.5, comparing induced XMLC2-Hand1 pups with littermate controls expressing solely cardiac-specific rTTA transcription factor (XMLC2-rTTA).
Prior to birth, XMLC2-Hand1 mice are present in expected Mendelian ratios (14 XMLC2-Hand1 out of a total of 30 pups recovered at e18 after 48 h of maternal doxycycline induction, expected n = 15). The 3D reconstruction of high-resolution episcopic datasets [19],[20] revealed no significant structural difference between XMLC2-Hand1 and control littermates (Figure S1A). XMLC2-Hand1 pups exhibit respiratory distress shortly after birth, whereas control littermates appear normal (Figure 2B and Movie S1). XMLC2-Hand1 hearts were significantly lighter than controls (10.1±0.62 mg versus 7.08±0.55 mg; two-tailed t test p = 0.007, n = 6 each group). Genotype analysis indicated that Hand1 up-regulation also led to high rates of neonatal death in the immediate neonatal period (six Hand1 up-regulating pups out of 56 expected from 12 litters, collected at p0.5) with left ventricular cardiac rupture detected in two out of the three dead neonates collected (Figure 2C,D). As neonatal death rates in XMLC2-Hand1 are so high, we performed caesarian section on gravid XMLC2-rTTA female mice pregnant with XMLC2-Hand1 and control pups. Neonates were fostered to wild-type females, and at 4 h postpartum, all pups were sacrificed and genotyped. Sectioning/reconstruction revealed that XMLC2-Hand1 hearts were smaller than controls, but with no gross anatomical derangement (Figure 2E and F and Figure S1B).
Although histological analysis revealed no obvious tissue derangements or increase in apoptosis in XMLC2-Hand1 mice compared with controls (Figure S1), PAS staining revealed a marked depletion of glycogen in the myocardium of Hand1-upregulating neonates, consistent with exhaustion of glycolytic substrate (Figure 2G,H). We went on to analyse glycogen levels in the hearts of XMLC2-Hand1 mice compared with controls. We found that glycogen levels in XMLC2-Hand1 mice 2 h after caesarian section were reduced compared with XMLC2-rTTA controls (0.17±0.056 versus 0.98±0.153 mmol glygogen-derived glucose per mg heart tissue, n = 6, p = 0.0003, two-tailed t test) (Figure 2I).
Glycogen depletion suggested significant changes in neonatal myocardial energy metabolism as a result of Hand1 up-regulation, and we therefore performed Affymetrix microarray analysis to monitor changes in the myocardial transcriptome (Table S1). Gene ontology analysis revealed overrepresentation of genes involved in fatty acid metabolism amongst those showing significant changes as a result of elevated Hand1 expression (Table S2).
Expression of several genes encoding enzymes involved in cardiac fatty acid metabolism is up-regulated at birth. Using RT-PCR, we found that XMLC2-Hand1 hearts failed to up-regulate a subset of these genes. We found that at birth, expression of several genes encoding enzymes involved in lipid and acylcarnitine metabolism are up-regulated in the wild-type heart at p0.5, including ACC (acetyl coA carboxylase), MCD (malonyl coA decarboxylase), and CPT isoforms (carnitine palmitoyl transferase) as previously reported [21]–[24], along with FABP (fatty acid binding protein), FATP (fatty acid transport protein), ACSL (acyl coA synthase long chain 1), HSL (hormone sensitive lipase), and ATGL (adipose triglyceride lipase) (Figure 3A,B). Prolongation of Hand1 expression prevented the postnatal increase in expression of MCD, ACC, FABP, HSL, and CPT1a (but not CPT1b or CPT2). We found no significant differences in expression of mRNA encoding PPAR isoforms (Figure S2), glycolytic genes, or mitochondrial electron transport complexes between Hand1 up-regulating and control neonatal hearts (Figure S2). We detected broadly similar changes in gene expression following up-regulation of Hand1 in adult transgenic XMLC2-Hand1 mice inducibly up-regulating Hand1 (Figure 3C) and in stably transfected HL1 lines (Figure S2). Reduction of Hand1 expression levels in siRNA stably transfected HL1 cells, and e14.5 αMHC-Cre::Hand1(fl/fl) embryos led to a general increase in expression of lipid metabolising genes (i.e., opposite to changes in Hand1 up-regulating cells) (Figure 3D and Figure S2).
Previously, it has been found that an elevation in PGC1-α expression drives cardiac mitochondrial biogenesis, as part of postnatal cardiac energetic remodeling [25]. We confirmed that levels of PGC1-α mRNA rise in the postnatal mouse heart, but there was no significant change in PGC1-α mRNA levels in Hand1 up-regulating hearts (Figure 3E). We found no elevation of Hand1 protein expression in HL1 cells transfected with PGC1-α (Figure 3F). Furthermore, we detected no difference in mitochondrial/nuclear DNA ratio in XMLC2-Hand1 hearts compared with controls, or in Hand1 transfected HL1 cell lines compared with lines stably expressing shRNA against Hand1 (Figure 3G and Figure S2), nor of genes expressing mitochondrial respiratory complex isoforms (Figure S2). We detected no significant change in expression of mRNA encoding either PGC1-α or its downstream target gene ERR-α in hearts from αMHC-cre::VHL(−/−) neonates (Figure S2). These data imply that Hand1 in the neonatal heart acts independently of PGC1-α.
We next carried out ChIP, assaying canonical E-boxes (CAnnTG) in the 5′ promoters of the ACC, MCD, ACBP, FABP4, FATP, HSL, and ATGL genes, using a Hand1 antibody (Figure 3H and Figure S2). This revealed that Hand1 binds to several (but not all) E-boxes in the 5′ promoters of ACAC, MCD, FABP4, ACBP, and HSL promoters in vivo (albeit with varying degrees of avidity). No binding of Hand1 to these sites was detectable in chromatin isolated from e14.5 Hand1 null hearts (αMHC-Hand1(fl/fl)) (Figure S2) [26]. This suggests that repression by Hand1, at least in part, reflects direct control of these genes.
We then went on to confirm that Hand1 directly regulates one of these putative target genes. We cloned a 1 kb fragment of the mouse HSL promoter and ligated it into the pGL3 luciferase vector (Promega), and performed site-directed mutagenesis to abolish the middle e-box site, which we found bound Hand1 protein in the chIP assay (Figure 3I). Mutation of this e-box site led to a 7-fold increase in luciferase expression over the e-box containing promoter in untransfected HL1 cells, and a 13-fold increase in HL1 cells stably transfected with Hand1. Lower basal levels of promoter activation were found in Hand1 expressing cells (Figure 3I).
We examined whether Hand1 has an effect on lipid metabolism, as suggested by our transcriptomic analysis. We found lower overall levels of triacylglycerides and reduced levels of malonyl CoA in neonatal Hand1-persisting hearts compared with controls (Figure 4A–C). Multivariate analysis confirmed an overall decrease in lipid incorporation into acylcarnitine metabolites, with significantly decreased levels detected of C6-, C14-, and C18-containing acylcarnitine species (Figure 4D,E and Table S3). We also found that uptake of the fluorescently labeled lipid substrate BODIPY-500/510C1, C12 is reduced by 37.1% in Hand1 transfected cells compared with controls (p = 0.034 two-tailed t test) (Figure 4F). Therefore, elevated Hand1 expression levels not only decreases incorporation of lipid into cellular metabolic processes via targeting expression of several enzymes of lipid/acylcarnitine metabolism, but also by cardiomyocyte lipid uptake.
To test the hypothesis that the reduction in cardiomyocyte lipid uptake and synthesis mediated by Hand1 will be reflected in decreased cellular oxygen consumption, we tested the effect of Hand1 expression on oxygen consumption of stably transfected HL1 cardiomyocyte lines. We found basal oxygen consumption reduced by 53.7% (p<0.0001 two-tailed t test), maximal respiratory capacity by 64.9% (p<0.0001), ATP production by 54.7% (p<0.0001), and spare respiratory capacity (a measure of unused respiratory capacity) by 82.1% (p = 0.0004) compared with controls, transfected with an empty vector (empty vector control used throughout) (Figure 5A). Incubation of HL1 cells with excess of palmitate or glucose has the effect of driving the cell towards either β-oxidation of lipid or glycolysis, respectively. Lipid oxidation is reflected by a significant increase in oxygen consumption (Figure 5B). A total of 1 µM of the CPT1 inhibitor etomoxir blocks the import of fatty acids to the mitochondrion [27] and abolishes this increase in oxygen consumption in nontransfected palmitate-treated control cells (Figure 5C). However, oxygen consumption in Hand1 transfected cells incubated with palmitate does not change in response to etomoxir, implying that no lipid oxidation is occurring in Hand1 transfected cells (Figure 5C). We went on to assay the rate of Palmitate oxidation directly in Hand1 transfected and control cells transfected with an shRNA construct directed against Hand1, by incubating cells in 3H-labelled palmitic acid and measuring generated 3H2O [28]. We found a reduction of 51% in the rate of generation of 3H2O in cells stably transfected with Hand1 compared with Hand1 shRNA transfected cells (Figure 5D). We performed 3H-labelled palmitic acid oxidation studies in primary cardiomyocyte cultures from XMLC2-Hand1 neonates and adults, and found that elevation of Hand1 expression also leads to a significant reduction in palmitate oxidation in these cell types. Furthermore, cardiomyocytes from e15 Hand1null αMHC-Cre::Hand1(fl/fl) exhibit significantly increased levels of lipid uptake compared with αMHC-Cre::Hand1(+/+) littermates (Figure 5D). Taken together, these results show that Hand1 reduces oxygen consumption in HL1 cardiomyocytes, by inhibiting mitochondrial β-oxidation of fatty acids.
In order to investigate the effect of Hand1 levels on myocardial energy generation, we went on to examine mitochondrial function. We examined p.05 neonatal hearts from XMLC2-Hand1 and control pups.
Mitochondrial membrane potential (Δψm) is an indicator of mitochondrial energetic state. HL1 cells stably transfected with Hand1 demonstrate significant reduction in Δψm, assayed by TMRM fluorescence analysis, to 81.4%±5.2% of controls (n = 26 cells; p<0.001; Figure 6A). However, Hl1 cells stably transfected with an shRNA construct down-regulating Hand1 showed a significantly increased Δψm (to 116.8%±8% of control; n = 23; p<0.05; Figure 6A). This implies lower mitochondrial ATP generation in Hand1 up-regulating HL1 cells.
The redox state of mitochondrial NADH is a function of respiratory chain activity and substrate turnover. We measured the resting level of NADH autofluorescence in HL1 cells, which was then expressed as the “redox index,” a ratio of the maximally oxidized and maximally reduced signals [29]. The dynamic range of the signals was defined by obtaining the maximally oxidized signal following the response to 1 µM FCCP (which stimulates maximal respiration and fully oxidises the mitochondrial NADH pool). In these conditions, mitochondrial NADH is taken as fully oxidised and defined as 0%. The maximally reduced signal was then defined as the response to 1 mM NaCN (which fully inhibits respiration), preventing NADH oxidation, and so promoting maximal mitochondrial NADH reduction. In these conditions, NADH is taken as 100% reduced. HL1 cells down-regulating Hand1 did not significantly change the NADH redox state (35.7%±2.89%; n = 23 compare to 32.6%±2.4% in control shRNA transfected cells). In contrast, Hand1 up-regulation significantly increased NADH redox state (64.9%±5.9%, n = 25; p<0.001; Figure 6B), suggesting inhibition of mitochondrial respiration. It should be noted that Hand1 down-regulation and Hand1 up-regulation had an opposite effect on the total mitochondrial pool of NADH (Figure 6C). We found that Hand1 down-regulation significantly increases the NADH substrate pool (to 156%±6.5% of control; p<0.001) and Hand1 up-regulation significantly decreases the NADH pool in mitochondria (87.5%±4.6% of control; p<0.05).
We then carried out glucose flux analysis on stably transfected HL1 cell lines, using uniformly labeled 13C-Glucose, incubating for 4 h and analysing 13C lactate levels with 1H NMR. We found significant elevation of lactate production in Hand1 overexpressing cells compared with control Hand1 shRNA down-regulating cells (Figure 6F,H), and significantly elevated 13C labeling of lactate in adult XMLC-Hand1 hearts compared with controls (Figure 6G). We also found that primary cultured cardiomyocytes from Hand1 up-regulating neonatal hearts acidified the extracellular medium in a seahorse XF assay faster than control cardiomyocytes (Figure S3), implying that these cells are more glycolytic.
Taken together, these results imply that the effect of Hand1 is to reduce mitochondrial energy generation, and to switch cellular metabolism from aerobic glycolysis and mitochondrial energy generation to anaerobic glycolysis.
Since up-regulation of Hand1 reduces cardiomyocyte oxygen consumption, we hypothesised that an increase in cardiac Hand1 levels may enhance tolerance to ischaemia. We therefore tested the effect of Hand1 up-regulation in an animal model of myocardial ischaemia. Hearts were removed from adult (2 mo old) up-regulating (XMLC2-Hand1) and control (XMLC2-rTTA) mice after 1 mo of doxycycline administration, and subjected to 35 min of global ischaemia during Langendorff perfusion, followed by 30 min of reperfusion and infusion with prewarmed triphenyltetrazolium chloride as previously described [30]. Overexpression of the Hand1 transgene resulted in a 47% reduction in tissue death compared with control mice (p = 0.004) (Figure 6I) consistent with elevated Hand1 levels reducing cardiomyocyte oxygen consumption. Following 30 min of global ischaemia of Langendorff-perfused adult wild-type hearts, we were able to detect increased levels of HIF1α protein, but not Hand1 protein, using Western blotting (Figure 6J). This implies that Hand1 is not involved in the acute response to hypoxia/ischaemia.
We have previously found that adult Hand1 up-regulating hearts display a heart-failure-like phenotype of low threshold for ventricular arrhythmia and a diastolic defect [9]. As energy generation is also remodeled in failing hearts [31],[32], we measured PCr/Cr in Hand1 up-regulating adult mouse hearts. We found significant reductions in PCr/CR ratio (1.25±0.25 for controls versus 0.88±0.23 in Hand1 overexpression; p = 0.037) and PCr/ATP (1.70±0.06 versus 1.37±0.126; p = 0.0425) in Hand1 up-regulating hearts (Table S4). Therefore, ischaemia protection in Hand1 up-regulating hearts occurs at the expense of ATP production—that is, the same strategy employed in the fetal heart.
The mammalian cardiomyocyte is exposed to a large range of oxygen concentrations during development and terrestrial life. The ability of the cardiomyocyte to function in extremely low levels of oxygen is lost after birth, with serious medical consequences in the ageing human. Here we propose that the transcription factor Hand1 is part of a novel metabolic pathway adapting the embryonic heart to varying levels of hypoxia during development, birth, and adulthood. This pathway may be of significance in the adult during heart failure, as Hand1 is one of the “fetal” genes up-regulated in the failing cardiomyocyte. The links between heart development and heart failure are becoming more apparent, and may provide clues to future therapies.
There is circumstantial evidence connecting Hand1 expression with hypoxia during development. Hypoxia inducible factor (HIF) signaling is essential for formation of the placental trophoblast [14],[15], embryonic heart [16],[17], and developing nervous system [33]. Strikingly, these areas overlap Hand1 function in the developing embryo [10],[11]. Interestingly, HIF signaling is thought to be activated in the failing heart [34], where Hand1 is up-regulated [9]. Therefore, our data fit with a model whereby Hand1 expression is under control of hypoxia signaling in both the fetal and adult heart. Cardiac overexpression studies of the type described in this report are unable at present to differentiate between effects on the left and right ventricle. Development of reliable chamber-specific transgene expression is awaited for these studies. We found that cardiac-specific Hand1 null hearts exhibit up-regulation of the genes encoding proteins involved in lipid metabolism that are down-regulated in Hand1 overexpressing neonates. We also found that lipid oxidation in these hearts is increased relative to controls. These mice die in utero around e16–17 (our unpublished data and Mcfadyen et al. [26]). It is possible that the cause of death is an increase in oxygen demand due to up-regulated lipid metabolism. More broadly, the contribution of metabolic regulation to control of normal embryonic development is not yet clear.
The key adaptation of the fetal heart to hypoxia is the generation of ATP from oxygen-sparing glycolysis rather than oxygen-expensive lipid oxidation [35], although the absolute rates of glucose oxidation versus glycolysis are as yet unclear. There are several lines of evidence suggesting that selection of energetic substrate is coupled to ambient oxygen levels, and regulation of cellular lipid oxidation is a key mechanism to determine oxygen consumption. Experimental inhibition of cellular lipid metabolism by etomoxir, a CPT1 antagonist that prevents mitochondrial long-chain fatty acid import, shifts energy metabolism to glycolysis, leading to lower myocardial oxygen consumption, and protects against myocardial ischaemia [36],[37]. Exposure of neonatal rats to hypoxia results in a decrease in overall lipid content and remodeled acyl-carnitine metabolism, resembling the effect of persistent Hand1 expression on lipid metabolism [38].
Our understanding of the molecular mechanisms linking lipid oxidation rates with ambient oxygen in the heart, specifically around birth, is incomplete. It is known that PGC1-α has a role in postnatal maturation of cardiac metabolism via regulation of mitochondrial number [25]. The effect of Hand1 on the postnatal heart seems to be independent of PGC1-α. This is supported by the finding that overexpression of PGC1-α has a limited effect in the adult mouse heart [39] in contrast to the effects of Hand1 overexpression [9]. The transcriptional control of PGC1-α remains mysterious. We found that PGC1-α mRNA levels were not altered in neonatal VHL null hearts (Figure S2), implying that expression of this gene is not affected by changes in cardiac HIF signaling levels at birth.
Our data show that the effect of Hand1 in the heart is to down-regulate mitochondrial metabolism as well as lipid metabolism, reflected by changes in mitochondrial morphology, membrane potential, and glucose flux. This mechanism leads to an additional layer of regulatory complexity, as several glycolytic enzymes are known to be directly regulated by HIF signaling [40]. HIF signaling is thus likely to regulate several metabolic pathways in the neonatal heart in parallel. This may lead to an increased degree of metabolic flexibility, as evidenced by the fact that the phenotype of cardiac Hand1 up-regulating neonates is more severe than that of αMHC-Cre::VHL(fl/fl) neonates, which are born with relatively normal cardiac morphology and die with cardiac arrhythmia by the second postnatal week [41].
The idea that the same pathways are active in the fetus and adult heart is attractive, as it goes some way towards explaining the basis of the re-expression of the “fetal gene expression programme” seen in heart failure [8]. Perhaps a more accurate term is “hypoxia adaptive gene expression programme.” This is adaptation of the cardiomyocyte to low oxygen via metabolic and contractile gene isoform expression switching, and occurs to preserve oxygen at the expense of ATP production and lower cardiac output, as occurs in persistently Hand1 expressing neonatal hearts. There is evidence that this occurs in healthy humans. Healthy volunteers have been shown to up-regulate cardiac glucose oxidation at altitude [42], and it was recently found that healthy, young volunteers suffered what is essentially a reversible cardiomyopathy involving decreased ATP production and diastolic dysfunction on ascent of Everest [43]. Indeed, protection of the mouse heart against ischaemia by etomoxir occurs at the expense of ATP production and decreased lipid oxidation [37]. Remodeling of energy metabolism may be adaptive in the short term to protect against hypoxia, but is associated with a poor long-term clinical outcome in human heart failure [31],[32]. Our data showing that Hand1 is not significantly induced by ischaemia in the Langendorff perfused heart suggest that a hypoxia–Hand1 pathway is not involved in the response to acute ischaemia. This pathway seems more likely to be important in the response of the myocardium to chronic or repeated hypoxia/ischaemia. Interestingly, “hibernating” myocardium, whose function is temporarily decreased by repeated hypoxia, has been shown to revert to glycolysis [44],[45].
Our Langendorff perfusion data suggest that the HIF1α/Hand1 pathway may be active in the adult heart. However, this must be regarded as preliminary evidence at the moment. While this assay has proved to be a robust, reproducible assay for ischaemia/reperfusion studies, there are some important caveats when extrapolating Langendorff data to whole-animal physiology. The fact that the heart is removed from the mouse and perfused is clearly a major factor in this. Furthermore, the perfusate substrates contained in the Krebs-Henseleit buffer used in this assay are predominantly crystalloid, and are designed to optimize performance of the isolated heart, rather than to recapitulate physiological conditions [46]. The supraphysiological glucose concentrations in the Langendorff perfusate would, in theory, push the hearts towards a more glycolytic metabolism, potentially leading to an underestimate of the effects of Hand1 with respect to ischaemia protection. While it is theoretically possible to gain some measure of cardiac contractile function with a ventricular balloon in the Langendorff assay, we did not measure “cardiac function” in our Langendorff assay. We have argued in the past that such measurements in the Langendorff system are prone to artifact [46]. In vivo models of myocardial infarction will be necessary to investigate formally a role for modulating Hand1 levels in myocardial ischaemia protection. However, we have previously shown that adult Hand1 up-regulating mouse hearts display a diastolic defect without significant systolic dysfunction at steady state [9]. Studies are now ongoing to investigate formally a potential role for Hand1 in myocardial infarction.
Finally, our finding that Hand1 activity forms part of the regulatory mechanism adapting the fetal heart to intrauterine hypoxia may have clinical relevance. There is a growing body of evidence suggesting that cardiomyocyte lipid metabolism is of importance in determining oxygen consumption and therefore susceptibility to ischaemia [37],[47],[48]. It has also been shown that tight control of lipid metabolism is important in modulating oxygen consumption; overexpressing VLDL receptors in transgenic mouse hearts increases mortality following experimental myocardial infarction, presumably by increasing oxygen consumption via increasing lipid substrate presentation to the cardiac mitochondria [49]. Our data on the cardioprotective effects of Hand1 expression in a model of ischaemia support the idea that therapeutic manipulation of lipid metabolism in ischaemic cardiomyocytes may be beneficial. We hypothesise that therapeutic strategies in cardiac ischaemia and heart failure could be based on the fetal model of hypoxia protection, whereby modulation of metabolic substrate preserves oxygen at the expense of ATP production.
All mouse experiments were carried out in compliance with institutional ethical and welfare standards and under Home Office regulation.
Doxycycline 3 mg/kg was administered as described [9]. RNA was extracted from hearts using Trizol reagent (Invitrogen) according to the manufacturer's instructions. Complementary DNA was made using Superscripts 3 kits (Invitrogen). RTPCR was carried out on an Applied Biosystems 7000 analyser with SYBRGreen (Thermo Scientific), using 18s RNA as a control. All PCR primers were purchased from Qiagen, or are listed in [9].
High-resolution episcopic microscopy was carried out as published [20].
A 0.9 kb fragment of the mouse HSL 5′ promoter was isolated by PCR using primers listed in Text S1. This was cloned into the pGL4 plasmid (Promega). Mutation of the e-box site was performed using a Quikchange SDM kit (Stratagene, Santa Clara). Luciferase activity was estimated using the Dual-Luciferase assay kit (Promega) and an Anthos Lucy spectrophotometer (Biochrom, Cambridge).
Chromatin was prepared from neonatal mouse hearts by previously published methods [50]. See Text S1 for details.
Lipids were extracted using the methanol/chloroform/water method as described [51]. See Text S1 for details.
Frozen hearts were homogenized in 6% perchloric acid to extract CoA esters, and homogenates were spun at 12,000×g for 5 min, 4°C. Malonyl-CoA concentration in the supernatant was measured using HPLC as described previously [47],[52].
Hand1 transfected or control HL1 cells were incubated in serum HEPES buffered saline at 37°C for 10 min (two washes), then incubated in 5 µg/ml BODIPY-palmitate (Invitrogen) for 2 min at 37°C, washed with cold HBS, then imaged on a Zeiss confocal microscope. Fluorescence was quantified by ImageJ and normalized to cell number (10 high power fields per well, three wells per genotype).
Hearts were snap frozen immediately after sacrificing neonatal mice 2 h following caesarian section, before access to milk. Frozen hearts were ground in liquid nitrogen, and glycogen extracted and quantified enzymatically using the Abcam Glycogen Assay Kit, according to the manufacturer's instructions (Abcam, Cambridge, UK). Glucose was measured in a nonhydrolysed aliquot of each sample, and subtracted from the hydrolysed value, to give glycogen-derived glucose values. Samples were analysed on an Anthos Lucy Spectrophotometer (Biochrom, Cambridge).
HL-1 cells were transfected with full-length Hand1pcDNA construct with FugeneHD (Invitrogen) accordingly to the manufacturer's instruction. At 48 h after transfection, growth medium was supplemented with 0.4 mg/ml G148 in order to select clones overexpressing Hand1. Individual clones were picked and expression levels of Hand1 were verified by qRT-PCR and Western blot. For oxygen consumption rate and mitochondrial function test, neomycin was removed several passages before.
A Seahorse Bioscience Instrument was used to measure oxygen consumption rate as per manufacturer's instructions [53]. See Text S1 for details.
3H Labeled Palmitate (Sigma) oxidation was measured as previously described, incubating cells for 10 min [28].
Hearts from adult HAND1 transgenic and control at 2 mo of age, following 1 mo of doxycycline dosage were collected under isoflurane anaesthesia and ventilation. Chest was opened and hearts were freeze-clamped in situ with aluminum clamps precooled in liquid nitrogen. Freeze dried hearts were extracted with 0.4 M perchloric acid, and extracts were neutralized with 2 M KOH. Metabolite levels were measured by HPLC using the procedure described by us previously. Malonyl-CoA concentration was measured in the same extracts using a previously published LC/MS procedure [54].
Hearts were removed from adult mice and perfused on a Langendorff apparatus as previously described [30]. The ischaemia-reperfusion protocol consisted of 30 min stabilisation followed by 35 min global normothermic ischaemia and 30 min reperfusion. Global ischaemia was achieved by switching off the perfusion and immersing the heart in nonoxygenated buffer at 37°C.
At the end of the reperfusion period, hearts were perfused through the aortic cannula, with 1% prewarmed triphenyltetrazolium chloride (TTC) and then immersed in TTC at 37°C for 10 min. Then they were weighed and frozen at −20°C for 24 h. While still frozen, hearts were sliced from base to apex at a thickness of ∼≈1 mm. The slices were fixed in 10% formalin for 12 h to better define the boundaries between alive and dead tissue.
Heart slices were then photographed on a Perspex mounting block using a digital EsKape (Eskape, NY, USA) fixed camera. NIH Image 1.63 software was used to calculate the volumes of the whole heart and infarcted zones. The results were expressed as a I/R% of the dead tissue (I, infarct) developed in the whole heart (R, myocardium at risk) and presented as means ± standard error of the mean (SEM). The differences between groups were considered significant when p≤0.05.
See Text S1 for details.
|
10.1371/journal.pbio.1001685 | Coevolution and the Effects of Climate Change on Interacting Species | Recent studies suggest that environmental changes may tip the balance between interacting species, leading to the extinction of one or more species. While it is recognized that evolution will play a role in determining how environmental changes directly affect species, the interactions among species force us to consider the coevolutionary responses of species to environmental changes.
We use simple models of competition, predation, and mutualism to organize and synthesize the ways coevolution modifies species interactions when climatic changes favor one species over another. In cases where species have conflicting interests (i.e., selection for increased interspecific interaction strength on one species is detrimental to the other), we show that coevolution reduces the effects of climate change, leading to smaller changes in abundances and reduced chances of extinction. Conversely, when species have nonconflicting interests (i.e., selection for increased interspecific interaction strength on one species benefits the other), coevolution increases the effects of climate change.
Coevolution sets up feedback loops that either dampen or amplify the effect of environmental change on species abundances depending on whether coevolution has conflicting or nonconflicting effects on species interactions. Thus, gaining a better understanding of the coevolutionary processes between interacting species is critical for understanding how communities respond to a changing climate. We suggest experimental methods to determine which types of coevolution (conflicting or nonconflicting) drive species interactions, which should lead to better understanding of the effects of coevolution on species adaptation. Conducting these experiments across environmental gradients will test our predictions of the effects of environmental change and coevolution on ecological communities.
| Recent studies suggest that environmental changes may tip the balance between species that interact with each other, leading to the extinction of one or more species. While it is recognized that evolution will alter the way environmental changes directly affect individual species, the interactions between species force us to also consider the evolution of species interactions themselves. We use simple models of competition, predation, and mutualism to evaluate the effect of coevolution on the abundance of interacting species when climatic changes favor one species over another. In cases where the species have conflicting interests (i.e., where selection on one species for increased strength of the interaction is detrimental to the other, such as an organism becoming more aggressive towards competitors), we show that coevolution reduces the effects of climate change, leading to smaller changes in abundances and reduced chances of extinction. Conversely, when the species have nonconflicting interests (i.e., where selection for increased interaction strength on one species benefits the other, such as an organism avoiding competition with other species), coevolution increases the effects of climate change. Thus, gaining a better understanding of the nature of the coevolution between interacting species is critical for understanding how communities respond to a changing climate.
| Climatic changes, or indeed any change in the environment, have the potential to cause the local extinction of species, and to alter community composition and ecosystem functioning [1]. Numerous models have been used to predict how the density and geographical range of species will be affected by climate change, with mixed success [2]. In part, success is limited by the need to understand how changes in the density of one species affect densities of other species through their interactions. For example, reduced pollinator densities resulting from global climate change have led to local extinction of several plant species [3]. Climate change may also have direct effects on the strength of species interactions, and these are sometimes difficult to predict [4]. For example, climatic change has been implicated in mediating extinctions of amphibian species by altering the epidemiology of their pathogens [5]. Thus, the effects of climate change on species densities and extinction risks depend both on the direct effects of climate change on focal species and on the indirect effects acting through interactions among species, making predictions of the net effect of climate change on extinction risk challenging [6]–[9].
Rapid evolution is also expected to influence species responses to climatic changes [10],[11]. In addition to evolution directly driven by the changing climate, coevolution between species may modify species interactions [12],[13]. For example, Zhang and Buckling [14] factorially manipulated the environment and a bacterium's ability to coevolve in a bacterium-phage virus system where the environmental changes reduced phage infection. The phage could persist in the presence of a changing environment or when the host was allowed to evolve, but not when both environmental change and host evolution occurred simultaneously. This phage extinction was likely caused by a combination of increased costs associated with the coevolutionary arms race and a reduced effective population size resulting from a deteriorating environment [14]. Thus, coevolution has the potential to alter species interactions to the point of reversing the fate of the interacting species and is therefore likely to be an important determinant of extinction risk and community composition [15]–[17].
The theory addressing the evolution of species in direct response to a changing climate is well known in the context of climate change (reviewed by [18]), and there is also a relevant theoretical literature addressing coevolution [19],[20]. Coevolutionary analyses of the effects of productivity on coevolving ecological communities give insights into expected community responses to climate-driven changes in densities of particular species. For example, Hochberg and van Baalen [21] used predator-prey coevolution models to show that increased prey productivity can lead to increased defense against predators and a stronger arms race. Similarly, Abrams and Vos [22] demonstrated that in some scenarios increased prey mortality can lead to increased predator density, as prey invest less in predator defense. Indeed, microcosm experiments have demonstrated that increased resource abundance for a prey species can lead to increased prey defense, resulting in lower predator-prey ratios [23],[24]. If the cost of predator defense is associated with reduced intraspecific competitive ability, selection against well-defended phenotypes is expected to be strongest when competition is strong and predation weak [25], and several empirical studies have demonstrated that this type of coevolution can drive population dynamics (e.g., [26]–[28]). Theoretical studies focused on nutrient availability and range expansion have suggested that coevolution of competitors may also alter the effects of climate change on communities. As resources decline, divergent coevolution has the potential to reduce the ratio of interspecific to intraspecific competition, leading to increased coexistence in the presence of low resource availability [29]. In cases where climate change leads to range expansion and sympatric competitor distributions, divergent coevolution can lead to increased coexistence [30].
A frequent conclusion in these and other studies is that coevolution should be stabilizing, reducing changes in population densities of interacting species (e.g., [31],[32]). Here, we examine this hypothesis in detail. Our goal is to develop a simple, general theoretical framework to organize and synthesize the ways coevolution could modify the outcome of changing environmental conditions that will likely be pervasive with climate change.
To evaluate the effects of climate change, we present three coevolutionary models describing competitive, mutualistic, and predator-prey relationships between two species. Spatial structure may influence the effects of climate change on coevolving species [19],[20]; intermediate dispersal levels may slow local adaptation by diluting locally adapted genotypes, while low dispersal levels may speed local adaptation by providing advantageous genotypes [33]. In addition, when the climate itself varies across space, intermediate dispersal levels could lead to a geographic mosaic of coevolution where selection pressures and species traits vary across space [34]. Nonetheless, to focus on local adaption, we assumed that each species is represented by a single, panmictic population.
We modeled species interactions in terms of population dynamics: how the density of one species affects the population growth rate of the other. For example, for predation a high density of the predator will lead to a decrease in the population growth rate of the prey, and a high density of the prey will lead to an increase in the population growth rate of the predator; note that this general definition of predator-prey interactions encompasses host-pathogen and plant-herbivore interactions. While there is only a single interaction between species in the model, it is modeled as two parameters, one for the effect of the interaction on each species. Thus, for competitive interactions, one interaction parameter measures the negative effect of the density of the first species on the population growth rate of the second, and another parameter measures the effect of the second species on the population growth rate of the first.
We further assumed that each species has a trait that affects the strength of these interaction parameters. For example, a prey has a defensive trait that simultaneously decreases the negative effect of predation it experiences and decreases the positive effect accrued by the predator; similarly, a predator has an offensive trait that increases the predation rate on prey and increases the benefits obtained by the predator. Note that, in contrast to many models of species coevolution [35]–[37], we did not assume that there is trait matching in which the strength of interaction depends on a match between the traits values of each species; in our model both species have traits that cause monotonic benefits to the species. These benefits, however, have a cost that is exacted by decreases in the intrinsic rate of increase of the species. For example, a prey might increase its defensive trait and as a consequence suffer a reduced reproduction rate. Finally, we modeled trait evolution using a quantitative genetics approach, so the rate of evolution depends on the strength of selection and the additive genetic variance of the trait, where the additive genetic variance is constant. While this assumption about evolution is unlikely to hold in the long term (when mutations will be needed to maintain genetic variation), under very strong selection (which will cause loss of genetic variation), and for small populations (that lack large initial genetic variation and experience genetic drift), it is a reasonable starting point to investigate the short-term (hundreds of generations) response of species to climate change [38].
We used the models to pose the question: If the environment changes in such a way that the intrinsic rate of increase of one species rises, how will coevolution affect the equilibrium densities of both species? By “equilibrium density” we mean the density that would be obtained if changes in population density occurred on a more rapid time scale than evolution, although as we describe, this assumption gives insight into the case of rapid evolution on the same time scale as changes in density. We made the simplifying assumption that only one of the species experiences a direct change in its intrinsic rate of increase caused by the environmental change; this just makes it easier to separate the evolutionary changes in one species that directly experiences environmentally driven demographic changes from the other species that only responds indirectly through its interactions with the first. There is no loss of generality with this assumption, however, since the net effect of environmental changes to both species would be, to a first approximation, the simple combination of environmental changes to each species separately (Box 1).
There is a rich history of studies that show the effects of environmental change on demographic factors that affect the intrinsic rates of increase of species. For example, higher temperatures often lead to increased development rates in ectotherms, a relationship that is easily quantified [39]. Similarly, increased environmental carbon dioxide generally leads to increased plant growth, although the strength of this effect varies from species to species [40]. In addition to broad-scale climatic changes such as these, our models have implications for environmental changes on a more local scale. For example, increased nitrogen and phosphorus runoff and land management regimes can each alter growth or mortality rates, and significantly degrade the structure of ecological communities [41]–[43]. We intentionally did not specify a particular type of environmental effect in order to retain the general applicability of the models, although we recognize that there are a myriad of different effects that environmental changes can bring, and climate change will likely affect multiple environmental factors that will directly impact species' population growth rates.
A key issue in our models is how changes in the trait value of one species affects the fitness of the other species. For example, suppose that selection on the trait of a competitor to decrease the strength of competition it experiences simultaneously decreased the strength of competition experienced by the second species. This could occur if the trait reduced competition by reducing the feeding niche overlap between competitors, so the second species would benefit from selection on the first. We refer to this case as nonconflicting coevolution. Conversely, if the trait were to make the first competitor more aggressive and hence better able to defend itself against the second competitor, then the second competitor would suffer from the evolution of the first. We refer to this as conflicting coevolution. As we will show, the consequences of coevolution for the abundance of species depend on whether changes in the trait of one species is beneficial or detrimental to its interacting partner—that is, whether coevolution is nonconflicting or conflicting.
Competitors and mutualistic partners could experience either conflicting or nonconflicting coevolution, and different types of models have been used to describe each coevolutionary pathway. For example, competition models where competitors can reduce competition by shifting traits away from competitors (e.g., [35]) assume nonconflicting coevolution. In contrast, models focused on competitive arms races (e.g., [44]) assume conflicting coevolution between competitors. Although coevolution of mutualists is traditionally modeled as nonconflicting (e.g., [45]), we might expect conflicting coevolution to be common in mutualists as well. For example, yucca moths pollinate yucca plants while ovipositing in yucca flowers, and evolution of increased egg production within each flower leads to greater benefit received by the yucca moth, while negatively impacting yucca plants [46]. Thus, conflicting coevolution will occur for mutualists whenever there is the possibility of one partner cheating and reducing the benefit it provides [7],[46],[47]. For predator-prey interactions, evolution of prey to decrease the predation rate will generally be detrimental to the predator, whereas evolution of the predator to increase the predation rate will likely be detrimental to the prey. Therefore, coevolution of predator-prey interactions will generally be comparable to conflicting types of competition and mutualism, although as we discuss later, this might not strictly be the case for host-pathogen interactions.
To illustrate our theoretical results that are shared by all interactions—competition, mutualism, and predation—we used simple simulation models that share the characteristics discussed above. To aid the illustration, we selected parameter values intentionally to give coexistence of species (at least under some environmental conditions) and simple dynamics with stable equilibrium points. A theoretically more general, yet conceptually more challenging, approach to the same type of model is presented in Box 1; this general approach confirms that the qualitative patterns illustrated by our simulations are in fact found much more broadly under the general assumptions we have described.
We modeled coevolution of two competitors using a discrete-time, modified Lotka-Volterra competition model. The density of species i at time t, Ni,t, is given by(1)in which Fi gives the per capita population growth rate or, equivalently, the fitness of species i. The trait values that govern the strength of competition experienced by each species at time t are denoted i,t and j,t. The parameter αi(i,t+dj,t) is the competition coefficient measuring the effect of species j on species i. We assumed αi(i,t+dj,t) = exp(−i,t−dj,t). The parameter d determines whether coevolution is conflicting or nonconflicting, and hence is key to the model. If d<0, then increases in j,t (which reduces competition experienced by species j) increases competition experienced by species i, thereby giving conflicting coevolution. Conversely, if d>0, then increases in j,t decrease competition experienced by species i, leading to nonconflicting coevolution. For simplicity, we assumed that the value of d is the same for both α1(1,t+d2,t) and α2(2,t+d1,t), so that evolution has symmetric effects on both species.
The trait value i,t affects not only competition experienced by species i but also its intrinsic rate of increase ri(E,i,t). Specifically, we assumed that ri(E,i,t) = Ri+biE−fi,t, where f describes the cost of increasing i,t; thus, there is a trade-off between reducing competition by increasing i,t in αi(i,t+dj,t) and reducing the intrinsic rate of increase, ri(E,i,t). Because we assumed that αi(i,t+dj,t) has an exponential form, there is the possibility for an optimal fitness to be achieved at intermediate values of i,t. Other forms for αi(i,t+dj,t) may lead to optimal fitness at either zero or infinite values of i,t; we did not consider this situation, however, because these traits experiencing disruptive selection will likely fix within a local population. Finally, we assumed that the unspecified environmental variable E enhances the intrinsic rate of increase of species 1 (b1>0), implying that E represents a more-favorable environment. For species 2, we assumed there is no effect of environmental change (b2 = 0).
In the model, i,t gives the mean value of a quantitative genetic trait whose distribution among individuals in the population is symmetric with additive genetic variance Vi. Provided the magnitude of the variance is not too large [38],[48],[49], selection for changes in the mean value i,t is equal to the derivative of fitness with respect to the trait divided by mean fitness [50]. For our model:(2)
The model for two mutualists has the same structure as the competition model (1). For mutualism, the coefficient αi(i,t, j,t) = −log(1+i,t+dj,t) is negative, and the logarithmic form allows the optimal fitness to be achieved at intermediate values of i,t. As in the competition model, d determines whether coevolution is conflicting or nonconflicting. The other components of the model are the same as described for the competition model, and evolutionary change is described by Equation 2.
For predator-prey interactions we used a discrete-time version of a model in which the predator attack rate is determined by traits of both prey, 1,t, and predator, 2,t [51]. Prey trait 1,t represents antipredator defense behavior, whereas predator trait 2,t represents the ability of the predator to overcome prey defenses. Changes in the densities of prey Nt and predator Pt are given by:(3)where r (E, 1,t) = R+bnE−f1,t is the intrinsic rate of increase of the prey and depends on the environmental variable E, as in the competition and mutualism models. The predation rate a(E, 2,t, 1,t) = q(E)exp(−2,t 1,t) depends on the environmental variable E and declines with increasing prey defense, 1,t, or the predator's susceptibility to the prey defense, 2,t. We considered two scenarios, one in which the predation rate q(E) = Q0+bpE increases linearly with E (bp>0) while prey growth rate is unaffected (bn = 0), and the other in which the predation rate remains constant (bp = 0) while the prey intrinsic rate of increase rises with E (bn>0). Although we assumed q(E) is independent of prey density for simplicity, preliminary analyses showed that incorporating a nonlinear type II functional response [52] does not qualitatively alter the results. The predator experiences a cost of trait 2,t in the form of increased mortality; specifically, m(2,t) = m0+g/2,t, where g governs the cost to the predator of being able to overcome the prey defense. Finally, if VN and VP are the additive genetic variances for prey and predators, respectively, evolution is given by:(4)
To illustrate the importance of coevolution—especially the contrast between conflicting and nonconflicting coevolution—for the response of populations to environmental changes, we conducted two types of simulations. For each type, we assumed that the populations begin at eco-evolutionary equilibrium (i.e., traits and densities are both at equilibrium), with identical genetic variances for the two species. For mutualism and competition models, the two species were initially identical in every way except in their response to environmental change. For the first type of simulation, we tracked the trajectories of population densities and traits through time as the intrinsic rate of increase of one of the species increases with the environment, E. We compared the trajectories for different levels of genetic variance, because the lower the genetic variance, the slower the rate of evolution. The second type of simulation involved evaluating how environmental changes alter the ecological and coevolutionary equilibriums. To find these equilibriums, after changing the environment we simulated the models for an additional 1,000 generations to allow population densities and trait values to stabilize. We did not find alternative stable states, and thus present the single equilibrium for each scenario. These two types of simulations proved to give the same conclusions, with the simulations of trajectories giving only one additional piece of information: that trait values and densities moved uniformly to the equilibriums given by the second type of simulations. The correspondence between the two types of simulations results from the fact that the level of genetic variance determines the rate of approach to equilibrium but does not alter the equilibrium itself, which is a joint optimization of fitness in each species. To avoid redundancy, we only present the trajectories for the conflicting competition case, and subsequently focus solely on the equilibrium simulations. We refer the reader to Box 1 for a full mathematical treatment that does not depend on the specific equations we used for the simulation models. Finally, although we only considered two interacting species here, we have found qualitatively similar results in simulations of larger communities (results not shown).
To illustrate the competition model, we began by simulating the consequences of raising the environmental quality for species 1 (increasing E) through time while varying the rate of coevolution. When the additive genetic variances for the traits expressed by both species, V1 and V2, are zero, evolution cannot occur, whereas increasing V1 and V2 increases the rate of evolution. For this illustration we assumed competition is conflicting. Increasing E increases the density of species 1 and decreases the density of species 2, yet allowing evolution moderates both effects (Figure 1A). As V1 and V2 increase, the rate of change of population densities and trait values more closely track their equilibrium values Ni* and i*, that is, the values at which, for fixed E, Ni,t+1 = Ni,t and i,t+1 = i,t in Equations 1 and 2 (Figure 1A,B). The effects of coevolution are largely driven by changes in trait values for species 2, with less change in species 1 (Figure 1B). This occurs because species 2 evolves to invest heavily in the competitive arms race, limiting the decline in investment by species 1.
For conflicting competition (d<0, Figure 2A,C), equilibrium species densities are less sensitive to environmental change when there is coevolution, whereas coevolution augments changes in species densities when there is nonconflicting competition (d>0, Figure 2B,D). This occurs because an increase in the density of species 1 with environmental change leads to a decrease in the density of species 2. Because selection pressure is positively correlated with the density of the other species, species 1 experiences relatively less selection pressure from competition with species 2 compared to the selection pressure on species 2 from species 1. When competition is conflicting (Figure 2A,C), the decreased selection on species 1 is beneficial to species 2, which acts to limit the decline of the population of species 2 and hence the decline of its effect on species 1. Also, the increased selection on species 2 increases its per capita competitive effect on species 1. These two sources of selective pressures combine to help species 2 and, in turn, are detrimental to species 1. When competition is nonconflicting (Figure 2B,D), the converse occurs; the decreased selection on species 1 caused by low densities of species 2 increases the effect of competition on species 2, and the increased selection on species 2 decreases its per capita competition effect on species 1. This selective pressure benefits species 1, further increasing its density.
In summary, conflicting competition sets up coevolution as a negative feedback, because selection on one species to reduce competition increases its competitive effect on the other species. In contrast, nonconflicting competition sets up coevolution as a positive feedback, because selection to reduce the impact of competition on one species also reduces the impact of competition on the other (Box 1).
As with competition, the effects of coevolution on mutualists depended on the type of coevolution. When there is conflicting mutualism (d<0, Figure 3A,C), coevolution diminishes the effects of environmental change on equilibrium densities, in contrast to the case of nonconflicting mutualism (d>0, Figure 3B,D). This effect occurs because the increase in the density of species 1 due to the environmental change increases selection pressure on species 2 for investment in the mutualism. When the mutualism is conflicting, this change is detrimental to species 1 and limits its increase, because the benefits of mutualism decrease with the investment of species 2 in the interaction. In contrast, in the case of nonconflicting mutualism, increased investment by species 2 is beneficial to species 1, further increasing the density of species 1. In summary, conflicting mutualism sets up coevolution as a negative feedback, whereas nonconflicting mutualism sets up a positive feedback (Box 1).
For competition and mutualism, interacting species might have either conflicting or nonconflicting coevolutionary feedbacks. In contrast, predator and prey interactions are generally expected to exhibit conflicting coevolution and hence generate negative coevolutionary feedbacks: prey coevolution of defenses that reduce predation will be detrimental to the predator, and predator coevolution to increase the predation rate will be detrimental to prey. To verify this expectation, we analyzed both the case in which climate change increases the prey intrinsic rate of increase and the case in which climate change increases the predation rate and hence the predator population growth rate.
When climate change enhances the prey intrinsic rate of increase, the resulting increase in prey density leads to increased predator density, and in the absence of coevolution the equilibrium predator density increases dramatically (Figure 4A). In contrast, the equilibrium predator density increases more slowly when predator and prey coevolve (Figure 4A). As with conflicting competition and mutualism, higher predator density strengthens selection pressure for prey investment in the coevolutionary arms race (Figure 4C). With increased prey investment, the predator density cannot increase as much due to heightened prey defense (Figure 4A,C).
When climate change increases the predation rate but has no effect on the prey intrinsic rate of increase, the predator density increases rapidly in the absence of coevolution, but this increase is slowed by coevolution (Figure 4B,D). Because prey selection pressure is positively correlated with predator density, prey evolve higher defensive trait values in the presence of higher predation rates, which in turn lowers the predation rate, increases prey density, and decreases predator density. Thus, coevolution sets up a negative feedback loop that reduces the decline in prey density and increase in predator density (Figure 4B). While these results pertain to specialist predators that have no other prey species, we found that coevolution also reduces the ecological effects of climate change in a model for generalist predators (Figure S1).
We have shown, using simple models, that coevolution may increase or decrease the effect of environmental change, depending on the form that coevolution takes between species. In cases where species have conflicting interests, coevolution reduces the effects of environmental change on densities, because coevolution acts as a negative feedback to the effects of environmental change. Conversely, when species have nonconflicting interests, coevolution sets up a positive feedback that increases the effects of environmental change on densities. Given these contrasts, is coevolution in nature likely to involve conflicting or nonconflicting interests of interacting species? Competitors and mutualists, in particular, have the potential to coevolve along either conflicting or nonconflicting pathways. Thus, determining the predominant type of coevolution will be critical to identifying the long-term effects of climate change on species.
Below, we first give brief discussions of classical studies and show that cases of both conflicting and nonconflicting coevolution are common. Therefore, no a priori prediction can be made for their relative importance when anticipating the effects of climate change. We then turn to coevolutionary studies that directly address climate change, using these to show how evidence can be obtained to make and test predictions about the coevolutionary effects on specific systems facing climate change.
It has long been recognized that coevolution can lead to increased asymmetries in competitive abilities [15], which is the hallmark of conflicting coevolution. But the idea that competition drives partitioning of food sources is even older [53]–[55], and this is the hallmark of nonconflicting coevolution. The effects of climate change for specific competitors hinge on which type of coevolution occurs. Evidence suggests that both are common.
Laboratory experiments that evaluate the effect of competitive interactions on trait evolution for each species have documented both conflicting coevolution in flies [15] and nonconflicting coevolution in E. coli strains [56]. Furthermore, conflicting and nonconflicting coevolution are not mutually exclusive; Colpoda protozoans with initially weak competitive abilities have been shown to evolve along both pathways [57]. While these types of experimental studies have the advantage of documenting coevolution as it happens, they are limited by the range of species and time scales that are amenable to experiments, and the magnitude of environmental heterogeneity that may affect coevolution [58],[59].
Alternatively, field studies can be used to infer the prevalence of conflicting versus nonconflicting coevolution. Research focusing on character displacement in natural populations attempts to identify the effects of coevolutionary processes based on species' phenotypes in solitary and sympatric populations [60]. This approach has documented both conflicting [61],[62] and nonconflicting coevolution [63]–[65].
There is a rich theory describing the evolution of mutualisms [66],[67]. Theoretical predictions often suggest that mutualistic interactions have the potential to break down into parasitic interactions [47],[68],[69]; this is an extreme form of conflicting interests between species. If mutualism breakdown into parasitism is common, then conflicting coevolution is likely, and this will likely diminish the effects of climate change.
Nonetheless, if mutualistic partners can enforce good behavior of their partners [68], then nonconflicting coevolution is expected. For example, the plant Medicago truncatula discriminately rewards the most beneficial mycorrhizal partners with more carbohydrates, and mycorrhizal partners form partnerships only with the roots that provide the most carbohydrates [7]. Thus, each partner constrains the selection pressure of the other to allow only nonconflicting coevolution. If nonconflicting coevolution is frequently imposed by mutualists, our results suggest that coevolution between mutualistic species will exaggerate, rather than diminish, the effects of climate change on species densities.
Conflicting coevolution is expected for most types of predator-prey or consumer-resource interactions, because increases in prey defenses will decrease benefits to predators, and increases in predator effectiveness will be detrimental to prey. Nonetheless, evolution of parasite virulence could be different [70],[71]. The conventional wisdom is that parasites should evolve to be less virulent, because this will increase their transmission among hosts; parasites are not transmitted by dead hosts, at least not for long [72]. Nonetheless, this ignores, among other things, the relationship between the production of large numbers of propagules (that generally harms the host) and transmission rates, and more-detailed analyses generally predict evolution of parasite virulence to represent a balance between higher virulence caused by selection for production of propagules and lower virulence caused by selection for lengthening the transmission period [73]. Therefore, evolution of the parasite may be nonconflicting with the host, even at the same time evolution of the host to limit infection is conflicting with the parasite. In models describing this interaction (results not shown), we found that when climatic changes directly affect the parasite, coevolution in the host fuels a negative feedback loop that mitigates the effects of climate change. In contrast, in some cases when climatic changes directly affect the host, coevolution can lead to a positive feedback loop that exaggerates the effects of climate change on the host density. Thus, when there are both conflicting and nonconflicting coevolution, the ultimate outcome will be determined by whether the host or parasite experiences greater evolutionary change.
Given the widespread occurrences of both conflicting and nonconflicting coevolution in competition and mutualism, and to a lesser extent in predator-prey interactions, systems will have to be studied on a case-by-case basis to predict and test the role of coevolution in modifying the effects of climate change. This could be done either using experimental studies or taking advantage of naturally occurring environmental gradients.
An example of an experimental study is given by Lopez-Pascua and Buckling [74], who performed an environmental manipulation of bacterial productivity by altering nutrient concentrations in the growth media. They showed that increasing bacterial productivity increases the rate of coevolution between bacteria and phages. They proposed that this is due, in part, to increased selection pressure on the bacterial population in environments with high productivity (high intrinsic rates of bacterial increase). This increased selection stems from increased encounters with phages, as phages numerically respond to increased bacterial density. The phages then evolve greater infectivity in response to bacterial evolution. This explanation is consistent with our theoretical expectations for conflicting evolution of prey and predators; increasing the prey intrinsic rate of increase leads to evolution of stronger prey defenses against the predator (Figure 4C).
In addition to experimental manipulations of environmental factors, it is possible to take advantage of natural environmental gradients similar to classical studies of character displacement. For example, in a field experiment, Toju et al. [75] documented a climatic gradient in a coevolutionary arms race between the camellia beetle (Curculio camelliae) and its host plant, Japanese camellia (Camellia japonica). Female beetles use their snout to pierce the camellia fruit pericarp and oviposit eggs into seeds, with oviposition success determined by the length of the beetle's snout and ovipositor relative to the pericarp thickness. Thus, plant defense is determined by pericarp thickness, and beetle snout and ovipositor lengths determine beetle ability to overcome this defense. The authors measured beetle and plant traits along a latitudinal gradient, and previous work had showed that plants exhibit faster potential for growth at lower latitudes [76]. Our analyses suggest that, because increases in prey growth should increase predator densities and, in turn, increase selection pressure on prey, the coevolutionary arms race should be “won” by prey under environmental conditions that favor prey population growth (Figure 4A,C). Thus, in the camellia-beetle arms race we expect that coevolution will favor plants more at lower latitudes. The authors indeed found this to be the case; plants in high latitude populations that experienced endemic predation by beetles had pericarp thicknesses similar to populations that did not experience beetles. In contrast, at lower latitudes plant populations that experienced beetle predation had thicker pericarps than populations that did not. There was thus an increase in plant defense along the environmental gradient. Furthermore, this plant defense increased with decreasing latitude at a greater rate than weevil ovipositor length, suggesting that plants exhibited a larger coevolutionary advantage in environmental conditions with increased prey growth [75]. These results support our theoretical predictions that higher prey intrinsic rates of increase should lead to a coevolutionary advantage to prey, thereby buffering the changes in predator densities driven by climate change.
The majority of coevolutionary studies involving environmental manipulations or environmental gradients have been conducted on predator-prey or herbivore-plant systems where conflicting coevolution is likely. Similar experiments that document changes in traits and density might help build a better understanding of coevolution in competitive and mutualistic relationships. Laboratory studies have suggested that coevolution can lead to a reversal of competitive hierarchy in just 24 generations [15], and can occur fast enough to drive population dynamics [16]. Therefore, experimental competition studies in which environmental factors are manipulated are possible for some types of organisms. Environmental gradient, rather than experimental, studies will be more practical for larger organisms with longer lifespans that operate at larger spatial scales. Using character displacement to infer conflicting versus nonconflicting coevolution is necessarily correlative, although it opens up the study of coevolution in the context of climate change to a much wider range of species under natural spatial and temporal scales.
Studies that evaluate coevolution over environmental gradients fit within the broader conceptual paradigm of geographic mosaic theory [77] in which differences in coevolutionary selection among spatially separated populations are analyzed as genotype by genotype by environment interactions. A key feature of geographic mosaic theory is that some local populations experience environmental conditions under which coevolutionary pressures are strong. These “coevolutionary hotspots” are characterized by fitness equations ([77], p. 100)(5)where the fitnesses W1,E and W2,E of species 1 and 2 depend on both phenotypes 1 and 2, and on the environment E. This pair of equations has the same general structure as that we have used for Equations 1–4. Thus, our results address the possible character of evolution within coevolutionary hotspots, and how coevolutionary outcomes might differ under different environmental regimes.
We have only considered local populations, explicitly ignoring gene flow among populations. Thus, we have ignored the large body of theoretical and empirical studies evaluating gene flow among populations under different selective forces [77]–[79]. For example, Nuismer et al. [80] used spatially explicit population genetics models to show that isolated populations of mutualistic species were likely to reach equilibrium quickly, while antagonistic populations were likely to oscillate in both density and phenotype. When interaction types vary spatially, however, both dynamic and equilibrium clines occur, and the presence of each depends on the levels of selection and gene flow across the landscape [80]. In an experimental bacteria-bacteriophage community, bacteriophages became locally maladapted in the absence of gene flow, but became locally adapted when gene flow occurred between bacteriophage populations [81]. The importance of gene flow in both theoretical and empirical studies gives a caution to our recommendation that natural environmental gradients be used to assess the character of coevolution—conflicting versus nonconflicting—and whether coevolution sets up positive or negative feedback loops to environmental changes. Gene flow and a geographic mosaic of selective pressure may dampen or otherwise modify the effects of local selection on coevolutionary traits.
While it is recognized that evolution will play a role in determining how climatic changes directly affect species [18], the interactions among species force us to also consider coevolution between species. Our models suggest that the effects of coevolution on population densities depend on the presence of conflicting versus nonconflicting coevolutionary interests. While we encourage future studies that experimentally manipulate both coevolution and environmental change, we acknowledge that experiments are likely to be difficult logistically for most study systems. It may be possible, however, to use character displacement across environmental gradients to distinguish whether conflicting versus nonconflicting coevolution is more likely, even when directly measuring coevolution is impossible.
Experimental [15] and environmental gradient [60] approaches to infer the nature of coevolution are both five decades old, and we hope that our theoretical results provide new impetus for these types of studies. They give needed information to anticipate whether coevolution will increase or decrease the effects of climate change on the densities of interacting species.
|
10.1371/journal.pcbi.1002365 | Novel Approach to Meta-Analysis of Microarray Datasets Reveals Muscle Remodeling-related Drug Targets and Biomarkers in Duchenne Muscular Dystrophy | Elucidation of new biomarkers and potential drug targets from high-throughput profiling data is a challenging task due to a limited number of available biological samples and questionable reproducibility of differential changes in cross-dataset comparisons. In this paper we propose a novel computational approach for drug and biomarkers discovery using comprehensive analysis of multiple expression profiling datasets.
The new method relies on aggregation of individual profiling experiments combined with leave-one-dataset-out validation approach. Aggregated datasets were studied using Sub-Network Enrichment Analysis algorithm (SNEA) to find consistent statistically significant key regulators within the global literature-extracted expression regulation network. These regulators were linked to the consistent differentially expressed genes.
We have applied our approach to several publicly available human muscle gene expression profiling datasets related to Duchenne muscular dystrophy (DMD). In order to detect both enhanced and repressed processes we considered up- and down-regulated genes separately. Applying the proposed approach to the regulators search we discovered the disturbance in the activity of several muscle-related transcription factors (e.g. MYOG and MYOD1), regulators of inflammation, regeneration, and fibrosis. Almost all SNEA-derived regulators of down-regulated genes (e.g. AMPK, TORC2, PPARGC1A) correspond to a single common pathway important for fast-to-slow twitch fiber type transition. We hypothesize that this process can affect the severity of DMD symptoms, making corresponding regulators and downstream genes valuable candidates for being potential drug targets and exploratory biomarkers.
| Comparison of gene expression in diseased and normal tissue is a powerful tool of studying processes involved in pathogenesis and searching for potential drug targets and biomarkers of the disease's progression and treatment outcome. We have developed a novel approach for systematic knowledge-driven analysis of gene expression profiling data, which can suggest the underlying cause of the observed differential expression by identifying which expression regulators might be involved. These regulators can not only be the promising subjects of further investigation, but also potential drug targets, as normalization of their activity might alleviate some of the disease's symptoms. The targets downstream of suggested regulators can be proposed as exploratory biomarkers in disease treatment and prognosis. We used our approach to analyze public gene expression datasets of Duchenne muscular dystrophy – a progressive inherited disease in males. Some of the regulators and biomarkers that we found were already investigated in the context of DMD, while some of them were not yet studied and may be of interest for biological and clinical studies.
| Microarray-based expression profiling is a widely used, quick and inexpensive method to obtain information about the specific diseases. A traditional approach when searching for drug targets or candidate biomarkers for a specific disease is to look for genes differentially expressed between the disease and appropriate “control” samples. Various techniques have been applied to find statistically significant differentially expressed genes, including classical statistical tests (e.g. t-test) and those specifically developed for microarray data analysis (Limma [1], SAM [2], shrinkage T-statistic [3] and other).
To get the deeper understanding of the disease mechanisms, the functional analysis of differential genes can be performed using a number of different methods [4]. Typically they rely on Gene Ontology (GO) – based annotation of genes. Common approach is to pre-select differentially expressed genes based on differential fold-change and/or p-value threshold, and find the statistically enriched GO groups using Fisher's exact test. More sensitive approaches are based on gene set enrichment analysis (GSEA [5], [6]) to avoid differential cut-off selection issue.
In addition to Gene Ontology, the protein-protein functional associations, regulatory or biochemical networks can also be used as a source of functional protein annotation in enrichment analysis [6], [7], [8]. More elaborated classification and functional annotation methods [9], [10] are usually applied to protein-protein networks only. The potential drawback of this kind of networks for the analysis of expression data is that they eventually skip the important transcriptional factors if they are not differentially expressed themselves. In this paper we used a proprietary literature-derived gene expression regulation network as a source of functional protein annotation. This global expression network consists of direct or indirect effects of a network node (protein) on expression of other genes [11]. Unlike conventional GSEA [5], [6], which uses predefined collection of gene sets, Sub-Network Enrichment Analysis (SNEA) algorithm, implemented in Pathway Studio® software [11], constructs comprehensive collection of gene sets from ResNet, a global literature-extracted protein-protein regulation network. The gene sets are constructed for each individual network node (“seed”) and consist of all its downstream expression targets only (star-like subnetworks).
The central idea of SNEA approach is that if the downstream expression targets of a “seed” are enriched with differentially expressed genes, then the “seed” is likely to be one of the key regulators of the differential expression changes, e.g. a transcription factor responsible for the observed changes in expression or an upstream member of signaling pathway [12]. This literature-driven approach connects differentially expressed genes to major implicated pathways and key expression regulators. In contrast to other methods that utilize the same idea of finding upstream network regulators using expression data [13], [14], SNEA allows identification of any potentially important protein (not obligatory a transcriptional factor) leading to the observed expression changes, even if its own expression doesn't change. It becomes possible because of the usage of ResNet database where all relations are taken from the literature only. Hence, there is no restriction on the protein type that can be considered as potential “seed”, provided that it is reported to influence each individual downstream gene expression.
We have applied this approach to study Duchenne muscular dystrophy (DMD) using publicly available gene expression profile datasets and identified a set of potential regulators and downstream biomarkers of DMD progression and severity.
Duchenne muscular dystrophy is an X-linked recessive muscular disorder, caused by mutations in the dystrophin gene (DMD) [15]–[17]. Affecting about 1∶3500 newborn males, it is the most common form of muscular dystrophies and the most common sex linked disease in males [18]. The underlying genetic cause of DMD is the presence of a variety of DMD gene mutations that result in dystrophin reduction/absence in skeletal muscle [17]. Lack of dystrophin has multiple unfavorable consequences to a muscle fiber (reviewed in [19]), leading to apoptosis or necrosis with subsequent inflammation and fibrosis at the site of damage. The process of muscle regeneration is also activated, but, in humans, with the course of the disease the repair capacity declines and becomes insufficient [20]. Muscle tissue is replaced with adipose and fibrous connective tissue [21].
The average life expectancy of DMD patients varies from late teens to early thirties, and can be improved by respiratory support [22], [23] and drug therapy [24]. Currently, there is no cure for DMD, but some treatments targeting the secondary consequences of dystrophin deficiency, such as muscle damage, necrosis, apoptosis and failure of regeneration, are already available for patients. Glucocorticoids, such as prednisone and deflazacort, are widely used to alleviate some of the disease's symptoms [25].
Several tests are used in diagnostics of DMD, including measurement of physical parameters, serum level of creatine kinase, genetic testing for DMD mutations and muscle biopsy to confirm the reduction in dystrophin content. More accurate, preferably non-invasive and biologically explainable markers are needed to predict prognosis, estimate disease's severity and progression. Also new biomarkers are required in treatment and clinical trials for DMD, where they can be used to monitor drug efficiency and choose optimal drug dose.
In order to identify potential drug targets along with corresponding biomarkers, we have searched for the consistent SNEA regulators and their downstream expression targets using publicly available differential gene expression profiles and literature-extracted expression regulation network from muscle biopsies of patients with DMD. Suggested workflow implies aggregation of the data from multiple datasets and elucidation of common mechanisms that underlie differential expression. Studying these mechanisms from the prospective of searching for new drug targets can provide valuable insights in both biological and medical research.
The overall analysis workflow is presented in Figure 1. Five NCBI GEO DMD-related microarray expression profiles from muscle biopsies were aggregated according to the procedure described in Methods. To ensure robustness of our analysis we constructed five leave-one-out datasets each time aggregating four distinct experiments and omitting one out of total five available experiments. We also constructed single large dataset (referred to as “aggregated dataset”), where all five available microarray experiments were aggregated. Additional dataset (referred to as “reference dataset”) was constructed on the base of published meta-analysis [26], see Methods.
We performed SNEA with default parameters for each of the six datasets (five leave-one-out datasets plus aggregated dataset) and obtained six lists of 100 significant regulators. Regulators common for all six datasets were combined with regulators obtained by SNEA of reference dataset. This resulted in the list of 76 unique regulators, which can be viewed as potential drug targets. We also performed permutation test to ensure that this overlap is significant.
Next, we turned to selection of differentially expressed genes. For each of the 6 datasets (five leave-one-out datasets plus aggregated dataset) we performed gene ranking using combination of different methods (see Methods section). Then we identified genes which were present in top-500 lists for all six datasets. Out of all these consistently differentially changed genes, we have selected only those which were expression targets of selected consistent significant regulators. This produced a list of 140 candidate genes (105 over- and 35 under-expressed). These genes (potential biomarkers) have been sorted using the combination of expression rank in the aggregated dataset and the number of significant regulators as a score (see Methods section). We also manually evaluated top-20 up-regulated genes and top-10 down-regulated genes in respect to the supporting evidences from the available literature.
All analytical procedures were applied separately to over-expressed genes and under-expressed genes to look individually at processes and pathways activated and repressed in DMD.
The significant regulators of up- and down- regulated differentially expressed genes from six datasets were cross-validated and only those identified in all datasets were selected for further analysis. They were combined with regulators obtained from the SNEA of the reference dataset to produce the final list of 76 unique significant regulators shown in Table 1 below. More information about these regulators can be found in Table S1.
We have selected genes, which were consistently differentially expressed in six datasets (one aggregated dataset and five leave-one-out datasets). The fold-change threshold was established by analyzing fraction of genes present in all six top-k rankings for varying k, Figure 3. As can be seen, fraction of common genes in top-k rankings for different types of gene expression reaches a plateau for k roughly equal to 500. This means, that adding more genes will not increase percentage of overlap between different gene rankings. Hence we limited our analysis to top-500 differentially expressed genes for different types of regulation. The percentage of consistent genes in top-k of all datasets is about 40% (Figure 3). It means that analysis of differentially expressed genes from a single dataset can potentially lead to 60% of false positives. To increase reproducibility of obtained results we focused on the genes, presented in all six top-500 rankings.
From the top 500 up-regulated genes in aggregated dataset we have selected 240 genes also present among top 500 up-regulated in all 5 leave-one-out datasets. Similarly, from the top 500 down-regulated genes in aggregated dataset we have selected 191 genes also present among top 500 down-regulated in all 5 leave-one-out datasets. These two lists were combined into a single list of 431 consistently up/down regulated differential genes. We performed Fisher exact test to find significantly enriched categories from Gene Ontology, corresponding to biological processes. Results, presented in Table 2, in general reflect known changes that take place in affected muscles: up-regulated genes are commonly associated with inflammation and immune response, apoptosis and wound healing; down-regulated genes – with metabolic processes and muscle contraction.
Genes were further analyzed in order to evaluate their quality as biomarkers. A promising biomarker should be easily detected and correspond to a DMD-related process (e.g. muscle biology, fibrosis, inflammation) or DMD-related condition (e.g. dilated cardiomyopathy). We used a proprietary Ariadne DiseasesFX Database, which contains literature-extracted information about various types of relations between genes and diseases as well as data on presence of gene products in biofluids and among secreted proteins. We also made use of Ariadne ResNet 7 and Muscle Biology Gene Ontology, see Methods. Associations between 431 consistently up/down regulated genes and DMD-related processes and conditions are depicted in Table S2.
Out of 431 consistently changed genes, we have selected only those which are expression targets of significant regulators, selected using the above procedure. This produced a list of 140 candidate genes (35 down-regulated, 105 up-regulated) that have been finally sorted using combination of rank in aggregated dataset and number of significant regulators (see Methods). Most of them correspond to the processes of development and regeneration, immune response, response to glucocorticoids, hypoxia and extracellular matrix organization.
Top-ranked 20 positive and 10 negative genes have been individually analyzed using biological information available from scientific literature (PubMed). Mainly they are connected to fibrosis, inflammation, energy metabolism and other processes known to be affected in DMD. It was found that 12 out of these 30 were previously reported as related to muscle processes/disorders, the fact that can be considered as a proof of concept, providing the possibility to suggest new possible biomarker candidates on the base of suggested procedure.
In summary, this study demonstrates the possibility to decipher regulatory mechanisms of the specific disease (Duchenne dystrophy here) along with corresponding exploratory biomarkers on the base of multiple microarray data meta-analysis only. A lot of predicted expressional regulators are known to be involved in DMD, suggesting that others will also be verified hereafter. This means that all of the proposed regulators can be considered for further drug discovery, whereas their consistently differentially expressed downstream genes can serve as exploratory biomarkers with implicated mechanistic models.
All available microarray datasets of human DMD with more than 10 samples (total 5 datasets, see Table 3) were downloaded from NCBI GEO database [http://www.ncbi.nlm.nih.gov/geo/]. For each probeset intensity values were log-transformed and normalized to zero mean and unit variance. Missing data were imputed using K-nearest neighbor method with k = 10.
We have also utilized data presented in [26], where the lists of up- and down-regulated genes were extracted from research papers, related to skeletal muscle development and pathologies. We limited this dataset to studies of DMD or mdx mice resulting in total 2227 genes which were reported to be differentially expressed in at least in one paper prior to December 2005. For these genes we generated a pseudo-expression dataset for further analysis similar to the standard microarray experiment. If gene was reported to be up-regulated, the gene was assigned a positive value equal to corresponding number of supporting studies; if gene was reported to be down-regulated, the assigned value was negative.
To combine the data from different datasets, we performed the following aggregation procedure. For each probeset we calculated within-dataset log-ratio, two-sample Welch's t-test, Wilcoxon rank sum test and area under ROC curve. If gene on a chip was represented by two or more probesets, we selected the probeset with the least p-value for Wilcoxon rank sum test. We also calculated several other statistics, using popular methods designed specifically for microarray data: limma, SAM and shrinkage T-statistic. Limma, Linear Models for Microarrays [1], [84], is based on a Bayesian hierarchical model for posterior odds of differential expression. SAM, Significance Analysis of Microarrays, was proposed in [2]. Shrinkage T-statistic stabilizes the variances in the denominator via a James-Stein approach [3].
Finally, we have combined the results from different experiments to generate the single “differential” rank for each gene. Separate gene rankings were obtained for nine measures: log-ratio, Welch's t-statistic and corresponding p-value, Wilcoxon's W-statistic and corresponding p-value, AUC, limma, SAM and shrinkage T-statistic. We used Fisher's method to combine p-values of the same type [85]; values of other statistics were averaged for each gene. The final gene rank R was calculated as mean of the ranks from all methods. Each gene was also assigned a single differential log ratio value calculated as an average differential log-ratio from 5 original gene expression datasets.
In order to ensure reproducibility of obtained results, we performed a procedure, analogous to leave-one-out cross-validation: we constructed additional datasets each time aggregating 4 distinct microarray experiments out of total 5 available experiments. Thus we obtained 5 leave-one-out datasets where each microarray experiment was omitted. We also built one large dataset, where all 5 available microarray experiments were aggregated. All subsequent analyses were performed for resultant 6 datasets and the results were cross-validated as further described.
For functional analysis of high-throughput data on the level of potential regulators we used Sub-Network Enrichment Analysis (SNEA) algorithm, implemented in Pathway Studio software [11].
SNEA is a variation of gene set enrichment analysis algorithm, but unlike GSEA [5], [6] that uses predefined gene sets, SNEA utilized sub-networks to construct gene sets on the go. Here, each subnetwork consists of a node (mainly protein or class of proteins – “functional class”) in ResNet and all its expression downstream targets which are automatically derived from the literature. Global expression network includes direct (i.e. transcriptional factor A1 is reported in the literature to regulate specific gene B1) and indirect (i.e. growth factor A2, that can activate specific signaling pathway results to the change of downstream gene B2 expression) relations Ai->Bi. For each subnetwork seed SNEA considers all its expression targets as a gene set that is used for the classical GSEA (Mann-Whitney or Kolmogorov-Smirnov statistical tests).
Thus, SNEA determines the activity of expression regulators based on the differential expression of its targets and favors (assigns lower p-value) those of them which have more significant expression changes downstream.
We performed the SNEA in Pathway Studio with the default parameters: Sub-Network type: gene expression, Mann-Whitney test, p-value<0.05, number of regulators <100 for all log-ratio values (DMD vs. control) from the 6 aggregated datasets. The consistency of default parameters has been tested using 10 permutation tests. It has been shown, that the rate of significant SNEA seeds accidentally found in SNEA results applied to randomized experiment is less than 5%, which is in agreement with default p-value cutoff 0.05. For the reference dataset we ran SNEA with the same parameters using number of studies which reported gene to be differentially expressed. All enrichment algorithms were applied separately to over-expressed and under-expressed genes.
The final sorting of the differentially expressed genes have been done using the following scorewhere N – number of significant regulators upstream of the i-th gene and R –gene rank in aggregated dataset resulted from expression data analysis only.
Most computations were done using R [http://www.r-project.org/] and BioConductor [http://www.bioconductor.org/]. Values of limma, SAM and shrinkage T-statistic were computed using GeneSelector package [86].
Sub-Network Enrichment Analysis was performed using Pathway Studio 7.1 from Ariadne Genomics along with ResNet 7, database storing literature-derived network of biological relations [http://www.ariadnegenomics.com/]. Proprietary Ariadne DiseasesFX database was used for evaluation of gene quality as disease biomarker [Table S2], and ChemEffect [12] was used for studying drugs, related to the regulators of interest.
Muscle Biology Gene Ontology [http://wiki.geneontology.org/index.php/Genes_Involved_in_Muscle_Biology] was used to select genes associated with muscle-related processes.
|
10.1371/journal.ppat.1002618 | Modeling of the N-Glycosylated Transferrin Receptor Suggests How Transferrin Binding Can Occur within the Surface Coat of Trypanosoma brucei | The transferrin receptor of bloodstream form Trypanosoma brucei is a heterodimer encoded by expression site associated genes 6 and 7. This low-abundance glycoprotein with a single glycosylphosphatidylinositol membrane anchor and eight potential N-glycosylation sites is located in the flagellar pocket. The receptor is essential for the parasite, providing its only source of iron by scavenging host transferrin from the bloodstream. Here, we demonstrate that both receptor subunits contain endoglycosidase H-sensitive and endoglycosidase H-resistant N-glycans. Lectin blotting of the purified receptor and structural analysis of the released N-glycans revealed oligomannose and paucimannose structures but, contrary to previous suggestions, no poly-N-acetyllactosamine structures were found. Overlay experiments suggest that the receptor can bind to other trypanosome glycoproteins, which may explain this discrepancy. Nevertheless, these data suggest that a current model, in which poly-N-acetyllactosamine glycans are directly involved in receptor-mediated endocytosis in bloodstream form Trypanosoma brucei, should be revised. Sequential endoglycosidase H and peptide-N-glycosidase F treatment, followed by tryptic peptide analysis, allowed the mapping of oligomannose and paucimannose structures to four of the receptor N-glycosylation sites. These results are discussed with respect to the current model for protein N-glycosylation in the parasite. Finally, the glycosylation data allowed the creation of a molecular model for the parasite transferrin receptor. This model, when placed in the context of a model for the dense variant surface glycoprotein coat in which it is embedded, suggests that receptor N-glycosylation may play an important role in providing sufficient space for the approach and binding of transferrin to the receptor, without significantly disrupting the continuity of the protective variant surface glycoprotein coat.
| The tsetse fly transmitted parasite that causes human African trypanosomiasis, or sleeping sickness, scavenges iron from the bloodstream of the infected individual so that it can live, multiply and ultimately cause disease. To do this, it places a glycoprotein (a protein with carbohydrate chains attached) called the transferrin receptor on its surface to capture circulating human transferrin, an iron transport protein. It then internalizes transferrin receptor/transferrin complex and digests the transferrin part, releasing the iron for its own use. By analyzing the parasite transferrin receptor, we have been able to describe the carbohydrate chains of the transferrin receptor and thus complete a molecular model of this important glycoprotein. We have further built models of how we expect this low abundance glycoprotein will sit in the surface coat of the parasite, which is made of millions of copies of another glycoprotein. The results provide a ‘molecule's eye view’ of how the carbohydrate chains of the transferrin receptor provide the space necessary for the transferrin to bind to it without disrupting the protective coat.
| The tsetse-transmitted Trypanosoma brucei group of parasites cause human African trypanosomiasis and nagana in cattle and constitute a serious health problem for people and livestock in 36 countries of sub-Saharan Africa. T. brucei exists in the mammalian host as the bloodstream form trypomastigote and in the midgut of the tsetse fly vector as the procyclic form. The major surface molecules of the bloodstream form parasite are the glycosylphosphatidylinositol (GPI) anchored [1]–[4] and N-glycosylated [3]–[6] variant surface glycoproteins (VSGs), 5×106 homodimers of which form a dense monolayer over the whole trypanosome [4]. The ability of individual trypanosomes to switch expression from one VSG gene to another gives rise to antigenic variation by which the parasite population survives the host acquired immune response [7]. Other less abundant glycoproteins are arranged either apparently randomly within the VSG coat, like the invariant glycoproteins ISG65 and ISG75 [8], [9], while others have specific surface locations, like Fla1 which locates to the flagellar adhesion zone [10] and the transferrin receptor which locates to the flagellar pocket [11]. Still other glycoproteins are located primarily in intracellular sites, like lysosomal p67 [12], Golgi and lysosomal tGLP1 [13], endoplasmic reticulum GPIdeAc [14] and endosomal TbMBAP1 [15]. The surface of the procyclic form parasite is dominated by 3×106 copies of the GPI-anchored and N-glycosylated procyclin glycoproteins [4], [16], [17], about 1×106 free GPI glycolipids [18], [19] and a high-molecular weight glycoconjugate complex [20], [21]. While this life cycle stage shares some glycoproteins with the bloodstream form, like p67, tGLP1 and Fla1, others are clearly bloodstream form specific, like ISG65, ISG75, TbBMAP1 and the expression site associated gene (ESAG) 6 and ESAG7 subunits of the heterodimeric T. brucei transferrin receptor (TfR).
Some of these glycoproteins are encoded by polygene families, causing sequence heterogeneity in the populations expressed by the trypanosomes. In the case of the TfR ESAG6/ESAG7 subunits, the ESAG6 and ESAG7 genes are associated with telomeric VSG expression sites such that one dominant ESAG6/ESAG7 pair dominates according to which site (and VSG variant) is being expressed [22]. However, there is also some transcriptional breakthrough from other expression sites, as the ESAG6 and ESAG7 genes are immediately adjacent to the expression site promoters, providing some sequence heterogeneity in all TfR preparations [23]. There is functional significance with respect to which ESAG6/ESAG7 pair is expressed due to their different affinities for transferrins from different mammalian species [24], [25]. While there are quite complete data on the GPI anchor and N-glycan structures and N-glycosylation site occupancies of specific VSGs and procyclins [1]–[6], [17], [26] and on the structures of the total N-glycan repertoires of the bloodstream form [27], [28] and procyclic form [29], [30] of the parasite, there is a paucity of data of the glycosylation status of other specific T. brucei glycoproteins. In this paper, we describe the N-glycosylation status of the ESAG6 and ESAG7 subunits of the transferrin receptor (TfR) and, together with our previous description of the GPI anchor of the ESAG6 subunit [31], provide a relatively complete description of the glycosylation status of this low abundance (approximately 3000 copies per cell [32]) but nutritionally essential [33] glycoprotein. The results are discussed in the context of proposed mechanisms of protein N-glycosylation [34]–[37] and endocytosis [38] in T. brucei. We also build a molecular model of the glycosylated ESAG6/ESAG7 transferrin receptor, surrounded by models of glycosylated VSG molecules, to visualize how this receptor sits in the VSG coat on the flagellar pocket membrane and how it might bind its transferrin ligand.
The transferrin receptor (TfR) was purified by affinity chromatography on immobilized transferrin, following the method first described by Steverding and Overath [39]. An aliquot was analyzed by SDS-PAGE and silver staining, which showed the characteristic ESAG6 and ESAG7 subunits (Figure 1A). The identities of the ESAG6 and ESAG7 components were confirmed by excision of the individual bands, in gel tryptic digestion and proteomic analysis (data not shown). Endoglycosidase digests confirmed that both ESAG6 and ESAG7 carry N-linked oligosaccharides. Thus, digestion with both peptide N-glycosidase F (PNGase F), an enzyme that cleaves essentially all types of N-linked glycan, and Endoglycosidase H (Endo H), an enzyme that cleaves only oligomannose-type N-glycans, reduced the apparent molecular weights of both proteins, as judged by SDS-PAGE and Western blotting with anti-TfR antibodies (Figure 1B). However, PNGase F reduced the apparent molecular weights of both proteins more than Endo H (Figure 1B, compare lanes 1 and 3), suggesting that both proteins contain a mixture of Endo H-sensitive (i.e., oligomannose) and Endo H-resistant (i.e., paucimannose and/or complex) N-glycans. The heterogeneity still apparent in ESAG6 following complete de-N-glycosylation with PNGase F is presumably due to the reported heterogeneity in the α-galactose side chains of the GPI anchor attached to this TfR subunit [31].
Aliquots of purified TfR were separated by SDS-PAGE, blotted onto nitrocellulose and probed with anti-TfR antibody (Figure 2, lane 1) and by lectins. Consistent with the presence of oligomannose N-glycans, both ESAG6 and ESAG7 subunits gave a positive reaction with concanavalin A (ConA) (Figure 2, lane 2), as did the bovine ribonuclease B positive control glycoprotein (Figure 2, lane 4). These reactions were abolished when α-methyl-mannose was included in the blotting buffer (Figure 2, lanes 3 and 5), demonstrating the carbohydrate specificity of the ConA blots. However, neither of the ESAG subunits gave a significant reaction with the poly-LacNAc-specific tomato lectin (Figure S1) or, more importantly, with the far more permissive [40] N-acetyllactosamine (LacNAc) specific lectin from Erythrina cristigalli (ErCr) (Figure 2, lane 6) or with the terminal β-galactose-specific lectin ricin (Figure 2, lane 10). These experiments were performed under conditions where a strong reaction was seen against the positive control glycoprotein bovine asialotransferrin (Figure 2, lanes 8 and 12) and where the reactions with the ErCr and ricin lectins against the positive control were abolished by the inclusion of lactose or galactose plus lactose, respectively, in the blotting buffer (Figure 2, lanes 9 and 13). These data suggest that the Endo H-resistant N-glycans of ESAG6 and ESAG7 are not of the poly-LacNAc-containing complex type nor, indeed, even of the LacNAc-containing complex type and are, therefore, most likely of the paucimannose type.
The lectin blotting experiments, described above, suggested that ESAG6 and ESAG7 contain oligomannose and paucimannose N-glycans. However, there remained the formal possibility that the Endo H-resistant N-glycan fraction might include complex N-glycans fully capped with terminal α-Gal residues, which could abrogate ricin and ErCr lectin binding to the sub-terminal β-Gal residues and LacNAc units, respectively, and for which there is precedent in some VSG N-linked glycans [5]. Therefore, to analyze the N-glycan structures further, total N-glycans were released from TfR with PNGase F, radiolabeled by reduction with NaB[3H]4 and analyzed by high-performance thin layer chromatography (HPTLC) alongside radiolabeled N-glycan standards (Figure 3A). A ladder of bands was observed, stretching from the position of Man9GlcNAc2 to Man5GlcNAc2, with two additional bands of higher Rf, possibly corresponding to Man4GlcNAc2 and Man3GlcNAc2 paucimannose species. Significantly, there were no bands with Rf values consistent with complex N-glycans capped with terminal α-Gal residues or with poly-LacNAc-containing N-glycans, like those found in VSG variant MITat1.7 [5] (Figure S2). The radioactive material at the origin of the TLC plate in (Figure 3A) is present in all NaB[3H]4-labeled samples, including commercial glycan standards (Figure S2). A sample of the mixture of labeled N-glycans was separated by Dionex high-pH anion exchange chromatography (HPAEC) and three major radioactive peaks were recovered (Figure S3). These were individually analyzed by HPTLC alongside authentic radiolabeled N-glycan standards and it was found that peak b and peak c co-migrated with Man5GlcNAc2 by HPTLC, while peak a migrated ahead of Man5GlcNAc2 and was assigned as a putative Man4GlcNAc2 structure (Figure 3B). Consistent with the latter assignment, digestion of the peak a material with the Aspergillus saitoi Manα1-2Man-specific α-mannosidase (ASαM) caused an increase in Rf equivalent to the removal of a single hexose (Figure 3C, compare lanes 1 and 2). In contrast, the majority of the material in the peak b fraction was resistant to ASαM (Figure 3C, compare lanes 3 and 4), suggesting that this is a tri-antennary Man5GlcNAc2 structure of the conventional oligomannose series. By inference, we assign the peak c material as the bi-antennary Man5GlcNAc2 structure of the paucimannose series and, indeed, a small component of the peak b material does digest with ASαM to lose two hexose residues, suggesting this is a small amount of bi-antennary Man5GlcNAc2 contamination from the adjacent peak c (Figure 3C, lane 4). Unfortunately, there was insufficient radiolabeled purified peak c material on which to perform a separate ASαM digest. The proposed structures of the main N-glycan species are shown in (Figure 3A). These structures are consistent with the data in (Figure 3A–C) and also draw on our prior knowledge of the structures of the oligomannose and paucimannose series in bloodstream form T. brucei [5], [6], [28].
The aforementioned endoglycosidase digestion results, lectin blots and N-glycan structural analyses strongly suggest that ESAG6 and ESAG7 contain both oligomannose and paucimannose N-glycans, but not complex N-glycans. Previous work has shown that bloodstream form T. brucei expresses two classes of oligosaccharyltransferase (OST] activity [34]–[37]: One that transfers Man5GlcNAc2 from Man5GlcNAc2-PP-Dol to N-glycosylation sequons in relatively acidic environments and another that transfers Man9GlcNAc2 from Man9GlcNAc2-PP-Dol to the remaining N-glycosylation sequons. These activities are encoded by the TbSTT3A and TbSTT3B genes, respectively [34]. We therefore subjected purified TfR to Endo H digestion followed by PNGase F digestion, resolved the double-digested ESAG6 and ESAG7 by SDS-PAGE and performed in-gel tryptic digestion and analyzed the resultant peptides by LC-MS/MS. Using this protocol [34], peptides encompassing Endo H-sensitive (oligomannose) N-glycosylation sites appear with a 203 Da shift, from the single GlcNAc residue left attached to the Asn residue by Endo H, and peptides encompassing Endo H-resistant (paucimannose) N-glycosylation sites appear with a 1 Da shift, from the conversion of Asn to Asp by PNGase F. Using this technique, we were able to positively identify three of the five N-glycosylation sites of ESAG6 as occupied, one (Asn94) with Endo H-resistant paucimannose N-glycans and two (Asn10 and Asn344) with Endo H-sensitive oligomannose N-glycans. The pI values of these Asn-Xaa-Ser/Thr sequons ±5 amino acid residues are consistent with their modification by TbSTT3A and TbSTT3B OST activities, respectively [34] (Table 1). Peptides encompassing the remaining two putative N-glycosylation sites, at Asn219 and Asn234, were not observed but their surrounding sequences would suggest that they are both modified by TbSTT3B OST and are likely to carry oligomannose structures (Table 1). In the comparable ESAG7 analysis, we positively identify one (Asn10) of the three N-glycosylation sites as occupied with Endo H-sensitive oligomannose N-glycans, consistent with its modification by TbSTT3B. Peptides encompassing the remaining two putative N-glycosylation sites, at Asn94 and Asn218, were not observed but their surrounding sequences would suggest that Asn94 is modified by TbSTT3A OST and likely to carry paucimannose structures and Asn218 is modified by TbSTT3B OST and likely to carry oligomannose structures (Table 1). Representations of the glycosylation of the ESAG6 and ESAG7 subunits of the TfR are shown in (Figure 4). The proteomics analysis of the TfR components (described above) also indicated that the principal ESAG6 and ESAG7 sequences present the purified TfR preparation corresponded to those deposited under accession numbers CAQ57442.1 and CAQ57441.1, respectively.
Nolan and colleagues that have reported that TfR can be isolated from a trypanosome lysate with tomato lectin-Sepharose [38]. However, we did not identify any tomato lectin (TL) binding poly-LacNAc-containing N-glycans in either subunit of trypanosome TfR. We therefore entertained the possibility that TfR binds indirectly to TL through interaction with other glycoprotein(s) that do bear poly-LacNAc-containing N-glycans. To investigate this, we took osmotically lysed cells, depleted of VSG and TfR by the action of endogenous GPI-PLC on their GPI anchors, and isolated the total ricin-binding glycoprotein fraction, that includes the TL binding glycoproteins as a significant sub-set [27], and separated and immobilized them by SDS-PAGE and Western blot. The presence of TL-binding glycoproteins was confirmed by probing the blot with TL (Figure 5, lane 3) and the carbohydrate-specificity of this signal was confirmed by inhibition with chitin hydrolysate (Figure 5, lane 4). Identical blots were probed with anti-TfR antibodies before and after pre-incubation with purified TfR. Without pre-incubation with purified TfR, the anti-TfR blots were devoid of significant signal (Figure 5, lane 1), whereas with pre-incubation with purified TfR the anti-TfR blots showed two clear bands at apparent molecular weights of around 55 kDa and 97 kDa. From these data we conclude that TfR is able to bind to other glycoproteins that, in turn, can bind to ricin and therefore possibly also to TL.
Based on the widely accepted assumption that T. brucei TfR has a similar tertiary structure and quaternary structure to the N-terminal domain of VSG [41], [42], for which there are crystallographic data [43], we have made a homology model of the ESAG6/ESAG7 heterodimer of TfR and added to this representative N-linked glycan structures, according to the data and predictions presented in this paper (Table 1 and Figure 4), and a GPI anchor [31]. VSG MITat1.2 was modeled based on the crystal structure of the N-terminal domain [43], the NMR structure of the C-terminal domain [44], and representative N-linked glycan and GPI anchor structures [2], [5], [35], [36]. The N-terminal and C-terminal domains were placed with relatively compact linkers between the two domains and between the C-terminal domain and the GPI anchor. With extended linkers the two domains could be displaced significantly further from the membrane. Human transferrin was modeled based on the structure of iron-bound human transferrin in complex with the human transferrin receptor [45] and representative N-linked [46] and O-linked [47] glycans. The comparison between the models of TfR and VSG MITat1.2 is shown (Figure 6A). A model of TfR surrounded by VSG molecules at their expected surface density [48] is also shown (Figure 6B). Into this model we have placed a model of glycosylated human transferrin, in the same orientation in which it docks to the human receptor [45] (Figure 6C). Although the TfR model is based on the specific ESAG6 and ESAG7 species found in our TfR preparation (accession numbers CAQ57442.1 and CAQ57441.1), the highly conserved amino acid sequences and glycosylation sites of the T. brucei brucei ESAG6 and ESAG7 families (Table S1 and Table S2) suggests that it would be reasonable to assume that this is a general model for all T. brucei brucei ESAG6/ESAG7 heterodimers.
As well as contributing to a three dimensional model of T. brucei TfR, the experimental data on N-glycosylation site occupancy for three of the five N-glycosylation sites of ESAG6 and one of the three for ESAG7 presented in this paper (Table 1; Figure 4) provide support for the model of a unique mechanism of protein N-glycosylation in T. brucei [34]–[37]. According to this model, T. brucei N-glycosylation sequons in relatively acidic environments (like Asn94 of ESAG6) co-translationally receive exclusively (Endo H-resistant) biantennary Man5GlcNAc2 through the action of an oligosaccharyltransferase (OST) encoded by the TbSTT3A gene whereas the remaining sites (like Asn10 and Asn344 of ESAG6 and Asn10 of ESAG7) are acted upon post-translationally by an OST encoded by the TbSTT3B gene and receive exclusively (Endo H-sensitive) triantennary oligomannose Man9GlcNAc2. Once transferred to protein, the biantennary Man5GlcNAc2 structure on the acidic sites may then be processed to paucimannose (Man4GlcNAc2 and Man3GlcNAc2) structures with the latter, in some cases, further elaborated to complex glycan structures. Apparently this further processing to complex glycans does not occur on ESAG6 or ESAG7, where Man4GlcNAc2 appears to be the predominant endo H-resistant structure. The triantennary oligomannose Man9GlcNAc2 structures at the non-acidic sites can only be maximally processed to the triantennary oligomannose Man5GlcNAc2 structure, which appears to be the predominant endo H-sensitive structure on ESAG6 and ESAG7.
Another recent analysis of the single N-glycan of VSG MITat.1.8 also supported the model in [34]. In this case, a single acidic N-glycosylation site at Asn59 was found to be occupied exclusively by a biantennery complex N-glycan structure of Galβ1-4GlcNAcβ1-2Manα1-3(Galβ1-4GlcNAcβ1-2Manα1-6)Manβ1-4GlcNAcβ1-4GlcNAc. Presumably, and in contrast to the VSG MITat.1.8 example, the acidic TfR N-glycosylation sites fail to be processed beyond the trimming of up to two α1-2-linked mannose residues due to steric constraints, reducing access by α-mannosidases and preventing subsequent access by βGlcNAc-transferases.
It was suggested by Nolan and colleagues that T. brucei TfR contains poly-LacNAc glycans because ESAG6 and ESAG7 in whole cell detergent lysates bound to tomato lectin (TL) beads [38]. These authors further suggested a tentative model for endocytosis in trypanosomes, postulating an interaction between poly-LacNAc N-glycans on TfR (and other receptors) and a protein in the flagellar pocket with an extracellular TL lectin-like domain and a cytoplasmic domain that interacts with the machinery of the endocytic pathway. This model was supported by an approximately 5-fold reduction in transferrin endocytic rate when trypanosomes were incubated in 15 mM each of tri-N-acetyl-chitotriose and tetra-N-acetyl-chitotetraose. However, our data show that TfR does not contain any poly-LacNAc structures. Therefore, a direct link between receptor-linked poly-LacNAc glycans and endocytic machinery can be ruled out. However, we have shown here that TfR is able to bind to immobilized ricin-binding glycoproteins. Since the TL-binding glycoproteins of T. brucei are a sub-set of the ricin-binding fraction [27], these data may explain why TfR was found in the TL-binding fraction [38], i.e., through the non-covalent association of TfR with other glycoproteins. It should be pointed out that the association seen in the TfR overlay experiment could be through protein-protein and/or protein-carbohydrate interaction(s) and that, relevant to possible protein-carbohydrate interactions, the glycoproteins in the ricin-binding fraction contain oligomannose and paucimannose glycans as well as conventional complex and poly-LacNAc containing N-glycans. The latter two classes of glycan bind directly to ricin while the former are present because many glycoproteins contain a mixture of both oligomannose and/or paucimannose and complex and/or poly-LacNAc glycans attached to different glycosylation sites in the same polypeptide. While the indirect association of TfR with other glycoproteins could still be relevant for poly-LacNAc-mediated endocytosis in theory, the normal in vitro growth rate of bloodstream form trypanosomes under TbSTT3A RNAi knockdown [34], when the synthesis of almost all complex N-glycans (including poly-LacNAc glycans) is abrogated, also brings the model of poly-LacNAc-mediated endocytosis into question. We therefore suggest that we should return to a null hypothesis: That transferrin, captured by the parasite TfR embedded in the VSG coat, is endocytosed constitutively in clathrin-coated vesicles and that the extremely rapid turnover of the flagellar pocket membrane in bloodstream form T. brucei [49], [50] provides a sufficient rate of uptake of this (and other) essential macromolecular nutrients from host serum.
The molecular modeling of TfR alongside VSG shows that TfR is predicted to sit low in the VSG coat. However, the N-glycans of TfR significantly increase the surface area occupied by TfR compared with VSG. This may be physiologically relevant since the TfR glycans may contribute to protecting the underlying plasma membrane from lytic host serum components while providing sufficient space to allow access of the 80 kDa bi-lobed transferrin glycoprotein, regardless of the relative orientations of the receptor and the ligand (which are currently unknown) when binding takes place. Thus, the widest diameter of transferrin is significantly larger than that of aglycosyl TfR but similar to that of glycosylated TfR. Previously, we had speculated that since the single dimyristoyl-GPI anchor of the ESAG6/ESAG7 TfR heterodimer (as compared to the twin dimyristoyl-GPI anchors of VSG homodimers) would lead to relatively weak association of TfR with the flagellar pocket membrane [51], this might allow TfR to leave the membrane and dock with transferrin in the fluid phase of the flagellar pocket [31]. While the molecular modeling presented here does not altogether rule that model out, it does suggest that the TfR does not de facto have to leave the membrane to dock with its ligand.
Finally, one would predict that transferrin/TfR accessibility at the flagellar pocket membrane is under extreme spatial constraint to prevent complement activation by the underlying plasma membrane. In other words, one would predict that TfR should be able make sufficient space within the VSG coat to allow transferrin to approach and be captured but without exposing significantly more underlying plasma membrane than found throughout the rest of the VSG coat. This tuning of the space occupied by TfR appears to be satisfied by N-glycosylation of both of its subunits and it may explain why TfR has so many N-glycosylation sites (eight) compared to the structurally-related VSGs, which generally have only two or four N-glycans per VSG dimer [4].
Rodents were used to propagate sufficient T. brucei parasites for the purification of sufficient transferrin receptor for high-sensitivity structural analyses. The animal procedures were carried out according the United Kingdom Animals (Scientific Procedures) Act 1986 and according to specific protocols approved by The University of Dundee Ethics Committee and as defined and approved in the UK Home Office Project License PPL 60/3836 held by MAJF.
The transferrin receptor was purified from blood stream form trypanosomes as previously described by Mehlert and Ferguson [31] using affinity chromatography with transferrin-Sepharose which was first described in [39].
Exoglycosidase digests were carried out using both N-glycanase F and Endoglycosidase H as described in Izquierdo et al [34]. The exoglycosidase digests were analyzed by reducing SDS-PAGE with 4–12% gradient gels (Invitrogen), using MOPs buffer and then Western blotting onto nitrocellulose (GE Healthcare) as in [28]. After blocking and incubating in rabbit polyclonal anti-transferrin receptor (kindly supplied by Dietmer Steverding) at the dilution of 1 in 1000 then washing several times in blocking buffer, the membranes were incubated in Anti-rabbit HRP at a dilution of 1 in 20,000. After further washing visualization of the bands was achieved using ECL reagents (GE Healthcare).
SDS PAGE and Western blotting was carried out as above and then the membranes were stained using lectins as described in [34]. All lectin-biotin conjugates were obtained from Vector laboratories. Concanavalin A conjugated to biotin was used at a dilution of 1 in 3,000 (with or without 0.5 M α-methyl-mannose). Ricin-biotin was used at a dilution of 1 in 3,000 (with or without 10 mg/ml galactose and 10 mg/ml lactose), tomato lectin-biotin conjugate diluted was used at a dilution of 1 in 10,000 (with or without chitin hydrolysate, Vector Laboratories, at a dilution of 1 in 10), ErCr lectin was used at a dilution of 1 in 3,000 (with or without 200 mM lactose). The blots were washed extensively after being incubated with the lectin solutions and were incubated in streptavidin-HRP obtained from Sigma Aldrich and diluted to 1 in 10,000. Bands were visualized using ECL reagents as above.
The N-glycans of the trypanosomal heterodimeric transferrin receptor were released by PNGase-F and labeled with sodium borotritiide following the method described in [28]. After extensive cleanup steps to remove any contaminating tritiated material [28] the 3H-labeled glycans were analyzed by HPTLC [Merck silica gel 60] and fractionated by HPAEC as described in [28] and fractions were pooled according to the amount of radioactivity after 10% was used for scintillation counting. Some of the pools were digested using the broad specificity alpha mannosidase extracted from Canavalia ensiformis (jack beans) (Sigma-Aldrich) and the α1-2 specific alpha mannosidase extracted from Aspergillus saitoi (Prozyme), as described in [2]. After digestion the samples were desalted using a mixed bed column as described in [2] and then analyzed again by HTPLC as above. The HTPLC plates were run 3 times in butanol ∶ methanol ∶ water, 4 ∶ 4 ∶ 3 (v/v), with drying between each run, then dried, sprayed with En3Hance (Perkin Elmer) and fluorographed with intensifying screens for up to 8 weeks at −80°C.
Samples of transferrin receptor were digested with Endoglycosidase H followed by PNGaseF digestion (Roche), then analyzed by SDS-PAGE as above. Following staining with Simply Blue (Sigma-Aldrich) the bands corresponding to ESAG6 and 7 were cut out and subjected to proteomic analysis. An aliquot of the tryptic digest was analyzed by LC-MS on an LTQ Orbitrap XL (Thermo) using a Dionex 3000 Nano-LC as in [34]. The resulting data were analyzed using Mascot and the T. brucei geneDB protein database using variable modifications of N-acetylated glucosamine modification of Asn, which would signify an Endoglycosidase H sensitive site, and deamidation of Asn to Asp, which would signify an Endoglycosidase H resistant site, as described in [34].
A ricin-binding total glycoprotein fraction from bloodstream form T. brucei [27] was subjected to SDS-PAGE and transfer to nitrocellulose. These blots were probed with tomato lectin (with and without chitin hydrolysate inhibitor), as described above, or with or without purified TfR (approximately 0.2 µg/ml in phosphate buffered saline). The latter blots were subsequently probed with anti-TfR antibody with ECL detection, as described above.
Molecular modeling was performed on a Silicon Graphics Fuel workstation using InsightII and Discover software (Accelrys Inc.,San Diego, USA). Figures were produced using PyMol (The PyMOL Molecular Graphics System, Schrödinger, LLC). Protein structures used for modeling were obtained from the pdb database [52].
The homology model of T. brucei TfR was based on crystal structure of VSG MITat1.2 (pdb code - 1vsg [43]). The sequence alignment between ESAG6, ESAG7 and VSG was based on [44], modified to take account of the protein tertiary structure and the additional disulphide bonds present. The formation of the disulphide bonds in ESAG6 and ESAG7 between residues equivalent to residues 62 and 286 in MITat1.2 required a distortion of the helix starting at residue 61 and a rearrangement of the loop-containing residue 286. The additional disulphide bond in ESAG7 between residues equivalent to residues 203 and 220 in MITat1.2 could be accommodated with no alteration in the secondary or tertiary protein structure. The model of VSG MITat1.2 was based on the crystal structure of the N-terminal domain (pdb code – 1vsg [43]) and the NMR structure of the C-terminal domain (pdb code – 1xu6 [44]). The C-terminal domain was placed directly below the N-terminal domain [44] to allow for the dense packing of the N-terminal domains on the trypanosome surface [48]. The linkers between the two domains and between the C-terminal domain and the GPI anchor were modeled as relatively compact random loops. The model of human transferrin was based on the structure of iron-bound transferrin in complex with the human transferrin receptor (pdb code – 1suv [45]), N-linked and O-linked glycan structures and GPI anchors were added to all models as appropriate. The structure of the glycans were generated using the database of glycosidic linkage conformations [52] and in vacuo energy minimisation to relieve unfavorable steric interactions. The Asn-GlcNAc linkage conformations were based on the observed range of crystallographic values [53], [54] the torsion angles around the Asn Cα-Cβ and Cβ-Cγ bonds then being adjusted to eliminate unfavorable steric interactions between the glycans and the protein surface.
The following GenBank protein sequence accession numbers were used in this study: CAQ57442.1 and CAQ57441.1. The following Protein Data Bank (pdb) files were used in this study: 1vsg, 1xu6, 1suv.
|
10.1371/journal.ppat.0030139 | Hepatitis C Virus Induces E6AP-Dependent Degradation of the Retinoblastoma Protein | Hepatitis C virus (HCV) is a positive-strand RNA virus that frequently causes persistent infections and is uniquely associated with the development of hepatocellular carcinoma. While the mechanism(s) by which the virus promotes cancer are poorly defined, previous studies indicate that the HCV RNA-dependent RNA polymerase, nonstructural protein 5B (NS5B), forms a complex with the retinoblastoma tumor suppressor protein (pRb), targeting it for degradation, activating E2F-responsive promoters, and stimulating cellular proliferation. Here, we describe the mechanism underlying pRb regulation by HCV and its relevance to HCV infection. We show that the abundance of pRb is strongly downregulated, and its normal nuclear localization altered to include a major cytoplasmic component, following infection of cultured hepatoma cells with either genotype 1a or 2a HCV. We further demonstrate that this is due to NS5B-dependent ubiquitination of pRb and its subsequent degradation via the proteasome. The NS5B-dependent ubiquitination of pRb requires the ubiquitin ligase activity of E6-associated protein (E6AP), as pRb abundance was restored by siRNA knockdown of E6AP or overexpression of a dominant-negative E6AP mutant in cells containing HCV RNA replicons. E6AP also forms a complex with pRb in an NS5B-dependent manner. These findings suggest a novel mechanism for the regulation of pRb in which the HCV NS5B protein traps pRb in the cytoplasm, and subsequently recruits E6AP to this complex in a process that leads to the ubiquitination of pRb. The disruption of pRb/E2F regulatory pathways in cells infected with HCV is likely to promote hepatocellular proliferation and chromosomal instability, factors important for the development of liver cancer.
| Persons infected with hepatitis C virus (HCV) are at increased risk for liver cancer. This is remarkable because HCV is an RNA virus with replication confined to the cytoplasm and no potential for integration of its genome into host cell DNA. While it is likely that chronic inflammation contributes to liver cancer, prior studies with HCV transgenic mice indicate that the viral proteins are intrinsically carcinogenic. In this study, we have examined the interaction of one of these, the RNA-dependent RNA polymerase nonstructural protein 5B, with an important cellular tumor suppressor protein, the retinoblastoma protein (pRb). pRb is a master regulator of the cell cycle, and altered expression of some of the many genes it regulates may lead to cancer. We show that the abundance of pRb is strongly downregulated in cells infected with HCV, and that nonstructural protein 5B targets pRb for destruction via the cell's normal protein degradation machinery. The E6-associated protein appears to play a role in this process, which is interesting as it also mediates the degradation of another tumor suppressor, p53, by papillomaviruses. The loss of pRb function in HCV-infected cells likely promotes hepatocellular proliferation as well chromosomal instability, factors important for the development of liver cancer.
| Among viruses that infect the human liver, hepatitis C virus (HCV) is a leading cause of morbidity and mortality worldwide [1]. Chronic infection with HCV is a major risk factor for the development of cirrhosis as well as hepatocellular carcinoma (HCC) [2,3]. The incidence of this cancer has increased dramatically in recent years in Japan and the United States, reflecting prior increases in the prevalence of HCV infection, and in Japan HCV has replaced hepatitis B virus as the leading infectious cause of liver cancer. The strong association between HCC and HCV infection is particularly notable in that HCV is a positive-strand RNA virus, classified within the genus Hepacivirus of the family Flaviviridae [4]. Its 9.6-kb genome replicates in association with membranes within the cytoplasm of infected cells, and encodes a single polyprotein that is processed by both cellular and viral proteases into ten individual structural and nonstructural viral proteins.
Although inflammation associated with chronic hepatitis C is likely to contribute to the development of HCC, there is strong evidence that one or more of the proteins expressed by the virus contribute directly to carcinogenesis. The HCV core protein, a component of the putative viral nucleocapsid, has been shown to modulate the hepatocyte cell cycle [5,6]. Other studies suggest that expression of the nonstructural (NS) proteins, NS3 (a serine proteinase/helicase), NS5A (a replicase-associated phosphoprotein of uncertain function), or NS5B (the viral RNA-dependent RNA polymerase) may also affect control of cellular proliferation [7–10]. Moreover, transgenic mice expressing a high abundance of the core protein develop steatosis and HCC [11]. Liver cancer also developed in transgenic mice expressing a much lower abundance of the entire viral polyprotein, but not in a companion transgenic lineage expressing a higher abundance of the structural proteins (core, E1, E2, and p7) only [12]. None of these transgenic mouse lineages had demonstrable hepatic inflammation in advance of the development of HCC. Together, these data suggest a direct role for both structural and nonstructural HCV proteins in oncogenesis.
At least four different pathways that regulate either cell proliferation or cell death, the retinoblastoma (pRb)/E2F, p53, transforming growth factor-β (TGF-β), and β-catenin pathways, are commonly altered in HCCs [2]. Among them, pRb plays a major role in controlling the G1- to S-phase transition and mitotic checkpoints through a repressive effect on E2F transcription factors [13]. pRb functions as a tumor suppressor, and the gene which encodes it (RB) is frequently mutated in various types of tumors, including retinoblastomas, small-cell lung carcinomas, and osteosarcomas [14]. In previously published studies, we demonstrated that pRb protein abundance is negatively regulated in cells supporting the replication of subgenomic and genome-length HCV RNA replicons [8]. Similar to the DNA virus oncoproteins E1A of adenovirus and E7 of human papillomavirus (HPV), we found that the RNA-dependent RNA polymerase of HCV, NS5B, forms a complex with pRb in these cells, targeting pRb for degradation and resulting in a reduction in its abundance. This leads to the activation of E2F-responsive promoters in cells containing HCV RNA replicons, and promotes progression of the cell cycle from G1- to S-phase in cells expressing NS5B [8]. While potentially important with respect to the development of HCC in people with chronic hepatitis C, the molecular mechanisms underlying these observations have not been characterized.
Here, we show that pRb is downregulated not only in cells bearing HCV RNA replicons, but also in human hepatoma cells infected in vitro with different genotypes of HCV. We demonstrate that the downregulation of pRb occurs via the ubiquitin-proteasome system, and that it is dependent upon the activity of a known E3 ubiquitin ligase, E6-associated protein (E6AP). E6AP forms a complex with NS5B and pRb, and pRb is ubiquitinated in an NS5B-dependent manner. Our findings reveal a novel mechanism that regulates pRb abundance in HCV-infected hepatocytes, and offer an enhanced understanding of the events leading to the development of HCC in chronically infected patients.
We demonstrated previously that the abundance of pRb is downregulated post-transcriptionally in cells supporting replication of both subgenomic and genome-length HCV RNA replicons derived from a genotype 1b strain of HCV, HCV-N [8]. However, these earlier studies did not determine whether pRb is also downregulated during the course of HCV infection, because genome-length HCV-N RNA is not capable of producing virus that is infectious in cultured cells [15]. To address this question, we used a genotype 2a virus strain, JFH1, that is capable of undergoing the complete viral life cycle in Huh-7 cells in vitro [16–18]. pRb abundance was reduced significantly within 48–72 h of infection with JFH1 virus (Figure 1A and 1B), with quantitative analysis of immunoblots indicating that the pRb abundance 120 h after infection was approximately 20%–30% that of uninfected cells (Figures 1C and S1). The activity of pRb is normally regulated through its phosphorylation by cyclin-dependent kinases, and we also observed an equivalent reduction in the abundance of phospho-pRb in JFH1-infected cells using an antibody specific for pRb phosphorylated at residues 807/811 (Figure 1D and 1E).
Coincident with the reduction in total pRb abundance, confocal microscopy demonstrated a striking cytoplasmic relocalization of pRb in JFH1-infected cells (Figure 2A). Our earlier studies demonstrated that expression of the NS5B RNA polymerase is responsible for the downregulation of pRb in replicon-bearing cells, and that NS5B interacts directly with pRb through a Leu-X-Cys-X-Glu homology domain (LH314–318) overlapping the polymerase active site [8]. Since we lack antibody capable of labeling the genotype 2a NS5B protein within infected cells, we labeled the viral replicase complex in these cells with a broadly reactive polyclonal antibody to NS5A. NS5A is known to colocalize with HCV RNA and other HCV nonstructural proteins within the cytoplasmic “membranous webs” that are thought to be sites of viral RNA synthesis [19]. These studies revealed numerous JFH1-infected cells with an abnormal cytoplasmic accumulation of pRb, and a clear colocalization with the HCV replicase complex as labeled with anti-NS5A (Figure 2A, frame iv). Although the sensitivity of confocal microscopy for detection of pRb renders it difficult to deduce quantitative differences in nuclear pRb abundance in infected versus noninfected cells, we noted similar cytoplasmic relocalization of pRb in cells infected with a cell culture–infectious genotype 1a virus (H77S) developed recently in our laboratory [20] (Figure 2B). Some H77S-infected cells demonstrated nearly complete relocalization of pRb to the cytoplasm (Figure 2B, frame ii arrow), while in most infected cells, there was partial cytoplasmic relocalization (Figure 2B, frame iv arrow). A quantitative analysis of multiple cells indicated that there was a significant increase in the proportion of total pRb present in the cytoplasm of infected versus noninfected cells (Figure 2C; p < 0.001). Confocal microscopy also confirmed a striking reduction in nuclear phospho-pRb in JFH1-infected cells (Figure 2D; compare frames i and iii). Phospho-pRb was found to accumulate within the cytoplasm following treatment with epoxomicin (Figure 2D, frame vi). These results are consistent with an interaction occurring between NS5B and pRb within the cytoplasm. While pRb is normally confined to the nucleus, it does undergo nuclear-cytoplasmic shuttling involving phosphorylation-dependent nuclear export mediated by exportin1 [21]. It is likely that NS5B interacts with and traps pRb in the cytoplasm prior to its initial transport to the nucleus, or perhaps after nuclear export.
Although our previous studies demonstrated that NS5B downregulates the abundance of pRb post-transcriptionally [8], the mechanism by which this occurs is not clear. We confirmed our earlier observations that the stability of pRb is reduced in cells supporting HCV RNA replication by carrying out additional pulse-chase labeling experiments in Huh-7 2–3 cells which contain a genome-length RNA replicon [22]. These results indicated a significantly shortened half-life for pRb in the replicon cells, compared with clonally related cells (2–3c cells) from which the HCV RNA had been eliminated by prior interferon treatment (Figure 3A).
Since the abundance of pRb is regulated through proteasome-dependent pathways in the absence of HCV protein expression [23,24], we considered it likely that HCV may also regulate pRb in a proteasome-dependent fashion. To test this hypothesis, we determined whether proteasome inhibitors [25,26] could restore the abundance of pRb in 2–3 replicon cells. As shown in Figure 3B, epoxomicin, a potent and selective synthetic inhibitor of multiple protease activities of the proteasome, caused a marked increase in pRb abundance, nearly to normal levels, in 2–3 cells (Figure 3B, top panels). In contrast, epoxomicin treatment caused a slight increase in pRb abundance in the cured, HCV-negative 2–3c cell line (Figure 3B, lower panels), consistent with the normal regulation of pRb abundance by proteasomal degradation [23,24]. We also observed similar, cell-type–specific restoration of pRb abundance in 2–3 cells after treatment with lactacystin, an irreversible inhibitor of the 20S proteasome, or MG115, a reversible inhibitor of 20S and 26S proteasomes (Figure S2). Immunofluorescence analysis confirmed the rescue of pRb expression following MG115 treatment of 2–3 replicon cells, but revealed that pRb was localized primarily within the cytoplasm, and not the nucleus, in most 2–3 cells following treatment with the proteasome inhibitor (Figure 3C, arrows). Importantly, pRb retained its normal nuclear localization in MG115-treated 2–3c cells that lack HCV protein expression (Figure 3C, bottom panels). The retention of pRb in the cytoplasm of MG115-treated replicon cells (Figure 3C) is consistent with the interaction of NS5B discussed above. This interaction appears to trap the tumor suppressor protein within the cytoplasm and prevent its translocation to the nucleus in advance of its degradation by the proteasome.
To determine whether inhibition of the proteasome would similarly rescue pRb expression in cells infected with HCV, we treated JFH1-infected cells with epoxomicin. As shown in Figure 3D, this resulted in a marked increase in the abundance of pRb (compare lanes 1 versus 2), as well as an increase in high-molecular-mass pRb-immunoreactive protein. While the identity of this high-mass pRb-immunoreactive protein is uncertain, it is likely to represent ubiquitinated pRb (see below). The abundance of phospho-pRb was also increased following epoxomicin treatment of JFH1-infected cells, although there we observed no discernable high-molecular-mass phospho-pRb species (Figure 3D, lanes 3 versus 4). Considered collectively, these results indicate that HCV regulates the abundance of pRb by promoting its proteasome-dependent degradation. Although overexpression studies have suggested that the NS5B polymerase itself may be regulated by polyubiquitination and proteasome-mediated degradation [27], epoxomicin treatment did not enhance, but rather reduced, the abundance of NS5B in both HCV-infected and replicon cells (Figures 3B and 3D).
Although pRb abundance is normally regulated through proteasome-dependent pathways, such regulation does not necessarily require the ubiquitination of pRb [23,24]. The increase we observed in the abundance of high-molecular-mass pRb-immunoreactive protein in lysates of JFH1-infected hepatoma cells prepared following treatment with epoxomicin (Figure 3D, lane 2) suggests that HCV infection might promote the polyubiquitination of pRb. To assess this possibility, we immunoprecipitated pRb from lysates of Huh-7.5 cells that were infected with the JFH1 virus, then analyzed the precipitates in immunoblots using antibody to ubiquitin. We similarly studied infected cells that had been treated with epoxomicin for 20 h prior to lysis. These results revealed that HCV infection induces polyubiquitination of pRb (Figure 4A, lane 1). A significant abundance of polyubiquitinated pRb was not detected in lysates from mock-infected cells, even following treatment with epoxomicin (Figure 4A, compare lanes 3 and 1).
We also demonstrated HCV-dependent polyubiquitination of pRb in the 2–3 replicon cells by transfecting the cells with a vector expressing Flag-tagged ubiquitin, followed by immunoprecipitation (IP) of lysates with anti-Flag antibody and immunoblotting with anti-pRb (Figure 4B). While a small amount of polyubiquitinated pRb was detected in lysates of the cured 2–3c cells following treatment with MG115 (Figure 4B, lane 6), this was readily detected in lysates of untreated 2–3 replicon cells, and increased by MG115 treatment (Figure 4B, lanes 7–8). We obtained similar results in cells treated with lactacystin (unpublished data). Importantly, the anti-Flag precipitates from the 2–3 cells also contained abundant NS5B, indicating that NS5B was associated with the ubiquitinated pRb in these cells (Figure 4B, lanes 7–8). We observed no high-molecular-mass forms of NS5B, suggesting that there is no appreciable ubiquitination of NS5B under these conditions. These results confirm that HCV infection and/or RNA replication induce polyubiquitination of pRb as a prelude to its degradation by the proteasome.
Since our prior studies revealed that NS5B binds to and induces the degradation of pRb [8], we asked whether ectopic expression of NS5B would also induce pRb ubiquitination. We transfected normal Huh-7 cells with vectors expressing Flag-tagged HCV nonstructural proteins (NS3-4A, NS4B, NS5A, and NS5B), and immunoprecipitated cell extracts with anti-pRb, followed by immunoblotting with anti-ubiquitin antibodies. Overexpression of Flag-NS5B, but not other HCV nonstructural proteins, reproducibly resulted in polyubiquitination of pRb (Figure 4C). In contrast, the ectopic expression of an NS5B mutant (D318N/D319N) containing substitutions within the LH314–318 domain that mediates the interaction of NS5B with pRb [8] resulted in only minimal ubiquitination of pRb (Figure 4C; compare lanes 5 and 6). These data indicate that the interaction of NS5B with pRb leads to targeted destruction of pRb via the ubiquitin-proteasome pathway, representing a striking parallel to the mechanism by which the HPV E6 protein mediates destruction of p53 [28].
A PROSITE (http://ca.expasy.org/prosite/) search indicated that the NS5B protein does not contain a RING finger, or a HECT or ubiquitin interaction motif, making it unlikely that NS5B itself possesses ubiquitin ligase activity [29]. Thus, NS5B is more likely to induce the ubiquitination of pRb by mediating its interaction with a cellular ubiquitin ligase. In searching for this protein, we focused on two recognized cellular E3 ubiquitin ligases: the human homolog of the murine “double minute 2” protein (MDM2), which is involved in ubiquitination pathways that normally regulate the abundance of pRb and p53 [23,24,30], and E6-associated protein (E6AP) which is recruited by the HPV E6 protein to mediate the ubiquitination of p53 [28]. Importantly, E6AP is also known to form a complex with ubiquilin-1 (hPLIC-1), a ubiquitin-like protein that has been shown to interact with NS5B [27,31]. Also, recent studies indicate that E6AP mediates the ubiquitination and degradation of the HCV core protein [32].
To assess the role of these two ubiquitin ligases in NS5B-mediated degradation of pRb, we used siRNA interference to examine the impact of MDM2 and E6AP knockdown on pRb abundance in the 2–3 replicon cells. Using pools consisting of four different specific siRNAs, we were able to achieve effective reductions of each targeted protein (Figure 5A). MDM2 knockdown resulted in a modest but variable increase in pRb abundance (Figure 5A; compare lane 8 with lanes 1 and 2), while knockdown of E6AP reproducibly restored pRb abundance to a level approaching that of the control 2–3c cells (Figure 5A; compare lane 9 with lanes 1 and 2). Two different control siRNAs failed to increase pRb abundance in the replicon cells (lanes 6 and 7). This was also the case with an siRNA pool specific for an E3 ligase sharing the C-terminal HECT domain of E6AP, neural precursor cell–expressed developmentally downregulated protein 4 (NEDD4) (lane 8) [33], indicating that the effect of E6AP knockdown was specific to that protein. None of the siRNAs tested significantly enhanced the abundance of pRb in the control, HCV-negative cells (Figure 5A, lanes 1–5).
To further assess the specificity of the E6AP knockdown, we transfected 2–3 and 2–3c cells with each of the four individual siRNA molecules present in the E6AP pool tested in Figure 5A. These results demonstrated robust enhancement of pRb abundance in cells transfected with three of the E6AP-specific siRNAs (E5, E6, and E7), while none of the siRNAs influenced pRb abundance in the control 2–3c cells (Figure 5B). Transfection of siRNA E8 reduced the abundance of E6AP, but had no apparent effect on pRb abundance (Figure 5B, lane 10). It is possible that this might reflect the involvement of an E6AP splicing variant, as the E5, E6, and E7 siRNAs all target exon 9 of the E6AP gene, while E8 targets exon 10 [34]. Alternatively, we observed kinetic differences in the rates of pRb restoration with these siRNAs, most likely reflecting different efficiencies of E6AP knockdown (unpublished data). Transfection of siRNAs E5 and E6 resulted in increased pRb abundance within 48 h, while this was not observed with siRNA E7 until 96 h or more after transfection. As a final proof of specificity, we mutated one of the E6AP-specific siRNAs (E5), altering the base at two consecutive positions (E5mut; Figure 5C, bottom). Compared with the wild-type E5 siRNA, transfection of E5mut resulted in neither knockdown of E6AP or restoration of pRb abundance (Figure 5C; compare lanes 5 and 6). Consistent with these results, overexpression of both E6AP and NS5B in Huh-7 cells enhanced the downregulation of pRb observed previously with ectopic expression of NS5B alone [8] (unpublished data).
A hallmark of E6AP and other HECT domain ligases is their ability to form a physical complex with the molecule undergoing ubiquitination. Since the results described above suggest that E6AP may play a role in the NS5B-induced ubiquitination of pRb, we sought evidence for an interaction between E6AP, pRb, and NS5B.
We immunoprecipated pRb present in 2–3 cell extracts using antibody to pRb, and demonstrated that the precipitate contained detectable E6AP (Figure 6A, lane 6). In contrast, similarly prepared precipitates from the cured 2–3c cells did not contain appreciable amounts of E6AP (Figure 6A, lane 5), nor did precipitates generated from extracts of the 2–3 replicon cells using antibodies to MDM2 or Flag (Figure 6A, lanes 4 and 8). These results suggest that HCV induces the formation of a complex involving pRb and E6AP. As we had previously shown that NS5B interacts with pRb and targets it for degradation [8], we considered it likely that NS5B was responsible for the E6AP–pRb complex observed in lysates of the replicon cells. We confirmed the interaction of NS5B with pRb by demonstrating the coimmunoprecipitation of NS5B with pRb in lysates of 2–3 cells (Figure 6B, lane 6). For reasons that remain unclear, the amount of NS5B present in the pRb precipitates was markedly reduced when the replicon cells were treated with epoxomicin prior to preparation of the extracts (Figure 6B; compare lanes 5 and 6). The overall abundance of NS5B was also reduced by epoxomicin treatment (Figure 6B; compare lanes 1 versus 2, and Figure 3A and 3B), possibly reflecting nonspecific cellular toxicity and a related reduction in viral RNA replication.
To demonstrate that the formation of the pRb–E6AP complex was dependent upon NS5B and not other HCV proteins expressed by the replicon RNA, we transfected normal Huh-7 cells with vectors expressing various nonstructural HCV proteins, as shown in Figure 4C. Extracts prepared from these transfected cells were precipitated with antibody to pRb, then immunoblotted using antibody to E6AP. Only expression of wild-type NS5B resulted in the formation of a complex between pRb and E6AP (Figure 6C, lane 5). Importantly, this was not observed in cells expressing an NS5B mutant, D318N/D319N, that fails to bind pRb [8] (Figure 6C; compare lanes 5 and 6), or in cells ectopically expressing other nonstructural proteins of the virus: NS3-4A, NS4B, and NS5A. Taken together, the data shown in Figure 6 indicate that NS5B forms a complex with pRb and E6AP. Consistent with the results of the siRNA knockdown experiments described above, these data provide strong support for a role for E6AP in the NS5B-dependent degradation of pRb.
To determine whether E6AP in fact functions as an E3 ligase for pRb, we assessed the ability of recombinant E6AP to direct the ubiquitination of pRb in a reconstituted in vitro ubiquitination reaction using recombinant E1 and E2 proteins. However, these experiments failed to demonstrate ubiquitination of pRb by E6AP, either in the presence or absence of NS5B (Figure S3). These results suggest that E6AP may not be responsible for the NS5B-dependent ubiquitination of pRb, or that this process requires another cellular or viral protein partner that was not included in the reconstituted in vitro ubiquitination reaction. To distinguish between these possibilities, we transfected the 2–3 replicon cells with vectors expressing either wild-type E6AP or a dominant-negative E6AP protein, E6AP-C840A, that contains a single amino acid substitution within the C-terminal HECT domain, ablating its E3 ligase activity [35]. Quantitation of immunoblots from three independent experiments demonstrated that the overexpression of E6AP-C840A resulted in a reproducible increase in the abundance of pRb in the 2–3 replicon cells (Figure 7A [compare lanes 4 versus 6] and 7B). This increase in pRb abundance was clearly apparent in immunoblots of serial 2-fold dilutions of the cell lysates (Figure 7C). In contrast, overexpression of wild-type E6AP caused either no change or a slight increase in pRb abundance in 2–3 cells, and did not appreciably alter pRb expression in the cured 2–3c cells (Figure 7A). These results provide additional evidence that E6AP is required for NS5B-dependent ubiquitination of pRb, and are consistent with the effects of siRNA knockdown of E6AP (Figure 5) and the presence of an NS5B–E6AP–pRb complex in lysates of the replicon cells (Figure 6). We conclude that E6AP is the E3 ligase responsible for NS5B-dependent ubiquitination of pRb in vivo.
We have shown here that pRb is ubiquitinated and degraded in a proteasome-dependent fashion in cultured human hepatoma cells infected with HCV. These observations enhance the biological relevance of prior studies showing that pRb is downregulated by the HCV polymerase protein, NS5B, expressed by autonomously replicating RNA replicons [8]. In showing that the cellular E6AP and viral NS5B proteins form a complex that regulates pRb abundance, we have also provided an enhanced mechanistic understanding of this process.
E6AP was identified originally as a ubiquitin ligase that downregulates p53 in cells expressing the HPV E6 protein [28,36]. The E6–E6AP complex targets additional proteins for ubiquitination, including a set of PDZ domain proteins [37,38] and NFX1–91, a repressor of the hTERT promoter [39]. Interestingly, E6AP has also been shown recently to ubiquitinate and regulate the stability of the HCV core protein [32]. The partial restoration of pRb abundance we observed in the 2–3 replicon cells following either siRNA knockdown of E6AP (Figure 5) or overexpression of a dominant-negative E6AP mutant (E6AP-C840A; Figure 7) provides strong evidence that the ubiquitin ligase activity of E6AP is required for HCV regulation of pRb. Importantly, neither E6AP knockdown nor E6AP-C840A expression altered the abundance of pRb in the clonally related 2–3c cells that had been cured of the HCV replicon by prior treatment with interferon. We also found that the ability of NS5B to form a complex with E6AP is dependent upon the NS5B LH314–318 domain [8] that mediates the interaction of NS5B with pRb (Figure 6C; compare lanes 5 and 6). These data thus lead us to propose a model in which NS5B interacts with hypophosphorylated pRb within the cytoplasm and recruits E6AP to the complex, thereby inducing the ubiquitination and subsequent degradation of pRb via the proteasome. Such a model is consistent with the cytoplasmic redistribution of pRb, as well as the colocalization of pRb and viral nonstructural proteins (specifically NS5A) that we observed by confocal microscopy in cells infected with HCV in vitro (Figure 2A and 2B).
Several possible explanations exist for the inability of E6AP to ubiquitinate pRb in an NS5B-dependent fashion in reconstituted in vitro reactions (Figure S3). First, it may be that additional cellular or viral proteins are required for ubiquitination. Strong evidence exists for functionally important interactions between the NS5B polymerase and several cellular proteins other than pRb, including vesicle-associated membrane-associated proteins A and B (VAP-A and VAP-B), cyclophilin B (CypB), and protein kinase C–related kinase 2 (PRK2), which putatively regulates NS5B by phosphorylation, as well as hPLIC-1, mentioned above [27,40–43]. The deletion of a 21–amino acid C-terminal hydrophobic domain in the NS5B used in our assays (which is required for its solubility) [44,45], could also have affected the in vitro assays, potentially compromising the ability of the polymerase to undergo the conformational changes required for efficient interaction of the LH314–318 domain with pRb [8]. This domain overlaps the Gly-Asp-Asp motif within the active site of the NS5B polymerase, and is sequestered within the interior of the protein in its fully folded form [46].
Prior studies indicate that pRb abundance is regulated by several different mechanisms, reflecting in turn its role as a master regulator of the cell cycle. The E3 ligase MDM2 plays a prominent role in its regulation in the absence of HCV infection [23,24]. While we observed a variable increase in pRb abundance following siRNA knockdown of MDM2 in the 2–3 replicon cells (Figure 5A; compare lanes 6–7 and 10), this effect was typically less than that observed with E6AP knockdown (compare lanes 9 and 10). Moreover, in contrast to E6AP, we were not able to demonstrate an interaction between NS5B and MDM2 (unpublished data). pRb is also targeted for proteasome-mediated degradation by the E7 oncoprotein expressed by high-risk HPVs [47]. Although recognized for many years, the mechanism underlying its regulation by E7 has remained for the most part obscure. However, an active cullin 2 ubiquitin ligase complex has been reported recently to be associated with the HPV type 16 E7 protein, and has been implicated in this process [48]. The Epstein-Barr virus latent antigen 3C mediates degradation of pRb by recruiting the SCFSkp2 ubiquitin ligase [49]. Thus, the steady-state abundance of pRb appears to be regulated through the actions of three distinct E3 ligases: by MDM2 in normal cells, by SCFSkp2 in Epstein-Barr virus–infected B cells, and (as our data suggest) by E6AP in HCV-infected hepatocytes. The human cytomegalovirus pp71 protein also promotes the degradation of pRb and its associated family proteins, but does so in a proteasome-dependent, ubiquitin-independent manner [50]. In addition, gankyrin, an oncogenic ankyrin-repeat protein that is overexpressed in most HCCs, associates with pRb and reduces its stability [51].
While the abundance of pRb may be regulated through a diversity of mechanisms, its activity in controlling E2F transcription factors is largely regulated by phosphorylation [14]. The sequential phosphorylation of pRb plays a pivotal role in the G1/S-phase transition. In its hypophosphorylated state, pRb sequesters and represses the E2F family of transcription factors [13]. Mitogenic growth factors induce phosphorylation of pRb by activation of cyclin D–Cdk4/6 and cyclin E/Cdk2 complexes [52]. This results in release of E2F proteins from pRb, and promotes transcriptional activation of genes required for S-phase and DNA replication. It also appears to promote the nuclear export of pRb, at least in some cell types [21]. Conversely, growth-inhibitory signals reduce cyclin levels or induce Cdk inhibitors, resulting in decreased cyclin/Cdk activity, hypophosphorylation of pRb, and, subsequently, repression of E2F-target genes. Importantly, some ubiquitin ligases are known to catalyze ubiquitination in a phosphorylation-dependent manner. For example, ubiquitination of p27Kip1 is triggered by phosphorylation on Thr187 by Cdk2–cyclin E/A kinase complexes [53,54]. Phosphorylated p27Kip1 is then recognized by SCFSkp2, which polyubiquitinates p27Kip1, targeting it for degradation by the proteasome. At present, we have no direct evidence that phosphorylation influences the ubiquitination and degradation of pRb. However, the relatively low abundance of cytoplasmic phospho-pRb in epoxomicin-treated hepatoma cells infected with HCV (Figure 2D, frame vi) suggests that NS5B may interact preferentially with hypophosphorylated pRb. pRb is also subject to acetylation and sumoylation, both of which regulate pRb function within the cell cycle [55,56], and could also influence how pRb is ubiquitinated. The effect of such modifications on the regulation of pRb by HCV will require further study.
Since the ectopic expression of NS5B stimulates the activity of E2F-responsive promoters, S-phase entry, and cellular proliferation [8], it seems likely that HCV regulation of pRb abundance may enhance the proliferative rate of infected hepatocytes. This virus–host interaction may have evolved because it favors viral RNA replication, which is known to be stimulated in proliferating cells in vitro [57,58]. A more important question, however, is whether the NS5B-mediated degradation of pRb could contribute to the development of HCC in patients with chronic HCV infection.
While the manner in which the expression of E7 by high-risk papillomaviruses contributes to cervical cancer [59] may hold some analogies for HCV, it is likely that HCV regulation of pRb promotes the development of liver cancer in a more indirect fashion. In addition to its role in controlling the G1- to S-phase transition, pRb regulates mitotic checkpoints that are critical controls in prevention of cancer, as they allow cells to repair chromosomal damage before DNA replication or cell division. The risk of oxidative chromosomal DNA damage is likely to be enhanced due to inflammation within the liver during chronic hepatitis C, increasing the importance of these mitotic checkpoints. However, to be fully functional, these checkpoints require competent p53 and pRb pathways [60,61]. Expression of the E6 and E7 proteins from high-risk HPVs can override DNA damage-induced G1 arrest [62], while E6 proteins from high-risk HPVs are also able to compromise the G2/M DNA damage checkpoint [63]. We have shown previously that NS5B-mediated degradation of pRb results in an upregulation of the activity of the E2F-responsive Mad2 promoter [8]. Importantly, deregulation of Mad2, an essential component of the mitotic spindle checkpoint, leads to aneuploidy and an increased risk of tumors, including hepatomas, in mice [64,65]. Thus, while it remains to be proven that HCV infection disrupts mitotic checkpoints through downregulation of pRb, it is a reasonable hypothesis. By downregulating pRb abundance, HCV infection would both stimulate hepatocellular proliferation within a local environment rich in reactive oxygen species and also impair the ability of the cell to respond appropriately to DNA damage. The net result would be increased chromosomal instability. Such a hypothesis is consistent with the diversity of chromosomal abnormalities found in HCC, as well as the typically lengthy period spanning the onset of HCV infection to the development of HCC, and should provide a useful framework for future investigations.
Human hepatoma cells Huh-7 and Huh-7.5 [66] were grown in Dulbecco modified Eagle medium (Cellgro, http://www.cellgro.com/) supplemented with 10% (v/v) heat-inactivated fetal bovine serum, 100 U/ml penicillin G, and 100 μg/ml streptomycin, at 37 °C in a humidified atmosphere with 5% (v/v) CO2. The NNeo/C-5B 2–3, Huh-7–derived cell line containing autonomously replicating, genome-length, dicistronic, selectable HCV RNAs derived with the genotype 1b HCV-N strain were cultured with 500 μg/ml G418, as described previously (Cellgro) [22]. Its companion, interferon-cured progeny cell line 2–3c, was generated and maintained as described previously, and contains no HCV RNA [58].
Cell culture–infectious genotype 1a H77S and genotype 2 JFH1 viruses were harvested from the supernatant fluids of cultures of RNA transfected Huh-7.5 cells, and stored at −80 °C until use [16,20]. For experiments with JFH1 virus, cells were inoculated at a multiplicity of infection (MOI) of 1–2, and virus was allowed to adsorb to cells for 6–12 h at 37 °C prior to replacement of media. pRb abundance and cellular localization was ascertained by immunoblotting and confocal microscopy 48–120 h after infection. H77S infections were carried out at an MOI of ∼0.01 due to the lower efficiency of virus production, and cells were examined by confocal microscopy only.
pCMV6-hE6AP, containing full-length cDNA of human E6AP cloned into the mammalian expression vector pCMV6, was purchased from OriGene. A dominant-negative mutant, E6AP C840A, was generated by PCR mutagenesis using pCMV6-hE6AP as a template with the primers 5′-GCC TTT AAT GTG CTT TTA CTT CCG G-3′ and 5′-AGT ATG AGA TGT AGG TAA CCT TTC-3′. Human ubiquitin B precursor cDNA was purchased from OriGene (http://www.origene.com/), and cDNA representing the mature ubiquitin was subcloned into the pGEM-T Easy cloning vector (Promega, http://www.promega.com/) after amplification by PCR using the primers 5′-CCG GAA TTC ATG CAG ATC TTC GTG AAA ACC CTT AC-3′ and 5′-GCT CTA GAT TAA CCA CCT CTC AGA CGC AGG ACC-3′ to generate pTM-047. After confirming the sequence of both strands of the insert, pTM-047 was digested with EcoRI and XbaI, and the 0.25-kb fragment containing the ubiquitin open reading frame was subcloned into pcDNA3.1/Zeo/IRES-3xFLAG to generate a 3xFLAG-tagged ubiquitin expression vector. pEGFP-C1, pORF9-hRB1, pCMV-tag4-NS3/4A, pCMV-tag4-NS4B, pCMV-tag4-NS5A, pCMV-tag4-NS5B wt, and pCMV-tag4-NS5B D318N/D319N were constructed as described previously [8].
Lactacystin, MG115, and epoxomicin (all from Calbiochem, http://www.emdbiosciences.com/) were prepared as solutions in DMSO. For treatment of replicon cells, 2–3 replicon and cured 2–3c cells were seeded into 6-well plates and grown to 50% confluence. Inhibitors were added to the culture media at the indicated concentrations, and cells were incubated for 10–12 h (lactacystin or MG115) or 20 h (epoxomicin), followed by preparation of cell extracts for immunoblots. For epoxomicin treatment of JFH1 virus–infected cells, the inhibitor was added 20 h prior to lysis of cells and preparation of cell extracts.
siRNA oligonucleotide SMARTpools, each containing four siRNA oligonucleotides specific for human MDM2 (M-003279–02), E6AP/UBE3A (M-005137–00), and NEDD4 (M-007178–01), and individual E6AP/UBE3A siRNAs, were purchased from Dharmacon (http://www.dharmacon.com/). Negative control siRNAs (4611 and 4613) were from Ambion (http://www.ambion.com/). For knockdown experiments, 2–3 replicon and cured 2–3c cells were grown to 30% confluence in 6-well plates, and transiently transfected with 80 nM of siRNAs using Lipofectamine 2000 (Invitrogen, http://www.invitrogen.com/) according to the manufacturer's instructions. Protein extracts were prepared for further analysis 72–120 h after transfection.
Cells were seeded into 6-well plates 24 h before transfection and grown to 50% confluence. Before transfection, the culture medium was replaced with fresh medium without antibiotics. For overexpression of E6AP or E6AP C840A, 2–3 and 2–3c cells were transiently transfected with 4 μg of pCMV6 (empty vector), pCMV6-hE6AP, or pCMV6-hE6AP-C840A, along with 0.25 μg of pEGFP-C1 (Promega), using FuGENE 6 reagents (Roche Diagnostics, http://www.roche.com/). Protein extracts were prepared for immunoblots at 48 h after transfection.
Cells were washed three times with chilled PBS, and incubated in chilled lysis buffer (20 mM Tris-HCl [pH7.5], 150 mM NaCl, 10 mM EDTA-2Na, 1% [v/v] Nonidet P-40, 10% [v/v] glycerol, and 2 mM DTT) supplemented with 1 mM PMSF and 2 μg/ml aprotinin, or complete protease inhibitor cocktail (Roche), for 30 min at 4 °C. Cell debris was pelleted by centrifugation at 13,000g for 30 min at 4 °C, and supernatants were used as soluble fractions. Protein concentrations were determined by the modified Bradford assay with BSA as a standard (Bio-Rad, http://www.bio-rad.com/). SDS-PAGE and subsequent immunoblotting were done as described previously [8], using mouse monoclonal antibodies against β-actin (AC-15; Sigma, http://www.sigmaaldrich.com/), GAPDH (glyceraldehyde-3-phosphate dehydrogenase; Ambion), Flag tag (M2; Sigma), MDM2 (SMP14; Santa Cruz Biotechnology, http://www.scbt.com/), pRb (G3–245; BD Biosciences, http://www.bdbioscences.com/), and ubiquitin (P4D1; Santa Cruz Biotechnology), and rabbit polyclonal antibodies against phospho-pRb 807/811 (Cell Signaling Technology, http://www.cellsignal.com/), E6AP (sc-25509; Santa Cruz Biotechnology), NEDD4 (sc-25508; Santa Cruz Biotechnology), and NS5B (A266–1; ViroGen, http://www.virogen.com/; or provided as a generous gift by Dr. Craig E. Cameron, Pennsylvania State University, State College, Pennsylvania, United States). Membranes were probed with appropriate secondary antibodies conjugated with horseradish peroxidase, visualized by ECL reagents (Amersham Pharmacia Biosciences, http://www.amersham.com/), and exposed to x-ray films.
Huh-7 2–3 and 2–3c cells were seeded into 8-well Labtek chamber slides and grown until 50%–60% confluent, with or without the addition of 20 μM of MG115. After washing twice with PBS, the cells were fixed in methanol–acetone (1:1 [vol/vol]) for 10 min at −20 °C, air-dried for 60 min at room temperature, washed twice with PBS, and incubated with blocking buffer (1% BSA in PBS) overnight at 4 °C. pRb was visualized by staining with mouse monoclonal antibody G3–245. After washing three times with PBS, slides were further incubated with a goat anti-mouse Ig secondary antibody conjugated with FITC for 1 h at room temperature. Slides were then washed three times with PBS, counterstained with diamidino-2-phenylindole 2HCl (DAPI), mounted in Vectashield mounting medium (Vector Laboratories, http://www.vectorlabs.com/), and examined with a Zeiss AxioPlan2 fluorescence microscope (http://www.zeiss.com/).
Huh-7 2–3 and 2–3c cells or JFH1-infected Huh-7.5 cells were cultured in Labtek chamber slides (http://www.labtek.net/) and fixed with 4% paraformaldehyde in PBS for 30 min. Cells were permeabilized with Triton X-100 (0.2%) for 15 min and blocked with 10% normal goat serum at room temperature for 1 h. Cells were then incubated with the appropriate dilutions of primary antibodies for 1 h followed by secondary antibodies for 1 h at room temperature. HCV antigen was visualized with rabbit polyclonal antibody to NS5A (a generous gift from Dr. Craig E. Cameron) followed by Alexa 594 secondary antibody conjugate, or with human polyclonal antibody and an FITC-labeled anti-human Ig secondary antibody; pRb was visualized with visualized by staining with mouse monoclonal antibody G3–245 (BD Biosciences), and phospho-pRb by rabbit polyclonal antibody against phospho-pRb 807/811 (Cell Signaling Technology) followed by secondary antibodies: goat anti-mouse Ig conjugated to FITC or goat anti-rabbit Ig conjugated to Alexa 594. Slides were washed and counterstained with DAPI, and mounted in Vectashield mounting medium, then sealed and examined with a Zeiss LSM 510 laser scanning confocal microscope within the Infectious Disease and Toxicology Optical Imaging Core at the University of Texas Medical Branch.
Pulse-chase labeling of endogenous pRb protein was done as described previously [8].
For analysis of the interaction between pRb and E6AP in normal Huh-7 cells, cells were grown to 50% confluence in 10-cm dishes and transfected with 5 μg of pCMV-tag4, pCMV-tag4-NS3/4A, pCMV-tag4-NS4B, pCMV-tag4-NS5A, pCMV-tag4-NS5B wt, and pCMV-tag4-NS5B D318N/D319N. At 48 h after transfection, 20 μM of MG115 was added to the transfected cells for 10 h. Cells were lysed in 1 ml of IP lysis buffer (20 mM Tris-HCl [pH7.5], 150 mM NaCl, 10 mM EDTA-2Na, 1% [v/v] NP-40, 10% [v/v] glycerol, 1 mM PMSF, and 2 μg/ml aprotinin), and extracts were prepared as described above. IP was performed with 500 μg of extracts using anti-pRb monoclonal antibodies, as described previously, and immunoprecipitated proteins were analyzed by immunoblot [8]. For analysis of the interaction between pRb and E6AP in HCV replicon cells, 2–3 replicon and 2–3c cured cells were lysed in IP lysis buffer, and soluble extracts were prepared. IP was performed with 500 μg of extracts using 1 μg of anti-FLAG (M2; Sigma), anti-pRb (G3–245; BD Biosciences), or anti-MDM2 (SMP14; Santa Cruz Biotechnology) monoclonal antibody, and immunoprecipitated proteins were analyzed by immunoblot.
For analysis of the interaction between NS5B and ubiquitinated pRb, NNeo/C-5B 2–3 and 2–3c cured cells were transfected with 3xFLAG-tagged ubiquitin expression vector, and treated with DMSO or 20 μM MG115 for the last 10 h. At 48 h after transfection, cells were lysed in IP lysis buffer, and soluble extracts were prepared. IP was carried out using 500 μg of extracts and anti-FLAG monoclonal antibody, and immunoprecipitated proteins were analyzed by immunoblot.
For detection of ubiquitinated pRb, 2–3 replicon cells and 2–3c cured cells were cultured to 50% confluence in 10-cm dishes and transfected with 5 μg of 3xFLAG-tagged ubiquitin expression vector. At 48 h after transfection, cells were treated with DMSO, 10 μM of lactacystin, or 20 μM of MG115 for 10 h, and then lysed in IP lysis buffer supplemented with 1 mM NaF, 1 mM Na3VO4, 4 mM N-ethylmaleimide, and 6 nM ubiquitin aldehyde (ubiquitination buffer), and soluble extracts were prepared. A total of 500 μg of extracts were used for IP with anti-FLAG monoclonal antibodies, and immunoblots were carried out with anti-pRb or antiubiquitin monoclonal antibodies.
For detection of NS5B-dependent ubiquitination of pRb, normal Huh-7 cells were cultured to 50% confluence in 6-well plates, and transfected with 2 μg pCMV-tag4, pCMV-tag4-NS3/4A, pCMV-tag4-NS4B, pCMV-tag4-NS5A, pCMV-tag4-NS5B wt, and pCMV-tag4-NS5B D318N/D319N. At 48 h after transfection, 20 μM of MG115 was added to the transfected cells for 10 h. Cells were lysed in ubiquitination buffer, and soluble extracts were prepared. IP was carried out using anti-pRb monoclonal antibodies, followed by immunoblotting with monoclonal antiubiquitin antibody.
Reconstituted, in vitro ubiquitination reactions were carried out using purified recombinant pRb (Abcam, http://www.abcam.com/), recombinant NS5B protein with a 21–amino acid C-terminal deletion (Replizyme, http://www.replizyme.com/), and recombinant E6AP produced in baculovirus, essentially as described [35].
The Entrez Protein (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=Protein) accession numbers for the proteins discussed in this study are as follows: pRb (NP_000312), E6AP (NP_061828), MDM2 (NP_002383), NEDD4 (NP_006145), ubiquitin (NP_061828), HCV-N full-length polyprotein (AAD44719), and JFH1 full-length polyprotein (BAB32872). |
10.1371/journal.pbio.2004538 | Functional characterization of adaptive variation within a cis-regulatory element influencing Drosophila melanogaster growth | Gene expression variation is a major contributor to phenotypic diversity within species and is thought to play an important role in adaptation. However, examples of adaptive regulatory polymorphism are rare, especially those that have been characterized at both the molecular genetic level and the organismal level. In this study, we perform a functional analysis of the Drosophila melanogaster CG9509 enhancer, a cis-regulatory element that shows evidence of adaptive evolution in populations outside the species’ ancestral range in sub-Saharan Africa. Using site-directed mutagenesis and transgenic reporter gene assays, we determined that 3 single nucleotide polymorphisms are responsible for the difference in CG9509 expression that is observed between sub-Saharan African and cosmopolitan populations. Interestingly, while 2 of these variants appear to have been the targets of a selective sweep outside of sub-Saharan Africa, the variant with the largest effect on expression remains polymorphic in cosmopolitan populations, suggesting it may be subject to a different mode of selection. To elucidate the function of CG9509, we performed a series of functional and tolerance assays on flies in which CG9509 expression was disrupted. We found that CG9509 plays a role in larval growth and influences adult body and wing size, as well as wing loading. Furthermore, variation in several of these traits was associated with variation within the CG9509 enhancer. The effect on growth appears to result from a modulation of active ecdysone levels and expression of growth factors. Taken together, our findings suggest that selection acted on 3 sites within the CG9509 enhancer to increase CG9509 expression and, as a result, reduce wing loading as D. melanogaster expanded out of sub-Saharan Africa.
| Much of the phenotypic variation that is observed within species is thought to be caused by variation in gene expression. Variants within cis-regulatory elements, which affect the expression of nearby genes within the same DNA strand, are thought to be an abundant resource upon which natural selection can act. Understanding the functional consequences of adaptive cis-regulatory changes is important, as it can help elucidate the mechanisms underlying phenotypic evolution in general and provide insight into the development and maintenance of biodiversity. However, functional analyses of these types of changes remain rare. Here we present a functional analysis of an adaptively evolving enhancer element of a D. melanogaster gene called CG9509, of previously unknown function. We show that 3 single nucleotide polymorphisms located within the enhancer of this gene are responsible for an increase in CG9509 expression in cosmopolitan populations (outside of south and central Africa) relative to sub-Saharan populations, which include ancestral populations. We further show that CG9509 is involved in the regulation of growth rate and body size determination and propose that the CG9509 enhancer underwent positive selection to reduce wing loading as the species expanded out of sub-Saharan Africa.
| Gene expression variation is extensive both within and between species and is believed to underlie much of the phenotypic diversity observed among species, as well as among populations of the same species [1–2]. Furthermore, expression variation is thought to provide an abundant source of material for adaptation, as alterations in gene expression are more easily fine-tuned on a temporal and tissue-specific scale than changes in protein structure [3–4]. In particular, cis-regulatory elements, which are adjacent to genes and directly affect their expression, are thought to be frequent targets of adaptive evolution [2–6]. Despite this prediction, examples of adaptive cis-regulatory changes remain comparatively rare, although the number of such examples continues to grow [7–16]. The discrepancy between the predicted abundance and actual instances of identified adaptive cis-regulatory divergence is likely in part due to the difficulty in detecting regulatory adaptation, as well as in determining its effect on an organismal phenotype that may be the target of selection. Even in some of the best-studied species, the function of many genes remains unknown, and alterations in those with known functions often have pleiotropic effects, making it difficult to determine the link between an expression change and an adaptive organismal phenotype. As more instances of adaptive cis-regulatory evolution are uncovered, it is important to identify the genetic and molecular mechanisms that underlie them, which can help to further our understanding of the mechanisms of phenotypic evolution and shed light on the origins of biodiversity [17]. However, studies performing in-depth functional analyses of individual adaptively evolving cis-regulatory elements remain even more rare than those documenting adaptive cis-regulatory divergence (e.g., [18–21]).
Transcriptomic methods have proven effective at identifying putatively adaptive alterations in gene expression within and between species [22–28]. CG9509 is a gene initially identified as a candidate for adaptive cis-regulatory divergence through one such study that compared expression between a derived, European and an ancestral, sub-Saharan African (henceforth sub-Saharan) population of D. melanogaster [24]. Until now, the function of CG9509 has remained unknown, although it has been predicted to have oxidoreductase activity [29] and/or play a role in ecdysteroid metabolism [30]. Adult CG9509 expression was found to be 2–3-fold higher in the European population than in the sub-Saharan population (Fig 1A) [12,24], and this expression difference extends to other cosmopolitan (here defined as populations outside of south and central Africa) and sub-Saharan populations [31]. Transgenic reporter gene experiments revealed that variation within a 1.2-kb cis-regulatory element upstream of the gene (referred to here as the CG9509 enhancer, Fig 2A) can account fully for the expression divergence and shows evidence of a selective sweep in cosmopolitan populations [12,31]. This suggests that positive selection acted on the CG9509 enhancer to increase CG9509 expression after D. melanogaster’s expansion out of sub-Saharan Africa, which is estimated to have occurred approximately 15,000 years ago [32–34], but before the separation of European and Asian populations approximately 2,500–5,000 years ago [31,35]. Within the CG9509 enhancer, there are 9 single nucleotide polymorphisms (SNPs) and 1 insertion/deletion (indel) polymorphism (Fig 2B) that show large frequency differences between the populations and are candidates for the target(s) of selection responsible for the expression divergence. In all cases, the cosmopolitan variant is inferred to be the derived state, while the sub-Saharan variant is inferred to be ancestral [31].
In this study, we use site-directed mutagenesis and transgenic reporter genes to determine the effect of individual SNP and indel variants within the CG9509 enhancer on gene expression. We also use RNA interference (RNAi) and a newly discovered CG9509 hypomorph allele to reveal some of CG9509’s previously unknown biological functions. We find that 3 SNPs within the CG9509 enhancer contribute to the expression divergence seen between cosmopolitan and sub-Saharan alleles. Interestingly, 2 SNPs that have a small effect on expression are fixed in cosmopolitan populations and appear to have been the targets of a selective sweep, while the SNP with the largest effect on expression is at intermediate frequency in cosmopolitan populations, suggesting that another type of selection may be acting on this site. We also show that CG9509 expression influences larval growth and thus plays a role in determining adult body size and wing loading (i.e., the ratio of body mass to wing area). Furthermore, the genetic variants influencing CG9509 expression are associated with variation in these phenotypic traits. Our results suggest that selection on the 3 SNPs within the CG9509 enhancer occurred in order to reduce wing loading outside of sub-Saharan Africa.
Previous studies focused solely on adult CG9509 expression variation [12,31]. To determine if the adult expression pattern is established earlier in development, we surveyed larval CG9509 expression. Because D. melanogaster developmental gene expression is highly dynamic with a high transcriptional turnover, even among stages that are only a few hours apart [36], we focused our expression analysis on a well-established larval stage in order to ensure that any observed expression divergence is due to population divergence rather than developmental stage variation. To this end, we surveyed CG9509 expression in late wandering third instar larvae of 3 cosmopolitan populations (the Netherlands, Egypt, and Malaysia) and 2 sub-Saharan populations (Zimbabwe and Zambia). Similar to adults (Fig 1A), larval CG9509 expression in cosmopolitan populations was significantly higher than in sub-Saharan populations by 3–5.5-fold (t test, P < 5 × 10−4, Fig 1B and S1 Fig).
To determine which variants in the CG9509 enhancer (Fig 2B) contribute to the expression divergence between cosmopolitan and sub-Saharan D. melanogaster (Fig 1 and S1 Fig), we created a series of transgenic reporter gene constructs that were introduced into the D. melanogaster genome. Because these reporter gene constructs were tested in a common genomic background, our results should be free from the confounding effects of trans-acting factors. Briefly, a cosmopolitan and a sub-Saharan enhancer allele were cloned in front of a LacZ reporter gene, and 6 sites of interest at positions 1174, 1155, 1063, 821–817, 765, and 67 (Fig 2B) were mutated individually and in various combinations (Fig 3). The tested sites were chosen as those at which the derived variant was fixed in all cosmopolitan populations, but at a frequency ≤10% in sub-Saharan populations. Position 67, at which the derived variant is at intermediate frequency in cosmopolitan populations but absent from sub-Saharan populations (Fig 2B), was also tested because it had previously been associated with CG9509 expression variation [31,37]. Mutations were first introduced into the cosmopolitan enhancer allele, changing the nucleotide(s) to the ancestral (sub-Saharan) state. Sites found to have an effect on expression in the cosmopolitan background were then mutated to the cosmopolitan state in the sub-Saharan background. Adult and larval expression driven by the cosmopolitan enhancer was 3–4-fold higher than that driven by the sub-Saharan enhancer (Fig 3), which is in line with the expression divergence found in natural populations (Fig 1 and S1 Fig). Thus, the CG9509 enhancer can account for nearly all of the expression divergence between cosmopolitan and sub-Saharan larvae and adults.
In order to elucidate the function of CG9509 and the effects of its expression on organismal phenotypes, we performed a series of functional and tolerance assays on flies in which CG9509 expression was disrupted. For this, we used both a newly identified hypomorph allele and RNAi. The hypomorph allele (CG9509del) was discovered as a spontaneous mutation in an isofemale line derived from Munich, Germany. It contains a frameshift-causing deletion within the CG9509 coding region and shows greatly reduced levels of CG9509 mRNA (S2 Fig). As a control, flies homozygous for CG9509del were compared to wild-type isofemale lines derived from the same population at the same time. RNAi knockdown of CG9509 expression was achieved by crossing a ubiquitous Act5C-GAL4 driver line to a transgenic line containing an RNAi hairpin construct specific to CG9509 (RNAi-CG9509) and flanked by a yeast upstream activating sequence (UAS). As a control, CG9509 knockdown flies were compared to flies of the host strain from which they were derived (UAS-).
To determine if there was an association between SNP variation in the CG9509 enhancer and variation in either body size or wing loading, we performed further analyses in the Dutch population, as well as in the reconstituted homozygotes, hemizygotes, and/or heterozygotes of the F2 generation from reciprocal crosses of individuals from either the Dutch or a Rwandan population (S1 Text). The genetic background of these reconstituted F2 individuals represents a mixture of the 2 original parental genomes, which allows us to better disentangle the effects of SNP variation within the CG9509 enhancer from the effects of other variants in the genome (S1 Text), including trans-acting variants. Thus, association of CG9509 enhancer SNP variants with body size or wing-loading variation within this shared background should represent true associations, rather than spurious associations caused by linked variation elsewhere in the genome.
We have shown that the between-population expression divergence of CG9509 occurs in both larvae and adults (Fig 1) and is driven by nucleotide polymorphism within the CG9509 enhancer (Fig 3). We identified 3 SNPs that can account for the majority of the expression divergence, 2 of which (at positions 1174 and 1063, Fig 2B) only affect expression in larvae and have a relatively small effect on larval expression (Fig 3C). The third SNP (position 67, Fig 2B) accounts for the majority of the expression divergence in both adults and larvae (Fig 3). We propose that selection on the CG9509 enhancer occurred in 2 phases. First, the derived variants at positions 1174 and 1063 were the targets of the previously identified selective sweep [12]. These variants are fixed in cosmopolitan populations but absent or at low frequency in ancestral, sub-Saharan African populations (Fig 2B). Thus, this sweep likely occurred during or shortly after D. melanogaster’s expansion out of Africa but before the separation of European and Asian populations. A single haplotype spans positions 1174 and 1063 in cosmopolitan populations, suggesting that both derived variants were fixed by a single selective sweep. The derived variants at these positions are also present at low frequency in central African populations [47], suggesting that selection acted on standing variation. In a second step, we propose that the large-effect derived variant at position 67 arose as a new mutation on the selected haplotype and rose to intermediate frequency more recently in cosmopolitan populations. Consistent with this view, the derived variant is absent from sub-Saharan and central African populations (Fig 2B) [47]. It has been proposed that advantageous regulatory mutations with large effects are likely to display overdominance and, thus, remain polymorphic within populations [48]. The large effect of the derived variant at position 67 on CG9509 expression (Fig 3 and S1 Fig) and its intermediate frequency in cosmopolitan populations (Fig 2B) are consistent with this model. However, other causes for its maintenance at intermediate frequency, such as sexual antagonism, temporally varying selection, or the interaction of alleles at multiple loci, are also possible [48–51].
Until now, CG9509’s function and, therefore, the organismal phenotype(s) affected by variation in its expression have remained unknown. Here, we used RNAi-mediated knockdown of CG9509 expression and a newly identified CG9509 hypomorph allele to show that increased CG9509 expression is associated with reduced wing loading (Fig 6). Wing loading in a Dutch population was associated with polymorphism at position 67 (Fig 6D), and we also found that flies from Zambia had greater wing loading than those from the Netherlands (Fig 6C). When we surveyed wing loading in the F2 offspring of crosses between fly strains containing either cosmopolitan or sub-Saharan CG9509 enhancer variants affecting expression, wing loading varied in the expected direction (Tables 2 and 3), but this variation was small (approximately 1%–3%) and nonsignificant (Tables 2 and 3). Some of this discrepancy between our findings in Dutch and Zambian populations versus F2 offspring may be due to trans-acting variation present among the inbred lines from different populations. This variation should be greatly reduced in the F2 offspring, which share a more homogenous trans environment after a generation of recombination. However, we did detect significant associations between these variants and both body weight and wing size (Tables 2 and 3). Our inability to detect significant associations with wing loading may be a result of this trait being a ratio of 2 measurements, which increases trait variance and reduces statistical power. It is also possible that the effect size is too small to be detected as significant in our experimental design. Previous studies have shown that even small changes in flight load can lead to differences in flight performance [52–53], especially at low temperatures [53], and therefore impact fitness.
Consistent with our findings, several studies have documented clinal variation in wing loading among Drosophila populations across multiple continents [39,53–56], with reduced wing loading at higher latitudes, and this cline is thought to be maintained by selection. In Drosophila, and indeed all flying animals, relative size is important for flight aerodynamics. Furthermore, previous microarray comparisons of gene expression have found overexpression of muscle-related genes (including flight muscle components) in flies from Zimbabwe relative to those from the Netherlands [24,25], suggesting that flies in the ancestral range require a greater investment in flight muscle. This suggests that improved flight ability may have been an important adaptation as D. melanogaster expanded its species range, and selection on the CG9509 enhancer likely favored the reduction in wing loading conferred by increased larval CG9509 expression as D. melanogaster expanded out of Africa. Drosophila wing beat frequency and power output decrease as temperature decreases, resulting in reduced flight ability at cooler temperatures [53], and previous studies suggest that reduced wing loading may help counteract this effect [39,57–58]. Thus, improved flight ability may represent an adaptation to lower temperatures in the derived species range; however, general improvement of flight ability could also be adaptive. The energy conserved by improved flight ability could be used for other processes or aid in survival when resources are scarce. Improved flight ability might also aid in predator evasion or dispersion, which may have helped facilitate D. melanogaster’s expansion to new territories. However, it is important to note that, although selection for reduced wing loading represents a plausible scenario for adaptive regulatory evolution at the CG9509 locus, we cannot rule out the possibility that selection acted on an unobserved, pleiotropic trait associated with variation in the CG9509 enhancer.
We have shown that increased CG9509 expression is associated with reduced larval growth (Fig 4) and adult body size (Fig 5) and were able to associate body size variation with CG9509 enhancer sequence variation (Tables 2 and 3), with the high-expression, cosmopolitan variants associated with decreased body size. We also showed that weight is reduced in a sub-Saharan population in comparison to a Dutch population (S5 Fig). However, these results are contrary to expectations if selection acted on the CG9509 enhancer to reduce body size in cosmopolitan populations but in line with well-documented, latitudinal body size clines that are thought to be maintained by selection [55,59]. We additionally showed that increased CG9509 expression is associated with increased levels of the maturation hormone ecdysone (Fig 7A and 7B, S6 Fig). Most likely, the increased active ecdysone levels result in the reduced larval growth rate and a subsequently smaller body size, since the antagonistic interaction of ecdysone with insulin signaling is known to suppress larval growth [43,60], which in turn reduces adult body size. However, the mechanism through which CG9509 expression adjusts larval growth to reduce wing loading remains unknown. The effect on wing loading (Fig 6) is at least in part due to CG9509’s effect on active ecdysone levels (Fig 7A and 7B, S6 Fig), as ecdysone plays a key role in regulating proportional growth and coordinating the growth of individual organs with each other as well as with the entire body [61].
CG9509 is expressed in the larval fat body [62], which acts as a coordinator of larval growth [40,41,44]. Ecdysone signaling specifically in the fat body antagonizes insulin signaling in part via down-regulation of the positive growth regulator dMyc and translocation of the negative growth regulator dFOXO to the nucleus, where it activates the expression of target genes [44,45,63]. When we knocked down CG9509 expression, we found a decrease in both dMyc and dFOXO expression in late wandering third instar larvae (Fig 7D and S6 Fig), which is the stage during which the peak of the final and largest larval ecdysone titer occurs, signaling the onset of pupariation [40,41]. We detected a similar decrease in the early wandering third instar larval stage, which coincides with another, smaller ecdysone peak [40,41], but only in CG9509del larvae (Fig 7D). The reduction in dFOXO transcript expression is interesting, as it is dFOXO protein localization that suppresses growth [44,45,63]. However, a study documenting insulin/TOR network transcriptional variation found strong covariance for dFOXO transcript abundance and the expression of dFOXO-affected genes [64]. Thus, the expression of genes downstream of dFOXO may also be affected. The reduction in dMyc expression (Fig 7D) is counterintuitive, since its up-regulation in the fat body is expected during CG9509 expression knockdown. However, the effect of ecdysone on dMyc expression is both stage- and tissue-specific [40]; thus, the expression decrease is likely in another tissue and may represent a part of the mechanism through which CG9509 expression adjusts proportional growth to affect wing loading. Interestingly, previous studies have documented negative correlations of the expression of growth-associated genes with stress tolerance [65,66], which we also found for CG9509 expression (Table 1). However, this correlation could simply be a by-product of body size, which has been shown to correlate with stress tolerance in Drosophila [67].
We identified 3 SNPs that account for the majority of CG9509 expression divergence observed between cosmopolitan and sub-Saharan D. melanogaster (Figs 1 and 3). Indeed, when we mutated these SNPs, we were able to recover 100% of this expression divergence in the cosmopolitan background (Fig 3), although we also found evidence that unidentified SNPs in the CG9509 enhancer have epistatic effects on expression in the sub-Saharan background in larvae and the cosmopolitan background in adults (Fig 3). However, these epistatic effects are small relative to the magnitude of expression divergence that could be attributed to the 3 SNPs of major effect. Furthermore, the context-dependent nature of these effects makes it unlikely that they have been targets of positive selection, which acts most efficiently on additive genetic variation [68]. While we assume that the identified SNPs exert their effects on gene expression through interactions with trans-acting factors, the specific trans-acting factors that are involved remain unknown. To identify potential transcription factors that might interact with the identified SNPs, we scanned representative cosmopolitan and sub-Saharan CG9509 enhancer sequences for predicted transcription factor binding sites (TFBSs) [69]. All of the identified SNPs overlapped with at least 1 predicted TFBS, and for each SNP, differential binding (absence or a lower binding score in 1 sequence) was predicted for 2–8 TFBS matrix models (S2 Table). The majority of the identified transcription factors are known to be involved in developmental regulation and morphogenesis, including several forkhead box factors, the Iroquois complex genes, hairy, Distal-less, slow border cells, and twist [70–74]. Several, such as fork head and the Broad-Complex [75,76], are also known to be involved in insulin and/or ecdysone signaling.
It is important to elucidate both the mechanisms behind and the selective forces driving the adaptive divergence of cis-regulatory elements, as these examples help us to understand the genetic basis of phenotypic evolution, which can give further insights about biodiversity. Our results provide evidence that in cosmopolitan populations of D. melanogaster, positive selection has acted on 3 SNPs within the CG9509 enhancer to increase CG9509 expression and thereby reduce wing loading. While 2 of these SNPs appear to be the targets of a completed selective sweep, the third, which has the largest effect on CG9509 expression, has been maintained at intermediate frequency, suggesting that it has been subject to another mode of selection. Using natural variation, a mutant allele, and RNAi, we provide the first experimental evidence of CG9509’s function. We show that its expression influences larval and adult body size, as well as the ratio of wing-to-body size. We propose that the reduced wing loading conferred by elevated CG9509 expression represents an adaptation to improve flight ability as D. melanogaster expanded out of Africa. Because of the remarkable body size increase seen in CG9509del larvae and adults, we propose that the gene be named fezzik (fiz) after the giant character in The Princess Bride.
All flies were maintained as inbred, isofemale lines under standard conditions (22°C, 14 hours light:10 hours dark cycle, cornmeal-molasses medium). The phiX-86Fb stock [77], containing a mapped attP site on the third chromosome (cytological position: 3R 86F), was obtained from the Bloomington Stock Center (Indiana, United States) and used for phiC31 site-specific integration.
In adults, CG9509 expression is highly enriched in the Malpighian tubule, while in larvae, it is enriched in the Malpighian tubule and fat body [62]; therefore, we surveyed expression in whole flies and larvae. Total RNA was extracted from 3–5 adult males (aged 4–6 days) or 1–3 early or late third instar wandering larvae, and a DNAse I digestion was performed using the MasterPure RNA Purification Kit (Epicentre; Madison, Wisconsin, US). Two biological replicates were performed for each line and/or stage. Using random hexamer primers and Superscript III reverse transcriptase (Invitrogen; Carlsbad, California, US), 3 μg total RNA for each replicate was reverse transcribed following the manufacturer’s protocol. TaqMan Gene Expression Assays (Invitrogen; Carlsbad, California, US) were then performed on the resulting cDNA using probes specific to CG9509 (Dm01838873_g1), dFOXO (Dm02140207_g1), dMyc (Dm01843706_m1), and/or E74B (Dm01793592_m1) as well as a probe specific to the ribosomal protein gene RpL32 (Dm02151827_g1), which was used as an endogenous control. The ΔΔCt method was used to calculate normalized gene expression [79]. Briefly, for each biological replicate, the average threshold cycle (Ct) of 2 technical replicates was measured, and ΔCt was calculated as the mean Ct difference between the probe of interest and the RpL32 probe. The fold-change difference in expression relative to the Zimbabwe population for population comparisons or the control lines for CG0509 hypomorph and knockdown comparisons was then calculated as 2–(ΔCtX–ΔCtY), where ΔCtX is the mean ΔCt value for each biological replicate of the line of interest and ΔCtY is the mean ΔCt value of either the Zimbabwe or control lines. Significance was assessed with a t test. When more than 3 comparisons were made using the same data, a Bonferroni multiple test correction was applied.
The CG9509 enhancer region, spanning coordinates 14,909,008–14,910,193 of the X chromosome (release 6), was PCR-amplified from 2 cosmopolitan strains and 1 sub-Saharan strain as described in [12] and cloned into the pCR2.1-TOPO vector (Invitrogen; Carlsbad, California, US). The effects of 6 sub-Saharan sequence variants (positions 67, 765, 821–817, 1063, 1155, and 1174; Fig 2B) in the cosmopolitan background were examined. The sub-Saharan African variants were introduced into the cosmopolitan sequence using either standard cloning techniques or site-directed mutagenesis (S1 Text) [80]. For sites shown to affect reporter gene expression, the cosmopolitan variants were introduced into the sub-Saharan enhancer, and constructs with all contributing sites were generated in both a cosmopolitan and sub-Saharan background using site-directed mutagenesis (S1 Text) for a total of 13 reporter gene constructs. The original and the mutated enhancer sequences were confirmed via sequencing (S1 Text). The Escherichia coli LacZ coding region was then inserted downstream of the CG9509 enhancer sequence, and both were introduced into the pattB integration vector [77] using standard cloning techniques (S1 Text). The pattB vectors containing the CG9509 enhancer and the LacZ reporter gene were microinjected into early-stage embryos of the phiX-86Fb (attP site at cytological band 86F) strain [77], which contains a stable source of phiC31 integrase on the X chromosome. After microinjection, surviving flies were crossed to a white- strain to remove the integrase source, and stable lines homozygous for each of the constructs were established. A subset of the microinjections was performed by Rainbow Transgenic Flies (Camarillo, CA, US).
In adults, CG9509 expression is highly enriched in the Malpighian tubule, while in larvae, it is enriched in the Malpighian tubule and fat body [62], and adult reporter gene in whole flies has been shown to be a good proxy for Malpighian tubule expression [12]; therefore, we surveyed reporter gene expression in whole flies and larvae. For each reporter gene construct, β-galactosidase activity was measured in groups of 15 adult 4–6-day-old males or females or 8 late wandering third instar larvae. Soluble proteins were extracted, and a β-galactosidase activity assay was performed as described in [81] with the following modifications: flies or larvae were frozen with liquid nitrogen and homogenized before the addition of 200 μl of the 0.1 M Tris-HCl, 1 mM EDTA, and 7 mM 2-mercaptoethanol buffer (pH 7.5). β-galactosidase activity was measured spectrophotometrically by following the change in absorbance at 420 nm at 37°C. Four to eight biological replicates were performed per stage or sex. Significance was assessed using a t test, and a Bonferroni multiple test correction was applied for each stage and sex. To better understand the effect of position 67 on reporter gene expression in adults, an ANOVA using sex, background, the variant at position 67, and the interaction between the variant at position 67 and background, sex, and the other tested sites within the CG9509 enhancer was performed.
Wing load index was calculated as the wet weight of a fly divided by the area of its right wing. Flies were lightly anesthetized with CO2 and placed in preweighed 1.5 mL Eppendorf tubes on ice for 5 minutes before being weighed on a Mettler H51 scale (d = 0.01 mg, error = 0.05 mg). The weight of a fly was then calculated as the weight of the fly and tube minus the weight of the tube. For each fly, the right wing (or the left wing if the right wing was damaged) was then dissected, and the wing area was estimated as described above. For each line and sex, wing loading was measured for 5 flies for population comparisons and 10–15 flies for RNAi-CG9509/Act5C-GAL4 and CG9509del lines as well as their respective control lines. For F2 offspring, wing loading was measured for 11–35 flies for each genotype and sex. Significance was assessed using a t test for CG9509del and RNAi comparisons. For population comparisons, significance was assessed using an ANOVA with sex, isofemale line, and population or variant at position 67 as factors. For F2 offspring comparisons, significance was assessed using an ANOVA with sex, cross, and SNP variant(s) as factors. The effect of the SNP variant(s) was assumed to be additive, with cosmopolitan homo- and hemizygotes assigned a value of 2, sub-Saharan homo- and hemizygotes assigned a value of 0, and heterozygotes assigned a value of 1.
To assess larval growth rate, larval volume was measured in the following stages: second instar approximately 48 hours after egg laying (AEL), early third instar (72 hours AEL), early wandering third instar (110 hours AEL), and late wandering third instar (116 hours AEL). Larvae were staged as described in S1 Text. Before imaging, larvae were placed on ice for at least 5 minutes. Larvae were photographed using a Nikon D5100 camera and a compound microscope, and images were analyzed in ImageJ [82]. A piece of millimeter paper was included in all images for scale. Larval volume was calculated as 4/3π(L/2)2(d/2), where L = length and d = diameter [60]. For each stage and line, larval volume was measured in 15–20 larvae for RNAi-CG9509 /Act5C-GAL4 and CG9509del lines as well as their respective controls. Significance at each larval stage was assessed using a t test, and a Bonferroni multiple test correction was applied.
As a measure of developmental timing, the time from the first instar larval stage to pupariation and the duration of the wandering stage were measured. As described in S1 Text, flies were allowed to lay eggs for 12 hours, and first instar larvae were collected. Larvae were transferred in groups of 50 to cornmeal-molasses medium and allowed to mature. In order to measure the duration of the larval stage (L1 to pupariation), pupariation was recorded every 2 hours for 25–110 larvae per line. In order to measure the duration of the wandering stage, larvae were screened for onset of wandering behavior every hour and transferred individually to a petri dish containing moistened filter paper. Pupariation was recorded every hour for 10–50 larvae per line. Both assays were performed at 25°C to prevent fluctuations in developmental timing due to temperature.
DDT, malathion, ethanol, and cold tolerance assays were performed using RNAi-CG9509/Act5C-GAL4 and UAS-/Act5C-GAL4 flies. For DDT, malathion, and ethanol tolerance assays, for each line, sex, and concentration, 6–8 tolerance chambers with 20 flies each were exposed to 4 concentrations of a compound, and mortality was measured as the number of flies dead or unable to move after 30 minutes (malathion), 2 hours (DDT), or 48 hours (ethanol). For ethanol tolerance assays, tolerance chambers consisted of a plastic vial (diameter = 25 mm, height = 95 mm) with compressed cotton at the bottom containing 2.5 ml ethanol solution supplemented with 5% sucrose and sealed with a cork. For DDT and malathion assays, tolerance chambers consisted of glass vials (h = 5 cm, r = 1.65 cm) in which 200 μl of DDT (Dr. Ehrenstorfer; Augsburg, Germany) or malathion (Dr. Ehrenstorfer; Augsburg, Germany) diluted in acetone was swirled until the acetone dried; the vials were allowed to dry an additional hour before addition of flies and were sealed with compressed cotton soaked in 5% sucrose solution. For all assays, 2–3 control chambers containing only 5% sucrose solution were also tested. The data for each assay were fit to a generalized linear model using concentration, line, and sex as factors (unless sex was not significant, in which case it was removed from the model) and a quasibinomial distribution using the glm function in R [83]. For cold tolerance assays, for each line and sex, 25 groups of 5 flies were exposed to an ice water bath for 5 hours, and the time in minutes until each fly had recovered from chill coma (able to stand upright again) was recorded. The mean recovery time for each vial was calculated, and a t test was applied to assess significance.
|
10.1371/journal.pmed.1002217 | Patient Safety Incidents Involving Sick Children in Primary Care in England and Wales: A Mixed Methods Analysis | The UK performs poorly relative to other economically developed countries on numerous indicators of care quality for children. The contribution of iatrogenic harm to these outcomes is unclear. As primary care is the first point of healthcare contact for most children, we sought to investigate the safety of care provided to children in this setting.
We undertook a mixed methods investigation of reports of primary care patient safety incidents involving sick children from England and Wales’ National Reporting and Learning System between 1 January 2005 and 1 December 2013. Two reviewers independently selected relevant incident reports meeting prespecified criteria, and then descriptively analyzed these reports to identify the most frequent and harmful incident types. This was followed by an in-depth thematic analysis of a purposive sample of reports to understand the reasons underpinning incidents. Key candidate areas for strengthening primary care provision and reducing the risks of systems failures were then identified through multidisciplinary discussions.
Of 2,191 safety incidents identified from 2,178 reports, 30% (n = 658) were harmful, including 12 deaths and 41 cases of severe harm. The children involved in these incidents had respiratory conditions (n = 387; 18%), injuries (n = 289; 13%), nonspecific signs and symptoms, e.g., fever (n = 281; 13%), and gastrointestinal or genitourinary conditions (n = 268; 12%), among others. Priority areas for improvement included safer systems for medication provision in community pharmacies; triage processes to enable effective and timely assessment, diagnosis, and referral of acutely sick children attending out-of-hours services; and enhanced communication for robust safety netting between professionals and parents. The main limitations of this study result from underreporting of safety incidents and variable data quality. Our findings therefore require further exploration in longitudinal studies utilizing case review methods.
This study highlights opportunities to reduce iatrogenic harm and avoidable child deaths. Globally, healthcare systems with primary-care-led models of delivery must now examine their existing practices to determine the prevalence and burden of these priority safety issues, and utilize improvement methods to achieve sustainable improvements in care quality.
| Children receive most of their healthcare in the community setting rather than the hospital setting, but very little is known about the safety of this care.
There are signs from previous research that the UK is providing poorer quality pediatric care than its similarly economically developed counterparts.
The purpose of this study was to identify what safety concerns there are involving children in primary care, in order to accelerate and inform improvement efforts.
We analyzed 2,191 reports from a national collection of patient safety incidents that involved sick children in primary care in England and Wales.
Of the incidents included in this study, 30% were reported as harmful.
Medication errors, particularly in the community pharmacy setting, were commonly reported.
Incidents that involved diagnosis, assessment, or referral of sick children were the most harmful of those reported: there were ten deaths, 15 reports of severe harm, and 69 reports of moderate harm.
Poor communication underpinned many of the safety incidents reported as harming children.
It is important to note that our findings are limited by the biased nature of incident report data (not all incidents get reported) and require further studies to confirm them.
However, the frequency with which certain incidents are reported clearly points to areas of care requiring improvement.
Safer and more reliable medication dispensing systems are needed.
Out-of-hours telephone triage systems are not fit for pediatric purpose and require improvement.
Mandatory pediatric training for all general practice trainees is essential.
We hope that this study acts as an impetus for long-overdue widespread improvement efforts in this area.
| The United Kingdom (UK) has one of the highest child mortality rates in Western Europe: the 2,000 excess child deaths that occur annually compare unfavorably with Sweden, which is the best performing country in this region [1–3]. Intercountry variability in rates of child mortality is a well-described global problem. Despite this, there has been a dearth of research on the contribution of unsafe care to these potentially preventable child deaths [4,5].
Primary care is responsible for the majority of healthcare encounters in high-income countries. The safety of care provided to children in this setting is not well understood [6]. For example, in the UK, deaths from meningitis and pneumococcal infection—conditions whose outcomes rely heavily on “first access” services—are considerably higher than in other European countries [3,7,8]. Yet, the avoidable causative factors have not been identified with sufficient clarity for planning action that will prevent the delivery of unsafe care. Furthermore, increasing rates of inappropriate hospital admissions and avoidable referrals to hospital pediatric services indicate that primary care is struggling to meet the demands and changing needs of the pediatric population [7,9–12].
To our knowledge, no systematic approach has been taken to studying the burden of iatrogenic harm in children [4,5,13,14]. Methods that have been used include analysis of vital statistics and case note reviews (some guided by trigger tools) [13–17]. These methods are seldom able to explain why incidents occurred, an essential prerequisite to designing interventions to mitigate future unsafe practice [4]. On the other hand, incident reporting systems can provide detailed descriptions of safety incidents and their underlying contributory factors. Analyses of national repositories of patient safety incident reports have enabled detection and mitigation of rare and serious healthcare safety risks [18–24]. These analyses, in turn, can inform recommendations for clinical process redesign [18–22,25].
This study aimed to characterize the nature and severity of patient safety incidents involving sick children in primary care, to identify potential priority areas requiring action, and to make recommendations for improvement.
The Aneurin Bevan University Health Board research risk review committee waived the need for ethics review given the anonymized nature of the data (ABHB R and D reference number SA/410/13), and we therefore did not require informed consent.
The National Reporting and Learning System (NRLS) is a national repository of voluntarily submitted patient safety incident reports from healthcare organizations in England and Wales. Patient safety incidents are defined as “any unintended or unexpected incidents that could have, or did, lead to harm for one or more patients receiving NHS care” [26]. The NRLS was established in 2003 and is the largest repository of its kind, receiving approximately 65,000 reports of patient safety incidents involving children each year [4].
Healthcare professionals submit reports to their local healthcare organizations, where the reports are first analyzed and anonymized, and then submitted in batches to the NRLS. Reports can also be submitted directly to the NRLS online [26–28]. Each report captures structured categorical information such as patient age, incident location, incident date, and severity of harm outcome (no harm, low harm, moderate harm, severe harm, or death) [26–28]. In addition, each report contains three unstructured free-text fields where reporters can describe what happened, why they think it happened, and how they think it could have been prevented [26–28].
All incident reports submitted to the NRLS between 1 January 2005 and 1 December 2013 from primary care and involving sick children less than 18 y old were included. Primary care refers to generalist care in the community including, but not limited to, care provided by general practitioners (GPs) (or family physicians), community nurses, and community pharmacists. Reports involving sick children were broadly defined as any reports with descriptions of diagnoses, signs, symptoms, or prescribed medications implying acute or chronic illness in a child. Reports involving children were identified through applying an age filter, and reports involving sick children were identified through free-text searches using key terms and related permutations (Fig 1; S1 Text).
A retrospective cross-sectional mixed methods study was conducted. This involved systematically coding data using multiple coding frameworks to describe the incident, quantitatively exploring coded data to identify important patterns, and thematically analyzing a purposive sample of reports containing new theoretical insights. This methodology has been accepted by the international literature [23,25,28].
Each incident report underwent data coding using multi-axial frameworks to describe incident types (primary and contributory), potential contributory factors, incident outcomes, and harm severity (S2–S4 Texts) [23,25,28]. Primary incidents included those proximal (chronologically) to the patient outcome, whereas contributory incidents included those that contributed to the occurrence of another incident. Multiple codes for incident type, contributory factor, and incident outcome were applied to each report where necessary. The codes were applied systematically and chronologically according to nine recursive incident analysis rules developed by the Australian Patient Safety Foundation (S1 Table) [29]. This permitted modeling of the steps preceding and leading to primary incidents, e.g., contributory incidents and factors, which, in turn, resulted in patient outcomes (S1 Fig). The incident type, contributory factor, and incident outcome frameworks were developed in house [28]. Each incident report in the NRLS comes with a reporter-allocated harm severity; however, where the free-text descriptions conflicted with the reporter-allocated harm severity, harm severity was reclassified using WHO International Classification for Patient Safety definitions (see Table 1 for WHO definitions of harm severity) [3,23,25,30]. The medications involved in medication incidents were recorded and classified using the British National Formulary for Children, and the types of conditions affecting these children were recorded and classified using the International Classification of Diseases (ICD-10) (S2 Table) [31,32]. A random 20% sample of reports was independently double-coded by P. R. and H. W.
We undertook exploratory descriptive analysis of coded data [33]. The relationships between codes were explored using frequency distributions and cross-tabulations, to identify prevalent patterns in associated incidents and contributory factors (S3 and S4 Tables) [34]. Priority areas were identified based on the frequency and associated severity of harm. Recommendations for addressing these priority areas were informed by the factors reported as contributing to incidents, by focused searches of the literature, and by consultation with subject matter experts [23,25,28].
A purposive sample of reports that corroborated or contradicted emerging theories or contained “new” insights was identified during data coding [35–37]. These reports were exported for qualitative data analysis (NVivo 9, QSR International), and reread for familiarization. New codes were created to capture additional semantic (descriptive and superficial) insights and latent (underlying or inferred) insights present in reports and the contexts in which incidents occurred [25,35,36]. These codes were grouped into themes and sub-themes (by P. R. and A. C-S.) to support our understanding of the data and the underlying reasons for certain incidents [25,35,36].
Of the 3,636 incident reports potentially involving sick children identified through free-text searches, 2,178 were included; excluded reports involved well children (n = 876), did not describe a patient safety incident (n = 398), or contained insufficient information for coding (n = 184) (Fig 1). Cohen’s kappa (k) statistic of inter-rater (coding) reliability for primary incidents was high, k = 0.72 (95% CI 0.68–0.75; p < 0.01).
The incident reports involved care from the UK national telephone triage service, NHS 111 (formerly NHS Direct) (n = 646; 30%), out-of-hours health centers (n = 604; 28%), community pharmacies (n = 401; 18%), and general practices (n = 218; 10%) (Fig 2). The 2,178 reports described 2,191 primary incidents (hence 2,191 incidents referred to henceforth), largely involving infants between 28 d and 1 y old (n = 491; 22%) and preschool children less than 5 y old (n = 542; 25%). The most frequently described conditions included respiratory conditions (n = 387; 18%), injuries (n = 289; 13%), nonspecific signs and symptoms such as fever (n = 281; 13%), and gastrointestinal or genitourinary conditions (n = 268; 12%) (Table 2). Included reports described harm to 30% (n = 658) of children, including 12 deaths, 41 reports of severe harm, 218 reports of moderate harm, and 387 reports of low harm (Table 1).
Eleven categories of incident types (see Table 1) were evident from included reports. We present a summary of findings related to the priority areas requiring improvement; these include incident types with the highest burden of reported harm in terms of frequency, clinical harm outcomes, and level of harm severity. These priority areas, in descending order of frequency include the unsafe provision of medication, inadequate diagnosis and assessment, and failure of communication with and about the patient (Table 1). Contributory factors for all incidents are summarized in Table 3.
The 674 medication-related incidents (primary and contributory; harmful and nonharmful) were described in the home (e.g., from NHS 111 service calls), general practice, and community pharmacy settings. Most incidents (n = 386; 57%) were related to dispensing errors in community pharmacies; other medication incidents were administration errors (n = 123; 18%) typically in the home setting, prescribing errors (n = 68; 10%) in the general practice setting, and clinical treatment decision-making incidents (n = 66; 10%) in the general practice or out-of-hours setting (Table 1).
Children less than 1 y old were most frequently (n = 131; 19%) involved in reported medication-related incidents, and these children were largely being treated for epilepsy, asthma, and infections (Table 4). As highlighted in Table 4, inhalers for asthma treatment were frequently involved in medication-related incidents: for example, children were dispensed the wrong dose inhaler (n = 27), the wrong brand (n = 18), or the wrong inhaler medication (n = 16). Children with epilepsy were frequently dispensed the wrong dose of anticonvulsant (n = 27) or dispensed anticonvulsants with the wrong instruction labels (n = 11). Errors involving antimicrobial treatment were related to dispensing the wrong dose (n = 13), the wrong medication (n = 22), or medications with incorrect labels (n = 13).
Harm resulted from about one-third (n = 215; 32%) of medication-related incidents, including two deaths, six reports of severe harm, 64 reports of moderate harm, and 143 reports of low harm (Table 1). Incident outcomes included harm necessitating a hospital visit (n = 49), which included admissions to intensive care, e.g., after receiving chlorpromazine rather than chlorphenamine, and deterioration in a child’s condition (n = 21), such as increased seizure frequency after dispensing the wrong brand of lamotrigine. In addition, patient inconvenience was a frequently described incident outcome (n = 108), such as needing to revisit healthcare professionals (n = 52) or experiencing delays in medical management (n = 27), e.g., as a result of being dispensed the wrong medication.
Contributory factors were described for most (n = 242; 63%) dispensing errors. Staff mistakes were described (n = 172), such as confusing medications with similar names or appearances (Examples 1 and 3 in Box 1), e.g., long-acting beta-agonist (LABA) inhalers and LABA/corticosteroid combination inhalers (Example 3). Mistakes occurred in combination with medication factors (n = 39), such as different formulations of the same medication having similar packaging, e.g., beclometasone nasal spray and beclometasone inhalers; organizational factors such as busy or distracting work conditions (n = 28); or both medication factors and poor working conditions (n = 10) (Examples 1, 3, and 4). Other contributing factors included staff failing to follow protocols (n = 31), such as preparing two patients’ medications concurrently, and patient age-specific factors (n = 23) such as weight-based dose calculation errors (Example 5).
Similar contributing factors also underpinned prescribing and administering errors, which often occurred in combination with dispensing errors. For example, most medication administration errors (n = 91; 74%) were described as being the result of other incidents, i.e., contributory incidents, typically other medication errors such as dispensing errors (n = 41), prescribing errors (n = 10), or both (n = 7) (see Examples 1 and 3).
The 659 incidents related to diagnosis, assessment, and referral typically occurred in combination and as a result of each other (S3 Table). These incidents occurred via NHS 111 (n = 400; 61%), during telephone assessments provided by out-of-hours general practice care (n = 158; 24%), or in the general practice setting (n = 55; 8%). The children involved were typically young, under 3 y old, and presented acutely with the following: nonspecific signs and symptoms (n = 150), particularly fever (n = 67) and altered consciousness (n = 51); injuries (n = 146), particularly head injuries (n = 84); and skin or musculoskeletal conditions (n = 87), such as rashes (n = 34) and skin discoloration (n = 33).
Incidents associated with diagnosis, assessment, and referral were the most harmful reported in terms of severity, involving 10 deaths, 15 reports of severe harm, and 69 reports of moderate harm (Table 1). The most frequently described incident outcomes were patient inconvenience (n = 179; 27%), particularly as a result of delayed management of conditions (n = 157; 24%), and clinical patient harm (n = 90; 14%), such as deterioration of a child’s condition (n = 43; 7%). Deterioration outcomes also included four cases of potentially fatal diabetic ketoacidosis.
Diagnosis and assessment incidents mostly involved inadequate triaging (n = 232; 52%) of acutely unwell children and delayed assessment (n = 88; 20%) of these children. Most referral-related incidents (n = 154; 73%) involved assessments over the telephone and in the general practice setting, and included delayed referrals (n = 115; 55%) and failure to refer a sick child for escalation of care or specialist input when appropriate (n = 42; 20%). Incidents contributing to unsafe assessments included the following: inadequate history taking (n = 112; 25%); failing to identify high-risk or vulnerable children (n = 51; 11%), e.g., those with a history of repeated self-harming; and communication failures, such as inadequate safety netting with parents and caregivers (n = 118; 26%). Safety netting is defined within healthcare as providing information (as a safety net) to educate patients, parents, or caregivers and make them aware of when to appropriately seek medical attention in the event of illness, failure to improve, or deterioration medically [38].
Key contributory factors underlying diagnosis, assessment, and referral incidents, particularly those involving inadequate telephone assessments, were related to “protocolized” medicine. Staff failing to follow protocols was frequently described (n = 196; 30%), e.g., GPs were described as failing to follow fever and diabetic management guidelines (Examples 6 and 7 in Box 1; Table 3). In the context of telephone assessments, this included non-clinically trained health advisors choosing the wrong protocol, e.g., selecting a “head wound” protocol rather than a “head injury” protocol, or not using the protocol correctly (Examples 8 and 9). Protocols were also described as inadequate (n = 35; 5%), e.g., when they led health advisors to underestimate the urgency of the child’s condition. In the context of staff failing to follow protocols, or the protocols failing to adequately assess the urgency of a child’s condition, staff were criticized for not using critical thinking (n = 84; 13%; Example 10), despite not having any clinical training.
Of the 177 communication-related incidents reported, 19% (n = 33) were harmful, including two reports of severe harm, 11 reports of moderate harm, and 20 reports of low harm (Table 1). Communication failures with patients, parents, and caregivers were described in a range of primary care settings; however, most communication-related incidents occurred either via NHS 111 (n = 103; 58%) or in out-of-hours settings (n = 39; 22%), and half involved children less than 3 y old (n = 90; 51%).
For sick children in primary care, communication failures (n = 207) were more commonly reported as contributory rather than as primary incidents. Communication failures frequently underpinned medication incidents, particularly administration errors in the home setting, where parents and caregivers are typically responsible for medication administration, which is influenced by prior communication and instructions from healthcare professionals (Example 3). Communication failures were also frequently implicated in diagnosis and assessment incidents (Example 11), e.g., through inadequate safety netting (Example 12), providing the wrong advice, or not clearly communicating the correct advice (Example 13), particularly with regards to fever management in the context of telephone assessments. The most frequent contributory factor (n = 50; 28%) was staff failing to follow protocols, such as those related to safety netting (Table 3).
Based on the burden of incidents in terms of their frequency and severity, and the relative contribution of each incident type to subsequent incidents, the primary-care-related priority areas requiring improvement to reduce iatrogenic harm to sick children are the following: medication provision in the community pharmacy setting; telephone assessment and subsequent referral of acutely unwell children; and communication with patients and their caregivers.
Medication-related safety incidents are widely reported as the most common medical errors, and are thought to be considerably more prevalent in children than in adults [39–42]. Children are more vulnerable to healthcare harm for numerous reasons, such as weight-based dosing; poor availability of certain pediatric formulations, therefore requiring extemporaneous preparation by pharmacists; and dependency on caregivers to advocate for them [5,7,43–45]. Several high-profile reports highlight serious failures in the management of chronic conditions such as asthma and epilepsy in the community setting [2,46–49]. Our study and previous reports highlight that organizational factors (rather than staff knowledge) underpin such failures, suggesting this issue would benefit from quality improvement interventions in healthcare organizations [50–52].
In the UK, children account for 20% of general practice consultations, and 40% of the 500,000 calls received by NHS 111 (formerly NHS Direct) each month [53–56]. Numerous reports in this study criticized telephone assessors for not using critical thinking to challenge inappropriate outcomes reached using clinical decision support (CDS) protocols, arguably due to poor situational awareness. Many have expressed concerns about the safety of telephone assessment of children [53,57–63]. These concerns exist due to the potentially fatal consequences of underestimating the urgency of a child’s condition, the nonspecific nature of many childhood illnesses, the speed with which children deteriorate, and the lack of face-to-face contact, forcing assessors to depend on caregivers to observe the child, interpret those observations, and communicate them effectively [57,59–61,64]. The safety of CDS software used to triage children over the telephone is unclear, particularly its sensitivity to detect signs of serious illness in children [53,60–62,65–68], although its purpose is to minimize risk by standardization and to reduce assessor autonomy—a factor underlying many incidents [61,62].
Despite a study funded by the World Health Organization that echoes our concerns about iatrogenic harm arising from communication failures in primary care, there is a paucity of evaluative studies on this topic, particularly in relation to pediatric telephone assessments [69]. Numerous communication incidents in our study were related to inadequate safety netting during telephone assessment, and this is a well-acknowledged problem in the literature [39,49,70,71]. NHS 111 safety netting protocols have also been described as generic and not child-specific, and there is currently limited evidence to evaluate their role in healthcare-associated harm [38].
This is the first national analysis of patient safety incidents focusing on children and young people in the primary care setting, to our knowledge. Exploring problems in primary care as a whole at a national level, and focusing on the combinations of incidents and contributory factors, provides insights into the interaction of factors between various primary care settings that underlie iatrogenic harm and the subsequent trajectory of harm in this heterogeneous setting.
We sought to achieve methodological rigor through independent double-coding of a random 20% sample of reports, weekly meetings to discuss coding, and keeping an audit trail to aid reflexivity [72,73]. Incident report data are limited by underreporting and variable data quality; thus, our findings are not likely to be generalizable. It is not possible to comment on variation in underreporting between incident types or settings, given the unknown true denominator of patient safety incidents in primary care; therefore, we cannot comment on the relative safety of different healthcare settings. However, it is important to note that incident report data provide a considerable body of granular information on incidents and contributory factors perceived to be important by front-line healthcare professionals and staff [41]. In light of this, given the nature of these data, it would be premature to conclude that medication safety is a bigger problem than diagnostic error, or that the GP’s office is a safer care setting than an out-of-hours health center. Longitudinal studies using case note review methods to assess the frequency and burden of unsafe primary care are required to support such claims.
Our recommendations to improve primary care for children are drawn from the literature and were chosen to ensure they specifically target not only the priority areas identified in our study as requiring improvement but also the specific factors described as contributing to incidents in these priority areas. We corroborated our recommendations with subject matter experts.
Community pharmacy dispensing errors could be reduced through electronic transmission of prescriptions from general practice to the dispensing community pharmacy, as this would prevent errors at the prescriber–dispenser interface [74]. We also recommend implementing a bar coding system for all medications (as is often done in hospital pharmacies), to reduce the potential for human error by acting as an additional safety check prior to medication dispensing [75–77]. Education and training of all pharmacy staff in human factors could enable staff to recognize weaknesses in their own practice [78–83]. In addition, building improvement capability among staff could prove an effective and efficient method of improving patient safety [84].
This study supports the UK Royal College of Paediatrics and Child Health’s call for a robust evaluation of the effectiveness of NHS 111 for children and mandatory pediatric training for all general practice trainees [85]. Monitoring the safety of CDS used to triage sick children is a necessity to target improvement efforts to effectively prevent iatrogenic harm to children. Such improvement may include earlier clinician involvement in the assessment of younger children, who are more difficult to triage safely [68,85]. The outcomes of children assessed using CDS should be reviewed, and the CDS software updated and amended to improve its sensitivity and specificity for this population [86–90]. In addition, CDS could be amended to reduce the potential for certain errors, e.g., reminders when triaging head wounds to double-check the absence of a head injury (which would require triaging with a different protocol).
A lack of critical thinking was described as a contributory factor in many telephone triaging incidents. This is a form of poor situational awareness, with situational awareness referring to sensitivity to operations or “knowing what is going on” [91,92]. Examples of how situational awareness could be improved among telephone triaging staff include human factors training, daily safety huddles to provide feedback on positive and negative cases, and encouraging staff to recognize and act when CDS protocols and their outcomes seem inappropriate [86,92–94]. Increasing situational awareness among telephone triaging staff could—in combination with CDS—increase identification of high-risk children and enable mitigation of risks and appropriate escalation of care.
This study’s findings point to a clear need for improved communication with patients, parents, and caregivers in the context of explaining treatment plans, telephone assessments, and providing safety netting via the telephone. Parents and caregivers should receive oral and written information (perhaps via email, text messaging, or smart phone applications, whichever mode they prefer) regarding treatment plans and for safety netting purposes [38]. This approach is currently being rolled out for epilepsy care in the UK in the form of the Epilepsy Passport. In the context of telephone assessments, adherence to safety netting protocols could be improved through the use of mnemonics or checklists [95–98].
In order to expand on our capability to learn from incident report data, higher quality data are needed from healthcare professionals and staff. This will require them to have an understanding of patient safety and human factors, and training to write incident reports [99]. However, to gain a handle on the frequency and burden of unsafe care in children and target improvement efforts, pediatric safety research must mirror the trajectory of ongoing longitudinal studies into the safety of adult care in hospitals and community settings [100,101].
This study has highlighted opportunities to improve the safety of primary care for children through identifying recurring healthcare failures and commonly reported problems underlying them. Safer, reliable medication dispensing systems, redesigned NHS 111 algorithms that are fit for pediatric purpose, improved situational awareness in triage systems, a deeper understanding of communication failures between parents and primary and secondary care practitioners, and mandatory pediatric training for all general practice trainees are priority areas for redress. Globally, healthcare systems with primary-care-led models of delivery must now examine their existing practices to determine the prevalence and burden of these priority safety issues in care provided to children, in addition to reflecting on our recommendations to address these issues in the context of their own practice.
|
10.1371/journal.pgen.1007418 | Yeast heterochromatin regulators Sir2 and Sir3 act directly at euchromatic DNA replication origins | Most active DNA replication origins are found within euchromatin, while origins within heterochromatin are often inactive or inhibited. In yeast, origin activity within heterochromatin is negatively controlled by the histone H4K16 deacetylase, Sir2, and at some heterochromatic loci also by the nucleosome binding protein, Sir3. The prevailing view has been that direct functions of Sir2 and Sir3 are confined to heterochromatin. However, growth defects in yeast mutants compromised for loading the MCM helicase, such as cdc6-4, are suppressed by deletion of either SIR2 or SIR3. While these and other observations indicate that SIR2,3 can have a negative impact on at least some euchromatic origins, the genomic scale of this effect was unknown. It was also unknown whether this suppression resulted from direct functions of Sir2,3 within euchromatin, or was an indirect effect of their previously established roles within heterochromatin. Using MCM ChIP-Seq, we show that a SIR2 deletion rescued MCM complex loading at ~80% of euchromatic origins in cdc6-4 cells. Therefore, Sir2 exhibited a pervasive effect at the majority of euchromatic origins. Using MNase-H4K16ac ChIP-Seq, we show that origin-adjacent nucleosomes were depleted for H4K16 acetylation in a SIR2-dependent manner in wild type (i.e. CDC6) cells. In addition, we present evidence that both Sir2 and Sir3 bound to nucleosomes adjacent to euchromatic origins. The relative levels of each of these molecular hallmarks of yeast heterochromatin–SIR2-dependent H4K16 hypoacetylation, Sir2, and Sir3 –correlated with how strongly a SIR2 deletion suppressed the MCM loading defect in cdc6-4 cells. Finally, a screen for histone H3 and H4 mutants that could suppress the cdc6-4 growth defect identified amino acids that map to a surface of the nucleosome important for Sir3 binding. We conclude that heterochromatin proteins directly modify the local chromatin environment of euchromatic DNA replication origins.
| When a cell divides, it must copy or “replicate” its DNA. DNA replication starts at chromosomal regions called origins when a collection of replication proteins gains local access to unwind the two DNA strands. Chromosomal DNA is packaged into a protein-DNA complex called chromatin and there are two major structurally and functionally distinct types. Euchromatin allows DNA replication proteins to access origin DNA, while heterochromatin inhibits their access. The prevalent view has been that the heterochromatin proteins required to inhibit origins are confined to heterochromatin. In this study, the conserved heterochromatin proteins, Sir2 and Sir3, were shown to both physically and functionally associate with the majority of origins in euchromatin. This observation raises important questions about the chromosomal targets of heterochromatin proteins, and how and why the majority of origins exist within a potentially repressive chromatin structure.
| In eukaryotic cells, efficient genome duplication requires the function of multiple DNA replication origins distributed over the length of each chromosome [1–5]. Origins are selected by a series of steps in late M- to G1-phase during which the origin recognition complex (ORC) binds directly to DNA and recruits the Cdc6 protein [6–8]. The ORC-Cdc6-DNA complex recruits Cdt1-MCM to form an MCM double hexamer (dhMCM) encircling double-stranded DNA [9,10]. Several kinases and loading factors then remodel the dhMCM into two active CMG helicases (Cdc45-MCM-GINS) that unwind the DNA bidirectionally from each origin to allow the initiation of DNA synthesis [11]. Thus, dhMCM loading is the event that ‘licenses’ the DNA to function as an origin of replication in S-phase (for recent comprehensive review of yeast replication see [12]). All replication origins exist in the context of chromatin, and origin function is significantly affected by local chromatin structure (reviewed in [13,14]). Typically, the most efficient origins (i.e. the origins that are used in most cell cycles) are found in euchromatin, while less efficient origins are associated with heterochromatin. While there is intense interest in the impact of chromatin structure on origin function, the relevant molecular features of origin-adjacent chromatin and the steps in origin function that they control remain incompletely understood.
Yeast heterochromatin is characterized by hypoacetylated nucleosomes and associated heterochromatin regulatory proteins that promote a compact chromatin structure that makes the underlying DNA inaccessible to protein-DNA interactions and processes such as transcription and DNA replication initiation (recently reviewed in [15]). The Sir2 (Silent information regulator) protein, the founding member of a family of conserved NAD-dependent protein deacetylases, removes an acetyl group from lysine 16 of histone H4 (H4K16) and is required for heterochromatin formation in budding yeast [16–19]. Heterochromatin domains form at only a few discrete regions in the yeast genome, rDNA, telomeres, and the HM-mating type loci, and at each of these loci, Sir2 both deacetylates H4K16ac and stably binds to nucleosomes [15]. Heterochromatin formation at the HM loci and telomeres also requires nucleosome binding by the Sir3 and Sir4 proteins.
While it is recognized that the function of normally efficient origins can be suppressed when they are engineered within heterochromatic regions of the genome [20,21], several more recent studies reveal that SIR2 can also affect the function of euchromatic origins [22,23]. However, depending on experimental context, SIR2 can be interpreted to exert either positive or negative effects on euchromatic origins. In particular, recent studies reveal that SIR2 acts indirectly as a positive regulator of euchromatic origins by suppressing the function of rDNA origins [24,25]. The tandemly repeated rDNA gene locus on chromosome XII contains ~200 replication origins, but ~80% of these origins are suppressed by Sir2-dependent heterochromatin [24]. Because the rDNA origins account for >30% of all yeast genomic origins, deletion of SIR2 results in activation of many rDNA origins, which then sequester limiting S-phase origin activation factors from euchromatic origins, delaying their activation time in S-phase [26,27]. Thus, the positive role for SIR2 in regulating euchromatic origin function in these reports is explained as a byproduct of a direct inhibition of rDNA origins by Sir2.
Interestingly, other studies establish that SIR2 can also exert a negative effect at euchromatic origins. This negative role was revealed by a classic genetic screen that isolated sir2 mutants as suppressors of the temperature-sensitive cdc6-4 allele [23,28,29]. Cdc6, a member of the AAA+ protein family, must bind ATP to load dhMCM [30,31]. The cdc6-4 allele encodes a mutant Cdc6 with a lysine to alanine substitution in the conserved ATP binding motif. Cells with the cdc6-4 allele are viable but have S-phase associated growth defects at the permissive temperature and arrest growth at the non-permissive temperature with a failure to load dhMCM, as assessed by origin-specific ChIPs. Like sir2Δ, sir3Δ is also a robust suppressor of the cdc6-4 growth defect, whereas a sir4Δ mutant is only a very weak suppressor [29]. Thus the suppression by sir2Δ or sir3Δ is not easily explained by defects in classic Sir-heterochromatic gene silencing of known loci because Sir2, Sir3 and Sir4 are each equally essential for HM- and telomere silencing and only Sir2 is required for rDNA silencing. In addition, multiple sir2 alleles specifically defective in rDNA silencing do not suppress cdc6-4 ts lethality, indicating that a loss of rDNA silencing is not sufficient to explain SIR2’s negative effect on euchromatic origins [29]. These findings support a distinct, rDNA- and classic heterochromatin-independent function of SIR2 on euchromatic origins. Importantly, a sir2 catalytic mutant or a histone mutant that converts H4K16 into a residue that mimics acetylated H4K16 (H4K16Q) suppresses cdc6-4, indicating that the Sir2 deacetylase function is critical for its negative impact on euchromatic origins [23,29]. However, these data do not address whether the negative role exerted by SIR2 is due to direct roles for Sir2 within euchromatic regions of the genome, or whether SIR2 and SIR3 affect euchromatic origins through a shared mechanism.
In this report we investigated the molecular mechanisms by which SIR2 affected euchromatic origins by performing MCM ChIP-Seq and MNase-H4K16ac ChIP-Seq experiments and also by analyzing several published high-resolution ChIP-Seq experiments [32–34]. Our results provide evidence for a direct role for Sir2 and Sir3 in forming a repressive local chromatin environment around most origins that exist within euchromatin. A sir2Δ mutation was sufficient to restore MCM binding at the majority of euchromatic origins in cdc6-4 cells even at the non-permissive growth temperature for this allele. In wild type cells, nucleosomes immediately adjacent to the majority of euchromatic origins were relatively hypoacetylated on H4K16 compared to non-origin control loci, a behavior unique to this histone acetylation mark. Moreover, this origin-specific H4K16 hypoacetylation was completely dependent on SIR2. In addition, Sir2 and Sir3 were physically associated with euchromatic origins but not non-origin control loci. The levels of these three distinct molecular features of yeast heterochromatin—hypoacetylation of H4K16, Sir2 and Sir3 binding—correlated with how strongly a SIR2 deletion suppressed the MCM loading defect in cdc6-4 cells. Based on these results we propose that the dhMCM loading reaction has evolved to work within a potentially repressive Sir2,3-chromatin environment that forms around most euchromatic origins. Consistent with this model, a screen for suppressors of cdc6-4 temperature-sensitive lethality identified several histone mutants known to disrupt the Sir3-nucleosome binding interface.
To examine the extent to which a SIR2 deletion (sir2Δ) suppresses defects in MCM loading caused by the cdc6-4 allele, we assessed MCM binding by ChIP-Seq in four congenic yeast strains: wild type (SIR2 CDC6); sir2Δ; cdc6-4; and sir2Δ cdc6-4. Because sir2Δ rescues the temperature-sensitive growth defect of cdc6-4 [29], the experiment was performed using chromatin from cells incubated at 37°C, the non-permissive growth temperature for cdc6-4. Cells were arrested in M-phase at the permissive growth temperature (25°C), shifted to 37°C, and then released into G1-phase to allow time for MCM loading ((Fig 1A and S1 Fig). MCM ChIP-Seq was performed for each strain using a monoclonal antibody against Mcm2 [23]. Examination of MCM ChIP-Seq signals across the genome, as illustrated for chromosomes III and VI, revealed that MCM distribution in wild type and sir2Δ cells was similar (Fig 1B and 1C, S2 Fig). Cdc6-4 cells failed to produce MCM ChIP-Seq signals above background levels, consistent with the essential role of Cdc6 in MCM loading (Fig 1B and 1C). In contrast, cdc6-4 sir2Δ mutant cells showed MCM association, albeit to varying degrees, at most origins. (Fig 1B and 1C, S2 Fig). Therefore, deletion of SIR2 rescued origin-specific association of MCM at the majority of chromosomal origins in cdc6-4 mutant yeast, consistent with SIR2 having a pervasive negative effect at the majority of euchromatic origins.
A previous study screened for origins on chromosomes III and VI (Fig 1, gray boxes) that, when cloned onto a plasmid, were functional in cdc6-4 sir2Δ cells growing at the non-permissive temperature for cdc6-4 [23]. Five origins were identified as SIR2-responsive in this screen: ARS317 (the HMR-E silencer origin), ARS305, ARS315, ARS603 and ARS606 (Fig 1B, 1C and 1D, black boxes). ARS317 was not used in subsequent analyses here because it did not produce a robust MCM ChIP-Seq signal in wild type cells, and because we eliminated all origins that exist within heterochromatic domains for deeper analyses of euchromatic origins, as described below. The plasmid-based study suggested that at least 20% of yeast origins were likely to be SIR2-responsive [23]. In contrast, the MCM-ChIP-Seq revealed that ~80% of origins showed MCM binding in cdc6-4 sir2Δ cells (Fig 1B, 1C and S2 Fig). We note that the plasmid-based assay demanded that origin function was rescued to a level in cdc6-4 sir2Δ cells that allowed for colonies to form at 37°C, a selection that might have been more stringent than the screen for an MCM ChIP-Seq signal as in Fig 1. In this regard, it is notable that the origins identified as SIR2-responsive in the plasmid screen were indeed among those that showed the most robust rescue of MCM ChIP-Seq signals in cdc6-4 sir2Δ cells (Fig 1D, black boxes). In addition, the plasmid-based screen assessed only small origin fragments that might not have recapitulated a nucleosome organization required to show SIR2-responsiveness. Indeed, plasmid-born ARS1005 (a top origin identified by MCM ChIP-Seq) exhibited SIR2-regulation only when a larger chromosomal region surrounding ARS1005 was present that could accommodate the adjacent chromosomally directed nucleosomes (S3 Fig).
In yeast, each rDNA repeat contains a single rDNA origin, and these repeats are present in hundreds of tandem copies per cell on chromosome XII (Fig 2A). Some mutants that affect origin function, such as orc2-1, can be suppressed by reducing the rDNA origin-load [35,36]. Therefore, we tested whether reduced rDNA copy number might account for suppression of cdc6-4 by sir2Δ by analyzing the rDNA locus in the strains used for the MCM ChIP-Seq experiment by qPCR (Fig 2B). While rDNA copy number varied from colony to colony even within a single strain, on average, wild type, sir2Δ and cdc6-4 cells had similar levels of rDNA (Fig 2C). However, cdc6-4 sir2Δ cells showed a 2-fold reduction in rDNA levels relative to cdc6-4 cells (P = 0.02), raising the possibility that suppression of cdc6-4 by sir2Δ was mediated, at least in part, by reductions in rDNA copy number (Fig 2C). To further address this issue, we performed two additional experiments. First, we exploited the previous observation that sir3Δ also suppresses cdc6-4 [29]. In contrast to SIR2, SIR3 has no role in rDNA silencing [37,38]. Cdc6-4 sir3Δ cells grew at the non-permissive temperature for cdc6-4, as expected [29] (Fig 2D). However, the rDNA copy number of cdc6-4 sir3Δ yeast was slightly higher compared to cdc6-4 yeast, indicating that reduced rDNA copy number could not explain this suppression (Fig 2E). Second, we asked whether a reduction in rDNA copy number was sufficient to suppress cdc6-4 by generating cdc6-4 cells with ~35 copies of the rDNA locus (rDNA-35) (Fig 2D and S4 Fig). These cells also contained a FOB1 deletion to help maintain rDNA copy number [39,40]. The cdc6-4 rDNA-35 fob1Δ cells failed to grow at the non-permissive temperature and grew no better than the cdc6-4 rDNA-180 fob1Δ cells, which harbor ~180 copies of the rDNA locus (S4 Fig). Thus, reducing rDNA copy number was neither necessary nor sufficient to explain SIR-mediated suppression of cdc6-4. Cells harboring a reduction in rDNA copy number in cdc6-4 sir2Δ populations likely arose because loss of Sir2 increases recombination frequency within the rDNA array, and reduced rDNA copy number is selected for over passaging in many mutants compromised for replication [35,41,42].
The pervasive yet origin-specific rescue of the MCM loading defect in cdc6-4 cells by sir2Δ prompted us to consider the possibility that Sir2 might function directly on origin-adjacent nucleosomes within euchromatin. Sir2 is a deacetylase with specificity for acetylated lysine 16 of histone H4 (H4K16ac) [19]. Therefore, we used a comprehensive genome-wide histone modification atlas generated by high resolution MNase-ChIP-Seq of yeast nucleosomes to examine the acetylation status of nucleosomes adjacent to euchromatic origins [32] (Fig 3). To perform this analyses, we focused on two distinct groups of euchromatic loci: (i) experimentally-confirmed origins and (ii) non-origin intergenic regions that contain a match to the ORC binding site [43]. At confirmed origins, ORC and MCM binding, as well as origin activity are experimentally documented, while at non-origin intergenic regions with ORC site matches, ORC and MCM binding are not detectable in vivo, and no origin function has been detected (Fig 3A non-origins n = 179; origins; n = 259). These non-origin intergenic regions with ORC site matches serve as controls for origin-specific as opposed to primarily sequence-directed chromatin signatures [43]. 1201 bp fragments from these two distinct groups were aligned with the T-rich strand of the ORC site on the top strand in the 5’ to 3’ orientation, so that three nucleosomes positioned on either site of the origin could be examined (Fig 3B). Previous genome-scale studies have not identified Sir binding or function at euchromatic origins [34,44,45]. Thus, we expected that if SIR2 was acting on nucleosomes adjacent to euchromatic origins, its effect would be weak compared to its effect on nucleosomes within known SIR2-heterochromatic regions. Therefore, to avoid H4K16ac depletion within known silent chromatin regions masking signals from weaker but potentially physiologically relevant Sir effects on euchromatin, the HM silencer, telomeric (as defined by origins within 15 kb of chromosome ends) and rDNA origins were excluded from these analyses. We then used the average nucleosome occupancy data from this recent nucleosome modification study [32] to confirm that we could recapitulate previously published results about nucleosome occupancy and positioning differences between origin and non-origin loci [43,46]. The results confirmed the conclusion that origin-adjacent nucleosomes show high occupancy at more defined positions around the nucleosome-depleted origins compared to the control non-origin ORC-site containing control loci (Fig 3C). Next, the relative H4K16ac status of nucleosomes surrounding the ORC site for both groups was determined and normalized to the average H4K16ac status from a collection of nucleosomes present in a distinct collection of euchromatin intergenic regions that contained neither origins nor matches to the ORC site (n = 239). This analysis revealed that nucleosomes adjacent to origins were relatively depleted for H4K16ac but that nucleosomes adjacent to the control non-origin nucleosomes were not (Fig 3D). H3K9ac is also a potential substrate for Sir2 but, an H3K9Q substitution fails to suppress cdc6-4 [19,23]. In contrast to H4K16ac, H3K9ac was depleted similarly from nucleosomes adjacent to the origin and non-origin ORC-site control loci (Fig 3D). In fact, H4K16ac was distinct among nucleosome acetylation marks in showing origin-specific depletion (S5 Fig).
If depletion of H4K16ac on origin-adjacent nucleosomes was relevant to SIR2-dependent inhibition of MCM loading in cdc6-4 cells, then the origins most responsive to deletion of SIR2 might be expected to show the greatest depletion of H4K16ac. We defined origin SIR2-responsiveness as the ratio of the MCM ChIP-Seq signal in cdc6-4 sir2Δ cells to that in sir2Δ cells; the most SIR2-responsive origins generated ratios near 1.0, suggesting that deletion of SIR2 substantially rescued the MCM loading defect of cdc6-4 (Fig 3E). Comparison of H4K16ac status of nucleosomes surrounding the low, medium, and high-SIR2 responsive origin quintiles revealed that as a group the most SIR2-responsive origins exhibited the greatest depletion of H416ac (Fig 3F). Thus the varying degrees of SIR2-responsiveness to MCM association generally correlated well with the degree of H4K16 hypoacetylation, with the exception of the -1 nucleosome for the low and medium SIR2-responsive origins.
In a separate analysis, H4K16ac status of nucleosomes adjacent to euchromatic origins was also examined relative to the genome-wide average level of H4K16ac so that a comparison to origins within established SIR-heterochromatin domains could be made (S6 Fig). This analysis, presented as box-and-whiskers plots to show the variation in H4K16ac at the relevant nucleosomes in each group, also indicated that the most prominent depletion of H4K16ac occurred on the -1 and +1 nucleosome positions for euchromatic origins, and weakened considerably by the -3 and +3 nucleosomes, as also seen in Fig 3F. As expected, the -1 and +1 nucleosomes of euchromatic origins were less depleted for H4K16ac than the analogous nucleosomes for origins within SIR-heterochromatin, with the median values for the two types of origins differing ~3-fold. The differences between H4K16ac status were more striking between the two types of origins (i.e. euchromatic versus SIR-heterochromatic) at nucleosomes more distal from the origin, consistent with the H4K16ac status at euchromatic origins being more localized than the H4K16ac status within SIR-heterochromatin. Analyses of H4K16ac status at a few individual origins also indicated that the -1 and/or +1 nucleosomes were the most likely to show the greatest depletion of H4K16ac, particularly for the most SIR2-responsive origins. (S7 Fig).
The data described above raised the possibility that Sir2 was deacetylating H4K16ac nucleosomes adjacent to euchromatic origins. To test this possibility, we performed H4K16ac MNase ChIP-Seq experiments on the exact same wild type (SIR2) and sir2Δ cells used for the MCM ChIP-Seq experiment described in Fig 1 (Fig 4A). Analyses of wild type cells recapitulated the published H4K16ac MNase ChIP-Seq results shown in Fig 3D. We note that the relative level of origin-specific H4K16ac depletion was slightly greater in these new experiments, possibly due to differences in strain backgrounds or growth conditions. Regardless, the key result was that, in contrast to wild type cells, depletion of H4K16ac was lost from origin-adjacent nucleosomes in the sir2Δ cells, while the behavior of non-origin ORC site control nucleosomes was unchanged. Importantly, these effects also correlated with SIR2 responsiveness (Fig 4B). Thus, origin-specific depletion of H4K16ac on nucleosomes adjacent to euchromatic origins required SIR2.
Sir2 and Sir3 are physical components of yeast heterochromatin that have been detected at rDNA (Sir2), telomeres and HM loci (Sir2 and Sir3) by ChIP experiments in multiple studies (reviewed in [15]). However, neither protein has been reported to associate with euchromatic origins [44,45,47]. Depletion of H4K16ac at euchromatic origins was estimated to be between 25–35% of that detected at heterochromatic origins for the +1 nucleosome (S6 Fig). This observation is consistent with a proteomic analysis of nucleosome modifications on a plasmid based origin [48]. However, given that Sir2 is an enzyme, stable binding by Sir2, or even Sir3 if it is required only transiently, might not be required to establish or maintain this modification state. To test this possibility, we used more recently published high-resolution ChIP-Seq Sir2 and Sir3 data sets and excluded all origins from known heterochromatin origins, as described above for Figs 3 and 4 [34,44]. This analysis detected Sir2 and Sir3 ChIP-Seq signals on nucleosomes adjacent to origins but not on nucleosomes adjacent to non-origin controls (Fig 5A and 5B). Because the Sir3 data was generated from a high-resolution MNase ChIP-Seq experiment, we also examined these data at nucleotide resolution across the 1201 bp origin and non-origin fragments described in Fig 3B. To generate the baseline for these analyses, the number of reads that contained a given nucleotide in the ChIP sample was divided by the number of reads that contained that nucleotide in the starting material, for every single nucleotide within the euchromatic genome (i.e. nucleotides from known heterochromatin regions were excluded to form the “0” binding baseline for Sir3), following the method in [49] (Fig 5C). These analyses revealed the strongest Sir3 ChIP-Seq signal at the most proximal origin-adjacent nucleosomes (-1 and +1), but signals above baseline were also detectable at the -3, -2, +2 and +3 nucleosomes. As was true for the depletion of H4K16ac, the Sir3 signals correlated with the degree of SIR2-responsiveness as defined in Fig 3E (except in the opposite direction) (Fig 5D and 5E). Randomizing the origins into three new groups of equivalent numbers prior to plotting the Sir3 ChIP-Seq signals eliminated the correlation (Fig 5F).
A separate analyses of Sir3-3xHA ChIP-Seq data from the same study indicated that excluding nucleotides from SIR-heterochromatin when establishing the baseline was required for the Sir3 ChIP-Seq signals at euchromatic origins to rise above the “0” baseline [50] (S8 Fig), in contrast to what was observed for H4K16ac (S6 Fig). Nevertheless, in this independent analysis, Sir3-3xHA (detected with anti-HA) signals at euchromatic origins were clearly greater at the -1 and +1 nucleosomes of origins relative to the comparable nucleosomes at the non-origin ORC-site containing loci that have been used as controls for origin specificity throughout this study, and Sir3-3xHA signals also correlated with SIR2-responsiveness (S8B and S8C Fig).We interpret these results to mean that strong Sir3 association with heterochromatic regions reduced the “0” baseline and obscured comparatively weak Sir3 signals at euchromatic origins. Thus, we conclude that Sir3 associated with euchromatic origins but more transiently and at levels less than Sir3’s association with SIR-heterochromatin regions [50].
In summary, several molecular hallmarks of SIR-heterochromatin–Sir2, Sir3 and Sir2-dependent H4K16 hypoacetylation–could be detected at euchromatin origins but not at euchromatic non-origin controls. In addition, the relative levels of each of these molecular hallmarks correlated with how strongly a SIR2 deletion restored an MCM ChIP signal to origins in cdc6-4 cells grown at temperatures that otherwise abolished MCM loading.
Histones H3 and H4 form a tetramer that binds both to dsDNA and to two dimers of histones H2A and H2B to form the nucleosome. Many conserved residues on the histone H3 and H4 N-terminal tails as well as within the globular core region (“core-modifiable” residues) can be post-translationally modified, and, as discussed above, an H4K16Q substitution that should mimic the acetylated form of this residue (i.e. the form predicted to phenocopy the effect of a sir2Δ), suppressed the temperature-sensitive growth defect of cdc6-4 [23]. To test whether we could identify additional alleles with similar suppressive behavior, we assessed a library of 43 histone H3 and H4 mutations affecting 15 distinct residues for suppression of the cdc6-4 growth defect (Fig 6A). Mutant plasmids harboring a single copy of the HHT2-HHF2 locus marked with TRP1 were transformed into wild type and cdc6-4 cells lacking both chromosomal copies of histone H3 and H4 genes, maintained by a HHT2-HHF2 URA3 ARS CEN plasmid, and colonies were subsequently plated on 5-FOA containing media to select against the wild type HHT2-HHF2 plasmid. Recovered cells were then examined for growth on YPD medium at 25°C and higher temperatures. In agreement with previous studies, none of the histone mutants, with the exception of H4-S47E, caused growth defects in wild type cells [51], and several mutants were identified that inhibited growth of cells harboring the cdc6-4 allele even at the permissive growth temperature (25°C). However, relevant to this study, specific substitutions at only two residues, H3K79 (H3K79A and H3K79Q) and H4K79 (H4K79A and H4K79R) suppressed the temperature-sensitive growth defect of cdc6-4 cells (Fig 6B and S9 Fig). Notably, these residues are important for Sir3 binding to the nucleosome and for Sir3-mediated transcriptional silencing at the HM and telomeric loci [52–54]. H3K79 is a substrate for methylation by Dot1 and Sir3 binds preferentially to unmethylated H3K79 [55,56]. Based on these results, additional substitutions were engineered by site-directed mutagenesis at H3K79 and at four adjacent residues (H3K79E, H3E73A, H3E73K, H374A, H3T80A, H3D81A) (Fig 6B). In addition to revealing that H3K79E was a stronger suppressor of cdc6-4 than H3K79Q, analyses of these mutants also revealed that substitutions of H3E73 and H3T80 also suppressed the temperature-sensitive growth defect of cdc6-4 cells (Fig 6B and S9 Fig). H3E73 and H3T80 are important for HM and telomeric silencing [52], and H3T80 directly contacts the LRS domain of Sir3 [53,54]. In summary, suppression of cdc6-4 temperature-sensitivity was achieved by single substitutions within only four residues (H3E73, H3K79, H3T80, and H4K79) of H3/H4. Each of these residues clustered within a small patch on the nucleosome surface important for binding Sir3 (Fig 6C, highlighted in red) [53], strongly suggesting that disruption of Sir3 binding to nucleosomes suppressed the cdc6-4 growth defect. Together with the ChIP-Seq data, these genetic data supported the conclusion that both Sir2 and Sir3 alter the molecular features of nucleosomes adjacent to euchromatic origins, thus generating a local chromatin environment that can negatively impact the MCM loading reaction.
Sir2 and Sir3 are non-essential proteins known best for their direct functions in heterochromatin-mediated transcriptional gene silencing in budding yeast (reviewed in [15]). This report described a new, pervasive and direct heterochromatin-independent role for these regulators at euchromatic DNA replication origins. The role was pervasive in that Sir2 exerted a negative effect on the MCM loading reaction at most origins in the genome. The evidence in support of this conclusion came from a comparison of MCM loading at origins by ChIP-Seq in cdc6-4 mutant cells and cdc6-4 sir2Δ cells. Cdc6 is essential for the MCM complex loading reaction, and the cdc6-4 allele weakens Cdc6 function such that, under non-permissive growth temperatures, detectable MCM association was abolished from all origins. However, under the same conditions, cdc6-4 sir2Δ cells exhibited MCM association at most euchromatic origins (~80%). Thus, the ability of sir2 inactivating mutations to robustly suppress the temperature-sensitive lethality of cdc6-4, as well as growth defects caused by other replication alleles that reduce or abolish MCM loading [29], is due to genome-scale enhancement of MCM loading at most euchromatic origins.
The ability of sir2Δ to enhance MCM complex loading at most euchromatic origins in cdc6-4 cells was striking, but it did not indicate whether this effect was due to Sir2 acting directly within euchromatin in general or at euchromatic origins in particular. In particular, recent reports establish that SIR2 can alter origin function within euchromatin in yeast indirectly because of its function in rDNA heterochromatin formation that suppresses many of the rDNA repeat origins [22,25]. However, analyses of the rDNA locus presented here suggested that a similar rDNA-mediated mechanism could not explain why sir2Δ or sir3Δ suppressed cdc6-4. Instead, the data provided evidence that Sir2 functioned directly at euchromatic origins, and that it was this direct function of Sir2 that made yeast so vulnerable to defects in the MCM complex loading reaction caused by cdc6-4. First, relative hypoacetylation of H4K16 was observed for nucleosomes immediately adjacent to euchromatic origins but not for the analogous nucleosomes at non-origin ORC-site containing control loci. Thus, in an asynchronous population of wild type cells, nucleosomes flanking euchromatic origins were relatively hypoacteylated on H4K16. Of the twelve histone acetylation marks examined [32], H4K16ac was the only one that showed this origin-specific depletion. Second, origin-adjacent nucleosomal H4K16 hypoacetylation required SIR2; a sir2Δ mutant completely lost the nucleosomal H4K16 hypoacetylation pattern at euchromatic origins, without having any obvious effect on nucleosomes adjacent to the non-origin control loci. Third, analyses of previously published high-resolution ChIP-Seq experiments [34,50] provided evidence that Sir2 and Sir3 physically associated with nucleosomes adjacent to euchromatic origins but not non-origin control loci. Sir2 and Sir3 exhibited the same pattern but in the exact opposite direction of H4K16ac, which is what is observed for these proteins and this histone mark at SIR-heterochromatic domains. Lastly, each of these defining molecular features of classic SIR-heterochromatin correlated with SIR2-responsiveness–defined as the relative extent of MCM association rescue at origins in cdc6-4 sir2Δ cells. Because of these observations, and in particular the clear depletion of ChIP-Seq signals around origins generated by multiple histone antibodies, we think it is unlikely that these outcomes result from the non-specific “hyper-ChIPability” that affects certain highly transcribed regions [45,57]. Therefore, to the best of our knowledge, these data provide the first evidence that the heterochromatin proteins Sir2 and Sir3 act directly on the local chromatin environment of euchromatic origins.
Despite the common molecular hallmarks shared between the Sir2-3-chromatin at euchromatic origins described in this study and classic SIR-mediated heterochromatin, these two types of chromatin domains are different. For example, unlike SIR-heterochromatin that functions in gene silencing at the HM loci, telomeres or rDNA, there is no evidence that origin-adjacent chromatin functions as a robust transcriptionally silenced domain [58], although we cannot rule out the possibility that Sir2,3-chromatin is having small effects on the transcription of annotated genes near origins or on the expression of non-coding RNAs in these regions. In addition, SIR-heterochromatin affects multiple nucleosomes, defining domains that encompass ~4–10 kb of contiguous chromosomal DNA (reviewed in [15]), whereas the Sir2-3-chromatin at origins encompassed <1 kb of chromosomal DNA adjacent to the origin and affected only four to six nucleosomes at most. While it remains to be determined whether the specific molecular interactions that recruit Sir2 and Sir3 to euchromatic origins are also used to recruit these proteins to heterochromatic regions, their functional and structural impact on nucleosomes at euchromatic origins was substantially attenuated relative to their impact within heterochromatin. Therefore, we suggest that the Sir2-3 chromatin at origins defines a role for these proteins that differs from their roles in canonical Sir-heterochromatin.
Acetylation of histone H3 and H4 N-termini at origin proximal-nucleosomes generally enhances origin function [20,59,60]. However, it has been difficult to assign specific roles of this acetylation to individual histone H3 or H4 lysines. Indeed, in terms of origin activation, while acetylation of histone H3 and H4 tail lysines is clearly important for origin activation during S-phase, no single lysine acetylation event is sufficient [48]. In addition, several different combinations of multiple lysine to arginine substitutions on the histone H3 and H4 N-terminal tails can substantially reduce origin activation, suggesting that some threshold level of nucleosome acetylation, or the concomitant charge neutralization, is what is important for origin activation. In contrast, the initial identification of a sir2 mutation, and subsequently a H4K16Q substitution, as robust suppressors of cdc6-4 temperature-sensitivity and origin-specific MCM loading, indicated that H4K16 might be unique among histone tail lysines in being particularly relevant to origin licensing [23,29]. This study revealed that H4K16 was also unique among histone H3 and H4 tail lysines because H4K16 exhibited SIR2-dependent, origin-specific hypoacetylation on nucleosomes adjacent to euchromatic origins. Consistent with these observations, proteomic analysis of nucleosomes adjacent to a plasmid-origin reveals that H4K16 behaves uniquely among histone H3 and H4 tail acetylation marks [48]. Specifically, in G1-phase, nucleosomes adjacent to a plasmid-based origin were relatively hypoacetylated at H4K16 compared to bulk nucleosomes, whereas the other histone tail lysines analyzed (H3-K9, -K14, K-23; H4-K5, -K8, -K12) showed similar levels of hypoacetylation on bulk and plasmid-based origin adjacent nucleosomes. Because Sir2 and Sir3 also showed origin-specific association with nucleosomes, the distinctive role for H4K16 acetylation in control of the G1-phase MCM loading reaction can now be explained by formation of a Sir2,3 chromatin structure directly at many euchromatic origins, suggesting that origins, perhaps via ORC, can specifically recruit Sir2,3.
The data presented here provided evidence that the MCM loading reaction has evolved to work within a naturally repressive chromatin environment established by Sir2,3. Indeed, an otherwise severely crippled licensing reaction caused by the cdc6-4 defect can quite effectively license most origins if the native repressive chromatin environment is abolished by inactivation of Sir2. While the precise mechanistic step(s) of the MCM complex loading reaction that are affected by the Sir2,3-repressive chromatin environment remain to be examined, recent successes in reconstituting DNA replication in chromatin contexts in vitro offer an ideal pathway forward [61–63]. For example, while reconstituted chromatin does not block MCM complex loading in vitro, it would be interesting to learn whether the addition of Sir3 to these biochemical reactions is sufficient for inhibition and, if so, requires Sir3’s ability to bind to nucleosomes [53].
However, in terms of yeast physiology, the major challenge moving forward is to understand why such a chromatin-structure capable of inhibiting MCM complex loading may have evolved to exist at so many euchromatic origins in the first place. While a sir2Δ or a sir3Δ have no obvious genome-scale effect on yeast cell division, MCM complex loading or origin activation in cells with wild type MCM complex loading reaction components, the Sir2,3-chromatin structure discussed here does indeed exist in wild type cells, and, based on the strong genetic suppression of cdc6-4 and other alleles that weaken MCM complex loading [29], is clearly capable of exerting a substantial inhibitory effect on this reaction. Indeed, the results presented here suggest that the levels and or activities of the MCM complex loading proteins have evolved to contend with a particularly inhibitory local chromatin structure around euchromatic origins; a temperature-sensitive allele that weakened MCM complex loading, such as cdc6-4, would never have been isolated in yeast cells lacking Sir2 or Sir3 as the mutant Cdc6-4 protein provides for robust function in such contexts. It is possible that the Sir2,3- chromatin at euchromatic origins is simply a byproduct of yeast cells evolving transcriptionally silenced heterochromatic domains, and, as a result of strong evolutionary selection for these regions’ existence, combined with fundamental protein dynamics, Sir2 and Sir3 simply fall off of heterochromatin at some frequency and, when they do, end up concentrating at origins as they stochastically bounce around the nucleus [64–67]. Sir2 and Sir3 may concentrate at origins over other euchromatic elements because of the intrinsic ability of origins to act as silencers, as suggested by some studies [68,69]. Regardless, the outcome is that the MCM complex loading reaction had to evolve more robustly than it would have otherwise. Additionally, and/or alternatively, the Sir2,3-chromatin at euchromatic origins may play an important role in normal yeast physiology, offering a fine-tuning of the MCM complex loading reaction at many individual origins that, while at the first-level of cell growth and origin function in bulk laboratory assays may be hard to measure, nevertheless has important repercussions at the population and/or evolutionary scale for yeast and other eukaryotic organisms. Interestingly, over two decades ago a study reported how over-expression of Sir2 and Sir3 promoted severe chromosome instability and yeast cell lethality [70]. Perhaps these effects were related, at least in part, to Sir2,3 effects on origins. Finally, a role for Sir2 in origin control may indeed be conserved as human SIRT1, the ortholog of yeast Sir2, suppresses the function of dormant origins under conditions of replication stress [71].
The strains used for plasmid stability measurements and Mcm2 ChIP-Seq were the W303-1A derivatives: M138 (MATa), M386 (MATa cdc6-4::LEU2), M922 (MATa cdc6-4::LEU2 sir2Δ::TRP1) (described in Pappas et al. 2004), M1014 (MATa cdc6-4) and M2126 (MATa sir2Δ::TRP1). The M1321 (WT) and M1345 (cdc6-4) histone shuffle strains were described previously (Crampton et al., 2008). The strains used for genetic suppression and/or rDNA copy number experiments (Fig 2) include, in addition to M138, M386, M922 and M2126 described above, CFY43 (MATa FOB1 CDC6 SIR2 sir3Δ::TRP1), CFY4584 (MATa fob1 Δ::HIS3 CDC6 SIR2 SIR3 rDNA-35), CFY4585 (MATa fob1 Δ::HIS3 CDC6 SIR2 SIR3 rDNA-180), CFY4603 (MATa fob1 Δ::HIS3 cdc6-4 SIR2 SIR3 rDNA-35), CFY4604 (MATa fob1 Δ::HIS3 cdc6-4 SIR2 SIR3 rDNA-180), CFY4613 (MATa FOB1 cdc6-4 SIR2 sir3Δ::TRP1). CFY4583 and CFY4585 were provided by Jonathan Houseley (Babraham Institute, Cambridge, UK) [40].
For the MCM ChIP-Seq experiments, yeast cells were grown in liquid YPD at 25°C from single colonies until they reached an A600 of 0.2, at which point nocodazole was added to a final concentration of 15ug/ml and the cultures incubated for 2.5 hours, to obtain a uniform G2/M-phase arrest (S1 Fig). Cultures were shifted to 37°C for 30 minutes, the non-permissive growth temperature for cdc6-4, and then released from the nocodazole arrest at 37°C. When the majority of cells had entered G1-phase (55-minutes post-release for CDC6 SIR2, sir2Δ and cdc6-4 sir2Δ and 110 minutes for cdc6-4), OD equivalents of each cell line were cross-linked with formaldehyde for 15 minutes and then harvested, and chromatin was prepared for IP, by sonication for ChIP-Seq. Mcm2 ChIP was performed using a monoclonal antibody raised against yeast Mcm2 (gift of Bruce Stillman, Cold Spring Harbor Laboratory). Three independent biological replicates and two technical replicates for each biological replicate were performed for each strain. The ChIP DNA was prepared for deep sequencing by the UW-Madison Biotechnology Center using http://www.biooscientific.com/Portals/0/Manuals/NGS/5143-01-NEXTflex-ChIP-Seq-Kit.pdf. Sequencing was done on the HiSeq2000 1x100. Quality and quantity of the finished libraries were assessed using an Agilent DNA1000 chip and Qubit dsDNA HS Assay Kit, respectively. Libraries were standardized to 2nM. Cluster generation was performed using standard Cluster Kits and the Illumina Cluster Station. Single-end, 100bp sequencing was performed, using standard SBS chemistry on an Illumina HiSeq2000 sequencer. Images were analyzed using the standard Illumina Pipeline, version 1.8.2. Sequencing data for this project are available at the NCBI BioProject ID PRJNA428768.
For the downstream analyses (Fig 1 and S2 Fig), all data for a given strain were combined as each replicate generated virtually identical reads. However, for the cdc6-4 strain, which produced extremely low signal-to-noise data, only two technical replicates from a single biological sample were used. For downstream analyses using MochiView the CG1 format data was collated into 25 bp bins (Fig 1). The combined data from each strain was run through CisGenome to identify the top 1000 peaks, divided into 100 peak bins and the percent of peaks that overlapped with at least one ARS (ARS coordinates used from oriDB) determined. The oriDB lists 410 yeast origins as confirmed (http://cerevisiae.oridb.org/). Based on these data and the above analyses we chose to examine and compare the top 400 peaks from each sample for our analyses (S2 Fig). The raw data for CDC6 SIR2, sir2Δ and cdc6-4 sir2Δ were normalized and scaled to the same cdc6-4 data for to generate the final MCM signal values for downstream analyses.
The H4K16-acetylation MNase-ChIP DNA was generated from yeast strains M138 (CDC6 SIR2 strain used in Fig 1) and M2126 (CDC6 sir2Δ strain used in Fig 1) following the protocol described in [32], except a Bio101 Thermo FastPrep FP120 was used to break cells after crosslinking. Purified DNA was submitted to the University of Wisconsin-Madison Biotechnology Center. DNA concentration and sizing were verified using the Qubit dsDNA HS Assay Kit (Invitrogen, Carlsbad, California, USA) and Agilent DNA HS chip (Agilent Technologies, Inc., Santa Clara, CA, USA), respectively. Samples were prepared according the TruSeq ChIP Sample Preparation kit (Illumina Inc., San Diego, California, USA) with minor modifications. Libraries were size selected for an average insert size of 350 bp using SPRI-based bead selection. Quality and quantity of the finished libraries were assessed using an Agilent DNA1000 chip and Qubit dsDNA HS Assay Kit, respectively. Libraries were standardized to 2nM. Paired end 300bp sequencing was performed, using SBS chemistry (v3) on an Illumina MiSeq sequencer. Images were analyzed using the standard Illumina Pipeline, version 1.8.2. Sequencing data for this project are available at the NCBI BioProject ID PRJNA428768. For the data shown in Fig 4, downstream analyses were performed as described (Weiner et al., 2015). For the data presented as histograms relative to a baseline value derived from 239 euchromatic intergenic regions lacking either origins or ORC-site matches (Figs 3D, 3F, 4A, 4B, 5A, 5B, 5E and S5 and S7 Figs), the average signal for six contiguous nucleosomes in these regions was used. Thus each nucleosome assessed was normalized to the same value, and the log2 value of this ratio was plotted.
Quantitative PCR reactions were carried out in sealed 200 ul microplates in a BioRad C1000 Thermocycler, CFX96 Real-Time System. The conditions were: 95°C-3’|95°C -30”, 58°C -15”|x40 cycles Standard Melt Curve Analysis. Each reaction contained: 0.04 ng/uL genomic DNA template, 0.016 nM of each primer, and 1X qPCR master mix (0.1 mM dNTPs, 1.25% formamide, 0.1 mg/mL BSA, 0.5 U Taq DNA polymerase, 5 mM Tris-pH 8.0, 10 mM KCl, 1.5 mM MgCl2, 0.75% Triton and 0.5X SYBR Green DNA stain.) DNA concentration and primer efficiency validation were performed by testing primer efficiency by titrating template concentrations. Serial dilutions of DNA were used to test a range of DNA concentrations from 0.1 ng/uL to 0.0016 ng/uL in technical triplicates. Efficiency values for RIM15, NTS2 and NL primer sets were calculated as 83%, 83%, and 81%, respectively, by the BioRad CFX96 software. The entire range of DNA concentrations was found to be within the linear range of instrument response. Melt curve analysis revealed a single product from each primer pair that was confirmed in each experimental run. The primer pairs were: For rDNA: NTS2: GGGCGATAATGACGGGAAGA-Fwd and TGTCCACTTTCAACCGTCCC-Rev; 35S (D1/D2 loop of 26S rDNA): GCATATCAATAAGCGGAGGAAAAG-Fwd and ACTTTACAAAGAACCGCACTCC-Rev; For single copy control: RIM15: GCCAGAACATTGGGTCAGAT-Fwd and CCGGATACTCGGATGTGTCT-Rev. For the single copy experimental (aqua bars in Fig 2B) ERV46: CACAGCTAGGACACCACCAA-Fwd and AGGGACAAGGATCATCCAAA-Rev. The enrichment values are equal to 2- ΔCt where ΔCt = Ct (rDNA locus (or ERV46))—Ct (single copy locus, RIM15) [72]. This value should equal the copy number of the target locus being assessed.
Strains M1321 (MATa ade2 ura3 leu2-3,112 trp1 hht1,hhf1::LEU2 hht2,hhf2::HIS3/pDP378 (HHT2,HHF2 CEN6 ARSH4 URA3) and M1345 (M1321, cdc6::ura3 LEU2::cdc6-4) were transformed with the TRP1 histone H3, H4 plasmids pH3H4-WT and mutant derivatives as shown in Fig 6A at 25°C. Multiple transformants were streak purified on FOA medium, recovered to YPD at 25°C and then screened for growth at various temperatures on YPD medium. Histone mutations were also constructed in pDP373 (pRS414 (TRP1) containing HHT2-HHF2 on a Spe1 fragment) using QuikChange. The entire H3 and H4 genes were sequenced for varification, and subsequently transformed in M1321 and M1345 for the yeast growth experiments in Fig 6.
|
10.1371/journal.pcbi.1007225 | β-Methylamino-L-alanine substitution of serine in SOD1 suggests a direct role in ALS etiology | Exposure to the environmental toxin β-methylamino-L-alanine (BMAA) is linked to amyotrophic lateral sclerosis (ALS), but its disease-promoting mechanism remains unknown. We propose that incorporation of BMAA into the ALS-linked protein Cu,Zn superoxide dismutase (SOD1) upon translation promotes protein misfolding and aggregation, which has been linked to ALS onset and progression. Using molecular simulation and predictive energetic computation, we demonstrate that substituting any serine with BMAA in SOD1 results in structural destabilization and aberrant dynamics, promoting neurotoxic SOD1 aggregation. We propose that translational incorporation of BMAA into SOD1 is directly responsible for its toxicity in neurodegeneration, and BMAA modification of SOD1 may serve as a biomarker of ALS.
| The environmental toxin β-methylamino-L-alanine (BMAA) has been linked to cases of amyotrophic lateral sclerosis (ALS), but the role of this compound in disease is unknown. We propose that BMAA becomes incorporated into the ALS-linked protein Cu,Zn superoxide dismutase (SOD1), destabilizing it and promoting formation of the protein aggregates characteristic of ALS. Using computational techniques focused on the structure of BMAA-incorporated SOD1, we demonstrate that the presence of BMAA changes SOD1 structure and dynamics to promote aggregation. We propose that BMAA incorporation in SOD1 in the mechanism of the compound’s link to ALS, and that BMAA modification may serve as a biomarker for environmentally-linked cases of ALS.
| Amyotrophic lateral sclerosis (ALS) is a motor neurodegenerative disease that affects 2–9 individuals per 100,000 every year [1]. More than 150 mutations to Cu,Zn superoxide dismutase (SOD1) have been associated with ALS. Misfolded and aggregated SOD1 has been found in motor neurons in both sporadic and familial ALS [2]. In recent studies, a non-native trimeric oligomer of SOD1 has been shown to be toxic in the hybridized motor neuron cell line NSC-34, suggesting a causative role of misfolded SOD1 aggregates in ALS etiology [3]. The phenomenon of SOD1 misfolding is puzzling due to the protein’s remarkable stability (ΔΔG >20 kcal/mol) [4]; the mild destabilization (<5 kcal/mol) caused by ALS-linked mutations [5] does not significantly reduce the stability of SOD1 from that of the average human protein (~5–15 kcal/mol [6,7]), and so does not explain SOD1 misfolding [8,9]. Previous studies have demonstrated that post-translational modifications of SOD1 can contribute to destabilization [10], and that glutathionylation of Cys111 promotes SOD1 dimer dissociation, the required initial step for SOD1 aggregation [11], by ~1,000 fold [12,13]. Environmental toxins that modify proteins have also been proposed to play a role in ALS etiology.
The indigenous Chamorro population on Guam have an ALS incidence 100 times larger than the worldwide average, which has been linked to an enrichment of the toxin β-methylamino-L-alanine (BMAA) in their diet [14]. The quest for the mechanism of BMAA toxicity resulted in the hypothesis that this amino acid is misincorporated into proteins [15], resulting in formation of inclusion bodies in neurons [16]. Studies have demonstrated synergistic toxicity of ALS-linked mutant SOD1 and BMAA [17], yet no reports of misincorporation have been presented. A large-scale proteomic study has identified multiple proteins that featured misincorporated BMAA [18]. However, the reported misincorporation rates were low. Despite the low misincorporation rates, Ackerman and colleagues have argued that even a rate of 1 misincorporation per 10,000 codons can lead to neurodegeneration in mice [19]. Hence, identification of BMAA misincorporation into SOD1 may have been overlooked due to sensitivity issues, and never reported.
We propose that misincorporation of BMAA into SOD1 destabilizes the protein, increases aggregation propensity, and thus promotes ALS onset and progression. We hypothesize that BMAA can directly modify SOD1 by incorporation in place of serine during translation. As a proof of principle, we perform a computational analysis predicting the effects on thermodynamic stability of substituting BMAA in place of each of the ten serines in SOD1. We find remarkable destabilization of SOD1 due to BMAA misincorporation at all sites, strongly suggesting a direct role of this toxin on the etiology of ALS. We perform molecular dynamics simulations of modified SOD1S107B to evaluate the structural impact of such substitution, and find significant dynamic changes to residues participating in metal-binding and the intra-monomer disulfide bond, key structural determinants of SOD1 stability. These findings suggest a mechanism for the toxicity of BMAA in ALS, and provide support for the candidacy of BMAA as a long-sought biomarker for ALS.
We evaluate the effects of replacing serine residues with BMAA in the SOD1 dimer. Because misincorporation is a rare event, more than one instance in the same molecule would be unlikely, and thus we study the scenario of BMAA misincoporation into only one monomer of the SOD1 heterodimer. We computationally substitute each individual serine residue in SOD1 (PDB ID: 1SPD) to BMAA, and estimate the resulting changes in free energy (ΔΔG) of the structure. To control for the effect of computational mutation, we also perform the same calculation while converting the given residue to lysine. Lysine, similar to BMAA, is also an unbranched, positively-charged amino acid. We find that while mutations of each serine to either BMAA or lysine generally destabilize the SOD1 dimer (Table 1), mutations to BMAA result in significant destabilization, while mutations to lysine result in minor (<2 kcal/mol) or negligible (<1 kcal/mol) destabilization, and in some cases ΔΔG is within error of zero. We conclude from these results that substitution of BMAA for serine in the SOD1 structure results in an unfavorable structural shift resulting in thermodynamic destabilization, likely due to steric effects from the larger BMAA molecule.
To obtain the thermodynamic melting curve of BMAA-SOD1, we perform replica exchange DMD simulations at a wide range of temperatures. As a demonstration of potential effects of BMAA, we choose substitution of S107, as the smallest predicted ΔΔG (Table 1) upon misincorporation of BMAA. Misincorporation of BMAA at a site with a larger predicted ΔΔG would be likely to have larger thermodynamic effects. We find that the incorporation of BMAA into the SOD1 structure in place of serine-107 shifts the melting temperature of the protein by only ~2°C (Fig 1A). However, we observe evidence of lower temperature localized unfolding events present in BMAA-SOD1 that are absent from the unfolding of WT-SOD1, which displays one dominant peak in CV representing coupled dimer dissociation and monomer unfolding [13]. Supporting this hypothesis, we find that BMAA modification increases the potential free energy of the low-energy “ground state” of the SOD1 dimer, decreasing the stability of the native state (Fig 1B). This destabilization makes BMAA-SOD1 more likely to undergo localized unfolding events that can expose toxic epitopes, as well as lead to the protein aggregation characteristic of ALS. This destabilization of the SOD1 dimer by BMAA substitution provides a mechanism for the linkage of BMAA poisoning to ALS etiology.
To test our conclusion that mutation of serine to BMAA results in a significant structural change in SOD1, we perform discrete molecular dynamics (DMD) simulations of SOD1 with BMAA incorporated into one monomer of the structure in place of Ser107, the site at which BMAA misincorporation was predicted to have the smallest thermodynamic effect. Misincorporation of BMAA at a site with a larger predicted ΔΔG (Table 1) would be likely to have larger structural changes. Upon building and equilibrating our model of BMAA-SOD1, we find rearrangement of the beta-barrel of the modified monomer, and resulting lengthening and twisting of the beta-strands that form the SOD1 dimer interface (Fig 1C), with a total root mean square structural deviation of 3.24 Å. Although metal ions are necessarily constrained to their ligands in our simulations, we note that the distortion of the SOD1 structure extends to shifts in the orientation of metal-binding residues, especially those coordinating Zn (Fig 1D), which would potentially affect the binding affinity of Cu and Zn in vitro and in vivo. Binding of metal ions, especially Zn, contributes significantly to the stability of SOD1, and destabilization and loss of bound metal ions is the second step in SOD1 aggregation [11], and metal-binding residues feature several known ALS-linked mutations.
To further investigate the potential effects of incorporation of BMAA into the SOD1 structure, we analyze the dynamics of the BMAA-modified protein in low-temperature steady-state simulations and compare with wild-type protein. Changes in root mean square fluctuation (RMSF) over the length of the protein (Fig 2, top) upon BMAA modification reveal increased flexibility in the metal-binding loop (residues 49–84) and the residues directly surrounding the BMAA modification, as well as flexibility differences caused by slight shifts in the residues included in β-strands 1, 2, and 3 due to the rearrangement in β-barrels discussed above.
While changes in RMSF indicate differences in local stability, correlated dynamics are a more informative measure of the effect of protein modification on overall structure, stability, and function because they reveal dynamic coupling between distal regions of the protein [20]. Changes in dynamic coupling across SOD1 due to BMAA misincorporation would change not only local stability, but also how local instabilities are propagated to other regions of the protein, potentially resulting in additional changes to structurally important features. In calculating the correlated motions of residue pairs [20], we find profound differences in the motions of residues corresponding to key structural features of SOD1 known to promote integrity of the properly folded structure (Fig 2): namely, both cysteines of the intra-monomer Cys57-Cys146; the Cu-binding histidines 46, 48, and 120; the Cu-Zn bridging ligand His63; the Zn-coordinating residues His71, His80, and Asp83; and the structurally important residue Asp124, which forms a crucial connection between Cu- and Zn- binding residues and whose mutation has been linked to ALS [21]. We also observe significant disturbances to large portions of both the electrostatic loop and the metal-binding loop, which contribute to enzymatic function, maintain structural integrity, coordinate the binding of the metal ions, and prevent protein aggregation [22]. Together, these findings strongly support the conclusion that the incorporation of BMAA into SOD1 causes both static and dynamic structural disturbances that result in local destabilization of the region surrounding the modification, including the nearby electrostatic loop, and propagation of those instabilities to important structural features of the protein, leading to increased propensity for misfolding and aggregation. This work supports an SOD1-linked mechanism for the toxicity of BMAA in environmentally caused cases of ALS.
SOD1 dimer dissociation has been shown to be the first step in the misfolding and aggregation of SOD1 [11]. Proctor et al. [3] recently demonstrated that the association of misfolded SOD1 monomers into a non-native trimeric oligomer results in cytotoxicity in hybridized motor neurons. The remarkable thermodynamic stability of unmodified wild type SOD1 protects against this first necessary step of dimer dissociation [5], thus also protecting against the formation of toxic oligomers. However, the addition of exogenous factors to the SOD1 structure, such as post-translational modifications, has been shown to have a profound destabilizing effect on dimer stability [10,12,23]; oxidative glutathionylation is a particularly severe example of such a modification [12,24]. Given the high fraction (90%) of sporadic ALS cases as compared to those with a known genetic link, we have long hypothesized that other post-translational modifications may similarly impact SOD1 stability. BMAA is a good candidate because, while not overly abundant, this cyanobacteria-produced neurotoxin has been linked to significantly increased occurrence of sporadic ALS in populations with frequent dietary consumption of food sources containing high levels of BMAA [14,16].
In this work, we present the hypothesis, based on others’ experimental and epidemiological observations [15,18], that BMAA can be incorporated into SOD1, and demonstrate using computational structural analysis and simulation that incorporation of BMAA would promote SOD1 dissociation, loss of metals, and misfolding. Misfolded SOD1 then aggregates to form oligomers that, through as yet unknown mechanisms result in motor neuron death, thereby contributing to the neurotoxicity of BMAA and its linkage to sporadic ALS in areas of environmental contamination (Fig 3). We speculate that BMAA incorporation into SOD1 may be rare, explaining why this modification has not yet been reported. However, even rare events may promote an avalanche of misfolding events; the initiating destabilization by BMAA incorporation may serve as a nucleating event for the misfolding and aggregation of SOD1 through the templating mechanism [25–29]. Our analysis suggests the need for a comprehensive study of SOD1 modification patterns in ALS patients in order to uncover mechanistic patterns of disease onset and progression, and aid in understanding of potential lifestyle and preventative interventions for sporadic ALS.
We determine the changes in free energy (ΔΔG) for mutations of each serine residue in the SOD1 dimer (PDB ID: 1SPD) to β-methylamino-L-alanine (BMAA) or lysine using Eris [30,31]. Reported ΔΔG values represent the mean ± standard deviation of 20 independent rounds of Eris calculation. Each round of Eris calculation produces an expected value of the ΔΔG of mutation from 20 independent simulations for both wild-type and mutant protein with each simulation consisting of 20 steps of Monte Carlo optimization.
The BMAA rotamer library was generated using the Rosetta MakeRotLib protocol [32]. We used the Gaussian 09 program (Gaussian, Inc.) to optimize the initial structure of BMAA at the HF 6-31G(d) level of theory with a polarized continuum model of the aqueous solvent, which appropriately shields the positive charge on the BMAA side chain. We generate a backbone-dependent rotamer library from the initial structure using 10° increments for both φ and ψ angles for a total of 1296 (= 36 × 36) φ/ψ bins, within which each of the two χ angles of BMAA were sampled at 30° increments. The MakeRotLib protocol was used to obtain mean angles and probabilities for all combinations of the three staggered conformations for the two χ angles in each φ/ψ bin. Lysine parameters for Ramachandran probabilities, χ angle standard deviations, and the reference energy were used for both BMAA and lysine, as both feature unbranched, positively-charged side chains. The residue type parameter file for BMAA was built using pre-existing atom types in the CHARMM-based Medusa force field [33].
DMD implements step function potentials to describe inter-atomic interactions, as opposed to the continuous potentials used in traditional molecular dynamics (MD) [34–36]. We utilize an all-atom protein model that explicitly represents all heavy atoms and polar hydrogen atoms. Bonded interactions are represented using infinite square-well constraints for bond lengths, bond angles, and dihedral angles. Non-bonded interactions are adapted from the continuous CHARMM-based Medusa force field [33], van der Waals interactions are modeled using the Lennard-Jones potential, and solvation interactions are modeled using Lazaridis-Karplus solvation [37], all discretized by multi-step square-well functions for use in DMD. We model hydrogen bonding interactions using the reaction algorithm [38]. The DMD simulation engine (πDMD, v1.0) with Medusa all-atom force field is available from Molecules In Action, LLC (free to academic users, moleculesinaction.com).
Using the known X-ray crystallographic structure of wild type SOD1 (PDB ID 1SPD) as a reference structure, we deleted serine 107 from one monomer and replaced it with BMAA, which was joined in the peptide chain of SOD1 using peptide bond constraints and equilibrated using the discretized Medusa force field [33] in DMD with an iterative relaxation and equilibration protocol as previously described [13].
We use the replica exchange method to construct a thermodynamic profile of BMAA-SOD1 unfolding [39]. Independent replicas of the simulation system of interest are run in parallel at 16 different temperatures: 0.48 (∼242 K), 0.495 (∼ 249 K), 0.51 (∼ 257 K), 0.525 (∼ 264 K), 0.54 (∼ 272 K), 0.555 (∼ 280 K), 0.57 (∼ 287 K), 0.585 (∼ 295 K), 0.60 (∼ 302 K), 0.615 (∼ 310 K), 0.63 (∼ 317 K), 0.645 (∼ 325 K), 0.65 (∼ 327 K), 0.67 (∼ 337 K), 0.69 (∼ 347 K) and 0.71 (∼357 K) kcal (mol kB)–1. Every 50 ps, replicas neighboring in temperature attempt to exchange temperature values according to the Metropolis criterion. The replica exchange method increases sampling efficiency by allowing energetic barriers to be overcome with exposure to higher temperatures. We note that temperatures used in MD simulations do not directly equate to physical temperatures, but are useful to evaluate relative differences between systems.
Replica trajectories were combined for the analysis of folding thermodynamics using the MMTSB tool [40] for weighted histogram analysis method (WHAM) [41]. WHAM computes the density of states by combining energy histograms from simulation trajectories with overlapping energies and calculates the folding specific heat at constant volume at a function of temperature.
|
10.1371/journal.pntd.0000238 | Trypanosoma brucei rhodesiense Transmitted by a Single Tsetse Fly Bite in Vervet Monkeys as a Model of Human African Trypanosomiasis | We have investigated the pathogenicity of tsetse (Glossina pallidipes)-transmitted cloned strains of Trypanosoma brucei rhodesiense in vervet monkeys. Tsetse flies were confirmed to have mature trypanosome infections by xenodiagnosis, after which nine monkeys were infected via the bite of a single infected fly. Chancres developed in five of the nine (55.6%) monkeys within 4 to 8 days post infection (dpi). All nine individuals were successfully infected, with a median pre-patent period of 4 (range = 4–10) days, indicating that trypanosomes migrated from the site of fly bite to the systemic circulation rapidly and independently of the development of the chancre. The time lag to detection of parasites in cerebrospinal fluid (CSF) was a median 16 (range = 8–40) days, marking the onset of central nervous system (CNS, late) stage disease. Subsequently, CSF white cell numbers increased above the pre-infection median count of 2 (range = 0–9) cells/µl, with a positive linear association between their numbers and that of CSF trypanosomes. Haematological changes showed that the monkeys experienced an early microcytic-hypochromic anaemia and severe progressive thrombocytopaenia. Despite a 3-fold increase in granulocyte numbers by 4 dpi, leucopaenia occurred early (8 dpi) in the monkey infection, determined mainly by reductions in lymphocyte numbers. Terminally, leucocytosis was observed in three of nine (33%) individuals. The duration of infection was a median of 68 (range = 22–120) days. Strain and individual differences were observed in the severity of the clinical and clinical pathology findings, with two strains (KETRI 3741 and 3801) producing a more acute disease than the other two (KETRI 3804 and 3928). The study shows that the fly-transmitted model accurately mimics the human disease and is therefore a suitable gateway to understanding human African trypanosomiasis (HAT; sleeping sickness).
| Sleeping sickness is caused by a species of trypanosome blood parasite that is transmitted by tsetse flies. To understand better how infection with this parasite leads to disease, we provide here the most detailed description yet of the course of infection and disease onset in vervet monkeys. One infected tsetse fly was allowed to feed on each host individual, and in all cases infections were successful. The characteristics of infection and disease were similar in all hosts, but the rate of progression varied considerably. Parasites were first detected in the blood 4–10 days after infection, showing that migration of parasites from the site of fly bite was very rapid. Anaemia was a key feature of disease, with a reduction in the numbers and average size of red blood cells and associated decline in numbers of platelets and white blood cells. One to six weeks after infection, parasites were observed in the cerebrospinal fluid (CSF), indicating that they had moved from the blood into the brain; this was associated with a white cell infiltration. This study shows that fly-transmitted infection in vervets accurately mimics human disease and provides a robust model to understand better how sleeping sickness develops.
| In human African trypanosomiasis (HAT), the use of animal models has contributed enormously to what is currently known about the relationships between disease duration, parasite invasion of different body systems and the potential of resultant host clinical and biological changes as diagnostic and disease staging markers. Several host-parasite model systems have been developed, based on infection of various hosts with the livestock pathogen Trypanosoma brucei brucei and to a lesser extent the human pathogens T. b. rhodesiense and T. b. gambiense. Characterisation of these HAT models shows that the disease occurs in two stages irrespective of host: an early haemo-lymphatic trypanosome proliferation, and a late central nervous system (CNS) infection, indicating that the basic pattern is similar to the disease in humans. This is evidenced by demonstration of trypanosomes, first in the haemo-lymphatic system and later in the CNS of the mouse model with subsequent cerebral pathology [1],[2]. Models based on larger mammals such as the chimpanzee T. b. rhodesiense model [3], the vervet monkey T. b. rhodesiense model [4] and the sheep T. b. brucei model [5], also follow a similar two-stage disease pattern. These, unlike rodents, allow collection of cerebrospinal fluid (CSF) that has been used to demonstrate elevation of white cell counts and total protein levels as indicators of CNS stage disease [6].
The KETRI vervet monkey model has been reported to closely mimic HAT clinically, immunologically and pathologically [4], [7]–[9]. However, these previous studies were limited in scope in three important ways. Firstly, infections were initiated by intravenous inoculation (syringe) of bloodstream form trypanosomes as opposed to the natural human disease, which begins via the bite of a tsetse fly, with the intra-dermal inoculation of metacyclic trypanosomes. The difference between the two routes of infection has the potential to affect trypanosome virulence and subsequent disease pathogenesis that has been little explored to date. Second, disease progression has been monitored mainly in terms of clinical symptoms, gross pathology, histo-pathology and antibody responses [4],[7], with little reference to the development of blood pathology. Third, only a single strain of trypanosomes, KETRI 2537 [9], has been adequately characterised even though trypanosome strains vary in the severity of pathogenesis and virulence [10]–[11].
The present study was designed to address these limitations and thus improve further the potential utility of the model and our understanding of pathogenesis in trypanosome infections. We characterised the pathogenicity of T. b. rhodesiense in vervet monkeys, following infection from the bite of a single tsetse fly, hence mimicking the natural route of infection in man. This allowed us to measure a range of parameters in blood and cerebrospinal fluid (CSF) including several that had not previously been studied. It was thus possible to measure the development of clinical complications of HAT infections, such as anaemia, more precisely. We describe the development of clinical pathology resulting from infection with four cloned strains of T. b. rhodesiense.
This study was undertaken in adherence to experimental guidelines and procedures approved by the Institutional Animal Care and Use Committee (IACUC), the ethical review committee for the use of laboratory animals.
Trypanosome isolates that were used in this study (Table 1) were all initially obtained through collection of infected blood from patients in the western Kenya/eastern Uganda focus of endemic T. b. rhodesiense sleeping sickness (historically known as the Busoga focus). All the isolates are maintained as cryo-preserved stabilates in the KARI-TRC (formerly KETRI) trypanosome bank. The isolates were included in the study on the basis of the year of isolation, to give a wide temporal distribution and the locality of isolation to give a wide spatial distribution within this sleeping sickness focus. The selected stabilates were cloned using the hanging drop method described by Herbert and Lumsden [12].
Male teneral tsetse flies (Glossina pallidipes) were obtained from the KETRI colony initially established with pupae from the Lambwe Valley of Kenya, which is part of the western Kenya/eastern Uganda focus of HAT. In order to initiate infection of tsetse flies with each trypanosome clone, four sub-lethally irradiated (600 rads, 5 minutes) donor Swiss White mice were each inoculated intraperitoneally with 0.2 millilitres of the thawed T. b. rhodesiense stabilates, diluted in phosphate saline glucose (PSG). At peak parasitaemia, typically approximately 108 trypanosomes per millilitre, a batch of 50 teneral flies were allowed to feed essentially as described [13], and maintained thereafter on clean bovine blood by feeding via a silicon membrane. Thirty days after the infective blood meal, all the flies were chilled briefly and separated into individual fly cages. The flies with mature trypanosome infections were then identified by xenodiagnosis using Swiss White mouse. We were repeatedly unable to find trypanosomes in salivary probes on warm microscope slides [14].
Nine vervet monkeys (Chlorocebus aethiops, African Green Monkeys) of both sexes weighing between 2.7 and 5.2 kg were acquired from the Institute of Primate Research (IPR) in Kenya. They were housed in quarantine for a minimum of 90 days while being screened for evidence of disease, including zoonoses as described by Ndung'u and colleagues [8]. They were also dewormed and treated for any ectoparasite infestations. During the quarantine period, the animals became accustomed to staying in individual squeeze-back stainless steel cages and human handling.
During quarantine and also while in the experimental animal wards, the monkeys were maintained on green maize, fresh vegetables (bananas, tomatoes and carrots) and commercial monkey cubes (Monkey cubes, Unga Ltd, Kenya), fed twice daily (9.00–9.30 am and 3.00–3.30 pm), and given water ad libitum. After the expiry of the 90 days quarantine, the study animals were then transferred to experimental wards and acclimatised for a further two weeks prior to commencement of pre-infection data collection.
The monkeys were randomly allocated into four experimental groups, each containing at least one male and one female, for infection with T. b. rhodesiense clones as follows: KETRI 3741 (three monkeys, #s. 476, 515, and 536), KETRI 3801 (monkey #s. 523, 579), KETRI 3804 (monkey #s 556 and 574) and KETRI 3928 (monkeys #s 554 and 555). Pre-infection (baseline) data was collected over a period of 14 days after which each monkey was infected by allowing one tsetse fly, confirmed trypanosome positive through mouse infectivity tests, to feed on a shaved part of its thigh, while the monkey was under ketamine Hcl (Rotexmedica, Trittau Germany) anaesthesia. Before and following the infective tsetse bite, the monkeys were monitored for activity, posture, demeanour and general clinical presentation on a daily basis. Appetite was assessed daily, by scoring the proportion of the daily feed ration consumed by each monkey on a scale of 0 (no food eaten), 1/4, 1/2, 3/4 and 1 (full ration eaten).
Parasitaemia was assessed daily using the method of Herbert and Lumsden [12], using heparinised capillary blood drawn from the ear vein, starting from the third day after infection. Every four days, the monkeys were sedated using ketamine hydrochloride (10–15 mg per kg body weight intramuscularly) after which a detailed clinical examination was carried out and 2 ml of venous blood (femoral) sampled for a full haemogram. Every eight days, a CSF sample was also collected through lumbar puncture for assessment of CNS parasitosis and white cell numbers. The experiment was terminated through humane euthanasia at extremis. An individual animal was judged to be in extremis when for three consecutive days it was either unable or reluctant to perch, had very low feed intake (<1/4 of daily ration), and in addition had signs of advanced late stage disease (e.g somnolence). Euthanasia was carried out using 20% pentobarbitone sodium (Euthatal, Rhone Merieux).
Cerebrospinal fluid white cell counts (WCC) and total trypanosome numbers were concurrently counted using a Neubert chamber as previously described [6],[8]. Immediately after every sampling session (not exceeding one hour), total red blood cell (RBC) and related indices, white cell numbers and differential, platelet (thrombocyte) counts and associated parameters were determined using an AC3diff T Coulter counter (Miami, Florida, USA).
Data was entered and managed using Microsoft Excel (Version 2003). Statistical analysis was conducted using Statview for Windows Version 5.0.1 (SAS Institute Inc, 1995–1998, Cary, NC). The behaviour of the four trypanosome strains was analysed and is presented as tables and or graphs representing time bound changes in individual infected monkeys' clinical, haematological and cerebrospinal fluid pathology data. In addition, descriptive statistics [mean (and the corresponding 95% confidence intervals, CI), or medians and range] were derived for the entire group of nine monkeys. In addition to derivation of descriptive data, haematology data was further analysed using repeated measures ANOVA. Finally, Spearman's correlation coefficients were determined to assess the strength of association between CSF trypanosome and white cell numbers.
In this study, infection of vervet monkeys was initiated by the bite of a single infected tsetse fly. To our knowledge, this represents the first time single fly transmission of T. b. rhodesiense clones in vervet monkeys (or any other primate model) has been achieved, hence establishing a model that more accurately mimics the transmission of sleeping sickness as it occurs in humans. Information on natural HAT relies on data provided in case reports [15]–[16] and sometimes re-analysis of retrospective clinical, epidemiological and pathology data [17]–[18]. Such data are naturally limited on the questions of disease onset and duration as there are only a small number of case reports in which the patient could accurately remember the exact time of being bitten by a tsetse fly [16],[18]. In any case this is not readily accessible information for inhabitants of endemic areas, where tsetse fly challenge is continuous. This study has allowed us to generate information on the pathogenesis of HAT in a manner that more accurately mimics human disease, and facilitates documentation of data from precise sampling points during the course of a tsetse transmitted infection.
All four T. b. rhodesiense clones that were selected for this study were successfully transmitted through tsetse flies in the initial step of the model development protocol, a result that is consistent with the very good vectorial capacity of G. pallidipes for both human and animal trypanosomiasis in eastern Africa. Tsetse flies carrying mature infections were identified by xenodiagnosis using Swiss White mice. The pre-patent period in these mice (data not shown), and subsequently in monkeys (Table 2), showed considerable variation from host to host, consistent with observations in HAT patients [18]–[19]. In contrast, mice and vervet monkeys that are infected with T. b. rhodesiense via syringe passage show less variation in pre-patent period [4], presumably because the inoculum in these is usually well-defined, while tsetse flies are estimated to inject 0–40,000 (mean, 3,200) infective metacyclics [20].
Trypanosomes were detected in 7 of the monkeys within 5 days after infection (Table 2), indicating that the movement of trypanosomes from the site of the fly bite to the systemic circulation occurred quickly. This is remarkable, because it must be associated with the transition from non-proliferating metacyclics to rapidly dividing long slender bloodstream forms, clearly a survival strategy for the parasites. This movement was independent of the development of chancre, which was only observed in 5 monkeys. Chancres have been observed in HAT patients in whom these swellings are estimated to occur within 5–15 days of an infective fly bite [20], consistent with our data (Table 2). During the formation of the chancre, the metacyclics transform to rapidly dividing bloodstream form trypanosomes while the tissue at the inoculation site mounts a reaction characterized by a marked infiltration with polymorphonuclear leucocytes [21]–[22]. The immune reaction generated at the chancre is responsible for development of specific immunity against the variable antigen type of metacyclics [23].
The finding that the severity of the clinical disease differed between individual monkeys that were infected with the same strain emphasized the likely role of host immunity on disease outcome. The ability/inability of the host to control parasitaemia and its effect on disease duration is further indicated by the observation that parasitaemia patterns showed more fluctuation (clearly marked waves) in individuals with longer disease duration than in those with shorter durations (Figure 1). A similar trend was observed in mice (data not shown), consistent with previous reports [11]. In our study, clone 3801 produced a more acute disease, while clone 3928 manifested the most chronic disease; the other two clones were intermediate. These observations suggest that the parasites have intrinsic properties that, in part, determine virulence. Host factors also contribute to the disease profile, as evidence from variations in animals infected with the same clone indicates. A study of a number of isolates from eastern Uganda by Smith and Bailey [24] in mice showed that distinct acute and chronic strains of T. b. rhodesiense circulate in the focus and each strain is related to a given zymodeme. However, apart from individual parasite variations there were no features that could distinguish the Ugandan from Kenyan isolates. This supports the view that the four strains used in this study belong to the same endemic focus characterized by pockets of specific zymodemes with distinct clinical manifestations [24]. Similar diversities in clinical manifestations have been observed in HAT patients infected with T. b. rhodesiense [25] and T.b. gambiense [26] showing that animal models accurately mirror the situation in humans
Haematology results showed that anaemia developed early in the monkey infections; the decline in relevant parameters was detected as early as 8 dpi. However, the rate of decline of RBC and associated parameters was much slower than in T. brucei infected mice in which the numbers of circulating erythrocytes can fall by up to 50% within a week after infection [27]. Anaemia is a common occurrence in both T. b. rhodesiense and T. b. gambiense forms of sleeping sickness [15]–[16],[26],[28], similar to the case in vervet monkeys. However, determination of the rates of decline of RBC and associated parameters, is not possible in humans since neither the date of infection nor the pre-infection values in individual patients are known. The type of anaemia reported in our study, microcytic hypochromic, was different from the normocytic anaemia observed in T. brucei infected mice [29] or Nigerian mongrel dogs [30] during the acute phase of T. b. brucei infection. Microcytic hypochromic anaemia has previously been associated with iron deficiency [31] and could perhaps be related to failure of iron incorporation into red cell precursors or inefficient recovery of iron from the phagocytosed RBC, features which are common during acute trypanosomiasis [32]. Determination of the type of anaemia found in infected humans is complicated by presence of concurrent infectious and nutritional conditions [29]. This is compounded by the lack of appropriate haematology analysers in endemic areas, and has therefore not been systematically determined to our knowledge.
The severe progressive thrombocytopaenia reported in our study mirrors that found in other T.b. rhodesiense animal models [30] and human cases of sleeping sickness [28]. These findings indicate that unlike in mild cases of iron deficiency anaemia that are accompanied by thrombocytosis, the anaemia of trypanosomiasis in both humans and animals is severe and could be related to a deficit in the production of thrombopoetin [33]. Similarly, leukocyte changes are broadly consistent with findings from other non–human primate studies [3]–[4] and humans [17]. However, the strong granulocyte response that coincided with the day of first detection of trypanosomes in peripheral blood (median = 4 dpi) has not been reported before, perhaps due to the lower frequency of sampling employed in other studies. Importantly, the presence of multiple peaks of white cells during the course of the infection suggests that in spite of the widely reported immunosuppressive effects of trypanosome infections, myeloid precursor cells retain the ability to proliferate in response to dominant parasite VSG's expressed during the course of the disease. This is in agreement with findings that some bone marrow stem cells survive the damage caused by trypanosomes and retain the ability to repopulate the animal [34] and may account for the observation of very late stage leucocytosis in some individuals but not others.
The first evidence of trypanosomes in the CSF was on day 16 (range 8–40) days (Table 2). This event is recognised by WHO [6] as a definitive marker for the onset of late stage infection. The timing of CSF parasitosis was largely similar to earlier observations in the syringe passage monkey infections where the blood-brain-barrier (BBB) was breached within 7–21 days [8],[35]. Clone 3928 produced the most chronic infection of all isolates and, in monkey 554, was only detected in the CSF on day 40 after infection. Some T. b. rhodesiense isolates from south-eastern Africa foci and some from eastern Uganda have been reported to cause a chronic HAT infection in humans, taking relatively long to invade the CNS [24],[36]. One of the recognized markers of CNS pathology is the presence of raised numbers of leucocytes in the CSF above the background (pre-infection) levels [6], [8], 37–38. Indeed, there was positive linear association between trypanosomes in the CSF and white cell changes, suggesting that both events are primarily determined by a single cause, possibly damage to the blood-brain barrier. The numbers of trypanosomes in CSF increased dramatically as disease progressed, and clinical symptoms of disease necessitated individuals to be removed from the study on ethical grounds, marking the terminal stage.
The results of this study establish a cyclic T. b. rhodesiense model that more closely resembles the East African form of HAT. Although T. b. gambiense causes a more insidious slowly developing disease, the essential features including fever, loss of appetite, headache, fatigue, weight loss, leg paresthesis, gait difficulties and daytime somnolence are similar to symptoms observed in patients infected with T. b. rhodesiense [18],[39]. Thus, this disease model in which the infection is induced using the bite of a single fly can better represent the complex pathogenesis of natural HAT. This model allows more precise timing of events, such as date of infection, and the clinical and haematology features that follow. Consequently, it is hoped that the new model will gain application and facilitate studies that require good precision. |
10.1371/journal.pbio.1000376 | A Neuromedin U Receptor Acts with the Sensory System to Modulate Food Type-Dependent Effects on C. elegans Lifespan | The type of food source has previously been shown to be as important as the level of food intake in influencing lifespan. Here we report that different Escherichia coli food sources alter Caenorhabditis elegans lifespan. These effects are modulated by different subsets of sensory neurons, which act with nmur-1, a homolog of mammalian neuromedin U receptors. Wild-type nmur-1, which is expressed in the somatic gonad, sensory neurons, and interneurons, shortens lifespan only on specific E. coli food sources—an effect that is dependent on the type of E. coli lipopolysaccharide structure. Moreover, the food type-dependent effect of nmur-1 on lifespan is different from that of food-level restriction. Together our data suggest that nmur-1 processes information from specific food cues to influence lifespan and other aspects of physiology.
| Work on the model organisms C. elegans and D. melanogaster has contributed important and often surprising insights into the factors that determine lifespan. One intriguing finding is that lifespan in both animals can be extended or shortened by interfering with the function of neurons that smell or taste food. Indeed, specific taste neurons in C. elegans are required for the lifespan extension due to the restriction of the animals' level of food intake, while certain olfactory neurons in Drosophila inhibit this effect. Here we provide evidence that the sensory system also alters lifespan in response to specific food types as opposed to different food levels. C. elegans that feed on different E. coli strains can have different lifespans, which is not only dependent on the activities of a subset of sensory neurons but can also occur independently of food level restriction. We also show that the neuropeptide receptor NMUR-1 acts with the sensory system to affect lifespan in a manner dependent on the bacterial lipopolysaccharide structure. Thus, we identify both a food-derived factor and a component of a signaling pathway involved in the food-type effects on worm lifespan.
| The sensory systems of Caenorhabditis elegans and Drosophila melanogaster have been shown to modulate the lifespan of these animals [1]–[4]. This sensory influence involves subsets of gustatory and olfactory neurons [2],[3] that either shorten or lengthen lifespan, which suggests that (i) some of the cues that affect lifespan are food-derived and that (ii) these cues can exert different effects on lifespan. Since a reduction in food levels can increase lifespan [5], it is possible that the sensory system influences lifespan by simply regulating the animal's general food intake, and, indeed, the sensory system has been implicated in the lifespan effects of food-level restriction in Drosophila [3]. On the other hand, the sensory influence on lifespan, at least in C. elegans, can be uncoupled from the sensory effects on feeding rate, development, and reproduction [1],[2]. Since the lifespan effect of food-level restriction has been linked to changes in feeding rates and decreased development and reproduction [5], this suggests that the sensory system also affects lifespan through other mechanism(s).
The C. elegans hermaphrodite has 60 sensory neurons with dendrites that terminate in ciliated endings [6]. These specialized structures contain dedicated sensory receptors [7],[8] and are thus the sites of recognition for different types of environmental cues, including gustatory, olfactory, thermal, and mechanical stimuli [9]. Within its natural environment, C. elegans encounters various types of bacteria that can serve as food sources. Similar to the sensory influence on lifespan, some of these food sources have been shown to alter lifespan independently of development and reproduction [10]. At the same time, not all but only a subset of food-sensing neurons influence the lifespan of C. elegans grown on the standard laboratory food source [2], Escherichia coli OP50 [11]. Together these data raise the possibility that sensory neurons promote the lifespan effects of different food sources through a mechanism distinct from that of food-level restriction.
In this study, we have investigated the role of the sensory system in the food-source influence on C. elegans lifespan and the signaling pathway(s) that might be involved in this process. We show that the C. elegans sensory system recognizes food types to affect longevity. We also identify (i) the neuromedin U receptor nmur-1 as a neuropeptide signaling pathway involved in this process and (ii) a food-derived cue, the E. coli lipopolysaccharide (LPS) structure, which elicits the nmur-1 response.
Wild-type C. elegans have altered lifespan on different E. coli strains (Figure 1A and 1B). Indeed, we found that at 25°C the mean lifespan of wild-type worms is shorter on OP50 than on HT115 (Figure 1A and 1B), another food source that is widely used [12],[13]. To test the hypothesis that sensory perception contributes to these food source-dependent effects, we measured the lifespan of sensory mutants on OP50 and HT115 at this temperature.
The gene daf-10 encodes an ortholog of an intraflagellar transport complex protein that is required for cilia formation in a subset of sensory neurons (Table S1) [14],[15]. We observed that the lifespan of daf-10 mutants is extended to the same extent (44% versus 46%; Table 1) compared to that of wild type when grown on either OP50 or HT115 (Figure 1C), which suggests that some sensory neurons shorten lifespan independently of these two food sources.
In contrast, worms that carry mutations in osm-3, which encodes a kinesin motor protein required for cilia formation in a different subset of sensory neurons (Table S1) [16],[17], live long relative to wild type only when grown on OP50, but not when grown on HT115 (Figure 1D; Table 1). This implies that at least some of the osm-3-expressing neurons sense the lifespan-influencing difference(s) between these food sources.
Since osm-3 functions in cilia structure formation rather than in directly sensing or translating food-derived cues, we searched for non-structural genes that would act with the sensory system to regulate the food source-dependent effects on lifespan. Candidate genes would include those encoding sensory receptors and downstream signaling molecules, like neuropeptides and their receptors, which help transmit or modulate sensory information. Unlike individual sensory receptors specific for single cues, a single downstream factor may affect the integration of several cues, which would make the effects of this class of genes more readily detectable.
The C. elegans genome has more than 75 neuropeptide-like genes and more than 1,000 G-protein-coupled receptors, some of which function as neuropeptide receptors [9],[18]–[21]. We focused on a subset of these genes based on the availability of mutations and on the evidence that their homologs in other animals regulate feeding and metabolism [19],[21]–[25]. We compared the lifespan of the different mutants on OP50 and HT115.
While most neuropeptide signaling pathways had no effect on lifespan on the two food sources tested (Table S2), we found that animals carrying the deletion mutation ok1387 within the gene C48C5.1 live long on OP50 but not on HT115 (Figure 2A and 2B; Tables 1, S2, and S3). C48C5.1 is predicted to encode a seven-transmembrane neuropeptide receptor (Figure S1) with homology to mammalian neuromedin U receptors (NMURs), whose peptide ligand, neuromedin U (NMU), has been shown to regulate food intake [24]. We renamed C48C5.1 as nmur-1, since our study makes it the first phenotypically characterized member of the worm NMUR family, of which there are at least three other members—nmur-2 (K10B4.4), nmur-3 (F02E8.2), and nmur-4 (C30F12.6). As a confirmation that the wild-type function of nmur-1 is to shorten lifespan in a food source-dependent manner, we were able to rescue the long-life phenotype of the nmur-1 mutation on OP50 with the wild-type nmur-1 genomic locus, without shortening lifespan on HT115 (Figure 2G; Tables 1 and S3).
Next, we asked whether sensory neurons regulate the food source-dependent effects on lifespan through nmur-1. We found that loss of nmur-1 still considerably increases the lifespan of daf-10 sensory mutants on OP50 (Figure 3A; Table 1), which indicates that nmur-1 acts in parallel at least to some daf-10-expressing neurons. Surprisingly, loss of nmur-1 extends the lifespan of daf-10 mutants also on HT115 (Figure 3B; Table 1), which may suggest that the lifespan of nmur-1 mutants becomes food source-independent in the absence of daf-10 activity. Thus, nmur-1 appears to be subject not only to activation by certain environmental cues but also to inhibition by others.
In contrast, animals that carry both nmur-1 and osm-3 mutations have a lifespan phenotype similar to that of nmur-1 single mutants on OP50 and HT115 (Figure 3C and 3D; Table 1). This suggests that nmur-1 acts with osm-3 either in a subset of osm-3-expressing sensory neurons or in downstream cells. We observed expression of a gfp reporter for nmur-1 in the spermathecae of the somatic gonad, in several different types of sensory neurons, some of which co-express osm-3 (Table S1) [16], and in interneurons (Table 2), some of which receive inputs from, or modulate the activity of, osm-3-expressing sensory neurons [6]. This expression pattern, together with the genetic interaction between the mutations in nmur-1 and osm-3, suggests that nmur-1 plays a role in the processing of sensory information derived by the worm from various food sources.
We then explored the possible differences between OP50 and HT115, which might be recognized by the worm. OP50 is derived from an E. coli B strain [11], whereas HT115 is from an E. coli K-12 strain [26],[27]. To determine whether nmur-1 affects lifespan only on B strains but not on K-12 strains, we measured the lifespan of nmur-1 mutants on other bacteria derived from these two lineages. Interestingly, we found that nmur-1 mutants live long consistently on the B strain BL21 [28] and on HB101 (Figure 2C and 2E; Tables 1 and S3), a K-12 strain that contains a large stretch of B strain genomic DNA [29]. In contrast, the nmur-1 long-life phenotype is absent on another K-12 strain, DY330 [30], and only occasionally present on the K-12 strain DH5α (Figure 2D and 2F; Tables 1 and S3) [31]. Together these data suggest that nmur-1 affects lifespan in a largely B strain-dependent manner.
Although the B and K-12 strains clearly would have many differences, one of the few well-characterized molecular differences between these strains lies in the LPS structures (Figure 4A) on their outer membranes [32]–[34]. Since the LPS of the K-12 strain [33],[34] has a longer outer core than the LPS of the B strain [32], we tested whether LPS structure influences lifespan. We compared wild-type and nmur-1 mutant worms on E. coli K-12 mutants that have truncated LPS to worms grown on the corresponding K-12 parent strain. We found that wild-type worms live shorter on the LPS truncation mutants CS2198 and CS2429 [35],[36] than on the isogenic parent strain CS180 (Tables 1 and S3), which expresses wild-type K-12 LPS [35]. On the other hand, nmur-1 mutants live long compared to wild type only on the LPS truncation mutants (Figure 4C and 4D; Tables 1 and S3), but not on the K-12 parent strain (Figure 4B; Tables 1 and S3).
To exclude the possibility that all changes to the LPS will elicit the nmur-1 response, we also measured the lifespan of worms grown on the K-12 strain CS1861 that expresses the Shigella dysenteriae 1 O Antigen fused to the end of the full-length K-12 LPS [36]. We observed no lifespan difference between wild type and nmur-1 mutants on this strain (Figure 4E; Table 1). Together our data suggest that a short E. coli LPS structure can shorten worm lifespan in an nmur-1-dependent manner.
In contrast to nmur-1, we found that osm-3 can affect lifespan independently of the E. coli LPS structure (Figure 4F and 4G; Table 1), which indicates that at least some of the osm-3-expressing neurons detect other food-derived cues. However, even on the CS180 and CS2429 bacterial food sources, nmur-1 and osm-3 appear to act together in influencing lifespan, since osm-3; nmur-1 double mutants have the same lifespan phenotype as osm-3 or nmur-1 single mutants (Figure 4F and 4G; Table 1).
Food type [37] and sensory neurons [3],[38] have been shown to mediate the lifespan extension induced by dietary restriction (DR), which is commonly studied through restriction of food levels. Thus, the food type-dependent effects on lifespan we observe might reflect different levels of DR experienced by wild-type and mutant worms on the various food sources. To address this possibility, we measured the feeding rates, speed of development, total progeny, and the rates of reproduction of wild-type and nmur-1 mutant worms on five different E. coli strains (Figures S2 and S3), since restricting food levels is known to change these parameters [5]. For comparison, we used a genetic model for food-level restriction [39], a mutation in eat-2 that impairs pharyngeal function [40], which leads to decreased feeding rates on both OP50 and HT115 (Figure S4A). Unlike the nmur-1 mutation, we found that the eat-2 mutation increases lifespan on both food sources (Figure S4C), which suggests that the food-type effects of nmur-1 are not the same as those of food-level restriction. Moreover, we observed no correlation between lifespan and feeding rates or lifespan and development of wild-type or nmur-1 mutant worms on the different food sources (Figures 5A, 5B, S2A, S2B, S2C, S2D, S3A, and S3B), which is also unlike the reported effects of restricting food levels [5].
As expected for a genetic model for food-level restriction, we found that the lifespan extension conferred by the eat-2 mutation is accompanied by a decrease in total progeny on OP50 and HT115 (Figures 5C and S4B). Surprisingly, we also found that wild-type worms grown on different food sources do exhibit an inverse correlation between lifespan and number of progeny but that nmur-1 mutants can still live long without a proportionate decrease in total progeny (Figures 5C, S2E, and S3C). This suggests that the food source-dependent effects on lifespan have reproduction-dependent and reproduction-independent components, the latter of which is uncovered by the nmur-1 mutation. Interestingly, we also observed that food sources that increase wild-type lifespan induce the animals to reproduce faster (Figure 5D), which not only differs from eat-2 mutants (Figure 5D) but is also the inverse of the effects shown for food-level restriction on rates of reproduction [5],[41]. At the same time, we again saw no correlation between the nmur-1 mutant lifespan and its rates of reproduction on the different E. coli strains (Figures 5D, S2F, and S3D).
Since our data show that the effects of eat-2 on C. elegans physiology differ from those of nmur-1 or the different food sources, this suggests that the effects of the food sources and nmur-1 on lifespan can be distinct from food-level restriction. Consistent with this idea, we observed that, unlike long-lived, food level-restricted animals that have decreased lipid storage [5], nmur-1 mutants do not exhibit gross changes in fat storage compared to wild type on either OP50 or HT115 (Table S4).
Next, we asked whether nmur-1 acts through the insulin/IGF-1 daf-2 pathway [42], which has been shown to mediate a large part of the sensory influence on lifespan [1]. For example, the increased lifespan of osm-3 sensory mutants on OP50 has been shown to be partly dependent on daf-16 [1], a FOXO transcription factor that acts downstream of and is negatively regulated by daf-2 [43]–[45]. Accordingly, we found that removing nmur-1 does not significantly increase the lifespan of insulin/IGF-1 receptor daf-2 mutant worms (Figure 6A and 6B; Table 1), but loss of nmur-1 can still extend the lifespan of worms carrying a null mutation in daf-16 (Figure 6C; Table 1). Thus, our data suggest that nmur-1, like at least some osm-3-expressing neurons, acts either with daf-2 but at least partly independently of daf-16, or in parallel to the daf-2/daf-16 pathway.
To identify other factors required for nmur-1 to affect lifespan, we tested how removal of nmur-1 would affect the short lifespan caused by mutations in genes proposed to act independently of daf-16 [46]–[48]. We found that loss of nmur-1 can still extend the lifespan of animals with a mutation in either (i) the AMP-dependent kinase aak-2 (Figure 6D; Table 1), which regulates energy metabolism [46]; (ii) the heat shock transcription factor hsf-1 (Figure 6E; Table 1), which regulates stress response [47],[49],[50]; or (iii) the p38 MAPK pmk-1 (Figure 6F; Table 1), which regulates innate immunity [48],[51]. Although none of these factors appears essential for nmur-1 function, we did observe partial suppression of the nmur-1 phenotype in the hsf-1 mutant background. This could suggest that nmur-1 affects lifespan by acting through several parallel pathways that include hsf-1 and/or daf-16.
Food is a complex environmental factor that affects many physiological processes, including lifespan. In the laboratory, C. elegans are grown on agar plates, on which the bacterial lawn that serves as the food source presumably provides a large part of the worm's chemosensory and mechanosensory inputs. Thus, the previous finding that some gustatory and olfactory neurons function either to shorten or lengthen C. elegans lifespan [2] makes it likely that food-derived cues affect longevity through the sensory system.
Some sensory neurons have been shown to be required for the prolonged lifespan conferred by DR [3],[38], i.e., under conditions of limited food availability. However, the fact that the sensory system also modulates lifespan when food is abundant suggests that the sensory influence on lifespan involves more than one mechanism, as illustrated in this and previous studies [1],[2].
If food-derived cues alter lifespan through the sensory system, then it is likely that impairment of a specific set of sensory neurons that detect a given set of cues would affect lifespan only on some food sources. In this study, we provide a detailed investigation of the interdependence between food and sensory perception in regulating C. elegans longevity. We show not only that wild-type lifespan is modulated by different E. coli food sources (Figure 1A and 1B) but also that three genes, which have been shown to be expressed and/or act in sensory neurons, have food source-dependent effects on lifespan.
Mutations in two of these genes—osm-3 and nmur-1—increase lifespan on OP50 but not on HT115 (Figures 1D and 2; Tables 1 and S3). Since the effects of these mutations are non-additive (Figure 3C and 3D; Table 1), this suggests that osm-3 and nmur-1 influence lifespan through a common mechanism. On the other hand, a mutation in the third gene, daf-10, not only extends lifespan on both OP50 and HT115 (Figure 1C; Table 1) but also alters the food type-dependence of the nmur-1 effect on lifespan (Figure 3A and 3B; Table 1).
Together with their requirement in the formation of the sensory cilia in subsets of neurons [15],[16], the osm-3 and daf-10 data are consistent with a role for sensory perception in the food source-dependent effects on lifespan. In addition, the identification of a neuropeptide receptor gene, nmur-1, that interacts with osm-3 and daf-10 (Figures 3, 4F and 4G; Table 1) suggests a mechanism through which the sensory system mediates the effects of specific food cues on lifespan. The nmur-1 expression in sensory neurons and interneurons (Table 2) suggests that nmur-1 modulates the transduction of signals downstream of the sensory receptors. Based on the observed interactions among these three genes, we propose the following model: (i) osm-3-expressing sensory neurons detect the presence of certain food-derived cues and transmit this information through an nmur-1-dependent pathway, and (ii) a different set of daf-10-expressing neurons detects other food cues, some of which inhibit nmur-1 activity.
According to this model, the expression patterns (Table S1) of osm-3 [16] and daf-10 [15] should help define the candidate sensory neurons that might recognize the food cues that shorten or extend lifespan through nmur-1. daf-10 is necessary for proper cilia morphology in the mechanosensory CEP neurons and some unidentified neurons in the head and tail sensory organs called the amphids and phasmids, respectively [15]. Several amphid neurons also express osm-3 [16]: these include two pairs of gustatory neurons, ASI and ASG, that have been found to shorten lifespan on OP50 [2], and two other gustatory neuron pairs that co-express nmur-1—ADF, which by itself has no lifespan effect on OP50 [2], and ADL. In addition, osm-3 is expressed outside of the amphid organs in the IL2 inner labial head neurons and in the phasmid tail neurons [16], all of which have been proposed to have chemosensory function [15],[52].
Our discovery of a food source-dependent function for the C. elegans nmur-1 gene is consistent with the known food-associated activities of other members of the NMUR signaling pathway in mammals [24],[53] and insects [54],[55]. In mammals, NMUR2, the receptor isoform expressed in the central nervous system, and its ligand, the octapeptide NMU-8, have been implicated in the regulation of food intake and energy expenditure [24],[53]. In Drosophila, the gene hugin encodes two of the peptide ligands, PK-2 and HUG-γ, recognized by two of four NMUR isoforms [54]–[56]. hugin regulates not only the food-seeking behavior and feeding rate of larvae but also affects the rate of food intake of adult flies in a food type-dependent manner [54]. Like hugin, we find that nmur-1 exerts food-type specific effects on feeding rate (Figure S2A and S2C), although the nmur-1 regulation of this process appears to be parallel to its regulation of lifespan (Figure 5A). Similar to the neuronal expression of nmur-1, Drosophila hugin is expressed in interneurons that appear to relay gustatory information [54]. At present, a potential role for the fly or mammalian NMUR signaling pathways in the regulation of lifespan has not been reported. However, the evolutionary conservation of several aspects of NMUR signaling leads us to speculate that the effects on lifespan by this system might also be conserved across species.
The Drosophila NMU signaling system also includes a second neuropeptide precursor gene, capability, that encodes three other peptide ligands, CAPA-1, CAPA-2, and CAPA-3 (also called PK-1), that can activate three of the fly NMUR isoforms [56]. The C. elegans homolog of capability, nlp-44, has recently been identified [57]. Like capability, it is predicted to give rise to three peptides, one of which activates the receptor encoded by nmur-2 [57]. A mutation of nmur-2 gives no lifespan phenotype on the food sources we have tested (Table S2), but it will be interesting to determine whether peptides derived from nlp-44 can also activate NMUR-1.
A role of nmur-1 in the sensory influence on lifespan is supported by its expression in a number of sensory neurons and interneurons (Table 2). However, it remains possible that sensory cues regulate nmur-1 activity at the level of the somatic gonad, which is the only non-neuronal tissue that expresses the nmur-1 reporter gene (Table 2). At the same time, the expression of nmur-1 in a relatively large number of cells also makes it likely that the parallel effects of nmur-1 on lifespan, feeding rate, development, and reproduction (Figures 5A–5D, S2, and S3) are mediated by its activity in different subsets of cells.
The food source-dependent activities of nmur-1 raise the possibility that other neuropeptide signaling pathways—many of which are associated with the sensory system [18]–[20],[25] —will also affect lifespan or other aspects of physiology only under specific conditions. Although most of the neuropeptide signaling pathways we have screened so far on two food sources show no effect on lifespan (Table S2), it remains possible that they will have effects on other food types. Thus, the large repertoire of neuropeptides and their receptors in C. elegans might serve to translate environmental complexity into appropriate physiological responses.
We find that wild-type worms live shorter on the E. coli B strains BL21 and OP50 than on K-12 strains, like HT115 and DY330 (Figure 1A and 1B). Conversely, the nmur-1 mutation causes reproducible lifespan extensions on the B strains but not on the K-12 strains (Figure 2; Tables 1 and S3). Since B and K-12 strains differ in their LPS structure, we have tested the lifespan effects of specific alterations in the K-12 LPS that mimic aspects of the B strain LPS (Figure 4A). Although the effect of LPS on wild-type lifespan is not large, wild-type worms do live longer on full-length than on truncated forms of the K-12 LPS (Tables 1 and S3). We also find that the nmur-1 effect on lifespan is LPS-dependent and suppressed by full-length K-12 LPS but not by its truncated versions (Figure 4; Tables 1 and S3).
Although the LPS experiments were carried out in isogenic bacterial backgrounds, the effects of the LPS alterations might be indirect since they could lead to secondary changes in bacterial metabolism or surface structure. Indeed, LPS truncations have been shown to interfere with the expression of outer membrane proteins, increase capsule polysaccharide levels, and redistribute phospholipids from the inner to the outer leaflet of the outer membrane ([58] and references therein). However, these secondary changes have only been observed with mutations that disrupt the inner core of the LPS, like the mutation present in the CS2429 strain (Figure 4A), and thereby compromise the integrity of the outer membrane [58]. No such effects have been reported for truncations that affect only the LPS outer core, like the mutation in CS2198 (Figure 4A). Thus, the observation that nmur-1 extends lifespan on both CS2429 and CS2198 argues for a direct effect of the bacterial LPS on worm lifespan. Direct recognition of LPS is biologically plausible: LPS is the predominant component of the outer membrane of gram-negative bacteria and is consequently used by multicellular organisms from diverse phyla to recognize bacteria in the context of defense against pathogens [59],[60].
Nevertheless, the LPS structure is clearly only one of potentially many food-derived cues that influence worm lifespan. This is most evident from the LPS-independent lifespan phenotype of osm-3 mutants (Figure 4F and 4G; Table 1), and from the fact that the lifespan extension by the nmur-1 mutation is greater on OP50 than on any other strain with a similar, short LPS (Figure 2; Tables 1 and S3). Thus, changes in lifespan are likely triggered by different sets of sensory neurons in response to a variety of food-derived cues, and loss of nmur-1 interferes with the detection of several of these cues.
The LPS dependence of the nmur-1 phenotype makes it conceivable that nmur-1 may regulate stress-related and innate immune responses elicited by different food sources. We find that nmur-1 can still affect lifespan in the absence of either of three genes, daf-16, hsf-1, and pmk-1, all of which have major roles in stress responses and innate immunity [47]–[51],[61]–[63]. However, the mutations in daf-16 and hsf-1 can partly suppress the nmur-1 lifespan phenotype (Figure 6; Table 1), which makes it possible that the nmur-1 influence on lifespan requires a combination of mechanisms that involve daf-16, hsf-1, and/or other factors.
We find that the food-source influence on wild-type lifespan is strongly correlated with reproductive effects (Figure 5C and 5D), in that increases in lifespan are accompanied not only by a decreased number of total progeny but also a faster rate of reproduction. One possible interpretation of these data is that the different reproductive profiles cause the food source-dependent differences in wild-type lifespan. Indeed, with the exception of BL21, the bacterial diets we have tested seem to affect initial survival more than late-age survival. This is supported by age-specific force of mortality plots (Figure S5A): the different food sources alter wild-type mortality primarily before day 10 of adulthood but have little effect thereafter. It is conceivable that damage inflicted on somatic tissues [64] or neglect of somatic maintenance and repair during reproduction [65] are important determinants of early mortality. In agreement with this idea, we find that long-lived glp-1 mutant worms [66], which are sterile because they generate few or no germ cells [67], have very similar lifespan, at least on OP50, HT115, CS180, and CS2429 (W. M., unpublished data). This suggests that the food type-dependent effects on wild-type lifespan are indeed germline-dependent.
Interestingly, a recent study [68] has shown that different E. coli food sources can differentially affect fat storage in C. elegans. Wild-type worms grown on HB101 or HT115 are found to have lower triacylglyceride (TAG) levels than wild-type worms grown on OP50 [68]. Although that same study and another report [68],[69] question the reliability of fat stains with vital dyes, we also observe a slightly reduced fat storage in wild type on HT115 compared to wild type on OP50, using lipid labeling with a lipophilic fluorophore (Table S4). Thus, a correlation may exist not only between lifespan and reproduction but also between lifespan and TAG levels of wild-type worms. Since germline signals have been proposed to regulate both lifespan and intestinal fat storage [70], the food-type and reproduction-dependent effects on wild-type lifespan may also be mediated by changes in TAG levels.
In contrast, we find that nmur-1 exerts an additional effect on lifespan that is largely independent of reproduction (Figure 5C and 5D) and also appears to be independent of glp-1 on OP50 and CS2429 (B. A. and W. M., unpublished data) and fat storage on OP50 and HT115 (Table S4). Accordingly, the nmur-1 mutation can affect mortality prior to day 10 of adulthood (OP50 and CS180; Figure S5B) on the food sources that significantly reduce the total progeny of nmur-1 mutants (compare OP50 and CS180 in Figures 5C, S2E, and S3C). At the same time, nmur-1 mutants show reduced mortality after day 10, but not past day 16, of adulthood on the short LPS strains OP50 and CS2429 (Figure S5B), the latter of which has no effect on the nmur-1 mutant number of progeny (Figure S3C). Thus, our findings imply that food sources affect lifespan through both reproduction-dependent and reproduction-independent mechanisms, with the second being uncovered by the nmur-1 mutation.
Unlike the longevity-promoting effect of food-level restriction [5],[41], the food type-dependent effects on lifespan that we observe not only have reproduction-independent and fat storage-independent components (Figure 5C and 5D; Table S4) but are also independent of alterations in feeding rate and developmental rate (Figure 5A and 5B). In addition, our data show that different food types and nmur-1 affect initial mortality without decreasing late-age mortality (Figure S5), again unlike food-level restriction, which decreases the slope of the mortality trajectory and thus slows the rate of aging [71]. These data lead us to propose that these two forms of dietary influence on lifespan employ distinct, but possibly overlapping, mechanisms.
Another recent study [72] has shown that different DR regimens for C. elegans require different signaling pathways to affect lifespan. However, some of these regimens altered not only food levels but also the nature of food sources. In fact, at least one of these protocols, which lowered protein levels, does not decrease but increase reproduction ([72] and references therein), which suggests that the lifespan effect of protein restriction, unlike that of other DR protocols, could be partly reproduction-independent. Our data might help explain some of these findings, if one assumes that the net consequence on lifespan of some DR protocols represents a mix of independent effects from food-level restriction and food-type dependence. In the future, it would be of interest to determine whether the food type-dependent effects on lifespan will also require the activities of genes, e.g., the NFE2-related protein skn-1 [38] and the FOXA transcription factor pha-4 [73], that have been implicated in the longevity-promoting effects of DR.
All worm mutant strains used in this study were backcrossed six times to our lab wild-type (N2) strain, with the exception of nmur-1(ok1387), which was backcrossed eight times, and eat-2(ad1116), which was outcrossed once, before generation of different mutant combinations and any phenotypic analysis. The different worm mutant alleles used are indicated within the figures, supplementary tables, and their legends. Worms were grown for at least two generations at 25°C on the same food source used in a given phenotypic analysis, unless otherwise stated.
The E. coli strains used were: OP50 [11], HT115 [rnc14::ΔTn10 λ(DE3) of W3110] [13],[26],[27], BL21(DE3) [28], DY330(DE3) [Δ(argF-lacZ)U169 gal490*(IS2) pglΔ8 rnc<>cat λcI857 Δ(cro-bioA) of W3110] [30], HB101 [29], DH5α [31], CS180 [rfa+] [35], CS2198 [rfaJ19::Tnlac Δlac pyrD+ of CS180] [35], CS2429 [rfaC− of CS180] [36], and CS1861 (CS180 transformed with a plasmid that confers chloramphenicol resistance and encodes the proteins required for the expression of Shigella dysenteriae 1 O Antigen fused to the parent strain K-12 LPS) [36].
We generated two independent rescue lines using standard methods: nmur-1(ok1387); jxEx12[nmur-1p::nmur-1+myo-3p::rfp] and nmur-1(ok1387); jxEx40[nmur-1p::nmur-1+myo-3p::rfp]. The rescue fragment, which is a 7.96 kb-long PCR fragment of the wild-type nmur-1 genomic locus (injected at 100 ng/µl), includes the 2.9 kb sequence upstream of the nmur-1 start codon and the 1 kb sequence downstream of the correct stop codon (see Figure S1). The myo-3p::rfp (gift of Cori Bargmann) was used as a coinjection marker (injected at 100 ng/µl). As controls, we also generated wild-type and nmur-1 mutant worms that carry the myo-3p::rfp coinjection marker alone.
We observed that the extrachromosomal array jxEx12 has a large number of arrested embryos and larvae, whereas the extrachromosomal array jxEx40 produces ∼13% arrested larvae (25 arrested worms/196 total worms). These additional phenotypes might be due to a hyperactive NMUR-1 pathway caused by overexpression of the gene from its extrachromosomal copies.
To determine the expression pattern of nmur-1, we generated a transcriptional gfp reporter construct (nmur-1p::gfp; based on the pPD117.01 vector; gift from A. Fire), in which the gfp is flanked by the 2.9 kb sequence upstream of the nmur-1 start codon and by the 1 kb sequence downstream of the correct stop codon, including the newly identified 3′ UTR (see Figure S1). In addition, sequences from the four largest introns, 1, 4, 8, and 10, which may contain regulatory sequences required for expression, were fused downstream of the 1 kb 3′ cis sequences. This construct was injected into wild-type worms at a concentration of 100 ng/µl, and two independent transgenic lines, jxEx36 and jxEx37, were recovered, which show identical patterns of gfp expression.
All bacterial strains were grown from single colonies in Luria-Bertani medium overnight at 37°C. However, the medium used to grow the chloramphenicol-resistant strain CS1861 was supplemented with 100 µg/ml chloramphenicol. Nematode-growth agar plates (6 cm in diameter; [11]) were seeded with 100 µl bacterial culture and were allowed to dry at room temperature (23°C). Seeded plates were stored at room temperature and used within 5 d.
The survival analyses of all worm strains on the different bacteria were initiated on the first day of adulthood and carried out at 25°C. Throughout their reproductive period, the worms were transferred daily to new plates to separate them from their progeny. We used the JMP 5.1 (SAS) software to determine Kaplan-Meier estimates of survival probabilities and mean lifespan, and for all statistical comparisons. p values were determined by the logrank and Wilcoxon tests. The logrank test, which places more weight on larger survival times, is appropriate when comparing differences between groups of animals whose ratio of hazard functions (ratio of mortality rates) stays approximately constant over time [74]. However, when the hazard ratios do not stay constant with time, as when one survival curve shows more early deaths than another (e.g., wild type on OP50 versus wild type on HB101 or HT115 in Figure 1A and 1B), the Wilcoxon test is more appropriate for comparing differences between groups [74]. We found that the Wilcoxon test is more sensitive to the lifespan differences we see in most of our experiments, since the nmur-1 mutation and most bacterial food sources clearly affect mean lifespan more than the maximum lifespan, which in fact violates the logrank test assumption of constant hazard ratios. Here we refer to a Wilcoxon p value of ≤0.01 as a significant difference between the various groups of animals. For comparison, we report both the Wilcoxon and logrank test results in all tables.
For mortality plots, the age-specific force of mortality was calculated as Fx = −ln(1−Dx), where Dx is the probability of death between day x−1 and x of adulthood [75]. At least five independent trials of a given lifespan experiment were used to calculate means and standard errors of Fx, which were plotted on a log scale against age.
Feeding rates were determined on the first and fourth days of adulthood at 25°C by measuring the animals' pharyngeal pumping rates, which reflect the rates at which they eat bacteria [76]. The pumps of the pharyngeal bulbs of individual worms were counted 3 to 5 times over periods of 30 s. Each resulting mean value was then doubled to get “pumps per minute.” A two-way ANOVA test was used to compare the different genotypes on different food sources and p values were calculated with the Tukey post-test.
Developmental rate differences were determined through a population-based assay at 25°C. First-stage (L1) larvae that had hatched within a 2-h time window were collected and allowed to develop for 36.5 h. At this point, the number of second-stage (L2), third-stage (L3), and fourth-stage (L4) larvae, as well as of young adult (YA) or gravid adult (GA) worms were counted. The chi-square test was used to compare the resulting stage distributions across food sources or worm genotypes.
Total progeny and temporal profiles of egg-laying were determined at 25°C by culturing L4 larvae singly on plates of the appropriate food source. The worms were then transferred to new plates regularly until they stopped laying eggs. The eggs were allowed to hatch and the larval progeny were then counted. Two-way ANOVA and the Tukey post-test were used to compare the total number of progeny of genotypes across food sources. To ensure that the data followed a normal distribution, it was necessary to incorporate a statistical censoring procedure to exclude outliers (worms with a very low number of progeny) from the data set before the ANOVA test. Briefly, this involved the tentative identification of outliers and calculation of standard deviation (SD) for the remaining set. Then, from the full data set, we excluded worms that had produced less progeny than the mean minus 2.5 times SD. In general, this procedure led to exclusion of worms with a progeny number smaller than 90, which corresponded to ∼4% of the total data set. The exception is nmur-1 mutant worms feeding on HB101, for which two classes of worms seem to exist: one with a large number of progeny and another with a small number of progeny. In this particular case, censoring caused 25% of worms to be excluded from the analysis and the remaining data set to be biased considerably towards a larger progeny number, as can be seen in Figure S2E.
The temporal profiles of egg-laying were determined from the same statistically censored populations of worms. The Hill function, P(t) = Pmax * tn/(tn+t50n), was used to fit the cumulative number of progeny over time, where t denotes time, Pmax is the total number of progeny, n the Hill coefficient, and t50 the time until half of the progeny is produced. In Figures S2F and S3D, the data were normalized to Pmax.
For statistical assessments of correlations between mean lifespan and feeding, development, or reproduction on different food sources, we used the Pearson Product Moment test. To determine the correlation between lifespan and development, the stage distributions of the original data were used to calculate “speed of development” values, which are the percentages of worms scored as either young or gravid adults in the corresponding assays. eat-2 mutants have a value of zero on this scale because no mutant worms reached adulthood within 36.5 h after egg-laying. To correlate lifespan and rate of reproduction, the t50 of the fitted temporal reproduction profiles was used.
|
10.1371/journal.ppat.1004934 | Pseudomonas aeruginosa ExoT Induces Atypical Anoikis Apoptosis in Target Host Cells by Transforming Crk Adaptor Protein into a Cytotoxin | Previously, we demonstrated that Pseudomonas aeruginosa ExoT induces potent apoptosis in host epithelial cells in a manner that primarily depends on its ADP-ribosyltransferase domain (ADPRT) activity. However, the mechanism underlying ExoT/ADPRT-induced apoptosis remains undetermined. We now report that ExoT/ADPRT disrupts focal adhesion sites, activates p38β and JNK, and interferes with integrin-mediated survival signaling; causing atypical anoikis. We show that ExoT/ADPRT-induced anoikis is mediated by the Crk adaptor protein. We found that Crk-/- knockout cells are significantly more resistant to ExoT-induced apoptosis, while Crk-/- cells complemented with Crk are rendered sensitive to ExoT-induced apoptosis. Moreover, a dominant negative (DN) mutant form of Crk phenocopies ExoT-induced apoptosis both kinetically and mechanistically. Crk is generally believed to be a component of focal adhesion (FA) and its role in cellular survival remains controversial in that it has been found to be either pro-survival or pro-apoptosis. Our data demonstrate that although Crk is recruited to FA sites, its function is likely not required for FA assembly or for survival per se. However, when modified by ExoT or by mutagenesis, it can be transformed into a cytotoxin that induces anoikis by disrupting FA sites and interfering with integrin survival signaling. To our knowledge, this is the first example whereby a bacterial toxin exerts its cytotoxicity by subverting the function of an innocuous host cellular protein and turning it against the host cell.
| We have previously demonstrated that ExoT is both necessary and sufficient to induce potent apoptosis in host epithelial cells in a manner that depends primarily on its ADP-ribosyltransferase (ADPRT) domain activity. However, the molecular basis underlying ExoT/ADPRT-induced apoptosis remains unknown. In this study, we demonstrate that ExoT/ADPRT by targeting the adaptor protein Crk, transforms this innocuous cellular protein into a cytotoxin that induces atypical anoikis by disrupting the focal adhesion sites which in turn interferes with the integrin-mediated pro-survival signaling.
| Pseudomonas aeruginosa is a Gram-negative opportunistic pathogen that targets immunocompromised individuals and those with injured epithelia, making it one of the leading causes of nosocomial infections and the leading cause of morbidity and mortality in cystic fibrosis patients [1–3]. P. aeruginosa boasts a large arsenal of cell surface-associated and secreted virulence factors [4]. Prominent amongst them is the Type III Secretion System (T3SS) which contributes to the virulence of a large number of Gram-negative pathogens [5,6]. This conduit allows P. aeruginosa to directly translocate a set of peptide virulence factors, termed effector proteins, into the eukaryotic host cell, where they subvert host signal transduction pathways to advance P. aeruginosa infection [7]. To date, four T3SS effectors have been identified in P. aeruginosa: ExoU, ExoT, ExoS, and ExoY. ExoU is a potent phospholipase that induces necrotic cytotoxicity in eukaryotic cells [8,9]. ExoS and ExoT are homologous bifunctional proteins with an N-terminal GTPase activating protein (GAP) domain and a C-terminal ADP-ribosyltransferase (ADPRT) domain [10,11]. The GAP domains of ExoT and ExoS inhibit RhoA, Rac1, and Cdc42, small GTPases [12–15], while the ADPRT domains of ExoS and ExoT modify non-overlapping host targets. The ExoS ADPRT domain targets many host proteins including Ras, Ral, Rab, and Rac, and the ADPRT activity of ExoT modifies CrkI/II (the two isoforms of Crk) adaptor proteins and the glycolytic enzyme PGK1 [16–21]. Finally, ExoY is an adenylate cyclase that functions as an edema factor [22,23].
Unlike exoS, exoU, and exoY which are encoded in subsets of clinical isolates, exoT is present in almost all P. aeruginosa virulent clinical strains studied thus far [24,25], suggesting a more fundamental role for this virulence factor in P. aeruginosa pathogenesis. Indeed, P. aeruginosa strains defective in ExoT exhibit reduced virulence and are impaired in dissemination in mice [11,18,26]. Moreover, Balachandran et al. recently demonstrated an elegant host defense mechanism involving ubiquitin ligase Cbl-b that specifically targets ExoT, but not ExoS or ExoU, for proteasomal degradation [26]. This finding further highlights the importance of ExoT in P. aeruginosa pathogenesis and host responses to this pathogen. We and others have demonstrated that ExoT alters actin cytoskeleton, causes cell rounding, inhibits cell migration, functions as an anti-internalization factor, blocks cell division by targeting cytokinesis at multiple steps, and inhibits wound healing [12,13,18,27]. More recently, we demonstrated that ExoT is both necessary and sufficient to induce apoptosis in HeLa cells in a manner that is primarily dependent on its ADPRT domain activity [28]. However, the mechanism underlying the ExoT-induced apoptosis in epithelial cells remains unknown.
In this report, we demonstrate that ExoT-induced apoptosis is mediated by the Crk adaptor protein. Our data strongly suggest that ExoT/ADPRT activity, by ADP-ribosylating Crk, transforms this innocuous cellular protein into a cytotoxin that causes atypical anoikis by interfering with integrin-mediated survival signaling.
Most ExoT or ExoT/ADPRT-intoxicated HeLa cells exhibited movement after cell rounding and prior to succumbing to death, as determined by the uptake of propidium iodide (PI) impermeant nuclear stain, which fluoresces red in dead or dying cells [28,29] (Fig 1A, S1 Movie). This type of cell death morphologically resembled an apoptotic programmed cell death known as anoikis, which occurs as a consequence of loss of cell adhesion and/or inappropriate cell/matrix interaction [30]. Depending on the cell line or the environmental cues, anoikis can be initiated and executed by different pathways, including the intrinsic and the extrinsic apoptotic pathways [30]. However, some common features have emerged. The common hallmarks of anoikis include: enhanced and persistent activation of p38β and JNK by phosphorylation, which is required for anoikis cell death; degradation of p130Cas and paxillin focal adhesion proteins; down activation of FAK, and down-regulation of integrin-mediated survival signaling [30–32].
We conducted a time course infection study to evaluate the possibility that ExoT/ADPRT activity induced anoikis in epithelial cells. HeLa cells were infected at a multiplicity of infection (MOI) of 10 with isogenic mutants of the PA103 strain, a clinical isolate which encodes and expresses ExoU and ExoT [24,25], including PA103∆exoU (∆U) which carries an in-frame deletion in the exoU gene but expresses ExoT; PA103∆exoU/exoT(R149K) (∆U/T(G-A+)) which carries an in-frame deletion in the exoU gene but expresses ExoT with a mutant GAP but functional ADPRT domain; or PA103 pscJ::gentR (T3SS mutant, unable to deliver ExoT into host cells) (S1 Table). To be able to focus on ExoT-induced cytotoxicity, we conducted these studies in the exoU-deleted PA103 genetic background (∆U), as we have previously described [27,28]. Moreover, because the T3SS alone induces necrotic cytotoxicity, which is completely abrogated by ExoT [28], we also left out the effectorless T3SS proficient strain (∆U∆T).
In line with our hypothesis, infection with ExoT or ExoT/ADPRT expressing P. aeruginosa strains resulted in substantial and persistent activation of both p38β and JNK by 4hr post-infection (Fig 1B). To ensure that ExoT/ADPRT activity was sufficient to activate p38β and JNK in the absence of other bacterial factors, HeLa cells were transiently transfected with pIRES2 mammalian expression vectors, harboring wild-type ExoT (pExoT), ExoT with functional ADPRT (pExoT(G-A+)), or the inactive form of ExoT (pExoT(G-A-)), all C-terminally fused to GFP, or empty vector (pGFP) (S1 Table). GFP fusion does not alter ExoT’s virulence functions [27,28]. JNK and p38β activation was evaluated by immunofluorescent (IF) microscopy of fixed cells ~24 hr post transfection. Despite reduced expression of ExoT and ExoT(G-A+) (Fig 1G) due to Cbl-b mediated proteasomal degradation of ExoT and ExoT(G-A+) [26], transient transfection with pExoT-GFP and pExoT(G-A+)-GFP resulted in significant increases in p38β and JNK activation as compared to pExoT(G-A-)-GFP or the pGFP empty vector (Fig 1C–1F), indicating that ExoT/ADPRT activity is sufficient to activate p38β and JNK.
Activation of p38β and JNK leads to disruption in integrin-mediated survival signaling, culminating in anoikis [30,31,33]. We conducted similar infection studies to evaluate the impact of ExoT/ADPRT on integrin-mediated survival signaling using Akt activation and β-catenin activity as readouts. The Akt/β-catenin pathway is an important integrin-mediated survival signaling pathway whose disruption results in anoikis [32,34–36]. Integrin interaction with extracellular matrix (ECM) leads to activation (phosphorylation) of Akt. Phospho-Akt (p-Akt) in turn activates β-catenin by blocking its inhibitor GSK-3β through phosphorylation. (Unphosphorylated GSK-3β is an inhibitor of Akt/β-catenin mediated survival signaling as it targets β-catenin for proteasomal degradation [32,34–36]). Once activated, the β-catenin transcription factor turns on the expression of pro-survival proteins [34]. In line with our hypothesis, infection with ExoT or ExoT/ADPRT-expressing strains (∆U and ∆U/T(G-A+) respectively) reduced Akt activation (p-AktS473) levels by 6hr and p-Akt-mediated GSK-3β phosphorylation (inactivation) by 8hr post-infection (Fig 2A). This resulted in substantial reduction in β-catenin levels (Fig 2A) and β-catenin activity, particularly by 8hr post-infection (Fig 2B, n = 3, p<0.001).
HeLa S3 cells are derived from HeLa cells through planktonic growth and are frequently used as anoikis resistant cells because their survival is adhesion independent [37,38]. We reasoned that if ExoT or ExoT/ADPRT induced anoikis in the target cells, HeLa S3 should be resistant to their cytotoxicity. To test this hypothesis, we infected HeLa and HeLa S3 cells with ExoT-expressing ∆U, ExoT/ADPRT-expressing ∆U/T(R149K), T3SS mutant pscJ, or PBS (Mock). Cytotoxicity was observed by time-lapse videomicroscopy and measured every 15 minutes using PI uptake as a marker for cell death [29,39]. Consistent with our previous report [28], HeLa cells were sensitive to ExoT and ExoT/ADPRT-induced cytotoxicity (Fig 3A and 3C and S2 Movie). Despite similar levels of ExoT intoxication (Fig 3E), HeLa S3 cells displayed resistance to ExoT and ExoT/ADPRT-induced cytotoxicity (Fig 3B and 3D and S3 Movie). Taken together, these data indicated that ExoT and ExoT/ADPRT induced anoikis by interfering with integrin survival signaling.
The ADPRT domain of ExoT ADP-ribosylates a conserved arginine residue in the SH2 domain of CrkI and CrkII isoforms of Crk, disrupting Crk SH2 interactions with its cognate substrates [17,40]. Crk has been implicated in cell death, although its role in cytotoxicity remains controversial. Depending on cell type or physiological condition, Crk has been found to be either pro-apoptotic [41–45] or pro-survival [46,47]. We wished to examine the possible role of Crk in the ExoT/ADPRT-induced apoptosis in HeLa cells.
We transfected HeLa cells with a mammalian expression vector harboring wild type CrkI (pCrkI) or the CrkI/R38K mutant form (in which the conserved arginine 38 residue in the SH2 domain was mutated to lysine), fused at their C-termini to GFP (S1 Table), and assessed apoptosis by IF videomicroscopy in the presence of PI, as we described previously [28]. Because of low transfection efficiencies, we were unable to assess the impact of CrkII or CrkII/R38K mutation on survival in HeLa cells. Of note, CrkI/R38K has been shown to act as a DN mutant, interfering with CrkI and CrkII associated cellular activities [48,49]. Despite similar transfection efficiencies (Fig 4D), transfection with CrkI/R38K SH2 DN phenocopied ExoT and ExoT/ADPRT-induced apoptosis as it significantly increased Z-VAD sensitive apoptotic cell death in HeLa cells (Fig 4A–4C and S4 Movie; n = 610; p<0.001). The time to death, defined as the start time of gene expression as determined by GFP expression, to the time of death as determined by PI uptake, in the presence CrkI/R38K SH2 DN mutant was 9.1 ± 1.1 hr which is also nearly identical to the observed time to death in the presence of ExoT (8.5 ± 1.3 hr) or ADPRT (8.6 ± 0.8 hr) ([28] and Fig 4E). Similar to ExoT and ADPRT, transfection with CrkI/R38K SH2 DN also resulted in activation of JNK and p38β in HeLa cells (S1 Fig, n = 122 and n = 124 respectively, one-way ANOVA, p<0.001). Moreover, HeLa S3 cells were also completely resistant to CrkI/R38K SH2 DN-induced apoptosis (S2 Fig). Combined, these data indicated that ExoT-mediated apoptosis is phenocopied by the CrkI/R38K SH2 DN mutant form.
So far our data indicate that ExoT/ADPRT induces anoikis in a manner that likely involves Crk. The two most likely scenarios are: 1) Crk is required for survival and its modification by ExoT/ADPRT prevents it from performing its survival function; or 2) Crk is not required for survival but when it is ADP-ribosylated by ExoT/ADPRT, it is transformed into a cytotoxin which disrupts survival functions inside the cell. We favored the second scenario because; although, Crk-null (Crk-/-) mice die shortly after birth [50], Crk-/- cells not only survive, they are actually more resistant to apoptosis [42]. These reports strongly suggest that while Crk is required for development, Crk function is not essential for cellular survival per se.
We reasoned that if ExoT or ExoT/ADPRT-induce apoptosis in epithelial cells by transforming Crk, Crk-/- cells should be resistant to ExoT and ExoT/ADPRT-induced cytotoxicity. In contrast, complementing Crk-/- cells with Crk should then restore their sensitivity to ExoT and ExoT/ADPRT-induced apoptosis. To address these possibilities, we complemented Crk-/- mouse embryonic fibroblasts (MEFs) with either CrkI-GFP or GFP alone by transfection. The pCrkI-GFP or pGFP-transfected Crk-/- cells were then infected with ExoT-expressing ∆U, ExoT/ADPRT-expressing ∆U/T(G-A+) or T3SS mutant pscJ and assessed for cytotoxicity by IF timelapse videomicroscopy. Although more cytotoxicity was observed in Crk-/- cells that were infected with P. aeruginosa, regardless of the strain, when compared to HeLa cells, Crk-/- cells complemented with CrkI-GFP were significantly more sensitive to ExoT or ExoT/ADPRT-induced cytotoxicity, as compared to GFP complemented Crk-/- cells (Fig 5B and S5 and S6 Movies. Representative frames are shown in Fig 5A). Additionally, the time to death, (defined as the time of infection to the time cell death manifested by PI staining), in the presence of ∆U or ∆U/T(R149K), was also significantly faster in Crk-/- cells complemented with CrkI-GFP, as compared to the GFP-complemented Crk-/- cells that succumbed to death in response to infection with P. aeruginosa (Fig 5C).
If CrkI SH2 domain modification by ExoT ADP-ribosylation or by mutagenesis (CrkI/R38K) renders CrkI into a cytotoxin, expression of CrkI/R38K SH2 mutant, but not CrkI, should also result in cytotoxicity in Crk-/- cells, phenocopying ExoT’s effect in CrkI-complemented Crk-/- cells. Consistent with this view, transient transfection with the CrkI/R38K SH2 mutant resulted in significantly more cytotoxicity in Crk-/- cells, as compared to CrkI (Fig 6 and S7 Movie). Transfection with CrkI/R38K, W170K, which harbors null mutations in both SH2 and SH3 domains of CrkI [48], did not result in cytotoxicity in Crk-/- (Fig 6), indicating that the SH3 domain of CrkI must be functional for CrkI/R38K mutant to function as a dominant negative (DN) and a cytotoxin. Corroborating this hypothesis, complementing Crk-/- cells with CrkI/R38K, W170K double mutant did not render Crk-/- cells sensitive to ExoT or ExoT/ADPRT-induced cell death (S3 Fig, S8 Movie), indicating that the SH3 domain of CrkI must be functional for the ADP-ribosylated CrkI to act as a cytotoxin. As expected, Crk-/- cells did not express CrkI/II isoforms of Crk (S4 Fig).
We next sought to determine how ADP-ribosylation of CrkI by ExoT/ADPRT domain activity or by mutagenesis (R38K mutation) could potentially transform this cellular protein disruptive to integrin-mediated survival signaling. We hypothesized that CrkI ADP-ribosylation by ExoT/ADPRT could disrupt integrin-survival signaling by destabilizing the focal adhesion (FA) sites. We based this hypothesis on the knowledge that FAK activation and FAK/p130Cas interactions at FA sites stabilize the integrin/extracellular matrix (ECM) interactions and are required for survival signaling [51–53]. During FA assembly, integrin/ECM interaction results in activation (phosphorylation) of focal adhesion kinase (FAK), which in turn phosphorylates p130Cas and paxillin [54,55]. Phosphorylated p130Cas and paxillin then recruit Crk to FA by directly interacting with its SH2 domain [56]. Although ExoT’s impact on FA has not been investigated, Crk interactions with p130Cas and paxillin have been shown to be disrupted by ExoT/ADPRT activity [40]. Therefore, while FAK, p130Cas, and paxillin localization to FA is upstream of Crk, disruption of Crk activity by ExoT or by CrkI/R38K mutation could potentially disrupt FAK, p130Cas, and paxillin subcellular localization, and/or their maintenance, and/or their activation at FA sites. This would in turn interfere with integrin/ECM survival signaling and would lead to anoikis [53,57].
To test this hypothesis, we performed the aforementioned time-course infection studies with ExoT-expressing ∆U, ExoT/ADPRT-expressing ∆U/T(G-A+), and the T3SS mutant pscJ, and assessed the impact of ExoT or ExoT/ADPRT on FA sites by determining the total number of FAK and p-130Cas positive FA puncta per cell and the intensity of FAK and p-130Cas stainings per FA puncta by IF microscopy (described in Methods and S5 Fig). Infection with ∆U or ∆U/T(G-A+) bacteria resulted in substantial reduction in FAK and p130Cas localization to FA sites by 4h post-infection, as manifested by reduced number of FA puncta (Fig 7A–7D) and reduced staining intensities of FAK and p130Cas per FA puncta (S6A and S6B Fig). As expected, transient transfection with expression vectors harboring ExoT (pExoT) or the ExoT/ADPRT domain (pExoT(G-A+)) also resulted in significant reduction in FAK and p130Cas localization to the FA sites (Fig 7E–7H and S6C and S6D Fig), indicating that ExoT/ADPRT activity is sufficient to disrupt FA sites. Interestingly, infection with ∆U or ∆U/T(G-A+) did not significantly affect the total cellular levels of activated (phosphorylated) FAK (p-FAKY397) or activated (phosphorylated) p-130Cas (p-p130CasY165) as measured by Western blotting (Fig 7I). These data indicated that FAK and p130Cas are able to get to the FA sites where they become phosphorylated but they are not maintained at FA sites in the presence of ExoT or ExoT/ADPRT activity. Of note, in HeLa cells transfected with pCrkI-GFP, CrkI co-localized with FAK and p130Cas at FA sites (Fig 8A–8F, pCrkI panels). In contrast, CrkI/R38K did not localize to FA sites and disrupted FAK and p130Cas localization to the FA sites in HeLa cells (Fig 8A–8F, pCrkI/R38K panels), phenocopying the adverse effect of ExoT and ExoT/ADPRT on FA and supporting our hypothesis that ExoT ADP-ribosylation of CrkI renders CrkI disruptive to FA sites.
It is generally believed that Crk is an essential component of FA sites [56,58]. However, our data strongly suggested that although Crk may be recruited to FA, its function may not require for FA assembly or its dynamics. Consistent with this view, we found FAK to localize to FA sites in Crk-/- cells (Fig 8G–8I), indicating that FAK subcellular localization to FA sites and its maintenance in that compartment does not require Crk in this cell line. Interestingly, in Crk-/- cells that were transfected with CrkI-GFP, CrkI co-localized with FAK at FA sites (Fig 8G) indicating that CrkI can be recruited and maintained at FA sites in this cell line as well. Similar to HeLa, CrkI/R38K SH2 DN mutant did not localize to FA sites and disrupted FAK localization to FA sites in Crk-/- cells that were transfected with CrkI/R38K-GFP expression vector (Fig 8G–8I). Of note, CrkI/R38K, W170K (SH2 and SH3 double mutant) did not localize to FA sites and failed to disrupt FAK localization to FA sites in Crk-/- and HeLa cells (Fig 8G–8I), indicating that the SH3 domain must be functional for CrkI/R38K to act as a dominant negative.
We next asked whether CrkI presence in FA sites affected FA structures and/ or their function in in Crk-/- cells. We evaluated the impact of CrkI on FA structure and function in Crk-/- cells by determining how CrkI affected: (i) the number of FA puncta per cell, (ii) the ability of Crk-/- cells to adhere to the surface, as determined by their surface area, and (iii) the ability of Crk-/- cells to migrate, as determined by distance these cells travelled within 24 hr after transfection with CrkI. The data indicated that CrkI presence in FA sites in Crk-/- cells, complemented with CrkI, did not affect their FA structures or functions (S7 Fig). Collectively, these data indicate that while CrkI can be recruited to FA, its function is not essential for FA assembly or survival. However, when modified by ExoT or by mutagenesis, it can disrupt FA sites, which would result in anoikis.
We have recently reported that Pseudomonas aeruginosa ExoT is both necessary and sufficient to induce potent apoptosis in target host cells in a manner that primarily depends on its ADPRT domain activity [28] but the mechanism underlying ExoT/ADPRT-induced apoptosis has not been determined. We now report that ExoT/ADPRT induces anoikis by disrupting FA sites. Our data demonstrate that within 4h of exposure to ExoT or ExoT/ADPRT, FA sites are destabilized (Fig 7). Concomitant with FA disruption, intoxication with ExoT or ExoT/ADPRT leads to persistent activation of p38β and JNK (Fig 1), which are known drivers of anoikis [33,59]. This in turn leads to dampened integrin-mediated survival signaling by 6–8 hr post-exposure, as manifested by reduced Akt activation and β-catenin-mediated survival signaling (Fig 2). ExoT/ADPRT-induced anoikis is atypical in that we did not observe degradation of p130Cas or down-activation of FAK (Fig 7I), both of which have been reported to occur during anoikis and be required for cytotoxicity [30–32,53,60]. Instead, ExoT or ExoT/ADPRT interfered with FAK and p130Cas subcellular localization to the FA sites (Fig 7A–7H). Our data suggest that FA sites may serve as cellular survival centers. Consistent with this notion, JNK localization to FA sites and its interaction with FAK and p130Cas in that compartment has been shown to be required for survival [57]. Of note, ExoT and ExoT/ADPRT also disrupted JNK localization at FA sites (S8 Fig and Fig 1C, compare JNK staining at FA sites in the untransfected and pGFP- transfected cells to the pExoT or pExoT(G-A+)-transfected cells which lack JNK at FA sites).
Crk is generally believed to be an essential component of FA [56,58], although it is not clear whether it is CrkI or CrkII or both that function in FA. Moreover, Crk has also been implicated in cellular survival and/or apoptosis but Crk’s role in these processes remain controversial in that it is found to be either pro-apoptotic [41–45] or pro-survival [46,47]. Our data strongly suggest that Crk function may not be essential for FA or for cellular survival per se, but when modified by ExoT or by mutagenesis, Crk can be transformed into a cytotoxin that interferes with survival signaling by disrupting FA (Figs 5, 6 and 8). How does modification of CrkI SH2 domain by ExoT/ADPRT-mediated ADP-ribosylation render CrkI disruptive to integrin-mediated survival signaling? We propose that ADP-ribosylation by ExoT transforms CrkI into a dominant negative mutant that is disruptive to FA structures. Our reasoning is that while these ADP-ribosylated CrkI and CrkI/R38K are unable to interact through the SH2 domain with their cognate substrates such as p130Cas and paxillin [40], they would be able to interact with a number of essential FA components, such as C3G [61,62] through the SH3 domain. This prevents their localization to FA sites, and destabilizes FA structures, thus culminating in anoikis as our data demonstrate (Figs 7 and 8). Consistent with this hypothesis, CrkI/(R38K, W170K) which harbors mutations in both SH2 and SH3 domains failed to disrupt FA sites (Fig 8), or induce apoptosis in Crk-/- cells (Fig 6), or renders Crk-/- cells sensitive to ExoT or ExoT/ADPRT cytotoxicity (S3 Fig, S8 Movie).
Crk has been shown to function in various host defenses against bacterial pathogens, such as inhibiting EPEC-induced actin-based pedestal formation, increasing bacterial clearance through phagocytosis of IgG-opsonized pathogens by Fcγ receptors, and potentially enhancing innate immune activation through pattern recognition receptors (PRR) by sequestering bacterial virulence factors, such as EPEC’s Tir, which interfere with PRR signaling [63–65]. Therefore, it is not surprising that pathogens have evolved mechanisms to target this central host protein in order to advance their own agenda. For example, EPEC induces phosphorylation of the major regulatory tyrosine in CrkII, preventing CrkII from sequestering Tir, thus freeing this virulence factor to promote pedestal formation [63]. Additionally, P. aeruginosa blocks host cell proliferation by inhibiting Crk’s essential function for cytokinesis [27].
In summary, we propose that ExoT by ADP-ribosylating Crk transforms this cellular protein into a cytotoxin, which induces anoikis by disrupting FA structures and interfering with the integrin survival signaling.
HeLa (ATCC), HeLa S3 (ATCC), and Crk-/- cells were cultured in complete DMEM (Life Technologies) containing phenol red supplemented with 10% FCS, 1% penicillin/streptomycin, and 1% L-glutamine at 37°C in the presence of 5% CO2. For transfection experiments, 0.4μg plasmid DNA was used with Effectene (Qiagen) according to the manufactures protocol. Antibodies were from Cell Signaling Technologies (CST) unless otherwise noted: pAKT (#9271); Crk (BD 610036); GAPDH (GenScript A00191); pJNK (#4668); JNK (#9252); p130Cas (BD 610271); p-p130Cas(Y165) (#4015); p38β (#9212); p-p38β (#9215); FAK (BD 610087); pFAK(Y397) (#3283).
Briefly, HeLa, HeLa S3, and Crk-/- cells were grown in DMEM without phenol red with (for transfection studies) or without antibiotics (for infection studies) for 24 hr. These cells were then transfected with indicated expression vectors or infected with indicated strains as described [28]. 1h after transfection or at time of infection, cells were given 7μg/ml propidium iodide (Sigma) and then placed on an AxioVert Z1 microscope (Zeiss) fitted with a live-imaging culture box (Pecon) maintaining 37°C in the presence of 5% CO2. Time-lapse videos were taken using AxioVision 4.2.8 software. Video analysis was performed with ImageJ 1.47 software (NIH).
Samples were assessed by Western blot as described previously [29,66]. Briefly, cells were lysed following infection with 1% TX-100 containing a protease inhibitor cocktail (Roche Diagnostics), 100mM PMSF, and 100mM Na3VO4. Lysates were mixed with 4X SDS loading buffer and loaded onto 10% SDS-polyacrylamide gels. After resolving, gels were transferred to PVDF membranes, blocked with 5% milk, and probed overnight with primary antibody at 4°C. After washing, blots were probed with HRP-conjugated secondary antibody (Cell Signaling Technologies). Blots were developed with ECL or ECL+ reagent (GE Healthcare). Films were developed with an autoprocessor.
All bacterial strains and expression vectors and their sources are indicated in S1 Table. These isogenic strains were in PA103 genetic background. For infection studies, bacteria were cultured overnight in Luria-Bertani (LB) broth at 37°C without shaking. Bacteria were added at M.O.I of 10.
β-catenin transcriptional activity assay was performed as previously described [67]. HeLa cells were transfected with 400ng TOPFlash plasmid (Upstate) using Effectene (Qiagen). After 24 hr, the cells were washed with PBS and given fresh media either with our without antibiotics. The cells were either mock infected (PBS) or infected with the aforementioned strains which were resuspended in PBS. Luciferase levels were measured according to the manufacturers protocol using the Luciferase Assay System (Promega) using a Moonlight 2010 Luminometer (BD Bioscience) and normalizing to Renilla luciferase’s baseline luminance. Experiments were performed in triplicate for each time point.
Immunofluorescent microscopy was carried out as previously described [28,66,68]. Briefly, coverslips were treated with poly-l-lysine and 40μg/ml human fibronectin (Millipore) before seeding cells. After 24 hr, cells were either transfected for 17h or infected for 4h. At each end-point, cells were fixed with 10% ice-cold TCA for 10min. Cells were permeablized with 0.2% Triton X-100 (Sigma) for 15min at RT, blocked with 3% FCS for 1H at 37°C before staining overnight with primary antibody. Next, cells were washed 3x with PBS before staining with conjugated secondary antibody, AF594 or AF488 (Life Technologies) for 1hr at 37°C. The coverslip was mounted on DAPI containing VectaMount (Vector Laboratories). Cells were imaged AxioVision 4.2.8 software using an AxioVert Z1 microscope (Zeiss) using a 63X objective.
Immunofluorescent microscopy images were processed using ImageJ v1.47 (NIH). Cell outlines were first saved using the selection manager. A background subtraction was applied to the channel containing the focal adhesion marker. Next, a threshold level was applied to that channel followed by particle analysis to identify the number of puncta per cell selection and staining intensity per puncta. Each image set was processed using equal threshold and particle analysis values.
Two-tailed Student t-tests, one-way ANOVA with Tukey post-hoc test, and Chi-squared analyses were used to assess significance with p<0.05 considered significant. Analysis was carried out with Prism 6.0 (GraphPad).
|
10.1371/journal.ppat.1006123 | EV71 3D Protein Binds with NLRP3 and Enhances the Assembly of Inflammasome Complex | Activation of NLRP3 inflammasome is important for effective host defense against invading pathogen. Together with apoptosis-associated speck-like protein containing CARD domain (ASC), NLRP3 induces the cleavage of caspase-1 to facilitate the maturation of interleukin-1beta (IL-1β), an important pro-inflammatory cytokine. IL-1β subsequently plays critical roles in inflammatory responses by activating immune cells and inducing many secondary pro-inflammatory cytokines. Although the role of NLRP3 inflammasome in immune response is well defined, the mechanism underlying its assembly modulated by pathogen infection remains largely unknown. Here, we identified a novel mechanism by which enterovirus 71 (EV71) facilitates the assembly of NLRP3 inflammasome. Our results show that EV71 induces production and secretion of IL-1β in macrophages and peripheral blood mononuclear cells (PBMCs) through activation of NLRP3 inflammasome. EV71 replication and protein synthesis are required for NLRP3-mediated activation of IL-1β. Interestingly, EV71 3D protein, a RNA-dependent RNA polymerase (RdRp) was found to stimulate the activation of NLRP3 inflammasome, the cleavage of pro-caspase-1, and the release of IL-1β through direct binding to NLRP3. More importantly, 3D interacts with NLRP3 to facilitate the assembly of inflammasome complex by forming a 3D-NLRP3-ASC ring-like structure, resulting in the activation of IL-1β. These findings demonstrate a new role of 3D as an important player in the activation of inflammatory response, and identify a novel mechanism underlying the modulation of inflammasome assembly and function induced by pathogen invasion.
| The immune system protects the infected host and clears the invading pathogens. An important part of the innate immune response is the activation of NLRP3 inflammasome, which is induced upon exposure to pathogens. Activated inflammasome subsequently regulates the maturation of IL-1β that plays an important role in inflammatory response. Enterovirus 71 (EV71) is a highly contagious virus causing hand-foot-mouth disease (HFMD), meningoencephalitis, neonatal sepsis, and even fatal encephalitis in children by inducing many pro-inflammatory cytokines. Although NLRP3 inflammasome plays important role in regulating host immunity and viral infection, the assembly of NLRP3 inflammasome induced by viral infection is not known. In this study, we demonstrate that EV71 3D RNA polymerase activates NLRP3 inflammasome by binding to NLRP3. More importantly, 3D was found to interact with NLRP3 to facilitate the assembly of inflammasome complex by forming a specific ring-like structure. Therefore, these findings demonstrate a new role of viral 3D polymerase in the activation of inflammatory response, and identify a novel mechanism underlying the regulation of inflammasome assembly in responding to pathogen infection, which would provide insights into the prevention and treatment of viral infection.
| The innate immune system is a highly conserved signaling network important for protection of the infected host and clearance of the invading pathogen [1]. Recognition of the pathogen-associate molecular patterns (PAMPS) is dependent on host pattern recognition receptors (PRRs), whose activation results in the production of interferons (IFNs) and pro-inflammatory cytokines. Several families of PRRs have been identified, including the Toll-like receptor (TLR) [2], the RIG-I-like receptor (RLR) [3], the NOD-like receptor (NLR) [4], and the C-type lectin receptor (CLR) [5].
An important part of the innate immune response is the activation of inflammasome, a cytosolic complex of proteins that activates caspase-1 (Casp-1) to produce the pro-inflammatory cytokine interleukin-1beta (IL-1β) [6]. One of the best-characterized inflammasomes consists of NLR family PYRIN domain containing-3 (NLRP3) that harbors an N-terminal PYRIN domain (PYD), a NACHT-associated domain (NAD), and a C-terminal leucine-rich repeat (LRR) [7]. The PYD domain of NLRP3 interacts with the PYD domain of the adaptor protein, apoptosis-associated speck-like protein with CARD domain (ASC). NLRP3 oligomerizes through homotypic interactions between NACHT domains following the detection of pathogen infection or cellular stress. The LRR domain has been implicated in ligand sensing and auto-regulation. NLRP3 inflammasome is activated upon exposure to pathogens, including bacteria (Listeria monocytogenes and Staphylococcus aureus) [8] and viruses (Sendai virus, adenovirus and influenza virus) [9] [10], by host-derived molecules, such as extracellular glucose [11], extracellular ATP [12], and hyaluronan [13], and also detects signs of metabolic stress and environmental irritants [7]. Together with ASC, NLRP3 promotes the cleavage of pro-Casp-1 to generate active subunits p20 and p10, which regulate the maturation of IL-1β [14]. IL-1β plays an important role in inflammatory response by recruitment and activation of immune cells as well as production of secondary pro-inflammatory cytokines [15].
NLRP3 can recognize many RNA viruses to regulate innate immunity and viral replication, including enterovirus 71 (EV71) [16]. EV71 is a highly infectious RNA virus causing hand-foot-mouth disease (HFMD), meningoencephalitis, neonatal sepsis, and even fatal encephalitis in children [17]. EV71 induces many pro-inflammatory cytokines that play important roles in the development of inflammation and associated diseases [18]. Although NLRP3 inflammasome plays important role in regulating host innate immune response and viral infection, the assembly and activation of NLRP3 inflammasome mediated by viral infection are poorly understood.
In this study, we identify a novel mechanism by which EV71 facilitates the assembly of NLRP3 inflammasome. EV71 was found to stimulate the cleavage of pro-Casp-1 and the secretion of IL-1β by modulating the components of NLRP3 inflammasome. More significantly, EV71 3D protein was shown to play a stimulatory role in the activation of NLRP3 inflammasome. 3D protein is a RNA-dependent RNA polymerase (RdRp) essential for viral replication [19]. It activates the inflammasome through direct binding to NLRP3. The interaction of 3D with NLRP3 facilitates the assembly of inflammasome complex by forming a “3D-NLRP3-ASC” ring-like structure. These findings demonstrate a novel mechanism underlying the regulation of inflammasome complex assembly in response to pathogen infection, which would provide insights into the prevention and treatment of the viral infection.
EV71 infection induces pro-inflammatory cytokines that play important roles in the development of inflammation and associated diseases [20]. Among the pro-inflammatory cytokines, IL-1β plays important roles in the induction of inflammation by recruitment and activation of immune cells and production of secondary pro-inflammatory cytokines [21]. Thus, we determined the effect of EV71 on the production and secretion of IL-1β. TPA-differentiated THP-1 macrophages were infected with EV71. The secretion of IL-1β (Fig 1A) and cleavage of IL-1β (p17) and caspase-1 (p20 and p22) (Fig 1B, upper panel) in supernatants, and the production of pro-IL-1β in the cell lysates (Fig 1B, lower panel, lanes 3–4) were induced by EV71 infection. EV71 3D protein was expressed and VP1 mRNA was detected (S1A Fig) during EV71 infection (Fig 1B, lanes 3–4), indicating that EV71 replicated well in the cells. In addition, THP-1 macrophages were treated with lipopolysaccharides (LPS) and Nigericin (a bacterial toxin that activates NLRP3 by causing potassium efflux in a pannexin-1-dependent pathway) positive control. The expression of IL-1β mRNA (Fig 1C) was stimulated by LPS+Nigericin and EV71. It is clear that EV71 induces IL-β in THP-1 macrophage.
Next, TPA-differentiated THP-1 macrophages were treated with LPS and Nigericin, or infected with EV71 at different multiplicity of infections (MOI), as indicated. The secretion of IL-1β (Fig 1D) and cleavage of IL-1β (p17) and caspase-1 (p20 and p22) (Fig 1E, top panel) in the cell supernatants, and the production of pro-IL-1β in the cell lysates (Fig 1E, bottom panel, lanes 3–5) were induced by EV71 in dose-dependent manners. Similarly, the expression of IL-1β mRNA was activated by LPS/Nigericin and EV71 virus (Fig 1F). EV71 3D protein (Fig 1E) and VP1 mRNA (S1B Fig) were detected in viral infected cells, suggesting that EV71 replicated efficiently in the macrophages. Therefore, these results indicated that EV71 induces the production and secretion of IL-1β in dose-dependent manners. The activation of IL-1β is regulated by two pathways in response to pathogen infection: transcription of pre-IL-1β mRNA and proteolytical processing of pre-IL-1β protein by caspase-1 [7]. Previous study has demonstrated that EV71 infection activates NF-κB in rat vascular smooth muscle cells and SK-N-SH cells [22, 23]. We also demonstrated that the endogenous genes regulated by NF-κB were activated by EV71 and by lipopolysaccharide (LPS), an NF-κB-activating stimuli (S1C Fig). ASC oligomers were formed by EV71 infection and Nigericin treatment, indicating that EV71 and Nigericin induce inflammasome activation [24] (S1D Fig). Taken together, our results demonstrate that the production and secretion of IL-1β are induced by EV71 in macrophages.
The mechanism by which EV71 activates IL-1β was investigated. As maturation of IL-1β is mediated by the NLRP3 inflammasome [6], EV71 may activate IL-1β through modulating the NLRP3 inflammasome. To determine this possibility, HEK293T cells were co-transfected with plasmids encoding the components of NLRP3 complex (NLRP3, ASC, and pro-caspsase-1) along with the substrate of NLRP3 inflammasome (pro-IL-1β), and then treated with ATP and infected with EV71. The secretion of IL-1β (p17) was up-regulated by ATP and EV71 in cell supernatants (Fig 2A), and the production of IL-1β (p17) and Casp-1 subunit (p20) were enhanced by ATP or EV71 in cell lysates (Fig 2B). These results suggest that EV71 facilitates the activation of IL-1β and Casp-1 in the presence of NLRP3 inflammasome components (Casp-1, NLRP3, and ASC) in HEK293T cells.
The roles of components of NLRP3 inflammasome complex in the production and secretion of IL-1β were further evaluated by short hairpin RNAs (shRNAs) mediated knockdown of gene expression. Four stable TPA-differentiated THP-1 macrophages cell lines were generated, which stably expressed short hairpin RNAs (shRNAs) targeting NLRP3 (sh-NLRP3), ASC (sh-ASC), and pro-caspase-1 (sh-Casp-1), respectively. In sh-NLRP3 stable cells, NLRP3 mRNA was down-regulated, whereas ASC and pro-Casp-1 mRNAs were not affected (S2A Fig). In sh-ASC stable cells, ASC mRNA was attenuated, but NLRP3 and pro-Casp-1 mRNAs remain the same (S2B Fig). In sh-Casp-1 stable cells, pro-Casp-1 mRNA was reduced, while NLRP3 and ASC mRNAs were unchanged (S2C Fig). In addition, NLRP3, ASC, and pro-Casp-1 proteins were attenuated in sh-NLRP3 stable cells, sh-ASC stable cells, and sh-Casp-1 stable cells, respectively (S2D Fig). These results indicated that shRNAs were stably expressed and specifically silenced their target gene expression. The stable cell lines were then treated with Nigericin. In the cell supernatants, secretion of IL-1β was induced by Nigericin, but its activation was significantly attenuated by sh-NLRP3, sh-ASC, and sh-Casp-1, respectively (Fig 2C). Similarly, the cleavage of IL-1β(p17) and caspase-1 (p20 and p22) were enhanced by Nigericin (Fig 2D, top panel, lane 5), but their levels were significantly down-regulated by sh-NLRP3, sh-ASC, and sh-Casp-1, respectively (Fig 2D, top panel, lanes 6–8). These results demonstrate that shRNA-mediated gene silence of the components (NLRP3, ASC, or pro-caspase-1) of NLRP3 inflammasome attenuates the production and secretion of IL-1β.
The role of NLRP3 inflammasome components in EV71-induced activation of IL-1β was further explored in the stable cell lines infected with EV71. In the cell supernatants, the secretion of IL-1β (p17) was increased by EV71, but significantly attenuated by sh-NLRP3, sh-ASC, and sh-Casp-1, respectively (Fig 2E). The cleavage of IL-1β (p17) and caspase-1 (p20 and p22) was enhanced by EV71 (Fig 2F, top panel, lane 5), but significantly reduced by sh-NLRP3, sh-ASC, or sh-Casp-1 (Fig 2F, top panel, lanes 6–8). In the cell lysates, the productions of matured IL-1β (p17) were stimulated by EV71 (Fig 2F, bottom panel, lane 5), but down-regulated by sh-NLRP3, sh-ASC, or sh-Casp-1, respectively (Fig 2F, bottom panel, lane 6–8). These results suggest that activation of IL-1β by EV71 requires the components of NLRP3 inflammasome complex. Taken together, our results demonstrate that EV71 activates IL-1β through regulating the NLRP3 inflammasome.
The mechanism involved in the modulation of NLRP3 inflammasome mediated by EV71 was investigated. The effect of EV71 replication on the activation of IL-1β was evaluated. TPA-differentiated THP-1 macrophages were treated with LPS/Nigericin, incubated with ultraviolet (UV)-inactivated EV71 or heat-inactivated EV71, or infected with infectious EV71, as indicated. In the cell supernatants, the secretion of IL-1β was significantly increased by LPS/Nigericin and infectious EV71 and modestly up-regulated by UV- and heat-inactivated EV71 (Fig 3A). The cleavage of IL-1β (p17) and caspase-1 (p20 and p22) were significantly enhanced by LPS/Nigericin and infectious EV71, but not affected by UV- and heat-inactivated EV71 in the cell supernatants (Fig 3B, top panel). EV71 3D was expressed only in the cells infected with infectious EV71 (Fig 3B, bottom panel, lane 5), but not detected in the cells inoculated with UV- and heat-inactivated EV71 (Fig 3B, lanes 3–4), suggesting that UV- and heat-inactivated EV71 failed to replicate in the cells. Therefore, these results indicate that the replication of EV71 is required for the activation of IL-1β in macrophages.
Additionally, the effect of EV71 replication on the activation of IL-1β in human primary peripheral blood mononuclear cells (PBMCs) was evaluated. PBMCs were infected with EV71, treated with LPS, and inoculated with UV-inactivated or heat-inactivated EV71, respectively. EV71 VP1 mRNA was detected in the cells infected with EV71, but not in the cells treated with LPS, or inoculated with UV-inactivated or heat-inactivated EV71 (S3A Fig), indicating that EV71 replicates in PBMCs, but UV-inactivated EV71 and heat-inactivated EV71 failed to replicate. The secretion of IL-1β was stimulated by LPS, activated by EV71 in an MOI-dependent manner, but not affected by UV-inactivated or heat-inactivated EV71 (Fig 3C). The expression of IL-1β mRNA was stimulated by LPS, activated by EV71, but not by UV-inactivated or heat-inactivated EV71 (Fig 3D). These results suggested that the replication of EV71 is required for the production and secretion of IL-1β in PBMCs.
It is well established that NLRP3 inflammasome mediated innate immunity to virus through recognition of viral RNA [25, 26]. To further determine whether EV71 RNA alone can initiate the activation of the NLR3 inflammasome, we stimulated THP-1 macrophages with EV71 viral genomic RNA, HCV viral genomic RNA, and poly(dA:dT) as a positive control. The results showed that IL-1β secretion was induced by HCV genomic RNA and poly(dA:dT), but not by EV71 genomic RNA in the cell supernatants of THP-1 macrophages (Fig 3E). The cleavage of IL-1β (p17) and caspase-1 (p20 and p22) were not detected in the presence of transfected EV71 RNA (Fig 3F). The levels of pro-IL-1β mRNA and TNF-α mRNA were enhanced by the treatment of EV71 viral genomic RNA and HCV viral genomic RNA (S3B Fig). These results demonstrated that, unlike HCV genomic RNA and influenza A virus RNA, EV71 genomic RNA is unable to activate NLRP3 inflammasome.
Previous study has demonstrated that rapid NLRP3 inflammasome activation is independent of “priming”, given that both NF-κB activation and new protein synthesis are not necessary [27]. CHX-pretreated cells produced IL-1β normally in response to extracellular ATP which indicated that NLRP3 inflammasome activation by ATP stimulation does not require de novo translation. In contrast, pretreatment of cells with cycloheximide (CHX), a protein synthesis inhibitor, significantly inhibited EMCV-induced IL-1β secretion. This indicated that virus-encoded proteins activate the NLRP3 inflammasome [28]. We further determined whether EV71 viral protein synthesis was also required for the activation of IL-1β. TPA-differentiated THP-1 macrophages were infected with EV71 for 2 h, treated with CHX for 1 h, and then grown for additional 22 h. In the absence of CHX, the secretion of IL-1β was stimulated by Nigericin treatment (Fig 4A, lane 4) or EV71 infection (Fig 4A, lane 7). In the presence of CHX, Nigericin-induced secretion of IL-1β was relatively unaffected (Fig 4A, lane 6), but EV71-activated secretion of IL-1β was significantly repressed (Fig 4A, lane 9). In the absence of CHX, the cleavages of IL-1β (p17) and caspase-1 (p20 and p22) were stimulated by Nigericin treatment (Fig 4B, lane 4) or by EV71 infection (Fig 4B, lane 7) in the cell supernatants. In the presence of CHX, Nigericin-activated cleavages of IL-1β (p17) and caspase-1 (p20 and p22) were relatively unaffected (Fig 4B, lane 6), whereas EV71-activated cleavages of IL-1β (p17) and caspase-1 (p20 and p22) were significantly reduced (Fig 4B, lane 9). The production of EV71 3D (Fig 4B, lanes 7–9) was significantly attenuated in the presence of CHX (Fig 4B, lane 9). Therefore, the activation of IL-1β mediated by Nigericin do not require de novo protein synthesis, whereas the activation of IL-1β induced by EV71 requires de novo protein synthesis. Taken together, we demonstrated that EV71 replication and protein synthesis are essential for the activation of IL-1β in macrophages and PBMCs.
The roles of EV71 nonstructural proteins in the regulation of NLRP3 inflammasome were further determined, as EV71 protein synthesis is required for such regulation. Initially, the activity of NLRP3 inflammasome was determined in HEK293T cells co-transfected together with plasmids encoding NLRP3, ASC, pro-Casp-1, and pro-IL-1β. In the cell supernatants, the secretion of IL-1β (p17) was not detected in the presence of NLRP3 alone (Fig 4C, lane 2) or ASC alone (Fig 4C, lane 3), detected at lower level in the presence of Casp-1 (Fig 4C, lane 4) or NLRP3 plus ASC (Fig 4C, lane 5), and significantly activated in the presence of all three components, NLRP3, ASC, and Casp-1 (Fig 4C, lane 6). In the cell lysates, the levels of pro-IL-1β and Casp-1 (p22/p20) were relatively unchanged in the presence of one or two components (Fig 4C, lanes 1–5), but significantly stimulated in the presence of all three components (Fig 4C, lane 6). These results indicated that all components of NLRP3 inflammasome are required for the function of the inflammasome, and suggested that NLRP3 inflammasome is functionally effective under these conditions.
HEK293T cells were co-transfected together with plasmids encoding NLRP3, ASC, pro-Casp-1, pro-IL-1β, and then transfected with each of the EV71 non-structure proteins, 2A, 2B, 2C, 3A, 3C, and 3D, respectively, as indicated. The secretion of IL-1β in the cell supernatants was inhibited by 2A, relatively unaffected by 2B and 3A, down-regulated by 2C and 3C, and activated by 3D (Fig 4D). Similarly, the production of IL-1β (p17) in the cell lysates was inhibited by 2A (Fig 4E, lane 2), unaffected by 2B and 3A (Fig 4E, lanes 3 and 5), enhanced by 2C (Fig 4E, lane 4), reduced by 3C (Fig 4E, lane 6), but activated by 3D (Fig 4E, lane 7). In addition, the cleavage of caspsase-1 (p20) was inhibited by 2A (Fig 4E, lane 2), unaffected by 2B and 2C (Fig 4E, lanes 3 and 4), reduced by 3A and 3C (Fig 4E, lanes 5 and 6), but significantly stimulated by 3D (Fig 4E, lane 7). Moreover, NLRP3 was inhibited by 2A (Fig 4E, lane 2), unaffected by 2B, 2C, and 3A (Fig 4E, lanes 3–5), repressed by 3C (Fig 4E, lane 6), and enhanced by 3D (Fig 4E, lane 7). These results revealed that EV71 2A and 3C repress the production and secretion of IL-1β through attenuating NLRP3 inflammasome, which is consistent with a previous report [16]. More interestingly, our results demonstrated that EV71 3D was the only viral protein that stimulates the secretion of IL-1β (Fig 4D and 4E) and the cleavage of pro-Casp-1 (Fig 4E). The correlation between the opposite functions of EV71 proteins was then evaluated in TPA-differentiated THP-1 cells infected with EV71 at MOI = 20 for different time. The results showed that the secretion of IL-1β was induced by EV71 at 6 h p.i. (Fig 4F, lane 3) and then increased as the infection time increased (Fig 4F, lanes 4–6). The cleavages of IL-1β (p17) and caspase-1 (p20 and p22) were stimulated by EV71 at 6 h p.i. (Fig 4G, lane 3). Similarly, EV71 3D protein was also detected at 6 h p.i. (Fig 4G, lane 3) and then increased as the infection time increased (Fig 4G, lanes 4–6). However, EV71 3C protein was detected at 12 h p.i. (Fig 4G, lane 4) and then increased as the infection time increased (Fig 4G, lanes 5 and 6). EV71 2A protein was not detect by western blot, likely due to the concomitant restriction on its expression from its inhibition effect on host gene expression [29]. Thus, EIF4G, a known substrate of 2A, was used as an indictor for the activity of 2A proteases. The results showed that EIF4G protein was reduced at 36 h p.i. (Fig 4G, lane 6). The temporal expression of EV71 2A, 3C, and 3D were further confirmed in RD cells infected with EV71 at different times. The results showed that 3D expression was initiated at 6 h p.i. (Fig 4H, lane 4), whereas EIF4G cleavage and 3C expression was started at 12 h p.i. (Fig 4H, lane 5). Taken together, these results revealed that the secretion of IL-1β or the activation of NLRP3 inflammasome was stimulated by EV71 3D and that 3D protein can overcome the inhibitory effects of EV71 2A and 3C proteins in the secretion of IL-1β. Therefore, we focused the rest of study on determining the role of EV71 3D in the activation of NLRP3 inflammasome and the mechanism underlying such regulation.
EV71 3D induces the cleavage of pro-Casp-1 and the secretion of IL-1β through suggesting that it may activate NLRP3 inflammasome, as the NLRP3 inflammasome is critical for pro-Casp-1 activation and pro-IL-1β procession. The effect of EV71 3D on the activation of NLRP3 inflammasome was determined. HEK293T cells were co-transfected together with plasmids encoding NLRP3, ASC, pro-Casp-1, and pro-IL-1β, along with plasmid encoding EV71 3D, as indicated. The productions of IL-1β (p17) and Casp-1 (p22/p20) were not detected in the presence of one or two components of the inflammasome (Fig 5A, lanes 1–3 and 5–7), activated in the presence of NLRP3, ASC, and pro-Casp-1 (Fig 5A, lane 4), and further facilitated by 3D (Fig 5A, lane 8). Similarly, the secretion of IL-1β was not detected in the presence of one or two components of the inflammasome (Fig 5B, lanes 1–3 and 5–7), stimulated in the presence of NLRP3, ASC, and pro-Casp-1 (Fig 5B, lane 4), and further enhanced by 3D (Fig 5B, lane 8). These results demonstrated that EV71 3D enhances the activity of NLRP3 inflammasome to facilitate the production and release of IL-1β.
In addition, the role of EV71 3D in regulating the activity of NLRP3 inflammasome was evaluated in TPA-differentiated THP-1 macrophages which infected with lentivirus expressing EV71 3D protein for the stable protein expression cell lines. The results showed that 3D activated the secretion of IL-1β (p17) in the cell supernatants (Fig 5C, top panel), as well as the production of IL-1β (p17) and Casp-1 (p22) in the cell lysates (Fig 5C, bottom). Furthermore, the effect of EV71 3D on ASC pyroptosome formation was evaluated in TPA-differentiated THP-1 macrophages infected with lentiviruses expressing 3D. We used the TPA-differentiated THP-1 macrophages as a negative control and the TPA-differentiated THP-1 macrophages which were treated with Nigericin as a positive control. ASC pyroptosome formation was activated by 3D and Nigericin, respectively (Fig 5D). In addition, the role of EV71 in regulating the activity of NLRP3 inflammasome was evaluated in TPA-differentiated THP-1 macrophages infected with lentiviruses expressing EV71 3D, treated with Nigericin or ATP, and infected with Sendai virus (SeV). The secretion of IL-1β (p17) was activated by Nigericin, ATP, and SeV in the cell supernatants, whereas such activations were significantly facilitated by 3D (Fig 5E). Similarly, the secretion of IL-1β (p17) was stimulated by Nigericin, ATP and SeV, whereas 3D further facilitated the secretion of IL-1β (p17) mediated by Nigericin, ATP, and SeV (Fig 5F).
Previous study has demonstrated that aurintricarboxylic acid (ATA) was able to inhibit the RdRp activity of EV71 3D protein, but not inhibit the protease activities of 2A and 3C [30]. Our results revealed that lentivirus expressed EV71 3D activated the secretion of IL-1β in THP-1 cells in the absence of ATA (Fig 5G, lane 2), but failed to induce the secretion of IL-1β in the presence of ATA (Fig 5G, lane 4). In the absence of ATA, the secretion of IL-1β (p17) was activated by ATP (Fig 5G, lane 5) and such activation was significantly facilitated by 3D (Fig 5G, lane 6). In the presence of ATA, the secretion of IL-1β (p17) was also activated by ATP (Fig 5G, lane 7), but EV71 3D had no affect on ATP-induced secretion of IL-1β (Fig 5G, lane 8). Similarly, EV71 3D protein stimulated the cleavage of IL-1β (p17) in the absence of ATA (Fig 5H, lane 2), but failed to induced the cleavage of IL-1β (p17) in the presence of ATA (Fig 5H, lane 4). In the presence of ATA, the cleavage of IL-1β was induced by ATP (Fig 5H, lane 7), but EV71 3D failed to facilitate ATP-induced secretion of IL-1β (Fig 5H, lane 8). In addition, 3D protein production was not affected by ATA (Fig 5H, lane 2 vs 4, and lane 6 vs 8), suggesting that the RdRp activity of EV71 3D may be required for the activation of NLRP3 inflammasome. Moreover, TPA-differentiated THP-1 macrophages were infected with EV71 and treated with ATA and ATP. The results showed that ATP-induced secretion of IL-1β (Fig 5I, lane 3) was slightly reduced by ATA (Fig 5I, lane 3), but EV71-induced secretion of IL-1β (Fig 5I, lane 5) were significant attenuated by ATA (Fig 5I, lane 6). Similarly, ATP-induced cleavage of IL-1β (Fig 5J, lane 3) was slightly reduced by ATA (Fig 5J, lane 4), but EV71-induced cleavage of IL-1β (Fig 5J, lane 5) were significant attenuated by ATA (Fig 5J, lane 6). These results demonstrated that RdRp activity of EV71 3D is essential for NLRP3 inflammasome activation. Taken together, we revealed that EV71 3D protein induces the activity of NLRP3 inflammasome and the productions of IL-1β (p17) and Casp-1 (p22/p20), and that the RdRp activity of EV71 3D is essential for EV71-induced activation of NLRP3 inflammasome.
The mechanism underlying the regulation of NLRP3 inflammasome mediated by EV71 3D was elucidated. Initially, we determined whether EV71 3D is interacted with the components of NLRP3 inflammasome. Interestingly, yeast strain AH109 was co-transformed with the combination of AD-3D and BD-NLRP3 inflammasome components or the three functional domains of NLRP3. We revealed that EV71 3D was interacted with NLRP3 LRR domain (S4A and S4B Fig). The interaction between EV71 3D and NLRP3 was verified by co-immunoprecipitation (CoIP) assays in HEK293T cells co-transfected with plasmid expressing Flag-NLRP3 or HA-3D, as indicated. CoIP results showed that 3D was associated with NLRP3 (Fig 6A) and NLRP3 was interacted with 3D (Fig 6B). The interaction between EV71 3D and NLRP3 was further determined in TPA-differentiated THP-1 macrophages infected with lentiviruses expressing 3D. CoIP results indicated that 3D was also interacted with NLRP3 in the treated macrophages (Fig 6C). These results demonstrated that EV71 3D protein can interact with NLRP3 protein.
NLRP3 contains several prototypic domains, including PYD, NACHT, and LRR domains. Four plasmids expressing NLRP3, PYRIN domain, NACHT domain, and LRR domain were constructed (S4C Fig). The domains of NLRP3 involved in the interaction with EV71 3D were then determined. HEK293T cells were co-transfected with plasmid expressing HA-3D along with plasmids expressing Flag-NLRP3, Flag-PYD, Flag-NACHT, and Flag-LRR, respectively. CoIP results showed that 3D was interacted with NLRP3 (Fig 6D, lane 2), NACHT domain (Fig 6D, lane 6), and LRR domain (Fig 6D, lane 8), but not PYD domain (Fig 6D, lane 4). Similarly, NLRP3 (Fig 6E, lane 2), NACHT domain (Fig 6E, lane 6), and LRR domain (Fig 6E, lane 8), but not PYD domain (Fig 6E, lane 4), were interacted with 3D. Therefore, we demonstrated that EV71 3D binds with NLRP3 through interacting with the NACHT and LRR domains.
We then determined whether 3D also interacts with other components (ASC and Casp-1) of NLRP3 inflammasome through interacting with NLRP3. Upon activation, NLRP3 is oligomerized, which leads to NLRP3 PYD domain clustering and presentation for interaction with ASC PYD domain, and ASC CARD domain subsequently recruits pro-Casp-1 CARD domain to permit the auto-cleavage and the formation of active Casp-1 p10/p20. Thus, we evaluated the ability of 3D in the interaction with ASC. CoIP assays showed that 3D was interacted with NLRP3 (S5 Fig, lane 2), but not ASC (S5 Fig, lane 4). The interaction between 3D and ASC was further investigated in HEK293T cells co-transfected with plasmids expressing NLRP3, PYRIN domain, ASC, and 3D, as indicated. 3D was associated with ASC only in the presence of NLRP3 (Fig 7A, lane 2), but not in the presence of PYRIN domain (Fig 7A, lane 4). Similarly, ASC was associated with 3D only in the presence of NLRP3 (Fig 7B, lane 2), but not in the presence of PYRIN domain (Fig 7B, lane 4). NLRP3 PYRIN domain interacts with ASC, but not with 3D, suggesting that 3D is associated with ASC through interacting with NLRP3. In addition, the interaction of 3D with NLRP3, ASC, and pro-Casp-1 was evaluated by GST pull-down assays. HEK293T cells were co-transfected with plasmid expressing Flag-NLRP3 and plasmids encoding Flag-pro-Casp-1, Flag-ASC, and GST-3D, as indicated. The results showed that 3D could interact with NLRP3 even in the absence of ASC and pro-Casp-1 (Fig 7C, lane 2), and could also interact with NLRP3 but not with pro-Casp-1 in the absence of ASC (Fig 7C, lane 4). However, 3D was associated with NLRP3, ASC, and pro-Casp-1 in the presence of all components of NLRP3 inflammasome (Fig 7C, lane 6). The purified GST-LRR could interact with HA-3D protein which also demonstrated the interaction between NLRP3 and EV71 3D protein (Fig 7D). These results indicated that 3D is associated with ASC or pro-Casp-1 through interacting with NLRP3. Moreover, the interactions of 3D with endogenous NLRP3 and ASC were explored in TPA-differentiated THP-1 macrophages infected with EV71. The results showed that 3D was co-immunoprecipitated with endogenous NLRP3 and ASC (Fig 7E). Thus, EV71 3D is associated with the components (NLRP3, ASC, and pro-Casp-1) of NLRP3 inflammasome through interacting with NLRP3.
As 3D interacts with NLRP3, the effect of 3D on the sub-cellular distribution of NLRP3 was examined under confocal microscopy. We evaluated the effect of EV71 on the sub-cellular distribution of NLRP3 in TPA-differentiated THP-1 macrophages infected with EV71 by immunofluorescence assays. In mock-infected macrophages (Fig 7Fa to 7Fd), NLRP3 was diffusely distributed in the cell cytosol (Fig 7Fa and 7Fd). However, in EV71-infected cells (Fig 7Fe to 7Fh), NLRP3 was co-localized with 3D (Fig 7Fe and 7Ff) to form distinct spots in the cell cytosol (Fig 7Fh). These results revealed that 3D interacts with NLRP3 to form a 3D-NLRP3 complex in the cytoplasm during EV71 infection. The role of EV71 on the sub-cellular distribution of ASC was also determined in TPA-differentiated THP-1 macrophages infected with or without EV71. In mock-infected macrophages (S6a to S6c Fig), ASC was diffusely distributed in the nucleus and cytosol (S6a and S6c Fig). In EV71-infected cells (S6d to S6f Fig), ASC was distributed as distinct small spots in the cytosol (S6d and S6f Fig). These results indicated that EV71 alters the sub-cellular distribution of ASC.
The effect of 3D on the sub-cellular distribution of NLRP3 was then determined in TPA-differentiated THP-1 macrophages infected with control lentivirus or 3D-encoding lentivirus. In the absence of 3D (Fig 7Ga to 7Gd), NLRP3 was diffusely distributed in the cytoplasm (Fig 7Ga and 7Gd). In the presence of 3D (Fig 7Ge to 7Gh), NLRP3 was co-localized with 3D (Fig 7Ge and 7Gf) to form large spots in the cytoplasm (Fig 7Gh). The association of 3D with the components of NLRP3 inflammasome was further examined in HEK293T cells co-transfected with plasmid expressing Flag-NLRP3 and plasmids expressing HA-3D, as indicated. In the absence of 3D, NLRP3 was diffusely distributed in the cytosol of HEK293T cells (Fig 7Ha and 7Hd); whereas in the presence of 3D, NLRP3 was co-localized with 3D (Fig 7He and 7Hf) to form large spots in the cytosol (Fig 7He). The association of 3D with the components of NLRP3 inflammasome was also determined in RD cells co-transfected with plasmid expressing Flag-NLRP3 and plasmids expressing GFP-3D, as indicated. Similarly, in the absence of 3D, NLRP3 was diffusely distributed in the cytosol of RD cells (Fig 7Ia and 7Id); but in the presence of 3D, NLRP3 was co-localized with 3D (Fig 7Ie and 7If) to form large spots in the cytosol (Fig 7Ie). These results revealed that 3D regulates the sub-cellular distribution of NLRP3 through interacting with NLRP3 in TPA-differentiated THP-1 macrophages, HEK293T cells, and RD cells. Taken together, we demonstrated that EV71 3D protein is associated with the major components (NLRP3, ASC, and pro-Casp-1) of NLRP3 inflammasome through interacting with NLRP3, and also revealed that 3D regulates the sub-cellular distributions of NLRP3 inflammasome by binding to NLRP3 and forming a 3D-NLRP3 complex. Thus, these results suggested that 3D may regulate the assembly of NLRP3 inflammasome.
In the absence of stimulation, NLRP3 is homo-oligomerized to form inactive preassembled complexes, which undergo conformational changes to form active inflammasome complexes containing ASC upon stimulation [31]. Since 3D protein is associated with the components of inflammasome through interacting with NLRP3, we speculated that 3D may facilitate the assembly of inflammasome by direct binding to NLRP3 after EV71 infection. To verify this hypothesis, HEK293T cells were co-transfected with plasmids expressing Flag-NLRP3, pcdna3.1(+)-ASC, and HA-3D, as indicated. ASC was recruited by NLRP3 (Fig 8A, lane 2), and NLRP3-mediated recruitment of ASC was significantly enhanced by 3D (Fig 8A, lane 3). The oligomerization of ASC is critical for Casp-1 activation and inflammasome function [24]. Therefore, we further investigated the effect of 3D on the oligomerization of ASC. HEK293T cells were co-transfected with plasmids expressing pcdna3.1(+)-ASC, Flag-NLRP3, and HA-3D, as indicated. The results revealed that oligomerization of ASC was detected in the presence of ASC (Fig 8B, lane 2), enhanced by NLRP3 (Fig 8B, lane 3), and not affected by 3D in the absence of NLRP3 (Fig 8B, lane 4). However, in the presence of NLRP3, oligomerization of ASC was significantly stimulated by 3D (Fig 8B, lane 5). Therefore, these results suggested that 3D facilitates the assembly of NLRP3 inflammasome through interacting with NLRP3.
It has been demonstrated that the localization of NLRP3 as spots is a character of the formation of inflammasome complex [6]. Thus, the effect of EV71 3D on the formation of NLRP3 inflammasome complex was explored in HEK293T cells co-transfected with plasmids encoding GFP, GFP-3D, Flag-NLRP3, and/or HA-ASC, as indicated. In the absence of 3D and ASC (Fig 8Ca to 8Ce), NLRP3 was diffusely distributed in the cell cytosol (Fig 8Cb and 8Ce). In the presence of 3D and absence of ASC (Fig 8Cf to 8Cj), NLRP3 was co-localized with 3D (Fig 8Cf and 8Cg) and distributed as large spots in the cell cytosol (Fig 8Cj). In the absence of 3D and presence of ASC (Fig 8Ck to 8Co), NLRP3 was co-localized with ASC (Fig 8Ck and 8Ci) and distributed as large spots in the cell cytosol (Fig 8Co). Interestingly, in the presence of both 3D and ASC (Fig 8Cp to 8Ct), NLRP3 was co-localized with 3D (Fig 8Cp and 8Cq) and ASC (Fig 8Cq and 8Cr) and distributed as large spots in the cytosol of transfected cells (Fig 8Ct). These results clearly showed that 3D, NLRP3, and ASC together formed a ring-like structure in the cytosol (Fig 8Cp, 8Cq, 8Cr and 8Ct). More interestingly, we observed that in the 3D-NLRP3-ASC complex, 3D (as indicated in green) was located inside in the ring-like structure, followed by the 3D-NLRP3 complex (green + red, as indicated in yellow) and the NLRP3-ASC complex (red + cyan, as indicated in white), and finally ASC (as indicated in cyan) was located outside the ring-like structure (Fig 8Cu, enlarged). Moreover, the formation of 3D-NLRP3-ASC complex was also observed in HEK293T cells co-transfected with plasmids expressing GFP-3D, Flag-NLRP3, and HA-ASC. The results confirmed that 3D, NLRP3, and ASC proteins together were indeed distributed as large spots in the cytosol and formed a ring-like structure (Fig 8De). In this ring-like structure, 3D was located inside (Fig 8Da), followed by NLRP3 in the middle (Fig 8Db), and ASC was located outside (Fig 8Dc). These results suggested that 3D facilitates the assembly of NLRP3 inflammasome, since it has been demonstrated that the inflammasome complex is assembled by the formation of a ring-like organization [32]. Taken together, we demonstrated that 3D facilitates the assembly of NLRP3 inflammasome complex by the formation of a “3D-NLRP3-ASC” structure through direct binding to NLRP3.
The assembly of inflammasomes is critical for the host innate immune response against pathogen infection [6]. The best-characterized inflammasomes is the NLRP3 inflammasome, which activates the maturation of IL-1β through promoting the cleavage of pro-Casp-1 to generate active subunits, p20 and p10 [14]. NLRP3 inflammasome regulates innate immunity in responding to several RNA viruses, including influenza A virus (IAV) [33], encephalomyocarditis virus (EMCV) and vesicular stomatitis virus (VSV) [34], measles virus (MV) [35], myxoma virus (MYXV) [36], adenovirus (Ad) [37], West Nile virus (WNV) [38], Rabies virus (RV) [39], Herpes simplex virus 1 (HSV-1) [40], Rift valley fever virus (RVFV) [41], human T-lymphotropic virus 1 (HTLV-1) [42], and EV71 [16]. However, the mechanisms underlying the assembly of NLRP3 inflammasome regulated by viruses have not been reported. In this study, we revealed a novel mechanism by which EV71 facilitates the assembly of NLRP3 inflammasome (Fig 9).
We showed that EV71 stimulates the production of Casp-1 and the release of IL-1β in macrophages and PBMCs. EV71 induces IL-1β secretion only in the presence of all components (NLRP3, pro-Casp-1, and ASC) of NLRP3 inflammasome, and knock-down of the components attenuates EV71-induced activation of IL-1βwhich demonstrated that EV71 activates IL-1β through regulating NLRP3 inflammasome complex. The activation of NLRP3 inflammasome is regulated by two pathways in response to pathogen infection [7]. The first pathway leads to the activation of nuclear factor-kappa B (NF-κB) that subsequently induces the production of pro-IL-1β, NLRP3, and ASC. The second pathway is initiated by the assembly of inflammasome, resulting in the activation of Casp-1 and the maturation of IL-1β. IL-1β acts as an important mediator of inflammation to stimulate the recruitment and activation of immune cells and the production of many secondary pro-inflammatory cytokines [15].
We further demonstrated that UV-inactivated EV71 and heat-inactivated EV71 fail to activate IL-1β and Casp-1, suggesting that the replication of EV71 is necessary for the activation of IL-1β. Unlike HCV genomic RNA and influenza A virus RNA, EV71 genomic RNA is unable to activate NLRP3 inflammasome. In addition, the protein synthesis inhibitor CHX blocks EV71-mediated release of IL-1βindicating that viral protein synthesis is also required for the activation of IL-1β. Previous study had demonstrated that several viral proteins inhibit NLRP3 inflammsome-mediated IL-1β secretion, including MV V protein [35], IAV NS1 and M2 proteins [43, 44], and encephalomyocarditis virus (EMCV) 2B protein [28]. Our detailed studies revealed that EV71 nonstructural proteins 3D are involved in the regulation of NLRP3 inflammasome. We demonstrated that EV71 3D plays a stimulatory role in the regulation of NLRP3 inflammasome. The 3D protein acts as a viral RdRp and plays an essential role in viral negative-strand synthesis and VPg uridylylation [45]. The VPg and uridylylated forms of VPg (VPg-pUpU) prime the initiation of RNA replication [46]. Many investigators have tried to inhibit the activity of 3D protein and thereby inhibit viral replication [30, 47, 48]. It is also involved in the regulation of cell S-phase arrest [49] and immune response [50]. However, the role of 3D in the regulation of NLRP3 inflammasome has not been reported. Thus, 3D function in the activation of NLRP3 inflammasome and the mechanism underlying such regulation were intensively investigated in this study.
Interestingly, we showed that EV71 2A and 3C repress the production and secretion of IL-1β through attenuating NLRP3 inflammasome, which is consistent with a previous report [16]. More interestingly, we demonstrated that EV71 3D is the only viral protein that stimulates the secretion of IL-1β. We further revealed that 3D can directly bind with NLRP3. We revealed that 3D interacts with NLRP3 LRR domain based on the result of yeast two-hybrid analysis, and further verified the interaction of 3D with NLRP3 by several approaches, including CoIP, GST pull down, and immunofluorescence assays. NLRP3 contains three important domains, PYD, NACHT, and LRR. We confirmed that 3D can interact with NACHT and LRR domains of NLRP3. Upon activation, NLRP3 is oligomerized that leads to the interaction of NLRP3 with ASC PYRIN domain, which subsequently recruits pro-Casp-1 through CARD domain to permit the auto-cleavage and formation of active Casp-1 (p10/p20) [7]. Although 3D was not interacted directly with ASC and pro-Casp-1, it is associated with NLRP3, ASC, and pro-Casp-1, through interacting with NLRP3. To our knowledge, there was no report describing a direct interaction between 3D and NLRP3. Thus, we at the first time demonstrated that a viral RdRp interacts directly with NLRP3 to activate the inflammasome. Moreover, 3D alters the sub-cellular distributions of NLRP3, ASC, and pro-Casp-1, and is co-localized with the inflammasome components to form large spots through binding to NLRP3. It has been demonstrated that distribution of NLRP3 as spots is a character of the formation of inflammasome complex [6]. Thus, our results suggested that 3D may regulate the assembly of NLRP3 inflammasome.
In the absence of stimulation, NLRP3 is homo-oligomerized, which undergo conformational changes to form active inflammasome complex by interacting ASC upon stimulation [31]. 3D up-regulates the association of NLRP3 with ASC, suggesting that it is involved in the formation of inflammasome complex. The oligomerization of ASC is critical for Casp-1 activation and inflammasome function [24]. In the presence of NLRP3, oligomerization of ASC is significantly stimulated by 3D, indicating that 3D may facilitate the assembly of NLRP3 inflammasome through interacting with NLRP3. More interestingly, 3D, NLRP3, and ASC together formed a ring-like structure, in which 3D interacts with NLRP3 that subsequently interacts with ASC. We observed that in the ring-like complex, 3D is located inside, followed by the 3D-NLRP3 complex and then the NLRP3-ASC complex, and finally ASC is located outside of the structure. It has been reported that the inflammasome complex is assembled by the formation of a ring-like organization [32]. Taken together, we demonstrated that 3D facilitates the assembly of NLRP3 inflammasome complex by the formation of a “3D-NLRP3-ASC” structure through direct binding to NLRP3.
In conclusion, we revealed a novel mechanism by which EV71 stimulates the activation of NLRP3 inflammasome by the virus-encoded 3D RNA polymerase. More importantly, 3D interacts directly with NLRP3 to facilitate the assembly of NLRP3 inflammasome complex by forming a “3D-NLRP3-ASC” ring-like structure. During the formation of the special structure, 3D binds to the LRR domain of NLRP3 that subsequently interacts with ASC through the PYRIN domain, ASC in turn binds to pro-Casp-1 by the CARD domain and activates Casp-1 (p20/p10), which finally stimulates the cleavage and release of IL-1β(p17). Thus, in this specific ring-like structure, the viral protein sites in the center, followed by NLRP3 in the middle and ASC then locates outside. IL-1β acts as an important mediator of inflammation by stimulating the activation of immune cells and the production of many secondary pro-inflammatory cytokines [15], which play important roles in the development of inflammation and associated diseases [51]. Thus, this study discovers a new role of 3D as an important regulator in the activation of inflammatory response, reveals a novel mechanism underlying the regulation of inflammasome assembly mediated by viral invasion, and would provide new insights into development of agent for the treatment and prevention of viral associated inflammation and diseases.
Blood samples of healthy donors were randomly collected from Wuhan blood donation center (Wuhan, China). To isolate peripheral blood mononuclear cells (PBMCs), blood cells were separated from blood samples and diluted in RPMI-1640 purchased from Gibco (Grand Island, NY, USA). Diluted blood cells (5 ml) were added gently to a 15 ml centrifuge tube with 5 ml lymphocyte separation medium (#50494) purchased from MP Biomedicals (California, USA), and centrifuged at 2,000×g for 10 min at room temperature (RT). The middle layer was transferred to another new centrifuge tube and diluted with RPMI-1640. The remaining red blood cells were removed using red blood cell lyses buffer purchased from Sigma-Aldrich (St. Louis, MO, USA). The pure PBMCs were centrifuged at 1,500×g for 10 min at RT and cultured in RPMI-1640.
The study was conducted according to the principles of the Declaration of Helsinki and approved by the Institutional Review Board of the College of Life Sciences, Wuhan University in accordance with its guidelines for the protection of human subjects. The Institutional Review Board of the College of Life Sciences, Wuhan University, approved the collection of blood samples for this research, in accordance with guidelines for the protection of human subjects. Written informed consent was obtained from each participant.
Human rhabdomyosacroma cell line (RD) and human embryonic kidney cell line (HEK 293T) were purchased from American Type Culture Collection (ATCC) (Manassas, VA, USA). Human monocytic cells (THP-1) was a gift from Dr. Bing Sun of Institute of Biochemistry and Cell Biology, Shanghai Institute for Biological Sciences.
THP-1 cells were cultured in RPMI 1640 medium supplemented with 10% heat-inactivated fetal bovine serum (FBS), 100 U/ml penicillin, and 100 μg/ml streptomycin sulfate. RD and HEK293T cell lines were cultured in Dulbecco modified Eagle medium (DMEM) purchased from Gibco (Grand Island, NY, USA) supplemented with 10% fetal bovine serum (FBS), 100 U/ml penicillin, and 100 μg/ml streptomycin sulfate. Cells were maintained in an incubator at 37°C in a humidified atmosphere of 5% CO2.
Lipopolysaccharide (LPS), ATP, phorbol-12-myristate-13-acetate (TPA), and dansylsarcosine piperidinium salt (DSS) were purchased from Sigma-Aldrich (St. Louis, MO, USA). RPMI 1640 and Dulbecco modified Eagle medium (DMEM) were obtained from Gibco (Grand Island, NY, USA). Nigericin was obtained from InvivoGene Biotech Co., Ltd. (San Diego, CA, USA). Antibody against Flag (F3165) and monoclonal mouse anti-GAPDH (G9295) were purchased from Sigma. Monoclonal rabbit anti-NLRP3 (D2P5E), monoclonal rabbit anti-IL-1β (D3U3E), monoclonal rabbit anti-caspase-1 (catalog no. 2225) and monoclonal rabbit anti-eIF4G (C45A4) were purchased from Cell Signaling Technology (Beverly, MA, USA). Monoclonal mouse anti-ASC (sc-271054) and polyclonal rabbit anti-IL-1β (sc-7884) were purchased from Santa Cruz Biotechnology (Santa Cruz, CA, USA). Monoclonal mouse anti-NLRP3 (ALX-804-818) was purchased from Enzo Life Sciences (Shang Hai, China) to detection endogenous NLRP3 in THP1 cells by immunofluorescence microscopy. Polyclonal rabbit anti-3D antibody was produced by ABclonal Technology (Wuhan, China). Lipofectamine 2000, normal rabbit IgG and normal mouse IgG were purchased from Invitrogen Corporation (CA, USA).
In this study we used the Xiangyang strain of EV71 (GenBank accession number JN230523.1), which was previously isolated by our group [52,53]. The virus was adsorbed at 37°C for 2 h and the unbound virus was washed away. Infected cells were cultured in fresh medium supplemented with 2% FBS. For the preparation of UV-inactivated EV71, the virus was dispersed in a tissue culture dish, and a compact UV lamp was placed directly above the dish for 30 min. For the preparation of heat-inactivated EV71, the virus was incubated at 65°C for 30 min to completely inhibit the activity of EV71. Virus titration was performed using RD cells in 96-well plates and expressed as the 50% tissue culture infectious dose (TCID50) per unit volume. Sendai virus (SeV) strain was a gift from Dr. Hongbing Shu of Wuhan University. HCV genotype 2a strain JFH-1 was kindly provided by Takaji Wakita.
RD cells were infected with EV71 virus at an infection of 0.1 PFU/cell. The supernatant was collected when cells showed maximal cytopathic effect from viral infection, centrifuged at 2,500 rpm for 30min, and then passed through 0.4μm filters. We used the E.Z.N.Z. viral RNA Kit for the isolation of EV71 virus RNA from the cell culture supernatant. The concentration of viral RNA was measured by NanoDrop 2000 which was purchased from Thermo scientific.
THP-1 cells were differentiated to macrophages with 60 nM phorbol-12-myristate-13-acetate (TPA) for 12–14 h, and cells were cultured for 24 h without TPA. The differentiated cells were then stimulated in 6 cm plates with EV71 virus, Lipopolysaccharide (LPS), Nigericin, or ATP. Supernatants were collected for measurement of IL-1β by ELISA. Cells were harvested for real-time PCR or immunoblot analysis.
The cDNAs encoding human NLRP3, ASC, pro-Casp-1, and IL-1β were obtained by reverse transcription of total RNA from TPA-differentiated THP-1 cells, followed by PCR using specific primers. The cDNAs were sub-cloned into pcDNA3.1(+) and pcDNA3.1(+)-3×Flag vector. The pcDNA3.1(+)-3×Flag vector was constructed from pcDNA3.1(+) vector through inserting the 3×Flag sequence between the NheI and HindIII site. The primers used in this study are shown in S1 Table.
To construct plasmids expressing EV71 proteins 2A, 2B, 3C, 3A, 3C, and 3D, corresponding fragments of EV71 cDNA were cloned into pEGFPC1 between the HindIII and SalI sites, resulting in green fluorescent protein (GFP) fusion protein. To construct pCAggs-HA-3D, the EV71 3D region was sub-cloned into pCAggs-HA vector using the EcoRI and KpnI sites. To construct pGEX6p-1-3D, the EV71 3D region was sub-cloned into pGEX6p-1 vector using BamHI and EcoRI sites. The PYRIN, NACHT, and LRR domain of NLRP3 protein was cloned into pcDNA3.1(+)-3×Flag vector using specific primers shown in S1 Table.
The targeting sequences of shRNAs for the human NLRP3, ASC, and caspase-1 were as follows: sh-NLRP3: 5’-CAGGTTTGACTATCTGTTCT-3’; sh-ASC: 5’-GATGCGGAAGCTCTTCAGTTTCA-3’; sh-caspase-1: 5’-GTGAAGAGATCCTTCTGTA-3’. A PLKO.1 vector encoding shRNA for a negative control (Sigma-Aldrich, St. Louis, MO, USA) or a specific target molecule (Sigma-Aldrich) was transfected into HEK293T cells together with psPAX2 and pMD2.G with Lipofectamine 2000. We using the 3*Flag sequence to replace the GFP protein in the pLenti CMV GFP Puro vector (Addgene, 658–5) for adding some Restriction Enzyme cutting site (XbaI-EcoRV-BstBI-BamHI) before the 3×Flag tag. Then the pLenti vector encoding EV71 3D protein was transfected into HEK293T cells together with psPAX2 and pMD2.G with Lipofectamine 2000. The primers were shown in S1 Table. Culture supernatants were harvested 36 h and 60 h after transfection and then centrifuged at 2,200rpm for 15 min. THP-1 cells were infected with the supernatants contain lentiviral particles in the presence of 4 μg/ml polybrene (Sigma). After 48 h of culture, cells were selected by 1.5 μg/ml puromycin (Sigma). The results of each sh-RNA-targeted protein and the lenti-3D protein were detected by real-time PCR and/or immunoblot analysis.
The concentrations of IL-1β in culture supernatants were measured by ELISA kit (BD Biosciences, San Jose, CA).
The supernatant of the cultured cells was collected for 1 ml in the cryogenic vials (Corning). The supernatant was frozen in -80°C for 4 h. The Rotational Vacuum concentrator machine that was purchase from Martin Christ was used for the freeze drying. The drying product was dissolved in 100 μl PBS and mixed with SDS loading buffer for western blotting analysis with antibodies for detection of activated caspase-1 (D5782 1:500; Cell Signaling) or mature IL-1β (Asp116 1:500; Cell Signaling). Adherent cells in each well were lysed with the lysis buffer described below, followed by immunoblot analysis to determine the cellular content of various protein.
The HEK293T whole-cell lysates were prepared by lysing cells with buffer (50 mM Tris-HCl, pH7.5, 300 mM Nacl, 1% Triton-X, 5 mM EDTA, and 10% glycerol). The TPA-differentiated THP-1 cells lysates were prepared by lysing cells with buffer (50 mM Tris-HCl, pH7.5, 150 mM Nacl, 0.1% Nonidetp40, 5 mM EDTA, and 10% glycerol). Protein concentration was determined by Bradford assay (Bio-Rad, Hercules, CA, USA). Cultured cell lysates (30 μg) were electrophoresed in an 8–12% SDS-PAGE gel and transferred to a PVDF membrane (Millipore, MA, US). PVDF membranes were blocked with 5% skim milk in phosphate buffered saline with 0.1% Tween 20 (PBST) before being incubated with the antibody. Protein band were detected using a Luminescent image Analyzer (Fujifilm LAS-4000).
The HEK293T whole-cell lysates were prepared by lysing cells with buffer (50 mM Tris-HCl, pH7.5, 300 mM Nacl, 1% Triton-X, 5 mM EDTA, and 10% glycerol). The TPA-differentiated THP-1 cells lysates were prepared by lysing cells with buffer (50 mM Tris-HCl, pH7.5, 150 mM Nacl, 0.1% Nonidetp40, 5 mM EDTA, and 10% glycerol). Lysates were immunoprecipitated with control mouse immunoglobulin G (IgG) (Invitrogen) or anti-Flag antibody (Sigma, F3165) with Protein-G Sepharose (GE Healthcare, Milwaukee, WI, USA).
TPA-differentiated THP-1 cells was cultured or infected by the EV71 (MOI = 20) virus for 24 h. Cells were fixed in 4% paraformaldehyde at room temperature for 15 min. After being washed three times with PBS, permeabilized with PBS containing 0.1% Triton X-100 for 5 min, washed three times with PBS, and finally blocked with PBS containing 5% BSA for 1 h. The cells were then incubated with the monoclonal mouse IgG1 anti-NLRP3 antibody (ALX-804-818-C100; Enzo life Sciences) and the polyclonal rabbit anti-3D antibody (ABclonal technology) overnight at 4°C, followed by incubation with FITC-conjugate donkey anti-mouse IgG and Dylight 649-conjugate donkey anti-rabbit IgG (Abbkine) for 1 h. After washing three times, cells were incubated with DAPI solution for 5 min, and then washed three more times with PBS. Finally, the cells were analyzed using a confocal laser scanning microscope (Fluo View FV1000; Olympus, Tokyo, Japan).
Total RNA was extracted with TRIzol reagent (Invitrogen), following the manufacturer’s instructions. Real-time quantitative RT-PCR was performed using the Roche LC480 and SYBR RT-PCR kits (DBI Bio-science, Ludwigshafen, Germany) in a reaction mixture of 20 μl SYBR Green PCR master mix, 1 μl DNA diluted template, and RNase-free water to complete the 20 μl volume. Real-time PCR primers were designed by Primer Premier 5.0 and their sequences were was as follows: VP1 forward, 5’-CCCTTTAGTGGTTAGGATTT-3’, VP1 reverse, 5’-CACCAGTTGGTTTAATGGAG-3’; NLRP3 forward, 5’-AAGGGCCATGGACTATTTCC-3’, NLRP3 reverse, 5’-GACTCCACCCGATGACAGTT-3’; ASC forward, 5’-AACCCAAGCAAGATGCGGAAG-3’, ASC reverse, 5’-TTAGGGCCTGGAGGAGCAAG-3’; caspase-1 forward, 5’-TCCAATAATGCAAGTCAAGCC-3’, caspase-1 reverse, 5’-GCTGTACCCCAGATTTTGTAGCA-3’; IL-1β forward, 5’-CACGATGCACCTGTACGATCA-3’, IL-1β reverse, 5’-AGCACTGAAAGCATGA-3’, TNF-α forward, 5’-GGGTTTGCTACAACATGG-3’, TNF-α reverse, 5’-AAGAACTTAGATGTCAGTGC-3’, IL-8 forward, 5’-ACTTCTCCACAACCCTCTGC-3’, IL-8 reverse, 5’-GTTGCTCCATATCCTGTCCCT-3’; GAPDH forward, 5’-AAGGCTGTGGGCAAGG-3’, GAPDH reverse, 5’-TGGAGGAGTGGGTGTCG-3’.
The construct pGEX6p-1-3D plasmid and pGEX6p-1-LRR were transfected into Escherichia coli strain BL21. After growing in LB medium at 3°C until the OD600 reached 0.6–0.8, Isopropyl β-D-1-thiogalactopyranoside (IPTG) was added to a final concentration of 0.1 mM and the cultures grew for an additional 16 h at 16°C for GST-3D protein. Isopropyl β-D-1-thiogalactopyranoside (IPTG) was added to a final concentration of 1 mM and the cultures grew for an additional 4 h at 37°C for GST-LRR protein. And then the GST protein, GST-3D protein and GST-LRR protein were purified from the E. coli bacteria. For GST-3D pull-down assay, glutathione-Sepharose beads (Novagen) were incubated with GST-3D or GST protein. After washed with phosphate-buffered saline (PBS), these beads were incubated with cell lysates from HEK293T which were transfected with plasmids encoding Flag-NLRP3, Flag-ASC and Flag-pro-caspase-1 for 4 h at 4°C. The precipitates were washed three times, boiled in 2×SDS loading buffer, separated by 10% SDS-PAGE, immunoblotted with anti-GST, anti-Flag, and ant-ASC antibody. It was the same for the GST-LRR pull down assay.
Saccharomyces cerevisiae strain AH109, control vectors pGADT7, pGBKT7, pGADT7-T, pGBKT7-lam, and pGBKT7-p53 were purchased from Clontech (Mountain View, CA, USA). Yeast strain AH109 was co-transformed with the combination of the pGADT7 and the pGBKT7 plasmids. Transformed yeast cells containing both plasmids were first grown on SD-minus Trp/Leu plates (DDO) to maintain the two plasmids and then were sub-cloned replica plated on SD-minus Trp/Leu/Ade/His plate (QDO).
The TPA-differentiated THP-1 cells were lysed by buffer (50 mM Tris, pH7.5, 150 mM Nacl, 1% Nonidetp40, 5 mM EDTA, and 10% glycerol) at 4°C. The transfected HEK293T cells were lysed by buffer (50 mM Tris-HCl, pH7.5, 300 mM Nacl, 1% Triton-X, 5 mM EDTA, and 10% glycerol). Lysates were centrifugated at 6000rpm for 15 min. The supernatants of the lysates were mixed with SDS loading buffer for western blot analysis with antibody against ASC. The pellets of the lysates were washed with PBS for three times and cross-linked using fresh DSS (2 mM, sigma) at 37°C for 30 min. The cross-linked pellets were then spanned down and mixed with SDS loading buffer for western blotting analysis.
All experiments were reproducible and repeated at least three times with similar results. Parallel samples were analyzed for normal distribution using Kolmogorov-Smirnov tests. Abnormal values were eliminated using a follow-up Grubbs test. Levene’s test for equality of variances was performed, which provided information for Student’s t-tests to distinguish the equality of means. Means were illustrated using histograms with error bars representing the SD; a P value of <0.05 was considered statistically significant.
|
10.1371/journal.pbio.1000349 | Pro-Survival Role for Parkinson's Associated Gene DJ-1 Revealed in Trophically Impaired Dopaminergic Neurons | The mechanisms underlying the selective death of substantia nigra (SN) neurons in Parkinson disease (PD) remain elusive. While inactivation of DJ-1, an oxidative stress suppressor, causes PD, animal models lacking DJ-1 show no overt dopaminergic (DA) neuron degeneration in the SN. Here, we show that aging mice lacking DJ-1 and the GDNF-receptor Ret in the DA system display an accelerated loss of SN cell bodies, but not axons, compared to mice that only lack Ret signaling. The survival requirement for DJ-1 is specific for the GIRK2-positive subpopulation in the SN which projects exclusively to the striatum and is more vulnerable in PD. Using Drosophila genetics, we show that constitutively active Ret and associated Ras/ERK, but not PI3K/Akt, signaling components interact genetically with DJ-1. Double loss-of-function experiments indicate that DJ-1 interacts with ERK signaling to control eye and wing development. Our study uncovers a conserved interaction between DJ-1 and Ret-mediated signaling and a novel cell survival role for DJ-1 in the mouse. A better understanding of the molecular connections between trophic signaling, cellular stress and aging could uncover new targets for drug development in PD.
| The major pathological event in Parkinson disease is the loss of dopaminergic neurons in a midbrain structure, the substantia nigra. The study of familial Parkinson disease has uncovered several disease-associated genes, including DJ-1. Subsequent studies have suggested that the DJ-1 protein is a suppressor of oxidative stress that might modify signaling pathways that regulate cell survival. However, because animal models lacking DJ-1 function do not show dopaminergic neurodegeneration, the function(s) of DJ-1 in vivo remain unclear. Using mouse genetics, we found that DJ-1 is required for survival of neurons of the substantia nigra only in aging conditions and only in neurons that are partially impaired in receiving trophic signals. Aging mice that lack DJ-1 and Ret, a receptor for a neuronal survival factor, lose more dopaminergic neurons in the substantia nigra as compared with aging mice that lack only Ret. Using the fruit fly Drosophila, we determined that DJ-1 interacts with constitutively active Ret and with its associated downstream signaling pathways. Therefore, understanding the molecular connections between trophic signaling, cellular stress and aging could facilitate the identification of new targets for drug development in Parkinson Disease.
| Specific and progressive loss of substantia nigra (SN) neurons is the central pathogenic event in Parkinson disease (PD), the most common movement neurodegenerative disorder, characterized by tremor, rigidity, and bradykinesia. A second pathological feature of PD is the presence of aggregated alpha-synuclein (Lewy Bodies) in the remaining SN neurons. In most PD patients the degree of dopaminergic axon degeneration in the SN target area, the striatum, exceeds that of SN cell body loss, suggesting a “dying back” model, whereby the axonal compartment is the first target of degenerative insults [1]. A major advance in PD research was the discovery of familial PD-associated genes and the characterization of their biochemical mechanisms [2],[3]. So far, transgenic mouse models that reproduce the genetic defects found in familial PD showed limited power in reproducing disease pathology, and most of them fail to exhibit degeneration of SN neurons [4] (see also [5]). This together with the fact that familial PD accounts for less than 10% of all PD cases (the rest being sporadic) suggests that multiple hits, including environmental factors, underlie selective neuronal death [6].
Access to neurotrophic factors is critical for maintenance of the nigrostriatal system in mice, and novel neurotrophic factors for SN neurons have recently been described [7]–[9]. We recently showed that genetic ablation of the receptor tyrosine kinase (RTK) Ret, the signaling receptor of glial cell line-derived neurotrophic factor (GDNF), led to adult-onset progressive and specific degeneration of the nigrostriatal system [10]. Consistent with a “dying back” model, Ret function was found to be important for striatal DA fiber maintenance, while its role in cell body survival was relatively moderate. Removal of GDNF in the adult brain led to more pronounced degeneration [8], suggesting that SN neurons in Ret mutant mice still have access to trophic support via Ret-independent pathways [11].
The PD-associated gene DJ-1 [12] encodes a small, dimeric, single domain protein that is thought to respond to oxidative stress and to protect neurons from environmental toxins [2],[13]–[16]. However, the molecular mechanisms underlying DJ-1 function are unclear. DJ-1 is localized to cytoplasm, nucleus, and mitochondria, and in each of these subcellular localizations DJ-1 may be neuroprotective [3],[17]. DJ-1 is an oncogene and was shown to synergize with the Ras/MAPK pathway in controlling cellular transformation [18]. It was suggested to negatively regulate the tumor suppressor PTEN, the major negative regulator of the phosphatidylinositol (PI)-3 kinase pathway [19]. DJ-1 ablation in mice alone did not affect the survival of SN neurons [20]–[25] but rendered SN neurons more sensitive towards the toxin MPTP [22]. A small (7%) population of ventral tegmental area (VTA) neurons requires DJ-1 during development for tyrosine hydroxylase (TH) expression [25]. Hence, it is currently not understood why loss of DJ-1 in humans causes specific loss of SN neurons.
Based on the capacity of DJ-1 to interact with pathways implicated in RTK signaling (PI3K/Akt and Ras/MAPK) [18],[19],[26], we investigated a possible cooperation between Ret and DJ-1 in regulating SN neuron survival in vivo. To this end, we generated double mutant mice lacking expression of Ret in midbrain dopaminergic neurons and DJ-1 in all cells of the body (DAT-Cre;Retlx/lx;DJ-1−/− mice, in short DAT-Ret;DJ-1 mice). Here we show that DAT-Ret;DJ-1 mice have significantly fewer nigral DA neurons than either single mutant, indicating that under conditions of trophic impairment, DJ-1 promotes survival of aging DA neurons. Remarkably, the loss is specific to GIRK-2 positive SN neurons, which project exclusively to the striatum and are more vulnerable in PD. Moreover, DJ-1 does not appear to promote target innervation, suggesting that DJ-1 acts at the level of the DA cell body, not in the axon. To understand how Ret-mediated trophic support relates molecularly to DJ-1, we used Drosophila whose genome contains two genes, termed DJ-1A and DJ-1B, which share significant homology with human DJ-1. Drosophila DJ-1 mutants have been shown to be sensitive to environmental toxins associated with PD [15] and to genetically interact with the PI3K/PTEN/Akt signaling pathway [19],[26]. Our genetic interaction studies in the eye system revealed that DJ-1A/B interact genetically with constitutively active Ret and associated Ras/MAPK, but not PI3K/Akt signaling. Paralleling our mouse results, we found that combined deletion of ERK and DJ-1 in Drosophila enhanced the developmental defects during eye and wing development caused by ERK deletion, providing evidence for an important role for the interaction between DJ-1 and RTK-related signaling during evolution.
We have previously shown that the Ret protein co-localizes with the dopaminergic marker TH [10]. Similarly, DJ-1 is expressed in SN and VTA neurons [27]. We determined by Western blotting that the expression of DJ-1 was not modified in the midbrain and striatum of aged DAT-Ret mice and vice versa (Figure S1). In cultured SH-SY5Y neuroblastoma cells, endogenous Ret expression was not modified when DJ-1 expression was downregulated by RNAi, nor were the levels of endogenous DJ-1 changed when these cells were stimulated with GDNF (Figure S1); thus, Ret and DJ-1 protein levels appear to be regulated by separate mechanisms. DAT-Ret;DJ-1 double mutant mice are viable and fertile. To detect morphological alterations in the nigrostriatal system, brain tissue sections of mutant and control mice were immunostained for TH and subjected to stereological quantification. In 3-mo-old DAT-Ret;DJ-1 double mutant mice, the numbers of TH-positive SN neurons were unchanged compared to age-matched controls (13,690±428 in control and 13,709±248 in DAT-Ret;DJ-1 mice, n = 3 mice/group, p = 0.95, student's t test) indicating that the nigrostriatal system developed normally in these mutants. When mutant mice were aged, however, the numbers of TH-positive SN neurons decreased significantly compared to age-matched controls (Figure 1A–C,G,H). In DAT-Ret;DJ-1 double mutant mice, the reduction was more pronounced (37% at 18 mo, 41% at 24 mo) than in DAT-Ret single mutant mice (24% at 18 mo, 25% at 24 mo). The difference between DAT-Ret;DJ-1 double and DAT-Ret single mutant mice was statistically significant (p<0.01, t test) and was not additive, since DJ-1 single mutants had normal numbers of TH-positive neurons (Figure 1G,H). Anti-Pitx3 immunostaining was used as an independent marker and revealed a similar reduction in SN neurons in DAT-Ret;DJ-1 double and DAT-Ret single mutant mice (Figure 1D–F,I). Because approximately one third of the neurons in the SN are non-dopaminergic, we also used the pan-neuronal marker NeuN to label all neurons in the SN and found that aged DAT-Ret;DJ-1 double mutant mice had significantly fewer neurons in the SN relative to DAT-Ret or control mice (Figure 1J). Since the analysis of TH-, Pitx3-, or NeuN-immunolabeled neurons yielded similar numbers of missing neurons in DAT-Ret;DJ-1 mice, we conclude that combined deletion of Ret and DJ-1 causes enhanced degeneration of SN neurons, relative to deletion of Ret alone. As we had previously shown for DAT-Ret single mutants [10], the observed defects were region specific: The nearby VTA region was not affected in DAT-Ret;DJ-1 double mutants (Figure 1L). The previously observed small (7%) decrease in TH-positive VTA neurons in DJ-1−/− mice [25] was not seen in this analysis, possibly because of a small shift in genetic background due to the presence of the Retlx allele. Finally, we excluded that the DAT-Cre transgene and the mutant DJ-1 allele somehow genetically interacted by comparing the numbers of TH-positive neurons in DAT-Cre;DJ-1−/− compound mice to DAT-Cre transgenics and littermate controls (wild-type and DJ-1+/− mice; Figure 1K). Together these results indicate a requirement for endogenous DJ-1 in maintaining SN neurons, when they are impaired in receiving Ret-mediated trophic signals.
Next we asked which SN subpopulation was affected by DJ-1: A9 neurons located in the ventral tier of the SN and projecting to the dorsal striatum are preferentially lost in PD [28]. They express the G-protein gated, inwardly rectifying potassium channel GIRK2 [29]. A9 neurons located in the dorsal tier of the SN and A10 neurons of the VTA project to different areas including limbic and neocortical regions. They express the calcium-binding protein Calbindin [30]. In 24-mo-old mice, removal of DJ-1 had no effect on the number of GIRK2-positive neurons as compared to littermate controls (Figure 2A,B,I). In contrast, removal of Ret alone caused a partial reduction of GIRK2-positive neurons (33% loss) and combined removal of Ret and DJ-1 had the strongest effect (51% loss; p<0.001 DAT-Ret;DJ-1 double versus CTRL; p<0.01 DAT-Ret;DJ-1 double versus DAT-Ret single mutants, t test; Figure 2C,D,I). Interestingly, the Calbindin-positive subpopulation in the SN was unaffected in all groups (Figure 2E–H,J). Our stereological quantifications revealed that approximately 9,600 SN neurons express GIRK2, while the remaining 3,700 neurons express Calbindin (Figure 2). If all 5,500 TH-positive neurons that were lost in DAT-Ret;DJ-1 mice were also GIRK2-positive, we would have expected a 57% loss of GIRK2 neurons (5,500 out of 9,600) and no loss of Calbindin-positive neurons. If, however, both populations had been equally vulnerable, we would have expected a 41% loss in both populations. The observed 51% loss in the GIRK2 subpopulation and no statistically significant loss of Calbindin-positive neurons suggest a much higher vulnerability of the GIRK2 subpopulation in DAT-Ret;DJ-1 and DAT-Ret mice. The quantification of soma sizes of surviving GIRK2-positive neurons revealed a small but significant reduction (9%) of the mean soma size in DAT-Ret single mutants compared to control littermates; this effect was not further enhanced in DAT-Ret;DJ-1 double mutants (Figure 2K–M).
We next evaluated the possibility that Ret and DJ-1 cooperate in maintaining target innervation of nigral DA neurons. The quantification of TH-positive fiber density confirmed a marked decrease in the dorsal striatum of 18- and 24-mo-old DAT-Ret single mutants compared to age-matched controls (Figure 3A,C,H,I; see also [10]). In contrast, no significant reductions of TH-positive fibers were observed in DJ-1 single mutant mice (Figure 3B,H,I). Interestingly, DAT-Ret;DJ-1 double mutants displayed reductions of TH-positive fibers that were not significantly different from DAT-Ret single mutants (46% at 18 mo and 52% at 24 mo, Figure 3D,H,I). Similar results were obtained when the dopamine transporter (DAT) protein was used as an independent marker for DA terminals (54% reduction in both mutant lines; Figure 3E–G,J). In this case the DAT-Cre knock-in mice were used as controls, since they have reduced levels of DAT protein (unpublished data) due to the loss of one functional copy of the DAT gene. These results indicate that DJ-1 is not required for maintaining target innervation in DA neurons that are partially impaired in receiving trophic support. To evaluate the motor performance of aged mutant mice, we followed their horizontal activity in an open field arena. Consistent with previous observations [21], DJ-1 null mice were found to be hypoactive, despite having normal numbers of SN neurons and normal target innervation (Figure 3K). Mice carrying the DAT-Cre transgene inserted into the 5′ UTR of the DAT gene were slightly hyperactive (Figure 3K), in agreement with previous reports [31]. Removal of Ret or Ret and DJ-1 function did not further modify motor behavior compared to DAT-Cre control mice (Figure 3K). We then measured the levels of total striatal dopamine in these mutants and found a significant increase in dopamine levels in mice carrying the DAT-Cre transgene, while removal of Ret or Ret and DJ-1 did not further modify these levels compared to DAT-Cre control mice (Figure 3L). The TH enzyme is a critical regulator of dopamine production in DA neurons, and our analysis of TH levels in the different aging mutants revealed no significant differences in TH levels (Figure 3M,N). Taken together, these results suggest the existence of compensatory mechanisms that maintain dopaminergic homeostasis in DAT-Ret and DAT-Ret;DJ-1 mice, despite the occurrence of partial neurodegeneration in the SN and the striatum.
Dopaminergic-specific deletion of Ret leads to enhanced astrogliosis, but not microglial recruitment in the striatum of 24-mo-old mice (Figure 4 and [10]). Using the microglial marker Ionized binding calcium adapter molecule (Iba-1) and the astrocytic marker glial fibrillary acidic protein (GFAP), we evaluated the occurrence of neuroinflammatory processes in the striatum of aged DAT-Ret;DJ-1 mice and corresponding controls. We found no enhanced recruitment of Iba-1-positive microglial cells in DAT-Ret;DJ-1 or DAT-Ret mice compared to controls (Figure 4A–D). The recruitment of reactive astrocytes in the DAT-Ret striatum was found to be significantly elevated after 24 mo, while additional removal of DJ-1 did not enhance this process (Figure 4E–N). These observations correlate with the above-mentioned histological, behavioral, and physiological measurements and suggest that removal of Ret and DJ-1 function does not exacerbate the structural and functional defects in SN axon terminals caused by Ret deprivation.
To obtain independent evidence for genetic interaction between Ret and DJ-1 and to begin characterizing the underlying intracellular pathways, we used the developing Drosophila eye system, which is very sensitive to dosage changes in RTK signaling and downstream components of the PI3K/Akt and Ras/Mapk pathways. While Drosophila DJ-1B is ubiquitously expressed, DJ-1A appears to be enriched in certain tissues such as testes [14],[15]. We used a DJ-1 specific antibody [15] and confirmed that DJ-1B is expressed at high levels in the adult head; DJ-1A expression was not detected in WB, but the presence of the DJ-1A transcript was confirmed by RT-PCR (Figure 5B) [14]; moreover, overexpression of constitutively active versions of Ret, Raf, ERK/rolled, or wild-type Akt1 did not modify endogenous DJ-1 levels (Figure 5A). Flies homozygous for DJ-1A and/or DJ-1B null alleles and flies overexpressing DJ-1A or DJ-1B in the eye (using the photoreceptor neuron-specific driver GMR-Gal4) displayed normal eye development and ultrastructure (unpublished data). Drosophila Ret (dRet) is highly homologous to mammalian Ret [32] and exhibits activities associated with human Ret both in tissue culture cells and during Drosophila eye development [33],[34]. A function for dRet has so far not been described in Drosophila, and in addition, dRet does not bind mammalian GDNF; we therefore utilized previously generated constitutively active forms of dRet (dRetMEN2A/B) that interact with the same pathways as WT Ret and were used to screen for novel Ret interactors [34]. Flies carrying the GMR driver fused to dRetMEN2B (GMR-dRetMEN2B) [34] develop with adult eyes of reduced size and rough morphology. Ommatidia sizes were increased by 35% and individual ommatidia were often fused together, had abnormal polarity, and had poorly patterned interommatidial spaces (Figure 5G,H,J). Despite an increase in ommatidia size, the overall eye size in GMR-dRetMEN2B flies was decreased by 30% compared to controls (Figure 5C,D,F), as a result of a late (pupal) pro-apoptotic wave induced by excessive proliferation and differentiation defects [34]. To determine whether DJ-1 is a dRet interactor, we crossed GMR-dRetMEN2B flies with flies carrying DJ-1A and/or DJ-1B microdeletions [15]. Remarkably, the defects in eye and ommatidia sizes induced by overactive dRet were completely rescued in flies with reduced DJ-1A/B levels (Figure 5E,F,I,J). Similar results were obtained with independent DJ-1A/B loss-of-function alleles and the GMR-dRetMEN2A gain-of-function allele (Figure S2). To test whether overexpression of both Ret and DJ-1 led to a more severe phenotype than the ones induced by active dRet alone, we overexpressed DJ-1A in flies with a moderate Ret-overexpression phenotype (GMR-Gal4/UAS-dRetMEN2A). The resulting flies displayed an enhanced eye phenotype (Figure 5K–N). Thus, both DJ-1A/B interact genetically with overactive dRet in controlling cell size and differentiation in the developing fly retina.
To gain insights into the mechanism(s) underlying the genetic interaction between Ret and DJ-1, we investigated the capacity of fly DJ-1A/B to genetically modify pathways that are known to mediate Ret function: PI3K/Akt and Ras/Mapk [33]. Strong overexpression of wild-type PI3K (GMR/PI3KWT at 30°C) led to a 25% increase in eye size and to a disorganized retina compared to controls (GMR-Gal4) (Figure 6A,B,D–F). These phenotypes were not rescued in a DJ-1B−/− background (Figure 6C,D,G). To test whether overexpression of DJ-1A/B could enhance the phenotype of increased PI3K/Akt signaling, we used the moderate eye phenotype induced by wild-type Akt1 overexpression (25% increase in eye and 20% increase in ommatidia sizes). The Akt1 overexpression phenotype was not further exacerbated by DJ-1A or DJ-1B overexpression, nor did the resulting eyes become disorganized (Figure 6H–Q). Similar results were obtained in flies expressing a constitutively active version of PI3K (PI3KCAAX, [35], Figure S3). Conversely, reduced PI3K activity (using the GMR-driven expression of a PI3K dominant negative version), leading to a moderate reduction in eye size and to loss of photoreceptors, was not further enhanced in a DJ-1B null background (unpublished data). Our findings are in apparent contrast to published reports indicating genetic interactions of Drosophila DJ-1 with PTEN, an inhibitor of Akt [19], and with mammalian DJ-1 being a modulator of PI3K/Akt signaling in cultured cells [19],[36]. We found that DJ-1A/B overexpression only mildly rescued the effects of PTEN overexpression (reduced eye and ommatidia size) and that a reduction in DJ-1A/B function did not visibly enhance the PTEN overexpression phenotype (Figure S3). Furthermore, our experiments using different cell lines failed to reproduce the previously reported modulation of Akt activation by DJ-1, in conditions of DJ-1 overexpression, RNAi knockdown, or in DJ-1−/− mouse embryonic fibroblasts (MEFs; Figure S4). Thus, our data suggest that DJ-1 does not synergize with PI3K/Akt signaling during eye development, nor does DJ-1 modulate the activation status of Akt under normal conditions. DJ-1 might interact with PTEN only in defined situations (e.g., in oncogenic conditions) but in a PI3K-Akt independent manner (see Discussion).
To investigate the interaction of DJ-1 with the Ras/ERK pathway, we used constitutively active versions of Ras and the Mapk ERK/rolled (rl), which impair eye development by promoting excessive proliferation and altered cell differentiation [37]. Overexpression of active Ras in R7 photoreceptor neurons with the sevenless promoter (Sev-RasV12) led to the induction of multiple R cells/ommatidium and to a rough eye phenotype (Figure 7A,B,F,G,U). This phenotype was rescued by reducing endogenous DJ-1A/B levels (Sev-RasV12/DJ-1A+/−/DJ-1B+/−) (Figure 7C,U). Sev-RasV12 flies displayed on average 8.5 R cells/ommatidium while in Sev-RasV12/DJ-1A+/−/DJ-1B+/− flies only 7.5 R cells/ommatidium were detected (Figure 7H,U). In addition, in Sev-RasV12 flies, 67% of ommatidia were abnormally fused with their neighbours, compared to only 5% in Sev-RasV12/DJ-1A+/−/DJ-1B+/− flies (Figure 7G,H and unpublished data). Conversely, overexpression of DJ-1A further enhanced the Ras-overexpression phenotype to 11 R cells/ommatidium in Sev-RasV12/GMR/DJ-1A retinas (Figure 7D,E,I–U). We next assessed the modulation of constitutively active rolled signalling (GMR/rlSEM) by DJ-1. Overexpression of rlSEM led to supernumerary photoreceptor neurons (9.5 R cells/ommatidium). DJ-1A/B overexpression or partial reduction of DJ-1B levels did not modulate this phenotype (Figure 7K–T,V). Moreover, in cultured cells, increasing or decreasing DJ-1 levels did not modulate the phosphorylation status of ERK1/2 under basal conditions or following stimulation by growth factors (Figure S4; see also [38]). These results suggest that DJ-1A/B function either between Ras and ERK or in parallel to the Ras/ERK pathway to control cell differentiation and proliferation induced by overactive Ras/Mapk signaling.
Our loss-of-function mouse experiments suggest that DJ-1 acts in parallel to Ret-mediated signalling to control dopaminergic neuron survival. The Drosophila interactions between DJ-1 and ectopic Ras signalling raise the possibility that DJ-1 acts in parallel to the Ret induced Ras/Erk pathway to control optimal activation of Mapk downstream targets. To test this possibility and to investigate whether DJ-1 interacts with endogenous Erk signalling in Drosophila, we performed double loss-of-function experiments. We chose to investigate this interaction in two places where Erk/rolled is known to play a crucial role during development: the development of photoreceptor neurons and wing venation. Flies carrying two hypomorphic rolled alleles (rl1) displayed a moderate eye phenotype caused by a mild reduction in the number of R cells/ommatidium (6.64; Figure 8A,C,E–G,Q,R). While control and DJ-1B−/− flies had a normal appearance and a normal complement of 7 R cells/ommatidium, eyes of rl1/rl1;DJ-1B−/− flies were significantly smaller, rough, and displayed on average only 5.34 R cells/ommatidium (Figure 8A–H,Q,R; p<0.05 CTRL versus rl1/rl1; p<0.001 rl1/rl1 versus rl1/rl1 DJ-1B−/−, t test). DJ-1B is thus required, as a rolled interactor to control photoreceptor neuron development. We then investigated the development of the wing and found that rl1/rl1 flies had a very mild defect in wing venation, the vein L4 being sometimes thinner (in about 20% of the animals; Figure 8I–P,S). While in control and DJ-1B−/− flies the L4 vein developed normally, in rl1/rl1;DJ-1B−/− mutants the thinning of the L4 vein was either short (in 33% of animals), long (37% of all cases), or the L4 vein was interrupted (in 25% of animals, Figure 8I–P,S). Such an enhanced phenotype was also seen in flies carrying a combination of rl1 and the stronger allele rl10 (a deficiency; [37]). DJ-1 is thus required, as a rolled interactor, to control the development of the wing. These results uncovered a novel DJ-1B−/− phenotype in the unchallenged fly and suggest that DJ-1B cooperates with Ras/Mapk signalling during photoreceptor neuron and wing imaginal disc development.
The mechanisms underlying the selective death of SN neurons in PD are at present unclear. Mutations in DJ-1 cause PD and, although a number of in vitro studies have suggested several functions for DJ-1 in controlling cell survival and stress response, it remained unclear whether DJ-1 plays at all a role in neuronal survival in vivo. To explore the survival function of DJ-1 in aging neurons, we took a genetic approach that in difference to cell culture models allowed us to study the relevance of DJ-1 during the life span of the mouse and within the physiological environment of the brain. Remarkably, we found that DJ-1 becomes critical for survival only under trophic deprivation situations and only during aging. Current clinical trials examine the effects of GDNF (and its relative Neurturin) delivery for PD, although with conflicting results, suggesting that improvements in both technology and biological understanding of GDNF action are required [7],[39]. We have previously shown that GDNF signaling via the Ret RTK is required for maintenance of aging SN neurons [10]. The effect of Ret inactivation on cell survival was relatively moderate compared to its effect on DA axons innervating the striatum. The identification of new interactors for Ret might better define the context in which GDNF delivery is most beneficial and might provide new clues for PD therapy [7]. We show here DJ-1 acts in parallel to Ret-induced pathways to control SN DA neuron survival, and complementary studies in Drosophila suggest that the interaction between Ret and DJ-1 involves primarily Ras/ERK signaling. Because Ret and DJ-1 show convergence of their pro-survival activities, we suggest exploring the possibility that GDNF delivery might be most effective in PD patients carrying DJ-1 mutations, i.e. by activation of DJ-1 downstream signaling via Ret activation.
Based on these results we propose a model in which DJ-1 primarily promotes the survival of DA neurons that have suffered from an independent hit (trophic insufficiency) and have greatly reduced target innervation. In aged Ret single mutant mice, we have previously shown that the loss of target innervation exceeded cell loss; hence these mice contained a fraction of cells that survived during aging but had strongly reduced target innervation [10]. The present study shows that in aged DAT-Ret/DJ-1 double mutants, additional cell loss occurs such that the degree of cell loss exactly matches the degree of target innervation loss. The simplest explanation is that additional DJ-1 removal primarily leads to loss of cells that have strongly reduced target innervation (due to Ret signaling loss), while the larger fraction of cells with functional connections to the striatum remains unaffected. Alternatively, DJ-1 removal may affect both cell populations; however, as the projections to the striatum do not decrease further in the double mutant, surviving neurons would have to resprout and innervate the vacated target area to compensate for the expected loss of innervation. Since our previous study [40] showed that Ret signaling controls DA resprouting after toxic lesions, we find the latter explanation less likely.
The fact that removal of DJ-1 in Ret-deficient mice only accelerated the loss of DA cell bodies but not axons suggests that DJ-1 might exert its pro-survival activities in the SN dopaminergic cell body. Neurotrophic factor receptors, such as Ret, are transported from distal sites to the cell body, where they signal to promote survival. Components of the signaling machinery including activated Ras are also transported to the cell body in signaling vesicles [41]. Recent work succeeded in genetically uncoupling the survival requirements for the axon and cell body compartments. Specific molecules primarily regulate maintenance of cell bodies but not axons (including Bax, Bcl2, and JNK), further supporting the notion that different survival mechanisms operate in these two neuronal compartments [42]. Understanding the differential vulnerability of the axonal and cell-body compartments to aging and degenerative insults might improve our understanding of neurodegeneration and open new therapeutic avenues.
Why is there a specific requirement for Ret and DJ-1 activity in the GIRK2-positive subpopulation of SN neurons, considering that Ret and DJ-1 are expressed by most if not all DA neurons in SN and VTA [10],[27],[43]? GIRK2-positive neurons appear to be more sensitive than calbindin-positive neurons to toxic insult [44],[45] and calbindin-positive SN neurons are specifically spared in PD [46]. The presence and activity of GIRK2 itself may be a cause of vulnerability, since elevating the levels of GIRK2 further sensitizes these neurons [44]. Remarkably, GDNF was found to acutely modulate the excitability of midbrain dopaminergic neurons by inhibiting A-type K+ channels, a function that specifically involves the Mapk pathway [47]. Although the effects of Ret signaling on GIRK2 have not been studied, it is tempting to speculate that the modulation of Ras/Mapk signaling by Ret and DJ-1 also affects GIRK2 function and vulnerability of dopaminergic neurons. Further studies focusing specifically on the GIRK2 subpopulation will better define the exact biochemical processes involved in their survival and their interplay with other age-related cellular changes.
Our results show that DJ-1 promotes survival of dopaminergic neurons only in conditions of aging and trophic insufficiency, suggesting that the function(s) of DJ-1 might only be uncovered in specific circumstances. The lack of a strong phenotype in DJ-1 null mutants has prevented the analysis of DJ-1 function in vivo. In vitro studies have suggested several functions for DJ-1 [18],[19],[26],[48],[49]; however, these proposed functions remain to be validated in vivo. Because fly genetics has previously uncovered PD-associated mechanisms [50], we chose to investigate genetic interactions between Ret and DJ-1 in the Drosophila eye. Constitutively active versions of Ret mediate excessive cell proliferation, abnormal increases in cell size, differentiation, and polarity defects. These defects induce a late-onset pro-apoptotic wave (in late pupal stages), resulting in adult eyes of reduced size with fewer ommatidia. Thus, even though Ret is a pro-survival regulator in mammalian systems, its excessive activation in the fly developing retina induces developmental abnormalities that indirectly lead to partial eye degeneration. Loss of endogenous Drosophila DJ-1A/B reduces the signaling output of activated Ret and largely alleviates these developmental defects. Photoreceptor cell size and the balance between photoreceptor proliferation and differentiation are returned to normal in a DJ-1A/B loss-of-function background. When we instead activated the PI3K/Akt pathway in the eye, endogenous DJ-1A/B were not required, nor was DJ-1A/B overexpression sufficient, to modulate this phenotype, indicating that DJ-1A/B do not interact with the PI3K/Akt pathway. This is in contrast to previous work that proposed DJ-1 to be a potent modulator of the PI3K/Akt pathway [19]. However, recent work by the same group suggests that DJ-1 may do so only in oncogenic (hypoxia) situations [36], and our results suggest that the interaction between DJ-1 and PTEN is PI3K/Akt-independent. Indeed, emerging evidence suggests lipid phosphatase-independent roles of PTEN [51], the best studied being a protein-phosphatase-dependent inhibition of Ras/MAPK signaling [52],[53], modulation of JNK signaling [54], and several nuclear functions, including control of cell cycle progression and maintenance of genomic stability [55].
We found that DJ-1 is necessary and sufficient to mediate the effects of activated Ras during the development of the eye. Because DJ-1 failed to modulate constitutively active ERK/Mapk signalling during eye development, we propose that DJ-1 acts in parallel to Ras and either upstream or in parallel to ERK. We then investigated whether endogenous DJ-1B has any physiological function during the development of the fly. We found that DJ-1B is required, as an ERK/rolled interactor to control the development of photoreceptor neurons and wing venation. These observations establish a novel physiological role for DJ-1B in the intact fly. In vitro experiments previously suggested that DJ-1 might interact with Ras signaling. DJ-1 was first defined as a Ras-pathway interactor during oncogenic transformation [18] and a recent study reported that DJ-1 regulates the activation status of the ERK kinase in vitro [38].
Mechanistic studies in the Ret/DJ-1 mouse model are difficult to pursue, because of the region-specific and late-onset phenotype. Studies in which constitutively active or dominant negative Akt was virally delivered into the mouse brain suggested that Akt regulates the survival, cell size, and target innervation of SN neurons [56],[57]. Ret is a potent activator of both PI3K/Akt and Ras/ERK, and the loss of cell bodies, axons, and reduced cell size in DAT-Ret mice suggest possible defects in PI3K/Akt and/or Ras/ERK signaling. Our finding that DJ-1 does not interact with Akt signaling in Drosophila suggests that Akt signaling might not be the major pathway that cooperates with DJ-1 to regulate SN survival. Mice over-expressing activated Ras (RasV12) in the nervous system have larger neurons and embryonic mesencephalic neurons derived from these mice are resistant to toxin-induced degeneration, suggesting that Ras signaling promotes survival of SN neurons [58]. A recent analysis of Ret knockin mice revealed a critical role for Ras/B-Raf/IKK signaling, but not for PI3K and ERKs, in the survival of sympathetic neurons [59]. It is therefore possible that DJ-1 cooperates with Ras-associated signaling to promote survival of aging SN neurons deprived of trophic support. It is also possible that DJ-1 and associated Ras signaling cooperate with PI3K/Akt signaling to control common downstream effectors and further studies will address this possibility.
Deletion of both Ret and DJ-1 leads to a presymptomatic parkinsonian state in aging mice characterized by the lack of alpha-synuclein deposits, behavioural alterations, and loss of total dopamine, raising the possibility that the mechanisms regulating cell survival and target innervations might differ from those regulating protein homeostasis and dopamine dynamics. Compensatory mechanisms are likely to exist in the nigrostriatal system that maintains dopaminergic homeostasis below a certain threshold of SN neurodegeneration [1]. Several potential compensatory mechanisms have been previously described [60] and Ret/DJ-1 mice could serve as basis for further investigations of these mechanisms. The low penetrance of PD and the variability of symptoms in family members who inherit PD-associated mutations have raised the possibility that several risk factors interact to promote SN neuronal demise (the multiple hit hypothesis of PD [6]). A combination of high cytoplasmic calcium, elevated levels of free cytoplasmic dopamine, and the presence of alpha-synuclein induces selective death of cultured DA neurons, and interference with any of these individual hits alleviated neuronal cell death [61]. We report here a chronic genetic mouse model in which the interplay between three factors (aging, trophic insufficiency, and increased cellular stress due to DJ-1 inactivation) synergize and cause the loss of approximately 50% of GIRK2 DA neurons in the SN. Although the relevance of this three-component network to PD remains to be demonstrated, our findings underscore the importance of higher-order interactions between “sub-lethal” dopaminergic insults in promoting cell death.
In summary, we propose that the tight integration between aging, trophic signaling pathways, and the signaling network defined by PD-associated genes, including DJ-1, is critical for DA neuron survival. Further research on aging mechanisms coupled to studies on trophic factors and oxidative stress regulation may identify common denominators of these three processes and uncover new cellular targets for drug development in PD.
The generation of Retlx [62] and Dat-Cre [63] alleles were described previously. The DJ-1−/− mice are described in a separate paper [25]. The GMR-dRetMEN2 and UAS-dRetMEN2 flies were generously provided by Ross Cagan (Mount Sinai). DJ-1A and DJ-1B knockout as well as UAS-DJ1A and UAS-DJ1B flies were a kind gift from Nancy Bonini (University of Pennsylvania). Wild-type, dominant negative (D954A), and constitutive active (CAAX) UAS-PI3K flies were from Sally Leevers (Cancer Research UK). Sev-RasG12V flies were from Marc Therrien (University of Montreal), and UAS-rlSEM flies were kindly provided by Jongkyeong Chung (KAIST, Korea). UAS-dPTEN flies were a kind gift from Tak Mak (University of Toronto). All other fly lines were from the Bloomington Stock Center.
Immunohistochemistry, stereology, and fiber density measurements were essentially performed as previously described [10]. Thirty µm-thick free floating sections were used for immunostainings. Primary antibodies were directed against: Tyrosine hydroxylase-TH (mouse monoclonal, 1∶2000, DiaSorin, Stillwater Massachusetts, USA), Pitx3 (rabbit polyclonal, generously provided by M.P. Smidt (Utrecht University), 1∶1000, [64]), GIRK2 (rabbit polyclonal, 1∶80, Alomone labs), Calbindin (mouse monoclonal, 1∶500, Sigma), Iba1 (1∶1000, rabbit polyclonal, Wako, Neuss, Germany), and GFAP (1∶500 rabbit polyclonal, DakoCytomation, Glostrup, Denmark). For immunofluorescence, sections were first premounted, and then the following primary antibodies were used: anti-TH antibody (mouse monoclonal, 1∶2000, DiaSorin, Stillwater Massachusetts, USA) or with anti-DAT (rat polyclonal, 1∶500, Chemicon/Millipore). For stereology, every sixth section spanning the ventral midbrain was used for measurements. To quantify the density of astrocytes in the dorsal striatum, one picture was acquired from every sixth section of the dorsal striatum. Six to eight sections were analyzed/animal and at least 4 animals were analyzed per group.
GIRK2 immunostained coronal sections were analyzed using a bright field microscope with a 40× objective. Random cells were selected with stereological methods using the StereoInvestigator software. Five to seven animals per group were analyzed by circling cell soma of 149–275 cells per animal.
Eighteen-mo-old mice were sacrificed, brains were removed, snap-frozen, and the striata were dissected. The tissue was homogenized in 0.1 M perchloric acid containing 0.5 mM disodium EDTA and 50 ng/ml, 3,4-dihydroxybenzylamine as an internal standard, centrifuged at 50,000 g for 30 min, and filtered through a 0.22 µM PVDF membrane. The samples were subjected to HPLC analysis as described previously [10].
To test general activity of aging control and mutant mice, mice were subjected to open field behavioral assessment. Eighteen-mo-old mice were housed individually in a room with 12 h/12 h reversed day-night cycle. All experiments were conducted during the night period in a quiet room by 12 lux light. Mice were placed into a 59 cm×59 cm large arena for 20 min, and their movement was followed using EthoVision Pro 2.2. (Noldus, Sterling, USA). The experiment was repeated on the consecutive day and the average distance each mouse travelled during the two trials was determined. Experimental protocols were approved by the government of Oberbayern, Germany.
Pictures of P1–P5 eyes and wings were acquired using a Leica MZ 9.5 stereomicroscope equipped with a Leica DFC320 digital camera (LeicaMicrosystems, Wetzlar, Germany). For toluidine blue stainings, heads from P1–P5 animals were dissected and post-fixed in 2.5% glutaraldehyde. After washing with PBS, heads were incubated in a 1% osmium tertaoxide solution (Science Services, Munich, Germany), then dehydrated in ethanol solutions of increasing concentrations, followed by a 10 min incubation in propylene oxide. Heads were then incubated overnight in a solution containing 50% propylene oxide and 50% durcupan epoxy resin, which contained 48% component A/M, 40% hardener B, 2.25% accelerator C, and 9% plasticizer D (Sigma-Aldrich). Then, heads were incubated overnight in 100% durcupan epoxy resin. The next day, heads and fresh durcupan resin were transferred to molds, oriented tangentially, and then cooked overnight at 60°C. The heads were then removed from molds and cut using a 2088 ultrotome (LKB, Bromma, Sweden). Three µm-thick sections were collected, mounted, and then stained using a pre-warmed toluidine blue solution that contained 0.1% toluidine blue (Serva Electrophoresis, Heidelberg, Germany) and 2.5% sodium carbonate. After a quick wash in water, sections were allowed to dry and were then covered with paraffin oil. Pictures at different retinal depths were acquired for each head. To determine ommatidium size and the number of photoreceptor neurons/ommatidium, at least 150 ommatidia/animal from at least 4 animals were analyzed.
Heads and bodies were separated from 10 WT and DJ-1−/− flies and snap frozen in liquid nitrogen. RNA preparation was performed using the RNAeasy kit and the QIAshredder spin column (Qiagen) according to the manufacturer's instructions. The RNA concentration was determined using a NanoDrop ND1000 spectrometer, and 10 ng per sample of total RNA was subjected to RT-PCR amplification with 25, 30, or 35 cycles using the Qiagen OneStep RT-PCR kit according to the manufacturer's instructions. The following exon-spanning primer pairs were used: 5′-CAAGCAAGCCGATAGATAAACA-3′ (GAPDH forward) 5′-CAAGTGAGTGGATGCCTTGT-3′ (GAPDH reverse) 5′-GGAAAGATCCTTGTTACCGTG-3′(DJ-1A forward) 5′-CCATCCTGGACCACAGTCTT-3′ (DJ-1A reverse).
A pCMV-myc-DJ-1 construct and the empty pCMV-Myc vector were acquired as a kind gift from Phillip Kahle (Hertie Institute, Tübingen, Germany). SiRNA oligonucleotides (stealth-siRNA, Invitrogen) had the following sequences: (DJ-1) AGGAAAUGGAGACGGUCAU-CCCUGU; (CTRL) ACAGGGAUGACCGUCUCCAUUUCCU. The sequences have been described previously and validated for off target effects [13],[48].
MEFs were isolated from E13.5 WT or DJ-1−/− embryos according to standard procedures. Experiments were performed at passage 4–6. MEFs, HeLa cells, and A549 cells were cultured in DMEM supplemented with 10% serum, 1% L-Glutamine, and 1% pen/strep. SH-SY5Y cells (ATCC #CRL-2266) were cultured in the DMEM/F12 (1∶1) supplemented with 10% serum, GlutaMAX, and 1% pen/strep. MEFs were transfected using the standard CaPO4-precipitate method. HeLa cells were transfected using Lipofectamine 2000 (Invitrogen) for plasmids or Lipofectamine RNAiMAX (Invitrogen) for siRNA by the forward transfection method, according to manufacturer's instructions. SH-SY5Y and A549 cells were transfected using Lipofectamine 2000 (Invitrogen) for plasmids or Lipofectamine RNAiMAX (Invitrogen) by the reverse transfection method, according to the manufacturer's instructions. Plasmid overexpression experiments were incubated for 24 h after transfections before analysis for all cell types. SiRNA knockdowns were incubated for 48 h after transfection for HeLa and A549 cells, while SH-SY5Y cells were incubated for 96 h after transfection.
Fly P1–P5 heads from at least 50 animals were collected and snap frozen in liquid nitrogen, then stored at −80°C. The lysis and detection were performed as previously described [15] using an anti-DJ1A/B antibody (rabbit polyclonal, 1∶500, kind gift from Leo Pallanck). Cultured cell lines were harvested in a lysis buffer containing 1% Triton X-100, 150 mM NaCl, 1 mM EDTA, 10 mM Tris-HCl (pH 7.5), 100 mM NaF, 1 mM NaVO3, 10 mM Na4P2O7, and Complete protease inhibitor mixture (Roche Diagnostics). Mouse brains were snap frozen, ventral midbrains and striata were dissected on ice, and homogenized in a buffer containing 150 mM NaCl, 50 mM Tris_HCl, pH 7.4, 2 mM EDTA, 1% Nonidet P-40, 1% SDS, and Complete protease inhibitor mixture (Roche Diagnostics) by 10 strokes in a dounce homogenizer. Cell lysates and brain homogenates were centrifuged at 1,000 g for 10 min, supernatants were saved, and the protein concentration was determined using the DC protein assay (BioRad). Samples were subjected to SDS-PAGE and Immunoblotting according to standard techniques. The following antibodies were used: anti-AKT (9272, Cell Signaling Technology), anti-phospho-AKT (Ser473, 9271, Cell Signaling Technology), DJ-1 (ab4150, Abcam), anti-phospho-p42/p44 MAPK (4376, Cell Signaling Technology), anti-p42/p44 MAPK (9102, Cell Signaling Technology), anti-Ret (70R-RG002, Fitzgerald), and anti-β-tubulin (T-8660, Sigma).
|
10.1371/journal.pntd.0000895 | Elimination of Active Trachoma after Two Topical Mass Treatments with Azithromycin 1.5% Eye Drops | Following an epidemiological study carried out in 2006 showing a high prevalence of blinding trachoma in the Far North Region of Cameroon, a trachoma elimination programme using the SAFE strategy was initiated: three yearly trachoma mass treatments were to be performed.
The entire district population (120,000 persons) was treated with azithromycin 1.5% eye drops in February 2008 and January 2009. To assess the effect of treatment on the prevalence of active trachoma, three epidemiological studies were conducted on a representative sample of children aged between 1 and 10 years. The first study was performed just prior to the first treatment, the second just prior to the 2nd treatment and the third one, one year later. The prevalence of active forms of trachoma (TF + TI) dropped from 31.5% (95%CI 26.4–37.5) before treatment to 6.3% (95%CI 4.1–9.6) one year after first treatment; a reduction of nearly 80%. One year after the second treatment, the prevalence decreased to 3.1% (95%CI 2.0–4.9), a total reduction of 90%. Furthermore, there were no more TI cases (only TF). There was no report of serious or systemic side effects. Tolerance was excellent.
Active trachoma mass treatment with azithromycin 1.5% eye drops is feasible, well tolerated, and effective.
| Trachoma is the leading cause of infectious blindness worldwide, accounting for 1.3 million cases of blindness. Although it has disappeared in many regions of the world, trachoma is still endemic in Africa, Eastern Mediterranean, Latin America, Asia, and Australia. The WHO has currently set a target of 2020 for controlling trachoma to a low enough level that resulting blindness will not be a major public health concern. Topical tetracycline was for a long time the recommended treatment for active trachoma, but compliance to the regimen is extremely poor. Azithromycin has properties that make it an ideal treatment for Chlamydia trachomatis: high efficacy, intracellular accumulation, and a long tissue half-life. There is now a new mass treatment of trachoma by azithromycin 1.5% eye drops which is as effective as the oral route. In the test health district of Kolofata, Cameroon, the prevalence of trachoma among children dramatically decreased from 31% to less than 5% after 2 treatments. A third treatment was performed in January 2010. An epidemiological surveillance is implemented to see if this removal will be permanent. It also avoids misuse of oral azithromycin and the eye drops are directly treating the site of the infection.
| Trachoma is caused by Chlamydia trachomatis and it is spread by direct contact with eye, nose, and throat secretions from affected individuals or by contact with objects, such as towels and/or washcloths, which have had similar contact with these secretions [1], [2]. Flies can also be a route of mechanical transmission [1]. Children are the most susceptible to infection but the blinding effects or more severe symptoms are often not felt until adulthood.
Infection is frequently passed from child to child but also from child to mother. Indeed, women are nearly two times [3] more affected than men by trachoma and trichiasis, probably because one of the primary activities of girls is taking care of their younger family members. This activity continues into adulthood, with women carrying the main responsibility of caring for children.
Trachoma remains the leading infectious cause of blindness in the world [4]. Mass oral azithromycin distribution has been used in several programs. At present, azithromycin is available in a 1.5% eye drop formulation in Europe, Maghreb and French speaking African countries. The National Blindness Prevention Programme and a Vision 2020 plan were established as a result of the prevention policy of the Republic of Cameroon., In December 2006, in the Kolofata Health District (Far North Cameroon), a study assessing the prevalence of active and scarring trachoma, signalled the presence of endemic trachoma with significant blinding potential [5]. The National Blindness Prevention Programme decided to plan an elimination program by implementing the SAFE (Surgery, Antibiotics, Facial cleanliness, and Environmental change) strategy [6], addressing the “A” (antibiotic) component by conducting mass treatment targeting the entire district population and using azithromycin 1.5% eye drops. The objective of this study was to assess the feasibility, tolerance and effectiveness of repeated topical mass treatment with azithtromycin 1.5% eye drops, used for the first time on a large scale to reduce the prevalence of active forms of trachoma in a population. The first year, mass treatment gave promising results with a decrease in trachoma prevalence of more than 25% (from 31.5% to 6.3%) [7]. This article presents results after two rounds of treatment.
In accordance with the WHO recommendations [8], the trachoma control programme in the Kolofata Health District called for one mass treatment per year for three years. The treatment consisted in one instillation of azithromycin 1.5% in both eyes in the morning and in the evening during three consecutive days. The study received authorisation from the Cameroon Ministry of Public Health in February 2008. The first round of treatment began on 23 February 2008 and ended on 10 March 2008, and the second one was undertaken between the 5th and 20th of January 2009 (Figure 1).
Each year, Théa Laboratories donated 120,000 complete treatments (720,000 single doses) of azithromycin 1.5% eye drops, and sent them by air from Europe to Yaoundé, Cameroon, and by train and truck from Yaoundé to Kolofata.
The target population was all residents of the Kolofata Health District [9]. The entire population was treated, but only children between 1 and 10 year olds were examined for active trachoma. During the 15 days preceding the beginning of treatment, the local community health workers, helped by a literate second-level community health worker, conducted an exhaustive door-to-door census of all residents of the Kolofata Health District. Each of the 250 local community health workers was assigned a village or neighbourhood of 400 to 500 residents.
The local community health workers then administered treatment by visiting each household morning and evening for three consecutive days. Ophthalmic nurses were supervising each day the mass treatment in the different villages. A briefing was organized in Kolofata hospital each evening where the supervisors reported day after day their assessment of the mass treatment.
As described by Huguet [7], among children aged between 1 and 9 year olds (up to their 10th birthday), three descriptive cross-sectional studies, to assess the effectiveness of treatment on the prevalence of active forms of trachoma in the population (trachomatous inflammation—follicular (TF) and/or trachomatous inflammation—intense (TI) [10], [11]), were conducted in the Kolofata Health District. The first was conducted prior to treatment in February 2008, the second prior the second treatment in January 2009, and the third in January 2010, one year after the second treatment (Figure 1). The standard WHO protocol for trachoma prevalence surveys was used [12].
The population studied was chosen at random and was based on the exhaustive list of villages and demographic statistics gathered in 2006 for the national census [9] and revaluate annually by a census at Kolofata sanitary district level only (118,617 inhabitants in 2009).
Assuming a prevalence of less than 5% at the end of the study (one year after the third treatment), it was necessary to include 2,400 children between 1 and 10 years in the study to obtain a precision of approximately 1.5% with a two-sided 95% CI and a cluster effect of 4 [13]. The 2,400 children were divided into 40 clusters, with 60 children per cluster chosen randomly.
All target-population residents (defined as children over 1 year and less than 10 years old [up to their 10th birthday] who had lived in the village for at least six months prior to the study date) in randomly selected households were registered and included in the population to be examined. When a family had left the community more than 6 months before the visit and the household remained empty, that household was replaced by the household nearest to it. A household that was only “temporarily” empty (less than 6 months) was not replaced. The research team returned up to three times to examine any subject absent during the preceding visit. If after the third visit the missing person was not found, that person was declared absent and not replaced.
Any family in the selected population whose head of household refused consent to participate in the study was not replaced [12]. This population corresponds to the enumerated population. The examined population corresponds to the population effectively examined.
A two-day training session helped assure standardization of procedures for conducting the census, examining subjects, and collecting and recording data. A post-training test was conducted on 50 trachoma patients to determine whether the trainee had mastered the WHO simple grading system and to confirm that each future examiner had a concordance of more than 80% with an expert examiner for each key sign. A pilot study was conducted in two villages not included in the current studies.
All children included in the study were examined by a senior nurse who everted the upper eyelid and examined the conjunctiva with a 2.5 magnifying glass and a torch held by an assistant who also recorded the data. The examiner changed gloves after the examination of each patient. Before examining the next person, the examiner verified that the assistant had filled out the study sheet in accordance with study protocol guidelines.
Data were compiled and analysed using EPIINFO 6 software. Estimated confidence intervals took into account the composition of sample clusters.
All subjects provided informed consent. As people were illiterate, informed consent was read to people and if they agreed to participate the participant or legally acceptable representative put their fingerprints on the informed consent and a literate witness signed on behalf of the participant. National ethics committee of Yaounde approved the way to collect consent and the study before the beginning of the study.
During the first annual treatment, azithromycin 1.5% eye drops were administered by the local community health workers to 111,340 of the 115,274 people counted in the census (coverage 96.6%) [7]. During the second round of treatment 105,802 people (45,288 adults and 60,514 children; 50,846 males and 54,956 females) received the full 6-dose treatment of azithromycin eye drops. In addition, 41,376 doses were administered to others who did not complete the full 6-dose treatment i.e. to people who were absent during at least one of the treatment administration visit.
The number of children examined during the study relative to the number counted in the random selected population is presented in Table 1.
During the study, age and sex distributions were similar in the sample populations before and after treatments (p>0.05) (Table 2).
In February 2008, before all treatment, the prevalence of active forms was estimated to be 31.5% (95%CI 26.4–37.5). One year after the first mass treatment (January 2009), this prevalence dropped significantly to 6.3% (95%CI 4.5–8.6) (p<0.0001). One year after two rounds of topical treatment (January 2010), prevalence dropped to 3.1% (95%CI 2.0–4.9) (p<0.0001) (Table 3), a decrease of 90%.
The prevalence of TF in the study sample was estimated to be 24% (95%CI 20.7–27.5) before treatment, it decreased significantly to 5.8% (95%CI 4.1–8) one year after first annual treatment (p<0.0001) and again significantly (p = 0.0001) to 3.1% one year after the second round of treatment (Table 3), either a decrease of 87%. The prevalence of TI was estimated to be 7.5% (95%CI 5.7–10) before treatment and disappeared after two annual treatments and (0.5% after 1st treatment (p<0.0001) and 0% after second one (p = 0.0005) (Table 3) (Figure 2).
Questionnaires concerning side effects of treatment were administered by community health workers during daily visits. According to WHO recommendations, if TF is 10% or more in children 1–9 years old, a mass treatment with antibiotic should be conducted throughout the district. Furthermore, as efficacy of the treatment was already assessed in a Phase III study, it was ethical that people of the communities were all treated, so there is no control group to assess the reliability of side effects. The few complaints recorded were local and brief (blurred vision lasting several minutes following instillation of eye drops or transient burning sensation in the eyes). There were no reported serious ocular or systemic side effects.
A statistically significant (p<0.0001) increase in trachoma prevalence was observed between the first prevalence study conducted in 2006 (26%) and the study conducted prior to the first treatment in 2008 (31.5%). It suggests that there was no secular decline trachoma in this area. However, according to the WHO recommendation, as it was unethical to not treat some people of this area, the mass treatment protocol do not planned to have a control group. Thus the prevalence reduction (from 31.5 to 3.1%) of active trachoma among children of this study is likely to be a result of the two mass treatments with azithromycin 1.5% eye drops and prevention campaigns.
By WHO definition, the current prevalence of 3.1% indicates that “trachoma as a blinding disease is being controlled” (TF<5% and TI<0.2%) [4]. The fact that no more TI grade (severity factor of the disease) was observed is particularly encouraging, since TI patients are those most likely to suffer blindness as the disease evolves [14], [15], [2], [12], [16], [17], [18], [19], [20].
Apart from minor complaints, treatment was accepted and well tolerated by both children and adults.
The importance of endemic trachoma in the district of Kolofata justified the mass treatment of the entire population with azithromycin 1.5% eye drops as part of the SAFE strategy and in accordance with WHO recommendations [8]. The apparent coverage decreased between 2008 and 2009 (97% to 89%) is an artefact produced by a change in reporting: in 2009, unlike 2008, only people who had completed all 6 doses were counted.
The most common drug currently used in trachoma mass treatment campaigns is azithromycin 20 mg/kg taken orally. In Niger from 2002 to 2005, SAFE strategy was implemented and three mass treatments using oral azithromycin were performed in 2 districts with 72 villages. Surveys were conducted 3 years apart (before and after program) The prevalence of TF among children decreased from 62.3% and 49.5% to 7.6% and 6.7% in three years [21], a reduction of 89% in one village and 85% in the other. At the same period, in Mali, mass treatment using oral azithromycin was conducted in 7 districts. The prevalence of TI among children decreased from 33% to 2.5% in 3 years [22], a reduction of 92.4%. In a study in Nepal where nearly 40% of children had active trachoma, three rounds of treatment directed at children age 1 to 10 years reduced clinically active trachoma to approximately 13% one year after the first treatment and to 4% one year after the second one. Furthermore, this three annual treatments were successful in reducing infection and disease in these children up to 6 months after the third last treatment (4% of children with active trachoma) [23]. However, is 6 months long enough to determine of the reduction of the disease in a population. To assess this, one study conducted in trachoma-hyperendemic communities in Tanzania determined, after two rounds of mass treatment with oral azithromycin spaced 18 months apart, the rate of trachoma and infection at 5 years, either 3.5 years last treatment. Results showed that 3.5 years after two rounds of mass treatment, trachoma was not eliminated but antibiotherapy appeared to be associate with lower disease prevalence: from 39.2–80.6% (according age groups) at baseline to 7.7–49.1% 5 years after baseline [24].
Considering oral azithromycin studies published, mass treatment with azithromycin 1.5% eye drops, with a reduction about 90% of active trachoma, is at least as effective as treatment with oral azithromycin. Moreover, a phase III clinical trial showed that topical azithromycin 1.5% twice a day for 3 days has a similar efficacy as a single oral 20 mg/kg dose of azithromycin for the treatment of active trachoma in children [25]. The prevalence study planned for 2011, one year after the third annual mass treatment with azithromycin eye drops, should be conclusive.
Finally, trachoma persists where people live in poverty without water, sanitation, [26], [1], [27] and proper waste disposal [26], [27]. Transmission of trachoma occurs where these conditions exist and should be expected to return after antibiotic treatment if the conditions are not changed. Improvements like construction of household pit latrines and hand-dug wells will bring about sustainable elimination of trachoma. However, the Ultimate Intervention Goal for Environmental improvements (UIG-E) defined by the WHO [4] is difficult to set up. Indeed, in 2008 and 2009, only four borehole pumps were installed by the Cameroon government, and two wells were built by OSF. Thus, we are far from UIG-E, i.e. easy access to safe water for 80% of the population [1].
Following encouraging results from the first two mass treatment campaigns with azithromycin 1.5% eye drops, one additional mass treatment campaign was planned. The third campaign took place in January 2010. A fourth study to track the evolution of active trachoma prevalence among children is planned for January 2011.
If the success of these first trachoma mass treatments with eye drops is confirmed, eye drops treatments would be a supplementary tool to fight trachoma in particular in young children and pregnant women.
|
10.1371/journal.pgen.1002302 | Insights into Hox Protein Function from a Large Scale Combinatorial Analysis of Protein Domains | Protein function is encoded within protein sequence and protein domains. However, how protein domains cooperate within a protein to modulate overall activity and how this impacts functional diversification at the molecular and organism levels remains largely unaddressed. Focusing on three domains of the central class Drosophila Hox transcription factor AbdominalA (AbdA), we used combinatorial domain mutations and most known AbdA developmental functions as biological readouts to investigate how protein domains collectively shape protein activity. The results uncover redundancy, interactivity, and multifunctionality of protein domains as salient features underlying overall AbdA protein activity, providing means to apprehend functional diversity and accounting for the robustness of Hox-controlled developmental programs. Importantly, the results highlight context-dependency in protein domain usage and interaction, allowing major modifications in domains to be tolerated without general functional loss. The non-pleoitropic effect of domain mutation suggests that protein modification may contribute more broadly to molecular changes underlying morphological diversification during evolution, so far thought to rely largely on modification in gene cis-regulatory sequences.
| Proteins perform essential regulatory functions, including control of gene transcription, a process central to development, evolution, and disease. While protein domains important for protein activity have been identified, how they act together to define the activity of a protein remains poorly explored. The predominant view influenced by prokaryotic transcription factors is that protein domains constitute independent functional modules, required for all aspects of protein activity. In this study, we used Hox proteins, evolutionarily conserved transcription factors playing key roles in the establishment of animal body plans, to examine how protein domains collectively shape protein activity. Results obtained using a broad range of biological readouts highlight a context-dependency in protein domain usage and interaction, revealing that protein domains are non-pleoitropic in nature. This suggests that protein modification may contribute more broadly to molecular changes underlying morphological diversity, so far thought to rely largely on modification of gene cis-regulatory sequences.
| How the diversity of animal body plans is established remains a central question in developmental and evolutionary biology [1], [2]. A key step towards understanding the molecular basis underlying diversity is to decipher mechanisms controlling proper genome expression, and how variations in these mechanisms have been at the origin of developmental and evolutionary diversity. While a large number of studies have focused on the impact of cis-regulatory sequences organization (reviewed in [3]), deciphering the intrinsic functional organization of trans-acting transcription factors remains largely unaddressed. Studies have identified functional domains ([4]–[9] and [7], [10], [11] for reviews), but how different protein domains jointly and collectively act for defining the overall activity has been poorly assessed. Yet, a recent study highlights that the synthetic shuffling of protein domains within proteins of the yeast-mating signaling pathway results in the diversification of the mating behavior, demonstrating the importance of protein domain interactions for functional diversification [12].
Hox genes, which encode homeodomain (HD)-containing transcription factors, provide a suitable paradigm to decipher how function is encoded within protein sequence, and how associated changes may constitute the origin of functional specification and diversification. Hox genes have arisen from duplication events of ancestral genes, followed by sequence divergence that promoted the emergence of up to 14 paralogous groups in vertebrates. Hox paralogue proteins display distinct regulatory functions, promoting axial morphological diversification in all bilaterian animals [13]–[17]. Previous work has established that sequence changes in the HD, the DNA binding domain, and a few additional protein domains, have played a major role in the diversification of Hox protein function [4]–[9], [18]–[21]. However, how protein domains functionally interact to shape overall protein activity remains elusive.
We focused on three protein domains from the Drosophila central Hox paralogue protein Abdominal (AbdA, Figure 1). These domains are related by their demonstrated or potential involvement in the recruitment of the Extradenticle (Exd) cofactor, homologous to vertebrate PBX proteins, known to have key roles in establishing Hox functional specificity. The first domain, known as hexapeptide (HX) or PID (Pbx Interacting Domain), with a core YPWM sequence, is found in all Hox paralogue groups, with the exception of some posterior Hox proteins. Biochemical, structural and functional studies have shown that this motif mediates interaction with the Exd/PBX class of Hox cofactors (collectively referred as PBC). The second domain, termed UbdA (UA) is specifically found in the central Hox proteins AbdA and Ultrabithorax (Ubx). This paralogue-specific domain was recently shown to be required for Exd recruitment in the repression of the limb-promoting gene Distalless (Dll) [8], [22]. The third domain (TD), similar in sequence (TDWM) to the YPWM motif, is also paralogue-specific. The TD motif retains the W that provides strong contact with the PBC class proteins, and matches the sequence of the HX motif in some Hox proteins (eg., Hoxa1). Evidence for an Exd recruiting role of the TD domain in AbdA however remains to be demonstrated.
To start unraveling how protein domains collectively shape Hox protein activity, the effect of single, combined double or triple domain mutations were analyzed using most known AbdA functions as biological readouts. The large functional window covered by the study allows identifying functional attributes of protein domains taken in isolation and collectively, and a quantitative analysis by hierarchical clustering highlights the functional organization of the Hox protein AbdA. Given the phylogeny of the studied protein domains, the work has also implication regarding the mechanisms underlying the evolution of AbdA protein function.
AbdA variants bearing single or all possible combinations of protein domain mutations (Figure 1A) were ectopically expressed through the binary UAS-Gal4 expression system [23]. Protein levels following induced expression were quantified and experimental conditions ensuring levels close to that of endogenous AbdA were selected (see Materials and Methods). Impact of AbdA variants on target gene control, phenotypic traits and locomotion behavior (Figure 1B), covering AbdA functions of increasing complexity in different tissues, were evaluated in the anterior region where the endogenous AbdA protein is absent. Quantified results (see Text S1) are presented as loss (and in few cases as gain) of regulatory potential.
Eleven functional assays were used to assess domain requirements for AbdA activity (Figure 1B). Four assays rely on the regulation of AbdA target genes, for which evidence of a direct regulation has been previously reported, including the regulation of Distalless (Dll) [8], [24], [25] and Antennapedia (Antp) [26] in the epidermis, and the regulation of wingless (wg) [27] and decapentaplegic (dpp) [28], [29] in the visceral mesoderm. Six assays rely on analysis of phenotypic traits. One of these phenotypic trait, oenocyte specification, results from the regulation of a single target gene [30]. Others, cerebral branch [31], somatic muscles [32], A2 epidermal morphology [33], [34], neuroblast [35], [36] and heart cell lineage specification [37] likely depend of the coordinated regulation of several target genes. Finally, we also used a behavioral trait, larval locomotion, thought to rely on integrated AbdA function in two distinct tissues, the somatic musculature and the nervous system [38].
In the somatic musculature, the abdominal specific pattern is characterized by the presence of muscle located ventrally and absent in thoracic segments, a feature that can be visualized by the expression of nautilus (nau) [32]. This distinction was previously shown to result, at least in part, from the activity of AbdA [32]. Accordingly, anterior ectopic expression of AbdA using the mesodermal driver (24B-Gal4) results in ectopic ventral expression of Nau in anterior segments (Figure 2). We found however that none of the AbdA protein domains under study, alone or in combination, was required to specify the abdominal specific features of the somatic musculature (Figure 2 and Figure S1). In the same conditions, a point mutation at position 50 of the homeodomain that impairs AbdA binding to DNA resulted in the loss of Nau inducing capacity (Figure 2 and Figure S1). The dispensability of the HX, TD and UA domains for specifying abdominal features of somatic muscle pattern is consistent with the fact that nau activation by AbdA is not dependent upon Exd activity [32], although results below argue that these domains assume other functions than Exd recruitment.
In the embryonic central nervous system, a subset of 30 neuroblasts (NB's) found in each hemisegment, including the NB5–6, generate a larger lineage in the thorax than in the abdomen. Recent studies demonstrated that posterior Hox genes, such as abdA, impose in the abdomen a smaller NB5–6 lineage by triggering an early cell cycle exit [39]. Misexpression of AbdA within NB5–6 in the thorax using ladybird(K)-Gal4 result in an early lineage truncation, mimicking the situation that normally occurs in the abdomen, ultimately leading to a smaller thoracic NB5–6 lineage size (Figure 3). Average number of NB5–6 cells in wild type thoracic and abdominal segments was previously estimated at 16 and 6 cells respectively: these values were considered as references for full (100%) or complete loss (0%) of repressive activities of AbdA variants on NB5–6 lineage. Intermediate repressive levels upon ectopic expression with ladybird(K)-Gal4 were deduced from the quantification of NB5–6 lineage cell numbers in thoracic segments T2/3 (see methods). Results obtained indicate that lineage truncation triggered by AbdA is similarly affected following UA, HX/UA, TD/UA and HX/TD/UA mutations (Figure 3), which can be best explained by a unique requirement of the UA domain for AbdA function.
In the visceral mesoderm, AbdA is expressed in parasegment (PS)8–12. The target genes wg and dpp are respectively activated (in PS8) and repressed (in PS8–12) by AbdA in the visceral mesoderm. Restricted (PS8) activation of wg by AbdA results from the action of the Dpp signal, locally produced by PS7 cells under the control of the Ubx protein [40]. Accordingly, anterior ectopic expression of AbdA only results in a mild activation of wg, as activation only occurs in cells experiencing partial repression of dpp [27]. Previous work has shown that the HX mutation results in a protein that activates dpp instead of repressing it, and consequently more efficiently activates wg [41].
AbdA variants were expressed with the 24B-Gal4 driver. Levels of regulatory activities were deduced following fluorescent in situ hybridization against dpp or wg in the visceral mesoderm of stage 14 embryos in PS1–PS7, ie anterior to endogenous AbdA expressing cells (PS8–12; Figure 4 and Figures S2 and S3). Arbitrary values have been assigned to regulatory activities of AbdA variants. For dpp (Figure 4A and Figure S2), no effect on dpp expression was scored by 0, normal repression of dpp expression in PS7 by 100 (partial repression was never observed) and ectopic activation (instead of repression) of dpp was scored by negative values (depending of the number of ectopic sites (see Text S1). For wg, in a manner similar to dpp, no effect was scored by 0, and positive and negative values were respectively assigned to normal (activation) or abnormal (repression) activities on wg expression (Figure 4B and Figure S3; see Text S1).
Results obtained allow two conclusions. First, single domain mutations result in strong modification of AbdA activity. Second, domain mutations often result not only in a quantitative, but also in a qualitative (neomorphic) modification of activity, changing AbdA from an activator to a repressor, or reversely from a repressor to an activator.
Oenocytes form under AbdA control in segments A1–A7. This occurs through AbdA-dependent activation of Rhomboid (Rho) in a chordotonal organ precursor cell called C1. Expression of Rho then enables the secretion of the EGF ligand Spitz that will instruct neighboring epidermal cells to differentiate into oenocytes [30]. In absence of AbdA, the EGF pathway is not locally activated and oenocytes are not specified [30]. Reversely, ectopic expression of AbdA induces oenocytes in thoracic segments.
AbdA variants were ubiquitously expressed with the armadillo (arm)-Gal4 driver. Oenocyte inducing potential of AbdA variants, visualised with the seven-up (svp)-lacZ enhancer trap reporter construct, was deduced from the number of thoracic segments that contain ectopic oenocytes (see Text S1). This inductive potential is reduced following single mutations of the UA domain and combined mutation of the HX/TD or TD/UA domains, and is abolished following HX/UA and HX/TD/UA mutations (Figure 5 and Figure S4). These observations suggest an additive contribution of the HX, TD and UA protein domains for oenocyte induction by AbdA, consistent with protein domains acting independently of each other, and contributing uniquely through additive contribution to protein activity.
The tracheal cerebral branch forms dorsally exclusively in the second thoracic segment T2, in response to repressive activities of Bithorax Hox proteins in T3-A8 segments [42]. This phenotypic trait can be followed by a breathless (btl) driven GFP reporter that extends posteriorly in the absence of Bithorax complex genes, and that is suppressed in T2 following Btl-driven expression of AbdA in the tracheal system (Figure 6A). Only full repression of cerebral branches was considered and repressive activities of AbdA variants thus correspond to either 0% (no repression) or 100% (full repression) (see Text S1). We found that the repression of the cerebral branch by AbdA is impaired following TD/UA and HX/TD/UA but not HX/UA or HX/TD mutations, revealing a functional redundancy between the TD and UA domains (Figure 6A, and Figure S5).
In the embryonic heart, abdominal segments are made of six pairs of cells, instead of four in thoracic segments [37]. This difference was shown to result from AbdA (and Ubx) promoting the six cell lineage in the abdomen [37], and in the thorax following AbdA ubiquitous expression in the mesoderm driven by the 24B-Gal4 driver ([37], Figure 6B). The visualization of the lineage is facilitated by a Dorsocross (Doc) staining, that labels two cells in each hemisegment, allowing to unambiguously identify each hemisegment. Effects of AbdA variants in cardiac cells specification were visualized by double fluorescent immunostaining against AbdA and Dorsocross (Doc). The six cell lineage inductive capacity of AbdA was scored by counting the number of cardiac cells in the T2 and T3 segments (see Text S1). Results showed that the six cell lineage inductive ability of AbdA is lost following HX/UA and HX/TD/UA mutations (Figure 6B and Figure S6). These observations again highlight functional redundancy, but between the UA and HX domains, instead of TD and UA domain as observed in cerebral branch specification. Additional examples of functional redundancy, yet in more complex pattern of interactions between protein domains were found in the biological contexts described below.
The limb-promoting gene Distalles (Dll) and Hox gene Antennapedia (Antp) are direct targets of AbdA [26], [43]. The ability of AbdA variants, following ubiquitous expression through the arm-Gal4 driver, to repress Dll (Figure 7A and Figure S7) and Antp (Figure 7B and Figure S8) was evaluated by examining the activity of a Hox responsive Dll enhancer (DME, [44]) and the expression of the Antp protein, respectively (see Text S1). Single domain mutations do not strongly affect repressive activities of AbdA on Dll and Antp, leading to a mean loss of 40%, with the exception of the TD mutation, which affects more (60%) the repressive activities on Antp. Combining domain mutations leads to stronger effects: in the case of Dll, simultaneous mutation of the HX and UA domains almost completely abolishes AbdA repressive activities, while in the case of Antp simultaneous mutation of the HX and UA domains or TD and UA domains results in a loss of 70% of AbdA repressive activity. More surprisingly, simultaneous mutation of the HX, TD and UA domains does not compromise further AbdA activity but instead restores a significant level of repressive activity, comparable to that of single domain mutated AbdA variants. This indicates that the three protein domains do not provide independent regulatory input, but likely act in interactive and mutually inhibitory ways.
A similar yet more complex pattern of domain interactions was observed in the specification of A2 epidermal morphology. In this tissue, AbdA promotes the formation of a stereotyped trapezoidal arrangement of denticle belts (Figure 7C). The potential of AbdA variants to specify A2 epidermal morphology was assessed following arm-Gal4 driven expression by scoring the denticle belts morphology and organisation in transformed A1 and thoracic segments (Figure 7C and Figure S9). Epidermal specification was not impaired by HX and slightly reduced by UA or TD mutations. Simultaneous mutation in two domains suggests functional redundancy between HX and TD, UA and HX but not between UA and TD domains. As noticed previously for the regulation of Dll and Antp in the epidermis, mutating the three domains simultaneously restores the activity, generating a protein that displays an activity close to the wild type protein.
In many animals including vertebrates, locomotion results from the coordinated action of regionally distinct sets of movements. Drosophila larvae crawl by means of three region specific movements [38]. The locomotion cycle starts by a contraction of the most posterior abdominal segments (A8/A9), followed by a wave of peristaltic movement in A1–A7, where each segment is transiently lifted up (D/V movement), pulled forward and lowered, starting from A7. When the wave reaches A1, the thoracic and head segments start moving by a telescopic type of movement (A/P movement), occurring through contraction of anterior segments [38]. It was established that AbdA is necessary and sufficient to specify the abdominal type of movement, namely abdominal peristalsis [38]. The potential of wild type and AbdA variants to promote abdominal peristalsis was evaluated following arm-Gal4 driven expression (Figure 7B), by scoring in the T3 thoracic segment D/V movements (see Text S1). Single domain mutations do not significantly alter promotion of abdominal peristalsis (Figure 7D and Figure S10). Again, two types of functional redundancy were observed: between the TD and UA domains, and to a lesser extent between the HX and UA domains. As in the case of Dll and Antp regulation and A2 epidermal morphology specification, triple domain mutation corrected the effects of double mutations, with a protein promoting abdominal peristalsis as efficiently as the wild type protein, providing an additional example of mutually suppressive activity of protein domains.
Previous studies have established that Exd is required for Dll [25] and wg [45] regulation, oenocytes [30] and epidermal morphology specification [46], and neuroblast lineage commitment [37], while dispensable for Antp [46] and dpp [47] regulation. In the case of cerebral branch specification, no conclusion could be reached since loss of Exd results in the absence of cerebral branch formation in the T2 segment [48]: this positive input of Exd hinders the assessment of a possible contribution for AbdA mediated cerebral branch repression in abdominal segments.
The potential implication of Exd in AbdA-mediated heart lineage commitment and larval locomotion is not known. Staining for Doc1 in embryos deprived for maternal and zygotic Exd showed that the abdominal hemi segments adopt the AbdA-dependent six cell lineage, showing the dispensability of Exd for this AbdA function (Figure 6C). The requirement of Exd for larval locomotion has been examined in homothorax (hth) mutant that impairs Exd nuclear transport and mimics exd maternal and zygotic loss [49]. The absence of peristaltic waves in this genetic context indicates a strict requirement of Exd for abdominal peristalsis (Figure S10).
Taken together with the protein domain requirement results, the exd dependency indicates that the HX, UA and TD domains, known (HX and UA) or candidate (TD) Exd recruiting domains, are also required for Exd-independent function. This is supported by the HX/UA requirement for heart lineage specification, by the HX and UA requirement for proper regulation of the dpp target gene, the HX/TD requirement for Antp repression and the requirement of TD for dpp target regulation. Collectively, this highlights that the HX and UA (and likely TD) protein domains are multifunctional, serving in some biological context Exd interaction function, while in others, they are used differently, for a molecular activity that still remains to be defined.
The complete set of quantitative data was analyzed using a hierarchical clustering method (Figure 8; see Materials and Methods). Clustering according to biological readouts does not reveal any clear grouping, regarding for instance developmental stage or tissue type, suggesting that the forces that govern domain usage and interaction between protein domains mostly reside in the regulated target gene. By contrast, clustering according to protein domains clearly reveals a hierarchical requirement of the domains for the various AbdA functions analyzed here. A bipartition of AbdA variants is observed, with the mutants for the HX, the TD and HX/TD domains on the one hand, and variants mutant for the UA domain, alone or in combination, on the other hand. Such bipartition suggests the existence of two functional modules that can be distinguished based on UA domain requirement. The first module, which relies mostly on the HX and TD domains, is used for a small subset of AbdA functions only. The second module relies on the activity of the HX, TD and UA domains, yet the requirements of the HX and TD domains are revealed only in UA deficient context. Thus, the driving force in this second functional module is the UA domain, as its mutation unmasks the requirement for the HX and TD domains, which is not revealed by their single or combined mutations. These results identify a prominent role of the UA domain in AbdA function.
Studies towards deciphering the mode of action of Hox proteins have so far essentially concentrated on how individual protein domains contribute to protein function. These focused approaches allowed in depth analyses, unraveling the intimate molecular and sometimes structural details of how protein domains contribute to protein function, providing decisive insights into how Hox proteins reach specificity. This work provides a different complementary approach towards deciphering the mode of action of Hox proteins. First it aims at studying protein domains in combinations, using combined and not only single protein domain mutations, considering that the overall protein activity is likely not a sum of the activity of individual protein domains, and that novel properties may emerge from interactions between protein domains. Second, it uses extensive in vivo biological readout, (most of the known AbdA functions), instead of a single or a few functions. While impairing the in depth analyses of previous focused approaches, the large functional window covered by this study allows the identification of features underlying the intrinsic functional organization of the Hox protein AbdA.
Although the approach taken relies on a gain of function strategy, special care was taken to select experimental conditions where proteins were expressed closed to physiological levels of expression. Biological readouts considered are functions that AbdA can sustain in ectopic places, suggesting that availability of AbdA protein partners is not a limitation of the experimental strategy chosen. Finally, the effects of expressing the AbdA variants (in all eleven biological readouts) were scored in regions anterior to the endogenous AbdA expression domain (ie in cells where the endogenous wild type gene product is not present), avoiding any further complexity that may result from competition with the endogenous AbdA protein.
Below, we summarize how results obtained shed light on the mode of action of the Hox protein AbdA and discuss the evolutionary implications.
This study identifies salient features underlying the intrinsic functional organization of the AbdA Hox transcription factor. Protein domains often display functional redundancy, with strong effects in most cases requiring simultaneous mutations of two or three domains. Redundancy was frequently observed between the HX and UA domains, or between the TD and UA domains, while redundancy between the HX and TD domains is less frequent (Figure 8). This indicates that redundancy does not necessarily rely on functional compensation through structurally related domains, since the HX and TD are closely related domains, while the UA domain is completely unrelated. Thus, functional redundancy rather reflects the potential to perform similar activities through distinct molecular strategies. This property likely confers robustness to Hox protein activity, accommodating mutations in protein domains without generally impacting on regulatory activities.
Protein domains within AbdA also generally do not act as independent functional modules, but instead display a high degree of interactivity, as demonstrated by the non-additive effects of domain mutations in the majority of the biological readouts studied. In addition, protein domains are often multifunctional, in the sense that they serve different molecular functions. This is illustrated by the fact that the HX and UA domains, previously described to mediate Exd recruitment, are also required for Exd-independent processes. Thus domain interactivity and multifunctionality are hallmarks of AbdA regulatory activity. These properties provide means to apprehend the bases underlying Hox functional diversity with a restricted number of functional modules, and therefore may account for the variety of Hox-controlled biological functions.
Protein domain usage and interaction between protein domains in AbdA strongly depends on the biological readout, suggesting that domain usage largely depends on the regulated target gene, and hence on the identity of the gene cis regulatory sequences. Recent reports support that DNA sequences impact on Hox protein activity: Hox binding site neighboring sequences are important for proper regulation of the reaper downstream target [50]; Sex combs reduced changes its conformation and activity depending on the cognate sequence [51]. Of note, a role for the target sequence in controlling the structure and activity of the glucocorticoid receptor has also been recently reported [52], indicating that this may generally apply for many DNA binding transcription factors.
Our results also have implication on how modifications in protein sequences are translated into changes in protein function during evolution. The HX domain, common to all Hox proteins, is ancient and found in all bilaterians, and provides a generic mode of PBC interaction (Figure 9). The UA domain, specific to some central Hox proteins (AbdA and Ubx in Drosophila), was acquired later, at the time of protostome/deuterostome radiation. It provides a distinct yet to be characterised PBC interaction mode, specific to some Hox paralogues only, allowing fine-tuning of Hox protein activity [22]. TD is found only in insect AbdA and not in Ubx proteins, suggesting that it arose after the duplication that generated Ubx and AbdA in the common ancestor of insects (Figure 9). Remarkably, within AbdA arthropod proteins, the HX domain has significantly diverged in some lineages like anopheles, while the TD domain has been strictly conserved.
Conceptually, two non-exclusive models could account for the evolution of protein function following the acquisition of a novel protein domain. In the first one, the acquisition provides a novel molecular and functional property, which adds to pre-existing ones. This is for example the case for the acquisition of the QA domain that confers repressive function to Ubx [6], and the acquisition/loss of HX or LRALLT domains by Futzitarazu (Ftz) from distinct insect species, which provides Ftz with the capacity to recruit either Exd or FtzF1 cofactors and switches its activity from a Hox to a segmentation protein [53]. In the second model, the acquisition of a novel protein domain interferes with the activity of pre existing domains, reorganizing the intrinsic functional organization of the protein. This view is supported by the predominant role of the UA domain and the widespread domain interactivity seen in this study.
Evolutionary changes in animal morphology is thought to mostly rely on changes in cis-regulatory sequences [1]. This is conceptually supported by the modular organization of cis-regulatory sequences, allowing subtle and cell specific changes in gene expression not deleterious for the animal. Experimentally, it is largely supported by the correlation between expression of key developmental regulatory genes and morphological changes (for example see [54]), and by changes in cis-regulatory sequences that impact on morphological traits [55]–[59]. Changes in animal morphology could also result from changes in protein sequence and function, as shown for Hox proteins in the morphological diversification in arthropods [4], [6]. However, changes in protein function are not believed to broadly contribute to morphological diversification during animal evolution, based on the assumption that changes in protein sequences are expected to have pleiotropic effects, which as such, do not provide a mean to convey subtle and viable evolutionary changes.
Our work grasps redundancy and selectivity in protein domain usage and as salient features of AbdA transcription factor intrinsic regulatory logic: even the HX domain, evolutionarily conserved in all Hox proteins, is essential for only one AbdA function, and often acts in a redundant way with the TD or the UA protein domains. Selective use of protein domains is also supported by findings of a few smaller scale studies of three other Drosophila Hox proteins: viable missense or small deletion mutations within the Scr protein coding sequences falls in different allelic series when examined for three distinct biological readouts [60]; deletion of C-terminal sequences of the Ubx protein, starting from an insect specific QA protein domain preferentially affects a subset of Ubx function [61]; dispensability of the HX was reported for the leg inducing capabilities of the Antp Hox protein, while required for other Antp functions [62]. This context dependent selective mode of protein domain usage, or differential pleiotropy, may be essential for the evolution of Hox protein functions, as it ensures developmental robustness of a Hox-controlled program while being permissive to evolutionary changes endowing novel functions to preexisting protein domains. In addition, our work also establishes that interactivity between protein domains is highly context dependent, suggesting that Hox protein function not only relies on selective mode of protein domain usage but also on selective mode of protein domain interactivity. Altogether, these observations challenge the view that changes in protein sequences necessarily have pleiotropic effects, giving more room for protein changes in the evolution of animal body plans.
24B-Gal4 and arm-Gal4 were used as embryonic mesodermal and ubiquitous drivers, respectively. Btl-Gal4 and lbe(K)-Gal4 for specific expression in the tracheal system and NB5–6 neuroblasts, respectively. The DME-lacZ and svp-lacZ lines are respectively from R. Mann (Columbia Univ., NY, USA) and S. Zaffran (IBDML, Marseille, France). exdXP11 and hthP2 alleles were used. Embryo collections, cuticle preparations, in situ hybridizations, and immunodetections were performed according to standard procedures. Digoxigenin RNA-labelled probes were generated according to the manufacturer's protocol (Boehringer Mannheim, Gaithersburg, MD) from wg and dpp cDNAs cloned in Bluescript. Primary antibodies used are: anti-Antp (4C3, dilution 1/100, Developmental Studies Hybridoma Bank (DSHB)); rabbit anti-AbdA (1/1000); guinea-pig anti-Doc2+3 (1/400) and rabbit anti-Dmef2 (1/2000) from L. Perrin (IBDML, Marseille, France); rabbit anti-Exd (1/1000) from R. Mann; rabbit anti-Nau (1/100) from BM Paterson (University of Texas Southwestern Medical Center, Dallas, TX); rabbit (1/500) or mouse (1/200) anti-GFP (1/500) from Molecular Probes; chicken anti-GFP (1/1000) from Aves labs; mouse anti-β-galactosidase (1/1000) from Promega; rabbit anti-β-galactosidase (1/1000) from MP Biomedical; anti-digoxigenin coupled to biotin (1/500) from Jackson. Secondary antibodies coupled to Alexa 488, Alexa 555 (Molecular Probes) or to biotin (Jackson) were used at a 1/500 dilution.
AbdA variant were generated by PCR. Domain mutations were YPWM→AAAA; TDWM→AVAI; KEINE→KAAAA. The homeodomain point mutation alleviating DNA binding is a mutation of position 50 (Q→K; [47]). Constructs were cloned in pUAST or pUASTattB vectors for transgenic line establishment. Lines were crossed with the appropriate driver, and collected embryos were stained with anti-AbdA to select the conditions (line and temperature) that result in expression levels similar (+/−15%) to AbdA wild type levels in A2 (see [22] for a detailed description of the procedure). Procedures used for quantification of biological readouts using at least 10 embryos of each genotype are provided in Text S1.
A matrix containing the values corresponding to the readout was built. The extreme values were given to the total loss of activity (value 0), and to the wild type activity (value 1 for 100% of activity). A hierarchical clustering algorithm (with Euclidian distance and average linking) was applied to the matrix using the MeV software suite [63]. The jacknife method was used for re-sampling the data and provides a statistical support for each tree node.
Boxplots drawn using the R-Software. Boxplot depicts the value distribution obtained for each tested genotype. Black points correspond to individual counts.
|
10.1371/journal.pcbi.1007090 | Connecting signaling and metabolic pathways in EGF receptor-mediated oncogenesis of glioblastoma | As malignant transformation requires synchronization of growth-driving signaling (S) and metabolic (M) pathways, defining cancer-specific S-M interconnected networks (SMINs) could lead to better understanding of oncogenic processes. In a systems-biology approach, we developed a mathematical model for SMINs in mutated EGF receptor (EGFRvIII) compared to wild-type EGF receptor (EGFRwt) expressing glioblastoma multiforme (GBM). Starting with experimentally validated human protein-protein interactome data for S-M pathways, and incorporating proteomic data for EGFRvIII and EGFRwt GBM cells and patient transcriptomic data, we designed a dynamic model for EGFR-driven GBM-specific information flow. Key nodes and paths identified by in silico perturbation were validated experimentally when inhibition of signaling pathway proteins altered expression of metabolic proteins as predicted by the model. This demonstrated capacity of the model to identify unknown connections between signaling and metabolic pathways, explain the robustness of oncogenic SMINs, predict drug escape, and assist identification of drug targets and the development of combination therapies.
| Complex and highly dynamic interconnected networks allow cancer to take different routes and circumvent chemotherapy. Therefore, understanding these context-specific networks and their dynamics of molecular interactions driven by different oncogenic signaling and metabolic pathways is very much needed to predict drug targets and the effect of therapeutics. We incorporated high-throughput transcriptome and proteome data into mathematical models to deduce properties of cancer cells through systems biology approach. Here we report the development, testing and validation of an integrated systems biology model of information flow between signaling and metabolic pathways to understand the regulation of the interconnection between them in cancer. Our model efficiently identified unique connections and key nodes important in signaling-metabolic information flow. We predicted some potential novel targets before performing actual drug tests. We have successfully applied this model to identify the interconnections altered in the constitutive signaling of the mutated EGFR by comparing EGF-dependent and wild-type EGFR signaling in glioblastoma multiforme.
| Diseases like cancer involve a large range of components that interact via complex and highly dynamic networks [1–3], and are interconnected with biochemical pathways [4–7]. These multipath interconnections may allow cancer and other diseases to take alternate routes and bypass the effects of therapeutic interventions. Traditional approaches to biological studies which focus on single molecules or pathways may not be able to capture and understand these complex networks of molecular interactions. To predict alternative or escape routes around blockades and to develop effective therapies [8], sophisticated mathematical and computational models are required [9–10].
Transforming traditional drug discovery approaches toward smarter therapeutic strategies, the field of systems biology is emerging [1, 9, 11–20]. Systems approach generally involve large-scale data collections, most often from high-throughput transcriptome or proteome analyses, incorporation of the data into mathematical models to deduce systems properties, model building and finally computational and/or experimental validation of model-derived hypotheses. Systems biology approaches may predict combination therapies for cancers driven by different oncogenic signaling and metabolic pathways.
Signaling and metabolic networks were studied using separate model systems [15, 21–26]. Mathematical models for signaling pathways had been based on logic models [27–30], kinetic models [31–33], decision tree [34], and other differential equation-based models [35]. Computational models of molecular signaling [36–41] have the potential to improve drug discovery and development [32,42–44]. Analyses of knockdown experiments [45] using mass spectrometry [46] and transcriptomics [47–49] are progressively refined and tuned towards specific physiological situations. While these studies have helped considerably to extend our understanding of tumor biology, they are still restricted to signaling pathways and do not integrate the metabolic pathways, which in some initial studies have been subjected to separate systems biology analysis.
Predicting the effects of multiple targeted drugs [8, 50] with modeling the information flow from new molecular interactions within pathways is challenging [51–55]. Here we report the development, test and validation of an integrated model for signaling and metabolic pathways in cancer using glioblastoma multiforme (GBM) as an example [47, 56–59]. GBM is the most prevalent and most aggressive brain tumor. In the majority of cases, tumor development is dependent on signaling via the epidermal growth factor receptor (EGFR) and requires EGF in lower-grade forms or is EGF-independent in the more aggressive forms. In most cases, the expression of EGFR is up-regulated, often related to the amplification of the EGFR gene. More than fifty per cent of EGFR-amplified GBM cases have in-frame deletions of exons 2–7 that code for the extracellular ligand-binding domain (EGFRvIII mutation) resulting in EGF-independent constitutive signaling and more aggressive tumor growth, higher invasiveness, increased resistance to treatment, and poor prognosis [60–62].
In this study, we choose the cell line U87MG expressing the EGF-dependent EGFRwt and its derivative U87MGvIII expressing the EGF-independent EGFRvIII-mutant, as models of low and high-grade GBM, respectively. Except for the EGFR mutation, the two cell lines have the same genotype but differ in growth behavior suggesting different metabolic requirements, and are experimental examples for the regulation of the interconnection of signaling and metabolic pathways, which is considered among the basic characteristics of cancer [63–64].
We implemented a probabilistic approach based on the Hidden Markov Model (HMM) utilizing the information of experimentally established protein-protein interactions (PPIs) [65–66] to extract novel paths and interconnections between signaling pathway proteins (S) and metabolic pathway proteins (M). To cope with the limitations of PPI identification, for example high error rates of the detection methods [67], incomplete data sets, ignorance of the physiological conditions in the cell or tissue compartments [68], technical problems and study biases [69], we collected information from curated sources (http://string-db.org) with high experimental score cut-off, which reduces false positive rates, and used transcriptome data from clinical samples to build more reliable and GBM context-specific PPI networks. To bridge the gap between transcriptome data and cell-biological processes, we incorporated proteome data from the GBM cell lines complemented with transcriptome data to solve the missing data problem caused by the failure of proteomics to capture all proteins in the cells due to sensitivity and reproducibility issues. Including experimental data, our dynamic model can make use of multiple weighted network properties to add biologically relevance and can extract novel paths of information propagation in networks. The results of the model were tested in rigorous in-silico perturbation experiments and experimentally validated in cell culture systems. Fig 1 depicts the overall strategy implemented with this study.
As a starting model, we constructed an integrated network where signaling (S) and metabolic (M) pathway proteins were connected through protein-protein interactors (PPIs). The signaling pathways were the apoptosis, Akt, EGFR, hedgehog (Hh), JAK-STAT, JNK, MAPK, mTOR, NF-kappa B (NF-κB), Notch, p53, Ras, TGF-β, and Wnt pathways. On the metabolic site, there were 81 pathways grouped into the six categories carbohydrate, lipid, amino acid, nucleotide, energy and xenobiotic metabolism (Fig 2).
To build an initial human protein-protein interactome network (HPPIN), the total of all human protein-protein interactions was extracted from the protein-interaction database STRING [70] for experimentally validated interactions including physical and functional associations. Then to structure the signaling-metabolic interaction network (SMIN), cross-connections between the fourteen signal transduction and six groups of metabolic pathway proteins selected from pathway databases (see Materials and methods) were constructed based on protein-protein interacting proteins (node) and cross-connecting links/paths. All possible connections between any given signaling pathway protein (S) and metabolic pathway protein (M) via protein-protein interconnectors (PPIs) were included (Fig 2A). To derive a simplified but informative network, the interactions were restricted to the second level of protein interactors, i.e. up to interactors of interactors. This led to four different types of cross-connected paths where signaling pathway proteins were either directly connected with metabolic pathway proteins or through one, two or three PPIs, respectively (Fig 2A). These paths connecting all above-mentioned signaling (S) and metabolic (M) pathway proteins were then converted into networks based on the protein-protein interaction status between the involved proteins (Fig 2B).
The resulting signaling-metabolic interaction network (SMIN) is shown in the left panel of Fig 3A with nodes in orange and edges (connection between two nodes) in blue,within the total interactome containing HPPIN network with grey nodes and edges. As a result of afore-said restriction criteria, Fig 3A shows the reduction of the network size (number of nodes/proteins and their interactions/edges) of SMIN (11,059 interactions formed by 2,785 proteins) from HPPIN (16,828 interactions formed by 5,703 proteins). As an example of network links, the detailed connections and interactions between one signaling pathway protein, CSNK2A1, and one metabolic pathway protein, NDUFA13, through different PPIs in the SMIN are shown on the right panel of Fig 3A. These two selected examples are indicated by asterisks. The SMIN was found to contain 158 direct (S-M) linking paths, 4,036 with one interactor (S-P-M), 91,847 with two interactors (S-P-P-M) and 2,110,205 with three interactors (S-P-P-P-M). These paths were formed between 158, 2,967, 22,307, 69,032 S-M pathway protein pairs, respectively (Table 1). Comparisons with respective random networks proved that our selected HPPIN and SMIN are non-random scale-free networks (S1 Fig and S1 Table).
To render the above condition-independent SMIN GBM-specific, quantitative comparative proteome analysis of the low- and high-grade GBM cell lines U87MG and U87MGvIII were performed and the data incorporated in the model to derive an enriched GBM-specific network. The use of cell lines helped to reduce the noise associated with the usual small size and heterogeneous cellular compositions of clinical samples. Proteomes was chosen as primary expression data to focus the models on the protein level, which is more closely related to the biological processes to be modeled than transcriptome data. Transcriptome date from clinical samples (see below) were subsequently added to reduce the missing data problem of proteomics and link the model to the clinical level. To decrease the complexity of the proteins in the cell extracts and minimize signal suppression by overabundant peptides, the proteins were first separated by SDS-PAGE. Then the gels were cut into one mm slices followed by separate in-gel digestion of the proteins with trypsin. The resulting fragments were extracted from the gel slices as individual samples, separated by reverse-phase nano-HPLC and analyzed on-line by ESI-Q-TOF mass spectrometry (S2 Fig). Quantification was done label-free by calculating the exponentially modified protein abundance index (emPAI) to avoid the drawback of mere signal intensity-based measurements. Only proteins with at least two tryptic fragments were identified by MS/MS with high confidence were considered, which further reduced the noise of protein identification although resulting in lower numbers of hits. The analyses were done with three independent replicates per cell line and the resulting data processed in two different ways. First, each of the three datasets for U87MG was compared to each of the three datasets for U87MGvIII resulting in nine pair-wise comparisons. Second, the average of three datasets of one U87MG was compared with the average of the three datasets for U87MGvIII.
A total of 907 unique proteins were identified, 771 from U87MG and 664 from U87MGvIII. Five hundred twenty-eight proteins were expressed in both cell lines, 243 only in U87MG (EGFRwt) and 136 only in U87MGvIII (EGFRvIII). These distinctively expressed proteins were considered as overexpressed in the respective cell lines in comparison to the other. Compared quantitatively, 458 of the 528 common proteins were expressed at similar levels, 70 proteins were either up (45) or down (25) regulated in U87MGvIII compared to U87MG (Fig 4A). Together, nearly half of the identified proteins (449) were differentially expressed, 268 down-regulated and 181 up-regulated in U87MGvIII versus U87MG (Fig 4B).
Over-representation analysis (ORA) based enrichment for cellular pathways, biological processes and molecular functions were performed using cell line specific and commonly overexpressed proteins. Fig 4C and 4D show the top 20 most enriched pathways for proteins exclusively overexpressed in U87MGvIII (EGFRvIII) and U87MG (EGFRwt), respectively. Similarly, Fig 4E and 4F provide the top 20 enriched pathways for commonly over and under expressed proteins, respectively.
The most significantly enriched pathways were found to be proteasome (p-value = 0.97763E-07) in U87MGvIII (EGFRvIII) and TCA cycle (p-value = 1.48369E-06) in U87MG (EGFRwt). However, Fructose and mannose metabolism (p-value = 6.7086E-04) and pentose phosphate pathways (p-value = 3.1503E-05) were found to be most significantly enriched for proteins commonly overexpressed and underexpressed, respectively.
Gene ontology (GO) based biological process and molecular function over-representation analysis was also performed using cell line specific overexpressed proteins. Most significantly enriched biological processes and molecular function were found to be Tricarboxylic acid metabolic process (p-value = 2.7304E-04) and Threonine-type peptidase activity (p-value = 0.001053394), respectively in U87MGvIII (EGFRvIII). In U87MG (EGFRwt), TCA metabolic process (p-value = 7.76842E-08) and pre-mRNA binding (p-value = 0.007857451) were found to be the most significantly enriched biological process and molecular functions, respectively (S3 Fig).
To uncover the signature of the reprogramming of global cellular processes by the EGF-independent constitutively active EGFRvIII in GBM, we mapped the data from the comparative proteomics of U87MGvIII (EGFRvIII) versus U87MG (EGFRwt) onto the above-described SMIN. All S-M interconnecting paths with at least one differentially expressed protein were extracted from the SMIN to generate a GBM-specific network (Fig 3B, left panel: highlighted with yellow nodes and green edges within the orange nodes and blue edges of the SMIN) with the assumption that those paths will have higher probability to be differentially active in mutant GBM condition. As an illustration of proteome data mediated extraction of a GBM-specific network, the interconnections between the aforementioned signaling pathway protein, CSNK2A1 and the metabolic pathway protein, NDUFA13 are highlighted after extraction from SMIN (Fig 3B right panel in comparison to Fig 3A right panel). Table 1 provides the details of the paths/pairs present at different stages of network development.
To make this network further enriched with potentially disease-relevant paths/pairs, weights specifying disease-related biological properties and expression states of the proteins were assigned to each node (protein/gene) and edge (interaction) of the GBM-specific network. The following three categories of proteins (nodes) were given additional weights. First, proteins cross-talking between different signaling pathways (signaling cross-talk, SC), second, rate-limiting enzymes (RLE) for their roles in regulating metabolic rates and pathways, third, EGFR mutation-specific differentially expressed proteins (dEXP) for their GBM-specific impact. The GBM-specific network included 446 dEXP, 349 SC, and 267 RLE proteins. Of these, 11 SC and 17 RLE proteins were up or down regulated suggesting their involvement in signaling-metabolic cross-connection in EGFR-mutated condition (S4A Fig). For systems-level interpretation and understanding the network property, local signaling entropy (Si) was introduced. Previous studies [71–73] showed that Si can be used as a measure of uncertainty in signaling information flow over a network and to identify important signaling pathways and genes/proteins in cancer. Effect-on-node (effs) of every protein (node) in the network provided significance of a protein based on SC, RLE and dEXP in its local network. To identify probable paths of information flow from a signaling to metabolic pathways, network entropy (Si) and effect-on-node (effs) properties were incorporated as node weights into the logic of the Hidden Markov Model (HMM). The edge weight of every two interacting nodes (gene/protein) were defined based on the principle of mass action (assuming that the probability of interaction of two genes in a given sample is proportional to the product of their expression values in the study samples) as probability of interaction (pij, where i and j are the two nodes) in GBM condition. To assign the expression value of each node present in the GBM specific network, the average expression value of each gene was calculated from the normalized transcriptome data from 239 GBM patients. Incorporating these transcriptome data as edge weights linked the network with biological information from GBM patients. It helped to assign an extra weight other than previously mentioned node weight for all connections made by two nodes based on their expression in clinical GBM patients. Furthermore, it helped to add another level of constraint on over-prediction of information flow for nodes, which got an extra weight based on SC, RLE, and dEXP but are not expressed at higher levels in GBM patients. This helped to incorporate the contribution of those nodes to the disease, which were identified by neither of the three before-mentioned node weights nor by proteomics. Moreover, the much broader coverage of gene expression by genome-wide transcriptomics compared to proteomics helped to overcome some of the missing-data-problem of proteomic datasets.
An HMM-based simple mathematical formalism was used to understand context-specific information propagation from signaling to metabolic pathways in the human biological network. Node weights and edge weights were used to define the two major parameters of the Markov model, emission (Ef) and transition (Tj) probabilities, respectively. Two model systems were implemented to apply HMM logic, Model 1 for SM pair identification (Fig 5A) and Model 2 for S-M linking path identification (Fig 5B). Model 1 emphasized source (Signaling Pathway Protein, S) and destination (Metabolic Pathway Protein, M) pairs i.e. SM pairs having higher chances of information flow for each type of connections (Figs 2A and 5A). Model 2 was applied to find S-M linking paths between those selected pairs from Model 1. Selection of SM pairs (Model 1) and S-M linking paths (Model 2) was based on Pathscore (see Methods for more details) a mathematical function of emission probability (Ef) and transition probability (Tj). For Model 1 positional emission probability was calculated considering the similar number of proteins because Model 1 was applied after grouping interconnecting links having the similar number of proteins forming the connection (Fig 5Ai–5Aiv). The calculated path scores of linking paths from the individual models were converted to statistical Z scores to identify the paths deviating from the mean. Based on the Z score under the individual models, the signaling-metabolic linking paths were classified as highly significant with Z score ≥3 (more stringent) or less significant with Z score ≥1 (less stringent) in EGFR-mutated GBM. The signaling and metabolic pathway proteins from the two ends of linking paths containing significant Z sores of each of the models were defined as the significant SM pairs. Multiple identifications of the same S-M pair from different models i.e. the formation of different types of significant linking paths involving different PPIs or different numbers of PPIs were nullified by considering them as a single. With that we identified 1, 2, 114 and 758 SM pairs meeting the more stringent cut-off (Z score ≥ 3) and 1, 8, 334 and 1,961 pairs with the less stringent cut-off (Z score ≥ 1) for the S-M, S-P-M, S-P-P-M and S-P-P-P-M linking types, respectively (Table 1, Model-1). In total 801 Z ≥ 3 and 2,055 Z ≥ 1 signaling-metabolic cross-connected SM pairs were identified between the 14 signaling and 6 groups of metabolic pathways as potentially important in EGFR-mutated GBM. These SM pairs were categorized according to the proteomic expression states of the source (S) and destination (M) proteins as UP-DOWN, UP-UP, DOWN-UP, and DOWN-DOWN. Including unidentified proteins in proteomics analysis (NA) of the cell lines, the couplet categories UP-NA, DOWN-NA, NA-UP and NA-UP, and unchanged expression states (NC) in the cell lines with SM categories UP-NC, DOWN-NC, NC-UP and NC-DOWN, and unidentified and unchanged SM pairs NA-NA and NC-NC were added (S4B Fig).
A number of cross-connections between signaling and metabolic pathways were identified with significant cutoff levels where either one or both pathway proteins (S and/or M) were not identified (NA) and/or unchanged (NC) by mass spectrometry indicating that the integrated network model can identify connections also where intermediate interactors are more important than SC, RLE or dEXP. The identification of pathway cross-connections is thus not dependent on proteomic identification of all constituent members but can be based on signaling crosstalk proteins and their expression status in GBM patients. It is important to include unchanged (NC) proteins in the model building as they might represent nodes in the paths that include other proteins that are differentially expressed. They might also play a role in the crosstalk between signaling paths and pathways or they might become important when known primary paths are blocked, e.g. by therapeutic intervention. The model could thus help to identify potential therapeutic targets for alternative therapies in cases of treatment failures or to design combination therapies that target primary together with potential escape pathways.
Mapping the pathway information of proteins in SM pairs showed which signaling pathways made a higher number of connections with which type of metabolic pathways in EGFR-mutated GBM. Six hundred one significant pairs with Z ≥ 1 were cross-connecting the MAP kinase pathway with all six groups of metabolic pathways (S2 Table). Similarly, the Ras, EGFR, AKT and p53 pathways were significantly connected to metabolic pathways through 570, 543, 549 and 179 SM pairs, respectively (S2 Table). As crosstalk between availabe signaling pathways is common whereas it is less common in between metabolic pathways (S5A Fig), some identified SM pairs and the respective proteins/genes may be shared. Analyzing the shared components in the five most connected signaling pathways revealed that MAPK pathway had the highest number of unique significant SM pairs (256) and genes/proteins (63) involved, followed by 194, 129, 116 and 95 unique significant SM pairs (S5B Fig) and 20, 30, 36, 19 genes/proteins (S5C Fig left) for the EGFR, AKT, p53 and Ras pathways respectively. Twelve cross-connected pathway protein pairs were common to all five pathways and 141 pairs shared by the Ras, EGFR, AKT and MAPK pathways indicating high connectivity between them, and their cross-connection with metabolic pathways suggesting important roles of the respective proteins in EGFR-mutated GBM (S5B Fig for pairs, C for genes left). In turn, the amino acid, carbohydrate, and nucleotide metabolic pathway groups were connected to all fourteen signaling pathways through 327, 289 and 326 cross-connected SM pairs (S2 Table, S5B Fig right)and 296, 260 and 268 genes in significant S-M paths, respectively (S5C Fig right). This indicates that altered cellular signaling related to the EGFR mutation and its constitutive activity affects most strongly these three metabolic pathway groups. As metabolic pathway enzymes interact via their substrates and products, there are few possibilities for interconnection between metabolic pathways except for the end steps, which is confirmed by the analysis of shared proteins. Twenty-three shared pairs were identified between the amino acid and the carbohydrate metabolic pathway, which relates to the low number of amino acids metabolites feeding into the tri-carboxylic acid cycle (S5B Fig right).
Multiple linking paths of different types (S-M, S-P-M, S-P-P-M, and S-P-P-P-M) or of the same type but through different PPIs were possible between SM pairs. Not all of these linking paths could be equally significant in EGFR-mutated GBM. To find the significant linking paths between the above-identified significant SM pairs, all possible paths between a single SM pair were considered under a single model (Model 2, Fig 5B). Since in Model 2 proteins forming interconnections between SM pairs vary, positional emission probabilities were calculated for these proteins. As an example, in Fig 5B the second position contained two proteins and third position one protein. Paths with path scores ≥80% of the highest path score for each SM pair were selected as significant from Model 2. Accordingly, all the significant paths were identified from high (more stringent Z ≥ 3) and less (less stringent Z ≥ 1) significantly specified SM pairs to identify the total of significant linking paths in the network. These analyses showed that 2, 21, 228, 625 and 876 significant paths were present with more stringent cut-off (Z ≥ 3) and 5, 84, 600, 1,564 and 2,253 paths with the less stringent cut-off (Z ≥ 1) for the four pathway types (Table 1). By these pathway-based analyses under less stringent condition (Z ≥ 1), 652 significant linking-paths were identified between 570 cross-connected SM pairs of the Ras pathway with all six groups of metabolic pathways (S2 Table). In addition, 668, 629 and 569 significant linking-paths were identified between 601, 549 and 543 cross-connected S-M pairs between the MAPK, AKT and EGFR pathways, respectively, and all 6 groups of metabolic pathways. Together, in EGFR-mutated GBM, these four signaling pathways were involved in the highest number of SM cross-connections with metabolic pathways: 368, 344 and 298 cross-connecting paths were found between the fourteen signaling pathways and 327, 326 and 289 cross-connected SM pairs involving the amino acid, nucleotide and carbohydrate metabolism, respectively (S2 Table).
Based on the identified significant (less stringent Z ≥ 1) SM pairs and the significant linking paths (path score ≥80% of the highest path score of each SM pair), we converted (as in Fig 2B) the significant paths into a network (Fig 3C left panel: highlighted with blue nodes and red edges within yellow nodes and green edges of the GBM network). This filtered network is more specific for EGFR-mutated GBM conditions. The filtration further eliminated non-significant or unimportant SM pairs and PPIs, which is shown, as an example, for the interconnections between signaling pathway protein CSNK2A1 and metabolic pathway protein NDUFA13 (Fig 3C right panel compared to Fig 3B right panel).
The GBM-specific network based on significant SM pairs and linking paths was restructured to implement the biological consequences as network properties i.e. color-coded proteomic expression states (up, down, no change and not identified), size of node symbols proportional to the numbers of connections passing through it, colors of the edges as connection formed between more (Z ≥ 3) or less (Z ≥ 1) stringently defined SM pairs, width of the edge as the probability of interaction or product of the average expression values of two interacting genes in GBM patients from the transcriptome data (Fig 6A). The resulting network showed the important signaling pathways and their interconnections with metabolic pathways in EGFR-mutated GBM with the significance of every protein (size of the node) and their interactions with interacting partners (width of the edge). Around the network, representative paths between 14 signaling to metabolic pathways are shown as examples (Fig 6A, S6 Fig as more details of RAS pathway). To explore the importance of signal-crosstalk proteins in signaling to metabolic pathway interconnections in EGFR-mutated GBM, the sub-network dependent on the top fifteen crosstalk protein-based interconnecting paths were extracted from the GBM-specific significant network (Fig 6B). These sub-networks showed which signaling pathways are mostly cross-talking and how they are connected with metabolic pathways. This information was used to extract candidate genes/proteins/paths of EGFR-mutated GBM for further analysis.
In silico perturbation analysis was performed for identification of paths that significantly change upon removal (e.g. by mutation or down-regulation) of a node (protein). To test the importance of the nodes/proteins in the final weighted network, each of the 654 nodes present in the Z ≥ 1 network (Table 1) was removed individually from the human interactome (HPPIN) and the node and edge weights were recalculated for the resulting networks and paths by recalculating Model 1 and Model 2 (Fig 2).
Accordingly, new significant SM pairs (Z ≥ 3 or 1) were identified on the basis of Model 1 and significant paths (path score ≥80% of the highest path score) between them on the basis of Model 2. We mapped the pathway details of the SM pairs and calculated the average path scores before and after perturbation for the 14 signaling pathways to all 6 groups of metabolic pathways and vice-versa. The difference of values (before vs. after perturbation) for the 654 proteins for all pathways were converted to Z-scores and plotted for each perturbed node for each signaling pathway (Fig 7A). The nodes for which the Z-scores deviated from the mean as -2 ≥ Z ≥ 2 were selected as effective for the respective signaling pathway to all metabolic pathway interconnections in EGFR-mutated GBM (Fig 7B). S3 Table lists the numbers of significant and effective proteins identified for the individual signaling pathways connected to all metabolic pathways and from all signaling pathways to the individual metabolic pathways. As a measure of its effect on signaling-metabolic interconnection, each perturbed node was ranked according to the difference between baseline and perturbed condition. This means that highly ranked proteins have an important role in the connections of the respective signaling pathway to all metabolic pathways or vice-versa.
The NOTCH pathway is shown as an example for the level of reduction in the network size when the GBM network is transformed to the significant GBM network to identify significant interconnecting paths and proteins (Fig 8). Fig 8A shows the interconnections of NOTCH pathway proteins with all metabolic pathway proteins present in the signaling-metabolic interaction network (SMIN) and Fig 8B shows only those NOTCH pathway proteins with interconnections to metabolic pathway proteins passing through effective nodes, i.e. nodes identified by the perturbation experiments (Fig 7B). Fig 8C presents the interconnections of the NOTCH pathway to all metabolic pathways in the GBM-specific network filtered on the basis of the weightage parameters for the nodes and edges, and Fig 8D shows the interconnections between the significant proteins only. Fig 8E shows the interconnections of the NOTCH pathway to all metabolic pathways in the GBM-specific significant network and Fig 8F the same only for significant nodes. The comparison of Fig 8B, 8D and 8F on basis of the effective node identified by the perturbation study (red colored) indicate the levels of filtration from the starting SMIN to the final GBM-specific significant network. We found 457,111 paths involving 1941 genes/proteins and 8047 interactions out of a total of 2,206,246 paths in the SMIN connecting the NOTCH signaling pathway to all metabolic pathways, of which 10% were found to have a significant (Z-score cut-off -2 ≥ Z ≥ 2) perturbation impact (PI) (Fig 8B). Comparable reductions of approximate 75% (Fig 8C) and 50% (Fig 8D) for both nodes and interaction were found when going to the GBM-specific condition. After Pathscore based filtration (Z-score ≥ 1), 146 paths involving 125 nodes and 166 interactions were identified in the NOTCH pathway (Fig 8E) of which 51 paths involving 59 nodes and 65 interactions formed by nodes with significant PIs.
The result of perturbation study proved the importance of the respective nodes in the final network for information flow from signaling to metabolic pathways. To validate these findings of interconnections between signaling pathway alterations and metabolic rearrangement, some of SM connections were selected from the final network for experimental validation (Fig 9). The selection was based on the effect of the removal of the respective node in the in silico perturbation experiments, the predicted effects on the metabolic pathways and the availability of specific small-molecule inhibitors, excluding transcription factors as their impact is obvious. The selected targets were calmodulin (CALM2), casein kinase II subunit alpha (CSNK2A1), 1-phosphatidylinositol 4,5-bisphosphate phosphodiesterase gamma-1 (PLCG1), the tyrosine-protein kinase ABL1 and B-cell lymphoma 2 (BCL2). The signaling-metabolic interconnecting paths including the selected proteins were extracted from the final GBM-specific significant network using the most stringent cutoff and the connected metabolic proteins linked to the above-mentioned signaling pathway proteins were identified (S1 File). As the translation of oncogenic signaling to metabolic adjustment is via the expression of metabolic enzymes, the changes of expression of the metabolic pathway proteins upon inhibition of the selected signaling proteins with small molecule inhibitors and cell viability were analyzed. The inhibitors were CGS-9343B for calmodulin (Fig 9A), Emodin for CSNK2A1 (Fig 9B), U73122 for PLCG1 (Fig 9C), Dasatinib for ABL1 (Fig 9D), and ABT199 for BCL2 (Fig 9E). U87MG (EGFRwt) and U87MGvIII (EGFRvIII) cells were incubated with the inhibitors and their effects on the expression of the predicted interconnected metabolic pathway proteins analyzed by quantitative RT-PCR, and the viability of the cells tested in MTT assays. The blockade of the signaling pathway proteins had significant impacts on the expression of the metabolic pathway proteins (Fig 9). In the case of the calmodulin inhibitor, the expression levels were reduced in both cell lines but for prostaglandin-endoperoxide synthase 2 (PTGS2) and O-linked β-N-acetylglucosamine transferase (OGT) much more pronounced in the EGFR-mutant cell line than in the wild type. The strongest effect was seen for phosphoglycerate kinase 1 (PGK1) with around 400-fold reduction of expression in both cell lines. The effects of the CSNK2A1 inhibitor are more differentiated. While the OGT and PGK1 expression is inhibited uniformly in both cell lines, 800-fold in case of PGK1, the inhibition of glyceraldehyde 3-phosphate dehydrogenase (GAPDH) is much stronger in U87MGvIII than in the wild type, the expression of ribonucleoside-diphosphatereductase 2 (RRM2) is decreased in the wild-type and enhanced in the mutant, and that of NADH dehydrogenase [ubiquinone] 1 alpha sub-complex subunit 13 (NDUFA13) is affected inversely. Inhibition of phospholipase C gamma 1 (PLCG1) results in enhanced expression of GAPDH and reduced expression of NDUFA13. The ABL1 inhibitor reduced the expression of pyruvate dehydrogenase kinase subunit 1 (PDK1) in both cell lines but much more in the mutant cells. In contrast, the expression of pyruvate dehydrogenase (lipoamide) beta (PDHB) is more reduced in the wild type, and that of pyruvate dehydrogenase subunit alpha 1 (PDHA1) is slightly enhanced in the wild-type and reduced in the mutant cells. Finally, the BCL-2 inhibitor reduces the expression of 6-phosphofructo-2-kinase/fructose-2,6-biphosphatase 3 (PFKFB3), GAPDH and RRM2 2-fold in both cell lines but enhances the expression of PGK1 in the wild-type 4-fold while reducing it in the mutant cells 10-fold.
All signaling pathway inhibitors have pronounced negative effects on the viability of both GBM tumor cell lines (Fig 9) but U87MGvIII is significantly more sensitive to the inhibition of casein kinase than the wild-type whereas the wild-type is more sensitive to the inhibition of PLCG1, ABL1 and BCL2. Half-maximal reduction of viability is seen for the PLCG1 and the ABL1 inhibitors in the single-digit, for the others in the double-digit micromolar range.
Further, the effect of the signaling molecule inhibitors on cell migration and invasion was tested. The results showed the similar negative effects of the inhibitors with a higher sensitivity of U87MGvIII, which reconfirms the potency of the signaling-metabolic interconnected network model (S7 Fig).
In all cellular condition, signaling, gene expression, and metabolic pathways must be coordinated to maintain cellular integrity and functions [74]. In this study, we propose an integrative systems biology model of information flow between signaling and metabolic pathways that use an interconnected scaffold based on the complete set of human interactome (HPPIN). The model successfully detects probable unique connections of genes/proteins involved in the interconnection of different signaling and metabolic pathways. The identified components of the network (i.e. genes/proteins as nodes and connections through edges) represent known physical and/or functional associations between proteins/genes. Depending on the oncogenic signaling, the interconnectors involved modulating the metabolic pathways change. We have successfully applied the model to identify the interconnections altered in the constitutive signaling of the mutated EGFR in glioblastoma multiforme (GBM) compared to EGF-dependent and wild-type EGFR (EGFRwt) signaling.
So far, the development of integrated models including all three or any two of signaling pathways, gene regulation and metabolic pathway in any cellular context is still very restricted. Difficulties of integration arise from the mode of information flow in these three processes significantly involving activation/inactivation, inhibition/induction and substrate/product respectively, as well as different extents and time frames or kinetics [74]. Attempts to integrate metabolism and gene regulatory networks were based on regulatory flux balance analysis (rFBA) [75,76], steady-state regulatory FBA (SR-FBA) [77], probabilistic regulation of metabolism (PROM) [78], integrative omics-metabolic analysis (IOMA) [79] and ordinary differential equation (ODE) [80]. There is a straightforward interconnection between signaling and gene regulation through transcription factor (TF) but the little informative interface and the dynamics of cellular localization of TF affect the integrative modeling. A few studies addressed the issues using the integrative logical, influence graph, Boolean and thermodynamics models [81–86]. Attempts to integrate signaling and metabolism is rarer because information in the two process areas flows in different ways via activation/inactivation and substrate/product interactions respectively. The few reports that address this issue are based on ODE, combined dynamic ODE and Boolean models [33, 80, 87]. Here we have taken a novel approach and integrated signaling and metabolic pathways by i) creating a weighted PPI network depending on the path of information flow from signaling pathway molecule to metabolic enzymes, and ii) applying the probabilistic framework of Markov processes to calculate information flow scores. We have used two major functions of the Markov model. First, to estimate the probability of a protein to be present at a particular point in the network to transmit information, transition functions (node weights) were implemented. Second, to estimate the connection strength between two proteins, emission functions (edge weights) were introduced. The two functions utilize different biological parameters to explore the integrated network and to calculate S-M Pathscore. In contrast, to other models for integration of signaling and metabolism pathways, the comparative proteomic expression of the signaling molecule, metabolic enzymes and interactors are effectively incorporated to construct a weighted PPI network (one criterion of node weights) and to filter out context-dependent (mutated vs. wild-type EGFR GBM) S-M interconnections. The second attempt toward integration of the two pathway types to make modeling information flow more accurate and render the model context-dependent based on patient-related disease-specific parameter, trancriptomic expression value from GBM patients were implemented for each node and, by using the principle of mass action, edge weight were defined. The implementation of these two layers of information distinguishes the model from others by integrating weighted network properties like network entropy (Si) and effect-on-node (effs). We believe that these have the potential to capture context-dependent properties of PPI networks in terms of information flow [88]. Genes/proteins having a high impact on the path formation are important in their local network as well as the global context. Unlike considering conventional network centralities (hubs, bottlenecks etc.) of nodes, biological state and expression profiles are used.
Integrating proteome and transcriptome data in the development of the model helps to mitigate problems in translating high-throughput omics into systems biology models. While proteome data more closely related to the biomedical and pathological systems properties of interest, they are less complete and more prone to variation and experimental errors then transcriptome data. The use of proteome data from cell lines helps to some extent to overcome the experimental problems, however, cell lines are selected cell culture-adapted models and may be quite remote from in vivo situations. Transcriptome data from clinical specimens are getting increasingly available and may add a close-to-clinic dimension to the systems biology model, and help to overcome the missing-data problem of proteomics. Furthermore, the two omics data sets were used at different steps of the network establishment and for different biological relevance. The proteome data of the EGFRwt and EGFRvIII-mutated cell lines were used first to filter out the GBM-specific network from signaling-metabolic interaction network (SMIN). On that GBM-specific network, we have incorporated the human GBM patient transcriptomic data as edge weights to make it more relevant to human GBM condition.
The resulting model is probabilistic, based on the assumption that the likelihood of signal flow relates to the expression levels of the nodes/proteins in the different paths/pathways. It is not computing the actual signaling status of the paths and their nodes. Incorporation of phosphoproteome data in this integrated network could provide additional information about the activation/deactivation state of the nodes thus adding elements of the actual information flow. This could help to add some rationale weights for nodes that remained unchanged (NC) or were unidentified (NA) in the proteomic study as well as in GBM patient transcriptome data. Accordingly, it could add to solving the missing-data-problem of proteomics. Although phosphoproteome data could provide relevant information about signaling flow it could be less advantageous in the model of interconnection between the oncogenic signaling and metabolic pathways that are required to sustain the oncogenic processes. It could make the model more complex as expression and phosphorylation of proteins are not correlated directly; phosphorylation states are highly context- and time-dependent and requires to take into account the ratio of phosphorylated and non-phosphorylated signaling proteins at any given position within the cell and the spatial arrangements of active/inactive signaling pathway molecules.
To test the network properties, the dynamic behavior and the robustness of the model, we performed in silico perturbation experiments. The in silico perturbation experiments were done with the final GBM-specific model with all interactome, proteome and transcriptome data incorporated in this order to render a general interactome map GBM-specific, and calculate node/edge weights and path score. In silico perturbation was done by silencing the signaling molecules one by one, and then redeveloping the model and recalculating node/edge weights and path scores to determine the relevance of the respective signaling molecule and the affected signaling path, to identify possible escape paths bypassing the blockade, and to identify targets for experimental validation of the model. Based on the perturbation analysis we postulated that proteins with higher impact as loss-of-function or gain-of-function or Pathscore difference between signaling-metabolic pathways and vice-versa have a higher ranking and chances to transmit information between the studied pair of pathways. The perturbation studies also captured the weighted topological changes in the dynamic network and the distinct sets of interactors that transmit information between proteins. Implementation of path score differences before and after perturbation adds additional accuracy to the model to identify the most important paths of information flow among all probable paths. Applying this model from 437,449 probable GBM-specific S-M paths, we identified 2,253 (Z≥1) (0.52%) and 876 (Z≥3) (0.20%) significant paths of information flow in mutated EGFR-dependent compared to wtEGFR EGF-driven GBM. To experimentally test and validate the interconnections predicted by the model, inhibition of signaling pathway proteins (S) present in the SMIN with small molecule inhibitor was performed, which resulted in alteration of metabolic pathway protein expression (M). We found that the performance of the model in predicting the dynamics of the large-scale signaling networks is comparable to state-of-art methods for extracting context-dependent information flow [89]. One node knockout at a time in the in silico perturbation experiments indicated, first, the importance of that protein in the signaling-metabolic interconnected paths in establishing low or high grade GBM condition, and second, alternative signaling routes that can become relevant as escape responses to therapy.
It is important to note that in biological networks including signal transduction, PPI is highly dynamic in nature and may undergo continuous changes [90]. These context-specific network dynamics need to be predicted by systems biology models. Our probabilistic integrated network model captures network dynamics using the topology of PPI networks based on dynamic data under different circumstances. The probabilistic network-based dynamic model uses experimental data and deals with the uncertainty of systems to predict drug targets and understand the effect of therapeutics. In our study, nodes forming paths significant to information flow in a particular biological context is the key to network inference. Nodes present in information flow path with high network weights and connected with higher edge weight can be considered for drug intervention and combinatorial therapy. Our model efficiently identified key nodes which are important in S-M information flow with confirmation of some known and prediction of the potential novel drug targets [89] that can be used alone or in combination to inhibit mutated EGFR-mediated GBM.
This model is now a useful tool and will be used to develop strategies and experiments to study causal relationships. Among the questions to be addressed will be to understand the molecular biology of high-grade GBM versus low-grade GBM, and mechanisms of therapy resistance. However, in its current form, this model is not engineered to measure signal flow in and across signaling pathway(s). It is also not applicable to other regulatory mechanisms such as transcriptional, post-translational and miRNA based modulations of biomolecular interactions. Nonetheless, the model should be robust enough to integrate and utilize large-scale pan-omics data including genomics, transcriptomics, proteomics, metabolomics and post-translational modification information to accurately model the cellular context under a given scenario. These issues are important for assessing systemic changes and we are addressing them in follow-up studies. The strategy for developing the model can also be applied to other cancers and to non-oncological conditions.
Interestingly, GAPDH and PGK1, commonly reported reference [91] and housekeeping genes [92], were identified (Fig 7B) as significant (~40% of 20 networks in the study) by our model and previously reported studies [47] have shown their association with cancer [93–95]. In this study, also significant expression change in two cell lines for GAPDH and PGK1 was found after inhibition of CSNK2A1 and BCL2, respectively.
To our knowledge, this has not been reported before. A study [96] with a mice model has shown increased expression level of bcl-2 under the control of pgk-1, which suggests a possible relation to our findings. Not surprisingly, we have also found FOXO, a key player in cell fate decision [97] as candidate target. In addition, the HSP90 family gene (HSP90AA1) and CREBBP were identified as important agents in forming interconnections between multiple signaling/metabolic pathways. Inhibitors of HSP90 (17-AAG) and CREBBP (ICG-001) have recently shown effects in glioblastoma and other cancer models/cell lines [98–103]. They have entered clinical trials for several cancer types [104, Clinical trial numbers:NCT01606579, NCT01764477, NCT02413853]. This confirms that our model can identify potential therapeutic targets before performing actual drug tests.
The presented network model can be used to explore novel therapeutic strategies against cancer including combination therapies. This work also represents a step towards finding alternative routes/pathways in cancer or other diseases and thereby predicting the potential path of therapy evasion that can be included in the development of new therapies that aim to prevent therapy resistance in cancer and other diseases.
The human grade-IV glioblastoma cell line U87MG expressing wild-type EGFR (U87MG) was purchased from ATCC (USA). The U87MG-derived genetically engineered U87MGvIII cell line with exons 2–7 deleted from the EGFR gene (EGFRvIII) cell was a kind gift from Professors Webster K. Cavenee and Frank B. Furnari, Department of Medicine and Cancer Center, University of California at San Diego, La Jolla, USA. The cells were grown in Iscove’s Modified Dulbecco’s Medium (IMDM; Gibco, Thermo Fisher Scientific Inc., Schwerte, Germany) supplemented with 10% heat-inactivated fetal calf serum (FCS; Biochrome, Berlin, Germany) and 1% penicillin/streptomycin solution (Gibco, Thermo Fisher Scientific Inc., Schwerte, Germany) in a humidified atmosphere with 8% CO2 at 37°C.
GBM cells (1 × 104) were plated in 96 well plates and treated with inhibitors of Calmodulin (CGS-9343B, Sigma), Casein kinase II subunit alpha (Emodin, Sigma), 1-phosphatidylinositol 4,5-bisphosphate phosphodiesterase gamma-1 (U73122, Sigma), Tyrosine-protein kinase ABL1 (Dasatinib, Santacruz) and B-cell lymphoma 2 (ABT199, Santacruz) for 24hr and subsequently incubated with 3-(4, 5-dimethylthiazol-2-yl)-2, 5-diphenyl tetrazolium bromide (MTT,100 μg/ml DMSO) in fresh culture medium for 3 hr. Optical density (OD) was taken at 550 nm with an ELISA reader (Thermo) as described elsewhere [105].
U87MG and U87MGvIII were plated separately in 6-well plates with>90% confluence. Scratch-wounds were made with a micropipette tip, washed thrice to remove the floating cells and treated separately with ABL1, PLCG1, BCL2, CALM2 and CSNK IIA inhibitors at their IC50 dose (5μM, 3μM, 25μM, 45μM and 40μM, respectively) in medium and incubating them for the indicated time periods. Images were taken at 0 hrs and 8 hrs. The wound width was measured for untreated and treated groups from at least five different fields of three separate experiments, and percentage wound healing was calculated from the width at 8 hrs versus the initial width at 0 hrs, all using ImageJ software.
U87MG and U87MGvIII cells (5x104) were seeded separately to the upper part of matrigel-coated invasion chamber in serum-free medium (200 μl) and the lower chamber was filled with medium (600 μl) containing respective inhibitors of ABL1 and PLCG1 at their IC50 doses. The cells on the lower surface of the insert were stained with the crystal violet after 24 hrs and the numbers of invaded cells were counted by inverted light microscopy from at least three different fields of three different experiments.
U87MG and U87MGvIII cells were treated with the indicated inhibitors at half of the IC50 dose for 24hr. Total mRNA was isolated by RNeasy Mini Kit (QIAGEN) as per manufacturer’s instructions. RNA (1.0 μg) was reverse-transcribed to cDNA with the Reverse Transcriptase kit (Promega). For RT-PCR, the cDNA was amplified with specific primers for the mRNA of the indicated metabolic proteins in a Perkin-Elmer DNA thermal cycler. Real-time PCR analysis was performed by mixing cDNA with 2× SYBR green master mix using Roche Applied Science light cycler 480.0 instruments with the software version 1.5.0. Relative quantification of each target gene was normalized to two housekeeping genes (18s rRNA and β-actin) and expressed as a fold change compared with untreated control using the comparative cycle threshold (CT) method [6].
For proteome analysis, total proteins were extracted with SDS-PAGE sample buffer from U87MG and U87MGvIII cells grown to 80% confluency. EGFR expression status of both cell lines was confirmed by western blot analysis. Proteins (100 μg) of each cell line were separated by SDS-PAGE (10% acrylamide/0.8% bisacrylamide). After staining with Coomassie blue, the lanes for both cell lysate were sliced into one-mm thick gel pieces. The slices were transferred into 96 well round bottom polypropylene plate and destained with 100 μl, 20 mM ammonium carbonate/acetonitrile (60%/40% v/v), dehydrated by adding twice acetonitrile (50 μl) and dried under vacuum in a speed vac. The dried gel slides were rehydrated in 20 μl trypsin solutions (10 ng/μl) in 30 μl NH4HCO3 (20 mM) for digestion by incubation for 16–18 hr at 37°C. The supernatants were collected in 1.5 ml reaction tubes and the remaining tryptic fragments extracted from the gel slices first with 50 μl 50% acetonitrile with 0.1% TFA for 15 min, and then with 5% acetonitrile with 0.1% TFA. The extracts from each gel slice were combined and dried in a speed vac. The peptides were re-dissolved in 12 μl 2% acetonitrile with 0.05% TFA.
For MS/MS, the peptides of the tryptic digests (10 μl) were loaded via a pre-column at a flow rate 20 μl/min (2% acetonitrile, 0.05%TFA) onto an Acclaim PepMap C18 nano-HPLC column (75μm inner diameter×15 cm length; Thermo Fisher Scientific Inc., Schwerte, Germany) using an Ultimate 3000 nano-HPLC system (Dionex, Darmstadt, Germany). The peptides were eluted with a gradient of 5–60% solvent B (0.1% formic acid in 95% acetonitrile/5% water) in solvent A (0.1% formic acid in 1% acetonitrile/water) over 60 min, followed by 60–90% solvent B over 5 min, and 90% solvent B for 5 min at a flow rate of 220 nl/min. The nano HPLC system was directly coupled to a Micro-TOF-Q I mass spectrometer (BrukerDaltonics, Bremen, Germany). Mass spectra were acquired in the m/z range 50–2500 at an acquisition rate of 1.3 per sec. MS/MS spectra were acquired in the data-dependent mode with the fragmentation of the 5 most intensive peaks (absolute threshold 3000) with argon as collision gas, and fragment masses in the range of 400–1400 m/z. The mass spectrometry was run with dynamic exclusion of a time interval established and tested beforehand to avoid or minimize signal suppression by over-abundant peptides. A dynamic exclusion of 1 min was set to avoid repeated fragmentation of the most abundant peptides.
The MS and MS/MS spectra were processed with Data Analysis 3.4 and Biotools 3.1 (BrukerDaltonics). MS/MS searches for peptide identification were done via Biotools on a local Mascot (version 2.2) server with a precursor mass tolerance of 50 ppm and a fragment mass tolerance of 0.2 Da. Trypsin was specified as an enzyme, and one allowed missed cleavage and oxidation of methionine as variable modification. Data searches were done in SwissProt databank for human proteins. Proteins were identified with at least two peptides with a mascot score of higher than 26. The false discovery rate with these search and filter parameters was below 5% as confirmed with a decoy database. Exponentially modified protein abundance index (emPAI) was employed for protein quantification using the equation
emPAI=10NobservedNobservable-1
(i)
where Nobserved is the number of experimentally observed and Nobservable the calculated number of observable peptides for each protein [106]. Comparative proteome analysis was done with proteins identified by at least two peptides based on the emPAI scores for differentially expressed proteins. The proteome analysis was done three times for each cell line as an independent biological replicates.
Comparative proteome analysis was done for each proteome dataset (batch) of U87MG with each batch of U87MGvIII using the in-house PROTEOMESTAT-12 software. For this, the emPAI values of all proteins were normalized as
emPAInorm=emPAIprotein[∑emPAIprotein]batch×100
(ii)
where emPAIprotein was the emPAI value of a protein in a batch.
Multiple identifications of a protein within a proteome dataset were excluded, only keeping the highest emPAI value for subsequent analysis. Intra-batch analysis was performed to determine the total unique proteins as well as the overlap between the three batches of each cell line. Total unique proteins identified from U87MG and U87MGvIII cells were compared to identify the common and the differentially expressed proteins. Proteins identified only for one of the cell lines were considered up-regulated in that cell line. Commonly expressed proteins were further analyzed in two different ways to identify differentially expressed proteins. In one approach, the emPAIave values for each protein were compared between the two cell lines (iii).
emPAIave=∑batch=13emPAIproteinNumberofbatcheswithidentifiedprotein
(iii)
where emPAIprotein was the emPAI value of a protein in a proteome dataset.
The percentages of up or down-regulation of a protein in U87MGvIII were calculated as % ExpU87MGvIII using the following the equation
%ExpU87MGvIII=(emPAIaveU87MGvIII-emPAIaveU87MG)emPAIaveU87MG×100
(iv)
where emPAIave U87MGvIII and emPAIaveU87MG were the average emPAI value of a protein among the three proteome datasets for U87MGvIII or U87MG. Proteins that were ≥ 50% up- or down-regulated with p values ≤ 0.2 were considered differentially expressed in the U87MGvIII cell.
In a second approach, emPAInorm of a protein of each batch of one of the cell lines were compared separately with the emPAInorm for the same protein in each of the three batches of the other cell line. Then these comparative data were normalized by the LOESS method. Up- or down-regulation of a protein in U87MGvIII were calculated using the Eq (iv). From this second approach ≥ 50% up- or down-regulated proteins with an SE ≤ 35% were taken as differentially expressed in U87MGvIII. Both the results were combined and duplicates were excluded to identify the differentially expressed proteins.
A signaling-metabolic cross-connected network was created using 14 signaling and 81 metabolic pathways and protein-protein interactors of the pathway proteins. Signaling pathway databases were created by integrating information from on-line resources including KEGG [107], Reactome [108] Signallink [109], NetPath [110] and Biocarta (http://www.biocarta.com) for extensive coverage of the steps involved in the pathways. The 14 signaling pathways were included based on their association with cancer in general according to KEGG pathway database.
Metabolic pathway datasets were collected from KEGG. Human protein interaction data were taken from the STRING interactome database using only protein-protein interactors [70] established by high-throughput analyses with the high confidence score (≥ 0.7) to generate a protein-protein interaction network (PPIN). PPI were restricted up to the second level interactors, i.e. interactors of interactors, which brings about S(PPI)3M, meaning S-PPI1-PPI2-PPI3-M paths. PPI1 is the first level interactor of signaling molecule and PPI3 is the first level interactor of metabolic enzyme and PPI2 the second level interactor of both. Thereby, there are three PPIs between the signaling and the metabolic pathway proteins. The resulting network consisted of 5,703 proteins with 16,828 experimentally confirmed interactions.
NetworkX (https://networkx.github.io/), a network analysis framework in a Python language, was used to calculate the degrees, Estrata Index and betweenness centrality of nodes (genes/proteins).
Gene expression data used for edge weight determination in the weighted-network were collected from five gene expression data sets for GBM patients taken from GEO (Gene Expression Omnibus)[111] GSE4290 (77 samples), GSE53733 (70 samples), GSE50161 (34 samples), GSE36245 (46 samples) and GSE15824 (12 samples) (http://www.ncbi.nlm.nih.gov/geo/). All had been acquired with the Affymetrix GPL570 platform [http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GPL570]. The raw data of the total of 239 samples were normalized across all arrays within a set using the RMA quintile normalization procedure of the Gene Pattern Expression File Creator module [112]. Averages of normalized raw expression values were used for the study.
To construct cross-connecting linking paths (CCLPs) between signaling pathway proteins (S) and metabolic pathways proteins (M), common interacting proteins from the HPPIN were used. All possible unique connections between S and M with S-M (direct S-M interaction), and S-P-M, S-P-P-M, and S-P-P-P-M where P is a common interacting protein were established for common interacting proteins up to the second level in HPPIN (Fig 3A). Detailed descriptions of all possible connecting links are provided in the Fig 3B and Table 1. Differentially expressed (i.e. up-regulated or down-regulated) proteins in EGFRvIII versus EGFRwt expressing cells obtained from proteome analyses were mapped onto the cross-connected links [Table 1]. To assess the relevance of signaling to metabolic pathway interconnections for cancer pathogenesis, connections with at least one up-regulated or down-regulated protein (U87MGvIII versus U87MG) were selected to create functional cross-connecting sub-network (CCsN) with, in all, 2,320 protein nodes and 8,914 interactions. This CCsN was used for network topology analysis.
The local entropy of a protein in CCsN was calculated on the basis of probabilities of an interaction of that protein with its interactors determined by using the principle of mass action. The calculation of the interaction probabilities is based on the assumption that two proteins known to interact will have a higher probability of interaction when they are highly expressed. Thereby, the interaction probability (Wij) of two proteins in a network is proportional to the product of expression values (E of the corresponding genes (Ei and Ej) [72]:
Wij∝EiXEj
(v)
with the expression value of CCsN genes calculated as an average expression of a gene in the data set of 239 GBM patients.
Based on Wij, a stochastic matrix of normalized interaction probabilities (Pij) in the network was created for signaling entropy calculation. Probability of interaction between node i and node j was calculated as
Pij=Wij∑k∈NiWik
(vi)
where Ni is interactors of node I with ΣPij = 1.
The local entropy of node i was calculated as
Si=−1logki∑j∈N(i)pijlogpij
(vii)
where ki is the degree of node i in the CCsN.
To incorporate the impact of the interactors of a particular protein in the cross-connected network, the node-weight of every node i was specified based on the categories signaling cross-talk protein (SC), rate limiting enzyme (RLE) and up- or down-regulation in U87MGvIII versus U87MG (dEXP):
Wi={1;if(dEXP=13,RLE=13,SC=13),0;else}
(viii)
Effect of interactors on a node s in CCsN was defined as effect-on-node (effs) depending on the node weight of the interactors up to the second level:
effs=∑jk(∑inWidegreei+Wjnj)
(ix)
where k is the degree of node s, nj is the degree of protein j, wj and wi are the node weights of protein j and i.
Rate-limiting enzymes (RLE) are the enzymes in metabolic pathways whose kinetics determines the overall kinetics of the entire pathway. It is usually the enzyme with the slowest enzyme kinetics. The identification of the RLE is done based on the Michaelis–Menten equation, and the established Michaelis constants (KM) and Vmax of all enzymatic steps in metabolic pathways are listed in the respective biophysics and enzyme kinetic textbooks and databases.
The CCLP for information flow from signaling pathway protein (S) to metabolic pathway protein (M) was defined to be one of the four types shown in Fig 5A, and path scores were calculated on the basis of node and edge weights of the proteins involved in a path. To select important S-M pairs, imaginary penultimate signaling and metabolic proteins are considered as starting and ending state, and path score was calculated based on a hidden Markov model (HMM) with a forward algorithm. Emission probability (Ej), i.e. the positional probability of a protein at the particular position in that state was calculated as
Ej=Sj+Effj∑ik(Si+Effi)
(x)
where k is the number of proteins in that state, Si and Sj are normalized local entropy of proteins i and j, Effi is effect-on-node for protein i. Within the S-M pairs of a path (Fig 5B) information flow is again scored by considering all types of paths formed between single S-M pairs and calculated as
Pathscore=I∏j=1nEjTj
(xi)
where n is number states in a path, I initial probability (in our case it is equal to one), Ej is emission probability at state j, Tj is transmission probability at state j (in our case it is the probability of interaction Pij). Pathscore is converted into Z-score as
Zscore=X-μσ
(xii)
where X is raw Pathscore, μ is mean, σ is the standard deviation. A cut-off of ≥1 is applied to select significant S-M pairs and their cross-connecting links.
To understand the significance of the path scores in comparison to randomly generated paths, we have performed permutations of node weights (Sij) and edge weights (Pij) and generated 20 random paths with the same number of proteins involved in the identified paths. Averages of 20 path scores are compared with the corresponding original path score. S8 Fig shows significant difference between the original and the randomly placed weighted path scores.
In silico perturbation analysis was done by removing nodes/proteins from the network using an in-house programme for nodes/proteins present in the final weighted sub-network. All nodes in this sub-network were removed individually from the SMIN, and the impact was studied by performing the same weighted network analysis with the sub-networks generated after node removal as the starting point. Perturbation score (Ps) was calculated in two steps. First, we calculated significant pairs using Model 1 shown in Fig 5A and significant paths using Model 2 shown in Fig 5B using SIN-Ni network where Ni is the interaction of perturbed node N. Second, path scores after perturbation (Pathscore') for significant paths identified from step 1 were calculated. Perturbation score (Ps) was defined as
Ps=Pathscore−Pathscore,
To identify nodes with significant impact after perturbation we converted perturbation scores (Ps) into Z-score for all signaling to metabolic pathways and vice-versa. In the perturbation analysis, node and edge weights were re-calculated for networks and paths generated after removing a node (protein) from the initial interactome. New weights in Model 1 and Model 2 were used to score the new paths and to compare the scores before and after perturbation thus to identify significant nodes in the network forming connections between signaling and metabolic pathway proteins.
KEGG pathway based over-representation analysis (ORA) was performed using 136 up-regulated and 243 down-regulated proteins (EGFRvIII vs EGFRwt) and 45 commonly up-regulated and 25 down-regulated proteins using ‘protein coding gene set’ as the reference gene set in WebGestalt [113] web tool.
Additionally, Gene ontology (GO) based biological process and molecular function over-representation analysis was performed for the same genes. Top 20 significant (p-values <= 0.01 and <= 0.05) categories were ranked based on the false detection rate (FDR) calculated using Benjamini and Hochberg procedure.
|
10.1371/journal.pcbi.1003162 | Viral Proteins Originated De Novo by Overprinting Can Be Identified by Codon Usage: Application to the “Gene Nursery” of Deltaretroviruses | A well-known mechanism through which new protein-coding genes originate is by modification of pre-existing genes, e.g. by duplication or horizontal transfer. In contrast, many viruses generate protein-coding genes de novo, via the overprinting of a new reading frame onto an existing (“ancestral”) frame. This mechanism is thought to play an important role in viral pathogenicity, but has been poorly explored, perhaps because identifying the de novo frames is very challenging. Therefore, a new approach to detect them was needed. We assembled a reference set of overlapping genes for which we could reliably determine the ancestral frames, and found that their codon usage was significantly closer to that of the rest of the viral genome than the codon usage of de novo frames. Based on this observation, we designed a method that allowed the identification of de novo frames based on their codon usage with a very good specificity, but intermediate sensitivity. Using our method, we predicted that the Rex gene of deltaretroviruses has originated de novo by overprinting the Tax gene. Intriguingly, several genes in the same genomic region have also originated de novo and encode proteins that regulate the functions of Tax. Such “gene nurseries” may be common in viral genomes. Finally, our results confirm that the genomic GC content is not the only determinant of codon usage in viruses and suggest that a constraint linked to translation must influence codon usage.
| How does novelty originate in nature? It is commonly thought that new genes are generated mainly by modifications of existing genes (the “tinkering” model). In contrast, we have shown recently that in viruses, numerous genes are generated entirely de novo (“from scratch”). The role of these genes remains underexplored, however, because they are difficult to identify. We have therefore developed a new method to detect genes originated de novo in viral genomes, based on the observation that each viral genome has a unique “signature”, which genes originated de novo do not share. We applied this method to analyze the genes of Human T-Lymphotropic Virus 1 (HTLV1), a relative of the HIV virus and also a major human pathogen that infects about twenty million people worldwide. The life cycle of HTLV1 is finely regulated – it can stay dormant for long periods and can provoke blood cancers (leukemias) after a very long incubation. We discovered that several of the genes of HTLV1 have originated de novo. These novel genes play a key role in regulating the life cycle of HTLV1, and presumably its pathogenicity. Our investigations suggest that such “gene nurseries” may be common in viruses.
| Modification of existing genes, such as by duplication or fusion, is a common and well-understood mechanism by which protein-coding genes originate [1], [2]. In contrast, we have shown that viruses generate many proteins de novo (hereafter called “de novo proteins”) [3], [4]. Preliminary observations indicate that these proteins play an important role in the pathogenicity of viruses [3], [5], for instance by neutralizing the host interferon response [6] or antagonizing the host RNA interference [7]. Strikingly, p19, the only de novo protein characterised both structurally and functionally, has both a previously unknown structural fold and a previously unknown mechanism of action [7]. Thus, protein innovation seems to be a significant, but poorly understood part of the evolutionary arms race between hosts and their pathogens [5], 8,9.
Studying de novo proteins should thus greatly enhance our understanding of host-pathogen co-evolution and our knowledge of the function and structure of viral proteins [3], [10]–[14]. However, a major bottleneck that prevents the study of such proteins is their identification, which is very challenging. Finding that a viral protein has no detectable sequence homolog does not reliably indicate that it has originated de novo, because viral proteins evolve so fast that they can easily diverge in sequence beyond recognition. To circumvent this problem, in our previous work [3], [4] and in the current study, we focused on a special case of de novo proteins: those generated by overprinting. Overprinting is a process in which mutations in a protein-coding reading frame allow the expression of a second reading frame while preserving the expression of the first one (Figure 1), leading to an overlapping gene arrangement [10]. It is thought that most overlapping genes evolve by this mechanism, and that consequently each gene overlap contains one ancestral frame and one originated de novo [10]. Because overlapping genes are particularly abundant in viruses [15]–[17], they constitute a rich source of de novo proteins.
Identifying which frame is ancestral and which one is de novo (the “genealogy” of the overlap) can be done, in principle, by examining their phylogenetic distribution (the frame with the most restricted distribution is assumed to be the de novo one). One can exclude the possibility that the phylogenetically restricted frame is in fact present in other genomes but has diverged beyond recognition, by checking that outside of its clade, the ancestral frame is not overlapped by any reading frame [4]. This approach is simple and reliable [3], [4], but is not applicable to cases where the homologs of both frames have an identical phylogenetic distribution. For instance, it could identify the de novo frame in only a minority (40%) of overlaps in our previous study [3]. Therefore, a new method is needed to identify the de novo proteins in most overlapping genes.
The approach we investigated is based on the hypothesis that the ancestral frame should have a pattern of codon usage (i.e. which synonymous codon(s) is preferred to encode each amino acid [18]) closer to that of the rest of the viral genome than the de novo frame [10]. Indeed, analyses of plant RNA viruses and animal DNA viruses [19], [20] have shown that, within a given viral genome, genes generally have a similar pattern of codon usage, which is thought to depend on the overall GC content of the genome [19]–[21]. In overlapping genes, the ancestral frame, which has co-evolved over a long period with the other viral genes, is expected to have a codon usage similar to that of the rest of the genome (Figure 1). On the other hand, the de novo frame, at birth, will have a codon usage in effect randomized by the shift and thus unlikely to be close to that of the genome. In addition, constraints imposed by the ancestral frame might prevent the de novo frame from adopting, later, the typical genomic codon usage. Consequently, the de novo frame is expected to have a codon usage less similar to that of the viral genome than the ancestral frame (Figure 1). This approach has been empirically used to try and identify the de novo frame in a number of cases, as have been related methods which rely on the frequency on nucleotides at some or all codon positions [10], [22]–[29]. However, the reliability or accuracy of these methods has never been tested. Here we gathered a reference (“benchmark”) dataset composed of overlaps with known genealogy, and used it to answer the following questions: do de novo frames have a codon usage distinguishable from ancestral frames? If yes, can codon usage be used to identify the de novo frame? What is the reliability of the method and its sensitivity? Finally, we applied this method to overlapping genes whose genealogy was undeterminable by the phylogenetic method.
As described in Material and Methods, we assembled a dataset of 27 independent, experimentally proven overlapping genes longer than 140 nt (Table 1). 16 of them have been described previously [3], as indicated by an asterisk in Table 1, and 11 additional overlaps were collected for this study. The 27 overlaps come from 25 genera, distributed in 16 viral families covering a wide range of viruses (Table 1). 18 overlaps involve one gene being completely overlapped by the other, while in 9 the overlap is partial (e.g. Figure 2). All overlapping genes are in the same orientation, i.e. there are no antiparallel overlapping genes in the dataset. To identify the genealogy of the overlaps, we used the same stringent criterion as in our previous study [3], selecting only cases in which one frame, predicted ancestral, had a much wider taxonomic distribution than the other frame, predicted de novo. To be confident about the taxonomic distribution of each frame, we carried out extensive searches involving the most up to date similarity search tools, supplemented by in-depth manual searches using contextual information (see Material and Methods). The taxonomic distribution of each frame, and the corresponding evidence, are presented in Supplementary Table S1. Our predictions of ancestry are supported by functional data: almost all proteins encoded by a frame identified as ancestral have a function central to the viral cycle (such as capsid or replication), while most proteins identified as de novo have a “secondary” function related to pathogenicity (such as silencing suppressor or apoptotic factor) (Table 1). Thus, the predicted genealogy of the overlapping genes of the dataset is highly reliable.
We needed to exclude from the dataset ancestral frames that have entered their genome by distant horizontal transfer since these frames are not expected to have the same codon usage as that of their new viral genome, and are thus not suitable for codon usage analysis. Performing a detailed recombination analysis on all ancestral frames of the dataset was out of the scope of this article, and thus we simply detected cases of taxonomic incongruence (see Material and Methods). We detected two cases in which the ancestral frame had originated from another viral genome by distant horizontal transfer. The ancestral protein p104 of Providence virus (genus alphacarmotetravirus, family Carmotetraviridae) has statistically significant similarity with the replicase of viruses from a different family, Tombusviridae. Also, the capsid protein of Maize chlorotic virus (genus machlomovirus, family Tombusviridae) has significant similarity to that of sobemoviruses, an unassigned genus unrelated to Tombusviridae [30]. We established that horizontal transfer took place towards alphacarmotetravirus and machlomovirus from the other families by analysing the phylogenetic distribution of homologs of the ancestral proteins (not shown). Our results agree with previously reported findings that Providence virus has originated through recombination between a Tombusviridae-like and a Tetraviridae-like virus [31], and that the machlomovirus capsid protein is taxonomically incongruent [32]. We excluded these two cases from our analyses, and the final benchmark dataset is thus composed of 25 overlaps (Table 1).
As a measure of codon usage similarity between a given frame and the rest of the viral genome, we used the Spearman's rank correlation coefficient (rs) between the number of occurrences of each codon in that frame and in the viral genome (see Materials and Methods). Accordingly, the higher the rs of a frame, the more similar its codon usage is to that of the genome. For all gene overlaps of the benchmark dataset, we evaluated rsA, rsN (the rs of the ancestral and the de novo frame, respectively), and the difference (d21) between rsA and rsN (d21 = rsA−rsN). They are listed in the left moiety of Table 2, ranked by decreasing value of t-Hotelling. rsA is higher than rsN in 21 cases (i.e. d21>0) and lower (i.e. d21<0) in only 4 cases. This distribution is not random (P<0.001, in accordance to the binomial proportion test), suggesting that ancestral frames have a codon usage closer to their genome than de novo frames. This conclusion is supported quantitatively, since the median rsA (0.42) is significantly (P<0.01) higher than the median rsN (0.19) according to the Wilcoxon signed rank test [33]. These findings support the hypothesis that codon usage can, in principle, be used to determine the ancestral frame.
We now needed a method to infer, given any pair of overlapping frames, whether one frame had a codon usage significantly closer to the rest of the viral genome than does the other frame. In principle, a suitable method to assess the significance of the difference between the rs coefficients of each frame is Hotelling's t-test [34], [35]. However, Hotelling's t-test is applicable to correlation coefficients estimated from independent data, whereas our data are clearly not independent (see Material and Methods). Therefore, we assessed whether Hotelling's t-test was robust to the violation of the non-independence of data by comparing the results of Hotelling's t-test with simulated codon usage data (see Material and Methods). Values of rsA, rsN and d21 for simulated frames corresponding to each overlap are presented in the right moiety of Table 2. We performed a McNemar test [33], which indicated that both methods provide equivalent results (McNemar chi-square = 0.6; P = 0.50). Therefore, Hotelling's t-test is reasonably robust to violation of independence and is applicable to our problem.
Having established the validity of Hotelling's t-test, we used it to predict the ancestral frame (and consequently the de novo frame) in our dataset. Given two overlapping frames 1 and 2, a frame (for instance frame 2) was predicted ancestral only if it matched the following two criteria:
The first criterion corresponds to our main biological hypothesis, whereas the second criterion avoids a scenario in which the first criterion gives results that are mathematically significant but not biologically meaningful. For instance, if one frame had an rs of −0.7 and the overlapping frame had an rs of −0.1, the difference would be significant. However, it would be unjustified to return a prediction that the second frame is ancestral, because the negative value of its rs contrasts with our central hypothesis that the ancestral frame has conserved traces of the genome's codon usage.
The overlaps are listed in Table 3 by decreasing value of t-Hotelling. We found that both criterions were fulfilled for 13 of our 25 overlaps, and in all these cases the ancestral frame prediction was correct, i.e. concordant with that established by phylogeny (Table 3). Consequently, the specificity of the codon usage approach is high (specificity = 1.0, 95% confidence interval [CI] 0.77–1.00), but its sensitivity is moderate (sensitivity = 0.52, 95% CI 0.31–0.72).
We examined several factors that could influence the ability to predict the de novo frame by its codon usage.
A first factor is genome segmentation: five overlaps of the dataset belong to viruses with segmented genomes (Aquabirnavirus, Begomovirus, Hordeivirus, Omegatetravirus, Orthobunyavirus). The calculations above were done by considering all genomic segments of such viruses as their “genome”. However, considering only the segment encoding the overlap under study yielded the same predictions, suggesting that genome segmentation is not a confounding factor.
Second, an extreme GC content could also, in principle, confound codon usage analysis. However, the GC contents of the genomes we analysed here are in a moderate range (35–57%), and thus are probably not a source of bias.
Third, in principle, the relative frame (+1 or +2) of the de novo region with respect to the ancestral region could influence the power of codon usage analysis to distinguish their genealogy. As can be seen in Supplementary Table S2, 16 de novo coding regions are in the +1 frame relative to the ancestral region they overlap, while the remaining 9 de novo regions are in the +2 frame. Among the 13 overlaps for which there was a significant difference in codon usage between the two overlapping regions, in 9 cases the de novo region was in the +1 frame relative to the ancestral region, while in 4 cases it was in the +2 frame (Supplementary Table S2). A chi-square test (chi-square = 0.023; P = 0.90) indicates that the sensitivity of our method does not change depending on the relative frame of the de novo region with respect to the ancestral region, and thus that the relative frame is not a confounding factor.
A fourth factor is the age of overlaps: as de novo proteins age, they may progressively impose increased constraints on the ancestral frames, which may change their codon usage, and make it difficult or impossible to distinguish them from de novo frames [4]. Precisely estimating the age of overlaps is not possible given the state of our knowledge of viruses. However, one can use the taxonomical distribution of de novo frames as a heuristic to obtain a very approximate idea of their relative ages. For instance, a de novo frame found in a single species of viral family A has almost certainly appeared more recently than a de novo frame found in a whole family B (provided there is a good sequencing coverage in both families). We have applied this idea to infer the age of overlaps of the benchmark dataset.
De novo frames found only in one species were considered “young” (provided there are several species in the genus considered); overlaps found in more than one species but less than one genus were considered of “Intermediate” age, and overlaps found in more than one genus were considered “old”. We have indicated these estimated relative “ages” in Supplementary Table S2 (the exact taxonomic distribution of de novo frames is in Supplementary Table S1).
There is insufficient taxonomic coverage to estimate the age of overlaps in two genera, for which only a single species is known (betatetravirus and mandarivirus). The remaining 23 overlaps cluster in the following way: 3 young, 13 medium, and 7 old (supplementary Table S2). By codon usage analysis we have (correctly) predicted the genealogy of 3 young, 6 medium and 2 old overlaps (supplementary Table S2). We have analysed these data by the chi-square contingency table test. The Chi-square value was 1.95 (P = 0.30). Therefore, the predictive power of codon usage to identify the de novo frames does not seem to be dependent on their taxonomic distribution, and by extension, on their relative ages.
A fifth potential confounding factor is that some de novo frames have a biased amino acid (aa) composition. This raises the possibility that the aa composition of de novo frames could be the major explanatory factor of our results, and that differences in codon usage would be a consequence of this biased aa composition. To empirically determine whether aa composition contains more information about frame ancestry than codon usage, we carried out a correlation analysis of the aa composition of overlapping frames with the same statistical test as for codon usage, and compared the predictive power of both methods. We performed the same analysis as on codon usage data but on the frequency of the 18aas that have a degree of codon-degeneracy >1. The median value of the Spearman correlation between the aa composition of the ancestral frame and that of non-overlapping regions was 0.62, while the median value of the Spearman correlation between the aa. composition of the de novo frame and that of non-overlapping regions was 0.50. Unlike for codon usage (see above), the difference was not significant (P = 0.35 in accordance to the Wilcoxon signed rank test). Therefore, aa composition does not have as much predictive power regarding the genealogy of overlaps as codon usage, and our results are unlikely to be explained by the difference in aa composition between ancestral and de novo frames.
Finally, to study whether recombination could be a confounding factor, we examined codon usage in the two cases in which the ancestral frame had arisen by recombination (see above), excluded from the above statistics. For machlomovirus, the difference between rsN and rsA was not significant (Table 3, bottom, t-Hotelling = 1.01, P<0.20). On the other hand, in the case of Providence virus (Alphacarmotetravirus), rsN (0.51) was significantly higher than rsA (0.00) (t-Hotelling = 2.94; P<0.005), and positive. Thus, ignoring the recombination event would lead to the erroneous prediction that the replicase is the de novo frame. It would be interesting to determine whether the codon usage of the Providence virus replicase gene is similar to that of its original genome. However, we could not find the species from which the recombination had occurred, since a similarity search based on the nucleotide sequence of the replicase found no similar viral (or cellular) sequence.
We applied the codon usage method defined above to seven pairs of overlapping genes (gathered from the literature), in which both frames have the same phylogenetic distribution. Table 4 presents the codon usage values for these overlaps by decreasing value of t-Hotelling, and the corresponding predictions of ancestry. The codon usage of overlapping frames was significantly different in only two cases (deltaretrovirus Tax/Rex and alphanodavirus replicase/B2). Deltaretrovirus Tax and the betanodavirus replicase, respectively, had a codon usage significantly closer to that of the viral genome than the other frames, suggesting that they are the ancestral frames and that the de novo frames are Rex and B2. We discuss these two overlaps in more depth below (case studies number 1 and 2).
In the five other overlaps analyzed in Table 4, both frames had a comparable codon usage, preventing prediction of the de novo frame.
We examined in more detail the deltaretrovirus genome, which contains a complex pattern of overlapping genes at its 3′ end, in the pX region [36]–[39]. In addition to Tax and Rex, the pX region of Human T-lymphotropic virus 1 (HTLV1) encodes two other proteins in the sense strand, p12 and p30, and a fifth protein, HBZ, from the antisense strand [36], [37], [40] (Figure 3). The resulting arrangement has two long (>80 aa) triple overlaps: the N-terminus of p30 overlaps both p12 and the N-terminus of HBZ, while the C-terminus of p30 overlaps the N-termini of both Tax and Rex (Figure 3). The phylogenetic distribution of the overlapping genes in deltaretroviruses is summarized in Figure 4. P30 is expressed only in HTLV1 [36]. p12 has only been reported in HTLV1 [36], and its coding sequence is interrupted by a stop codon in HTLV2, while it has no equivalent in bovine leukemia virus. HBZ is present in HTLV1 but also in HTLV2, 3 and 4, where it is called respectively APH2, APH3 and APH4 (these proteins have statistically significant similarity with HBZ, indicative of homology). In the bovine leukemia virus genome, no equivalent of HBZ is expressed from the antisense strand in the region between the Env and Tax genes (Luc Willems, personal communication); instead microRNAs are expressed from the sense strand [41], [42]. Taking into account this phylogenetic distribution, and our codon usage predictions, the most likely evolutionary scenario (Figure 4) is that HBZ has originated in the common ancestor of HTLV1 to 4, after its divergence from bovine leukemia virus; p12 has originated de novo in HTLV1 by overprinting HBZ; and p30 has originated de novo in HTLV1 by overprinting both HBZ (in the N-terminus of p30) and Tax and Rex (in the C-terminus of p30). It is not possible to conclude whether p30 or p12 originated first, nor how Tax or HBZ originated (de novo or by horizontal gene transfer).
We made two additional observations regarding codon usage. First, the fact that Tax and Rex are involved in a triple overlap with a short region of p30 (Figure 3) constitutes a potential confusing factor in our prediction of ancestry by codon usage above. Nevertheless, the region of p30 overlapping Tax and Rex has a codon usage significantly more distant to that of the genome than that of Tax (t-Hotelling = 2.16; P<0.025). Therefore, the codon usage of Tax is close to that of the genome over the entire length of its overlapping region, which further suggests that Tax is the ancestral gene. Second, genes expressed from an antisense strand are not expected to have a similar codon usage to genes from the sense strand. Accordingly, the codon usage of HBZ is not correlated to that of the rest of the genome (rs = 0.00 for the entire HBZ gene, rs = 0.06 for the region of HBZ overlapping p30).
The existence of triple overlaps poses severe constraints on the sequence of the proteins encoded by the pX region, and we thus examined whether they had an unusual sequence composition, or were predicted to be structurally disordered [3] (see Material and Methods). We found that all proteins encoded by the pX region, with the exception of Tax, contained long regions with low sequence complexity (as defined by SEG [43]) over most of their length (dashed lines in Figure 3; see Supplementary Table S3), indicating that they were unlikely to adopt a typical globular structure [43], [44]. Tax has no specific region of low sequence complexity, but both its N-terminus, overlapping Rex and p30, and its non-overlapping C-terminus have a highly biased composition, being enriched in hydrophobic residues (P<0.005) and depleted in polar and charged residues (P<0.005).
In addition, HBZ and Rex were predicted to be mostly disordered, at least in the absence of binding partners, while p30 contained several long regions predicted disordered (see Supplementary Table S3). Only p12 and Tax were predicted to be mostly ordered. These results suggest that sequence constraints imposed by triple overlaps may lead to proteins with a highly biased sequence composition and/or structurally disordered [3], and further highlight the fact that Tax is different from the other proteins encoded by the pX region.
Finally, it may seem extraordinary that triple overlaps exist at all, given the sequence constraints they impose; in that light, we note that the relative frame arrangement that would impose the highest constraint, called “−2” [45], is not used for the overlap involving HBZ. (In this arrangement, codon positions 1 and 2 of a frame overlap respectively codon positions 2 and 1 of the antisense frame, with the result that the sequences of each frame are almost fixed by each other). As can be seen in Figure 3, the frame that is in the −2 arrangement relative to HBZ is the non-coding frame 0, rather than the p12 or p30 frames.
In the second case, the codon usage of alphanodavirus B2 (a suppressor of RNA silencing [46]) suggests that it has originated de novo by overprinting the disordered C-terminal extension of the polymerase domain (Table 4). However, several observations cast a doubt on the reliability of this prediction. A similar genomic arrangement occurs in a closely related genus, betanodavirus (though there is no detectable sequence similarity between either the C-terminal extensions of the replicases or the B2 proteins of both genera) (Figure 5). A parsimonious scenario would demand that the overlaps of both genera have the same origin and thus presumably a similar codon usage pattern. Yet this is not the case: in betanodavirus it is B2 that has a codon usage closer to that of the genome (though not significantly so). This discrepancy might be due to horizontal transfer (see below).
Intriguingly, a very similar arrangement occurs in two genera (ilarviruses and cucumoviruses) of another family of positive-strand RNA viruses, Bromoviridae, in which a silencing suppressor called 2b overlaps a C-terminal extension of the polymerase (Figure 5). Like in Nodaviridae, neither the overlapping regions of the replicases nor the 2b proteins of the two genera have any similarity. The codon usage of the 2b proteins of ilarviruses and cucumoviruses is indistinguishable from that of the region of the replicase they overlap (Table 4), making a prediction of ancestry impossible. In fact, the 2b proteins of ilarviruses might have a different origin from those of cucumoviruses, since these genera do not form a monophyletic clade [47]. Despite their similar genomic location, size and function, alphanodavirus B2 and cucumovirus 2b have different structural folds and different modes of binding to RNA, both previously unknown [46], [48]–[50], clearly indicating an independent origin. We think that together, these observations indicate that the overlaps have a complex evolutionary origin; the ancestral protein could differ in each genus (for instance it might be the C-terminal extension of the replicase in alphanodaviruses and the B2 protein in betanodaviruses), and in some genera the ancestral proteins might have entered their genome by horizontal transfer (see below).
We have shown that de novo frames originated by overprinting have a pattern of codon usage distinguishable from ancestral frames, which can be used to predict the de novo frame with a good specificity but intermediate sensitivity (working in around half the cases).
How do our results compare with previous empirical studies of codon usage? The codon usage of six of the overlaps presented here has been studied previously using a different method, the “codon similarity index” [4]. The qualitative trends reported were similar to the ones we observe. Four of the overlaps presented here were also analysed previously, by Pavesi et al [24] who studied their information content and their codon usage. Again, the numerical values they reported for codon usage are in very good agreement with those obtained here, as are their general conclusions. However, our improved statistical analysis allowed us to draw more powerful conclusions. For instance, Pavesi et al reported that both the tymovirus replicase and movement genes had a codon usage correlated with that of their genome, and concluded that it was consequently not possible to determine the ancestral gene [24]. In the present article, the use of Hotelling's t-test to compare two dependent correlation coefficients [51] allowed us to determine that the replicase gene had a codon usage significantly closer to its genome than the movement gene (Table 3), indicating (correctly) that it was the ancestral frame. Another study, on the VP2/VP5 overlap of avibirnavirus (homologous to the aquabirnavirus overlap studied herein), showed that VP5 had an unusual nucleotide usage and predicted that it was the de novo frame [27]. Our quantitative analysis is in agreement with these findings: VP2 has a codon usage significantly closer to the viral genome than VP5 (Table 3). Finally, a previous analysis of the cucumovirus replicase/2b overlap predicted that 2b was the de novo frame, based on its uridine content at the third codon position [22]. In contrast, our analysis detects no statistically significant difference between the codon usage of the overlapping region of the replicase and that of 2b (Table 4).
Why are ancestral and de novo frames distinguishable by their codon usage in only half of the overlaps? We examined in the Results several confounding factors, such as the relative frame of the overlapping regions, their sequence composition, and the estimated age of overlaps. They did not appear to have a significant impact on the predictive power of codon usage analysis. One note of caution is that we used a very crude estimate of the relative ages of overlaps (i.e. their taxonomic distribution). We could not use a more precise estimate, unlike a previous study [4], because our dataset contains both RNA and DNA viruses, which have no protein in common that could be used as a molecular clock, and because the proteins we studied often have more than 50% sequence divergence, preventing the determination of reliable phylogenies.
During the revision of this manuscript, following the suggestion of a reviewer, we tested a distance measure based on information theory approaches: the modified Kullback-Leibler (KL) distance [52]. Since dinucleotide frequency is an important genome signature [53], we have re-analysed our dataset by calculating the KL distance (based on the frequency of the 16 dinucleotides at codon positions 1-2, 2-3, and 3-1) between the ancestral frame and the non-overlapping coding regions of the genome (KLA), and between the de novo frame and the non-overlapping coding regions of the genome (KLN). The ancestral frame had a KL distance to non-overlapping regions lower than that of the de novo frame in 23 out of 25 overlaps. In contrast, in our approach, the rs of the de novo frame was lower than the rs of the ancestral frame in 21 out of 25 overlaps.
We could not evaluate by analytical methods whether the KL distance between the ancestral frame and the non-overlapping regions (KLA) was significantly smaller than that of the KL distance between the novel frame and the non-overlapping regions (KLN), because KL distances are gamma-distributed [52] and there is no generic analytical solution for the distribution of the difference in gamma distributed variables. Therefore, instead, we performed, on each pair of overlapping genes from our dataset, a permutation test to estimate whether the observed (KLN−KLA) was significantly higher than the null distribution of (KLN−KLA) derived from 10,000 random permutations of the nucleotide sequence of the ancestral and the novel frame. We found that, on our dataset, this permutation test on the KL distance has the same specificity as the t-Hotelling test (1.00) and a slightly better sensitivity (0.60) than the t-Hotelling test (0.52), although the performance of the two methods is not significantly different (McNemar chi-square = 0.16; P = 0.70). We hope that our publicly available dataset of overlapping genes with known genealogy (expected to grow) will encourage others to continue exploring these methods and others.
Our new method allowed us to make predictions of ancestry for two overlaps in which both frames have the same phylogenetic distribution (Table 4). In the alphanodavirus replicase/B2 overlap (case study 2), several elements suggest that horizontal transfer might have taken place and thus that the codon usage prediction is not reliable. In the deltaretrovirus Tax/Rex overlap (case study 1), our prediction that Rex has originated de novo by overprinting Tax is consistent with the function of Tax, which occurs upstream of that of Rex in the viral cycle [37], [39], [54]. It is also coherent with the fact that Tax has a higher sequence complexity than Rex or p30, and is under stronger selection pressure than Rex [55], [56], which is generally the case of ancestral frames [3], [4]. Our prediction is in agreement with that of a previous work, reached by analyzing the substitution rates of Tax and Rex, their nucleotide composition and their amino acid composition [55]. Tax and Rex are encoded by the same mRNA but have different start codons [57] and thus Rex presumably originated by the appearance of a new ATG upstream of Tax. Both Tax and Rex are present in all deltaretroviruses and only in those viruses, which suggests that Tax originated first in the common ancestor of deltaretroviruses, and that Rex originated by overprinting it “rapidly” afterwards (by biological timescales), before the divergence of deltaretroviruses. Rex must have then undergone a rapid functionalization, since it is indispensable for the viral cycle, like Tax [37], [39], [54].
An alternative scenario is possible but appears much less parsimonious: Rex might have originated in another organism with a different codon usage, and entered the genome of the ancestor of deltaretroviruses by horizontal transfer. It would then have diverged in sequence beyond recognition, and have been overprinted by Tax (which would have a codon usage similar to that of the genome by coincidence).
The pX region encodes five genes unique to deltaretroviruses [36]–[39], at least three of which (p12, p30 and Rex) have originated de novo, while the two others (Tax and HBZ) have either also originated de novo too (although earlier), or by horizontal transfer (Figure 3). The pX region thus constitutes a hotspot of gene origination, or gene “nursery” [58]. Strikingly, the two genes that have overprinted Tax, Rex and p30, play roles that are respectively complementary and antagonistic to Tax [38], [39], [59], while HBZ plays a role antagonistic to that of Tax [60], [61]. This suggests that the function of Tax was gradually controlled and refined by the appearance of new genes encoded in the same genomic location. Interestingly, other gene nurseries are found in a similar genomic position in other Retroviridae, such as lentiviruses or spumaviruses [62]. As seen above, the 3′ end of the replicase gene of some positive-strand viruses may also favour the origination of gene encoding silencing suppressors (Figure 5).
Such hotspots of origination of genes coding for proteins involved in the same pathways, and combining horizontal transfer and de novo origin, may be common in viruses. For instance, the movement proteins of Alphaflexiviridae and Betaflexiviridae are encoded in the same genomic position (downstream of the replicase gene) but belong either to the Triple Gene Block type [63], [64] or to the 30K type [65], indicating that at least one or possibly both types of movement proteins have entered these families by horizontal transfer [66].
The mechanisms that presumably favour the appearance of gene nurseries are unknown, but obviously of great interest. In the case of the deltaretrovirus pX region, we note that the high constraints imposed by the triple overlaps severely restrict the evolution of p12, p30 and Rex, and that consequently their present-day sequence composition is probably rather similar to the one they had when they first originated. We speculate that the pattern of origination seen in the pX region, in which de novo genes regulate the function of an ancestral protein, may be facilitated by the fact that low sequence complexity (and/or structural disorder) is compatible with a range of regulatory functions [67]–[69]. Thus, at birth, despite having a very “simple” sequence not honed by natural selection, these proteins may have had, by chance, a regulatory function and provided the virus with a fitness advantage that lead to their fixation.
Retroviridae encode numerous short, accessory genes, often accessed by alternative splicing or complex mechanisms leading to partially overlapping genes, and no doubt many remain to be discovered [62]. Yet at the time this article was submitted, almost none of these genes were annotated in the NCBI reference genomes [70]. This poor annotation is prejudicial to the study of these viruses. It is important that researchers who discover, or have discovered new genes, contact the NCBI viral genomes team to ensure that they are annotated properly.
Another, more general implication for genome annotation is that long, triple overlaps may have the potential to yield functional proteins relatively easily. Therefore, triple overlaps might be more abundant than previously thought (we know only two triple overlaps outside of deltaretroviruses, involving the P, V, and D or W proteins in Paramyxovirinae [71]–[74]). We thus recommend re-investigating known overlapping gene pairs to find whether in some cases a third overlapping frame might be expressed.
It has been proposed that the GC content of a genome was the main, though not the only, determinant of codon usage [19]–[21]. Our results confirm that it cannot be the unique determinant, otherwise the de novo and ancestral frames (which have the same GC content) would necessarily have a similar codon usage. Therefore, a significant evolutionary constraint(s) on codon usage must operate in addition to the GC content, and this constraint must be greater on ancestral frames than on de novo frames. Belalov et al. recently reported that the frequency of the dinucleotide CpG was an important factor of viral codon usage, and that CpG was significantly rarer at codon positions 2-3 than at positions 3-1 [75]. CpG is known to be underrepresented in RNA viruses [76], perhaps to avoid recognition from an antiviral CpG sensor [77]. However, the difference in frequency of CpG at different codon positions suggests that a second type of pressure exists, imposed by the translational apparatus. The authors thus suggested the existence of an evolutionary constraint on the genome deriving from a hypothetical cellular CpG sensor coupled (by an unknown mechanism) to the translational machinery. The existence of such a constraint would be coherent with our results, and we speculate that it might cause the difference in codon usage between ancestral and de novo frames.
Very little is known about de novo protein origination, although it is by now clear that this mechanism plays an important role in viral pathogenicity. Our method should allow the identification of more de novo proteins, and thus enhance our understanding of host-pathogen co-evolution. It will be of particular interest to apply it to gene “nurseries” such as the ones we have identified here, and to elucidate the pressures that shape them. Finally, we note that recent experimental and computational reports suggest that de novo origination of genes by overprinting may not be confined to viruses but on the contrary, be a much wider phenomenon than previously thought, both in eukaryotic [78]–[82] and bacterial genomes [83].
We retrieved all sequences from the NCBI viral genome database [84]. We used MAFFT [85] for multiple sequence alignment, HHpred [86] and HHblits [87] for remote homology detection, Phylogeny.fr [88] for phylogenetic analyses, and METAPRDOS [89] for prediction of protein structural disorder, respecting the guidelines of [44]. We used Composition Profiler [90] for analyses of protein global compositional bias with respect to Swiss-Prot (release 51), and SEG for analyses of protein local compositional bias [43]. SEG analyses were obtained from the web server ANNIE [91] with parameters 45/3.75/3.4 in order to identify long regions with a composition bias indicative of non-globular proteins [44].
We searched the NCBI genome database [84] for viruses that infected eukaryotes, with a genome shorter than 30,000 nucleotides, and which contained overlapping genes longer than 120 nucleotides. The cut-off of 30,000 nucleotides was chosen because curation of larger genomes is impractical [3]. We included an overlapping gene into the benchmark dataset only when two criteria were fulfilled: 1) the expression of both overlapping reading frames was experimentally verified; 2) the genealogy of the overlapping reading frames could be determined with good support by using the very stringent criterion described previously [3], regarding the taxonomic distribution of both overlapping frames. According to this criterion, one reading frame can be considered ancestral only if it has homologs in at least two viral families whereas the other, overlapping frame had in at most one viral family. Since viral proteins diverge very fast, identifying viral proteins conserved in at least two families requires powerful similarity search techniques, which are described below. The final dataset, presented in Table 1, contains 27 independent (non-homologous) overlapping genes whose genealogy is reliably established. The dataset contains no antiparallel overlapping genes because we could not find any whose existence had been convincingly proven experimentally in the genomes of short or medium size considered (<30 kb).
We used the following conventions to define the precise boundaries of the overlapping regions on which we performed calculations of codon usage. There are two types of overlaps: in internal overlaps, one overlapping gene is contained entirely within the other gene whereas terminal overlaps involve only the 3′ end of one gene and the 5′ end of another [92]. In the case of internal overlaps, for the longest frame, the first codon counted as overlapping was the most upstream codon that overlaps the start codon of the internal frame, and the last codon counted as overlapping was the most downstream codon that overlaps the stop codon of the internal frame. In the case of terminal overlaps, for the upstream frame, the first codon counted as overlapping was the most upstream codon that overlaps the other frame, and for the downstream frame the last codon counted as overlapping was the most downstream codon that overlaps the stop codon of the other frame.
In order to obtain a highly reliable genealogy of the overlaps, we needed to identify as distant homologs as possible for each protein of the dataset. However, not all homologs of a protein can be detected by conventional sequence similarity searches even if they have retained some sequence identity with the query, for a number of reasons [93], including the fact that databases of protein domains are underrepresented for viruses (our observations). We thus exploited “contextual” information available for viral proteins, such as taxonomy and genome organisation, to identify distant homologs overlooked by conventional searches [94]. We proceeded in the following way (the procedure is the same as in our previous article [3] but had not been described in detail). We first identified “straightforward” homologs of the query protein in the NCBI nr database (release 1st April 2012), by using HHpred [95] and HHblits [87] and selecting hits whose E-value was below the standard cut-off of 10−3. We then examined subsignificant hits (i.e. hits with an E-value superior to 10−3) up to E-values of 1000, looking for viral proteins that came from a taxonomically related virus, and/or occurred in the same position of the genome. Such subsignificant hits, which have weak similarity with the query protein and occur in a similar genomic context, constitute potential homologs. In order to test whether they were actually homologous with the query, we gathered homologs of these subsignificant hits (with E≤10−3), and used HHalign [96] to compare homologs of the query protein (obtained above) with homologs of the subsignificant hits. We considered that an HHalign E-value inferior to 10−3 indicated homology between the subsignificant hit and the query, but performed additional checks, such as verifying that the secondary structure and function (when available) of the hits were compatible with that of the query.
Whenever the structure of a protein from the dataset was available, we also performed structural similarity searches to identify structural homologs, using DALI [97] and FATCAT [98].
Because overlapping genes are not systematically recognised [16], [99] there is a theoretical possibility that some homologs of an overlapping frame might exist in related genomes but not be annotated, and therefore missed by similarity searches. For each overlap, we thus systematically checked that the genomes of other taxonomically related viruses did not contain conserved, unannotated open reading frames, as in [4].
We present in Supplementary Table S1 the taxonomic distribution of the homologs detected by our searches, together with the corresponding PFAM family (or clan) identified in the process.
Genes that have entered their genome by horizontal transfer can be identified by the fact that their phylogeny is discordant with the rest of the genome. A robust measure of this discordance is taxonomic incongruence, e.g. the existence of close homologs in a distant taxon. To detect taxonomic incongruence, we collected homologs of the protein products of each ancestral reading frame using blastp [100] on the Refseq database [101] with a cutoff E-value of 10−3. Hits to proteins from a different viral family than that of the query indicated taxonomic incongruence. To infer the direction of horizontal transfer, we analysed the phylogenetic distribution of homologs of the ancestral protein, both from the same family and from the distant taxon detected, and applied a parsimony criterion: the clade that has the wider phylogenetic distribution of the gene was most likely to be the clade of origin.
In the genetic code, 18 amino acids (aas) are degenerate, e.g. encoded by more than one codon, and they are encoded by 59 “synonymous” codons in total. For each viral genome sequence, we measured the number of occurrences of the 59 synonyms in the non-overlapping coding regions and in each of the two overlapping reading frames (Figure 2). For clarity we will refer to the ensemble of the numbers of occurrences of the 59 synonymous codons of a given reading frame as its “codon usage”. The codon usage of non-overlapping regions will be called the “codon usage of the genome”.
In some overlapping reading frames (generally short, i.e. less than 400 nucleotides), the number of occurrences of the synonymous codons for a given aa was smaller than the degree of degeneracy of this aa (for instance only 3 synonyms for arginine, a 6-fold degenerate aa). In these cases, we restricted the analysis to synonymous codons whose number of occurrences was at least equal to the degree of degeneracy of the encoded aa. We indicated in Table 2 the number of synonymous codons on which the analysis was carried out.
We wanted to utilize codon usage as a method to predict the genealogy of overlapping genes, and not simply to characterise the behaviour of overlapping genes. Therefore, we needed a method to assess whether the codon usage of ancestral frames was closer to the rest of the genome that the codon usage of de novo frames, and to assess whether this difference was statistically significant.
We have examined various canonical methods to evaluate codon usage bias: the Effective number of codons (ENC), [102] Codon Adaptation Index (CAI) [103], and Dmean index [104]. We found that ENC and Dmean had poor predictive power on the genealogy of overlaps (not shown). Initial tests suggested that CAI may have been more sensitive, but we could not easily test the statistical significance of the difference between the observed CAI distances. Therefore, we developed a new method, that had a good predictive power and that could yield estimates of statistical significance.
Our hypothesis was that, in overlapping reading frames, the ancestral frame could be identified by having a codon usage that was more similar to the codon usage of the genome than that of the other frame. Thus we designed a measure of the similarity of codon usage of each frame with that of the genome, and a method to assess whether one frame had a codon usage significantly closer to that of the genome than the other frame.
In order to quantify the similarity between the codon usages of two given reading frames, we used as a measure the Spearman's rank correlation coefficient (rs) [33] between the number of occurrences of the 59 synonymous codons of these two frames (i.e. between their “codon usages”, see above). Each viral genome was divided into three sets: a) the overlapping region of the reading frame 1; b) the overlapping region of the reading frame 2, and c) non-overlapping regions of the genome, composed of the sequences of non-overlapping genes, and, in cases where some genes of the genome partially overlapped, of their non-overlapping regions (Figure 2). For viruses with segmented genomes, all segments were included in the calculations. For simplicity, the codon usage of the third set, i.e. non-overlapping regions, will be referred to as the “codon usage of the genome”. In all viral genomes, we calculated the rs between the codon usage of the genome and that of each of the two overlapping frames under consideration (rs1 and rs2). The reason we collected the non-overlapping coding regions of a virus genome into an integrated set (as opposed to studying individual non-overlapping genes and analyzing their variance) is because the individual non-overlapping genes (or their non-overlapping regions, in cases of genes that partially overlap) are often short, which would have restricted correlation analysis to 2 or 3 dozens of synonyms.
Determining if a given frame “1” has a codon usage closer to that of the genome than the other frame “2” is equivalent to determining whether rs1 is significantly greater than rs2, i.e. whether the correlation between the codon usage of the first frame and that of the genome is significantly greater than the correlation between the codon usage of the second frame and that of the genome. This comparison involves two correlations coefficients that refer to a common variable (the codon usage of the genome), a situation categorized as “dependent correlation” [51] or as the study of “correlated correlation coefficients”, which can be addressed by the Hotelling t-test [34], [35]. The conventional Hotelling formula involves comparing Pearson correlation coefficients rp, but can be used with Spearman's correlation coefficients rs by converting them into their Pearson equivalents: [105].
The Hotelling t-value was calculated as follows:where n is the number of the compared codon frequencies, rp1 and rp2 are respectively the Pearson equivalents of rs1 and rs2, and rp12 is the Pearson equivalent of rs12 (codon usage correlation between the overlapping frames). We assess the Hotelling t-value according to the one-tailed Student's t-test.
The Hotelling's t-test is designed for correlation coefficients estimated from independent data. However, the data we examine in this study (the number of occurrences of synonymous codons) are clearly not independent, since the sum of the numbers of synonymous codons encoding a given aa is fixed. Consider, for example, a reading frame containing 28 Glutamine codons (an aa encoded by two synonyms, CAA and CAG). If the number of occurrences of CAA is 11, that of CAG will inevitably be 17 (i.e. 28−11), i.e. the number of occurrences of CAA and CAG are not independent. Therefore, we assessed whether Hotelling's t-test was robust to non-independence of data by comparing it with a simulation-based exact test. For each pair of overlapping frames of the dataset, we generated two simulated overlapping frames with an aa composition identical to that of the two original frames, and used the actual non-overlapping regions of the genome as a reference set.
One round of simulation was performed as follows: we randomly generated a number (n) of occurrences for each of the 59 codons encoding the 18 degenerate aas, keeping the sum of the occurrences of codons encoding each aa equal to that of the original frame (e.g. if there were 28 Glutamine codons in the original frame, the simulated frame could have any number of CAA and CAG totalling 28). We calculated the correlation coefficients rs1 and rs2 between the number of occurrences of all synonyms in both simulated frames and that of the actual genome. We repeated the same process 10,000 times, thus simulating the distribution of d21 expected assuming that the reading frames are randomly generated and that codon usage is not related to ancestry (i.e. the null distribution). We then tested whether the observed d21 (Table 2) was significantly larger than this null distribution.
Finally, we used the McNemar's non-parametric test [33] to determine whether the Hotelling's t-test and the simulation provide equivalent results (which would indicate that the Hotelling's t-test is robust to non-independence of data).
|
10.1371/journal.ppat.1002143 | The Cost of Virulence: Retarded Growth of Salmonella Typhimurium Cells Expressing Type III Secretion System 1 | Virulence factors generally enhance a pathogen's fitness and thereby foster transmission. However, most studies of pathogen fitness have been performed by averaging the phenotypes over large populations. Here, we have analyzed the fitness costs of virulence factor expression by Salmonella enterica subspecies I serovar Typhimurium in simple culture experiments. The type III secretion system ttss-1, a cardinal virulence factor for eliciting Salmonella diarrhea, is expressed by just a fraction of the S. Typhimurium population, yielding a mixture of cells that either express ttss-1 (TTSS-1+ phenotype) or not (TTSS-1− phenotype). Here, we studied in vitro the TTSS-1+ phenotype at the single cell level using fluorescent protein reporters. The regulator hilA controlled the fraction of TTSS-1+ individuals and their ttss-1 expression level. Strikingly, cells of the TTSS-1+ phenotype grew slower than cells of the TTSS-1− phenotype. The growth retardation was at least partially attributable to the expression of TTSS-1 effector and/or translocon proteins. In spite of this growth penalty, the TTSS-1+ subpopulation increased from <10% to approx. 60% during the late logarithmic growth phase of an LB batch culture. This was attributable to an increasing initiation rate of ttss-1 expression, in response to environmental cues accumulating during this growth phase, as shown by experimental data and mathematical modeling. Finally, hilA and hilD mutants, which form only fast-growing TTSS-1− cells, outcompeted wild type S. Typhimurium in mixed cultures. Our data demonstrated that virulence factor expression imposes a growth penalty in a non-host environment. This raises important questions about compensating mechanisms during host infection which ensure successful propagation of the genotype.
| Pathogenic bacteria require virulence factors to foster growth and survival of the pathogen within the host. Therefore, virulence factor expression is generally assumed to enhance the pathogen's fitness. However, most studies of pathogen fitness have been performed by averaging the phenotypes over large pathogen populations. Here, we have analyzed for the first time the fitness costs of virulence factor expression in a simple in vitro culture experiment using the diarrheal pathogen Salmonella enterica subspecies I serovar Typhimurium (S. Typhimurium). TTSS-1, the cardinal virulence factor for eliciting Salmonella diarrhea, is expressed by just a fraction of the clonal S. Typhimurium population. Surprisingly, time lapse fluorescence microscopy revealed that ttss-1-expressing S. Typhimurium cells grew at a reduced rate. Thus, the pathogen has to “pay” a significant “price” for expressing this virulence factor. This raises important questions about compensating mechanisms (e.g. benefits reaped through TTSS-1 driven host-interactions) ensuring successful propagation of the genotype.
| The ability to infect a host and elicit disease is dictated by the virulence factors expressed by a given pathogen. This may include, but is not limited to, protective factors neutralizing antibacterial defenses, enzymes involved in nutrient acquisition within the host, regulators of virulence factor expression and toxins or secretion systems for subverting host cell signal transduction. The coordinated expression of such virulence factors enhances colonization, growth/survival within the host and transmission. However, most studies of virulence factor function and pathogen fitness have been performed in bulk assays, averaging the phenotypes over large pathogen populations of genetically identical cells. In contrast, little is known about the potential advantages, costs or burdens arising from virulence factor expression by an individual cell of the pathogen population. Therefore, single cell analyses might be of significant interest, in particular if virulence factors, which are expressed in a bistable fashion by some but not all members of a pathogen population, e.g. the ttss-1 system of S. Typhimurium [1], [2], [3], [4], [5], as described in this paper.
Bistable gene expression is genetically encoded. In most cases, one particular genotype expresses one predictable phenotype in a given environment. However, in some cases, two different phenotypes are expressed by isogenic organisms living in the same environment. This is termed phenotypic variation, bimodal gene expression or bistability and represents a special case of gene expression [6]. The importance of bistability for pathogenic bacterial fitness and evolution is just beginning to be understood.
Like other cases of gene expression, bistability is generally observed in response to particular environmental cues. The response is driven by a dedicated (set of) regulator(s), which responds to environmental signals (operon model of Jacob [7]). This response is subject to stochastic fluctuations. In particular in the case of regulators expressed in a few copies per cell, this can significantly affect the active regulator concentration thus randomizing the corresponding phenotype in a population [8], [9]. In combination with non-linear responses (e.g. regulator multimerization, feedback loops), this can lead to formation of phenotypically distinct and stable subpopulations of isogenic bacteria [6], [8], [9], [10], [11]. In terms of evolution, two models may explain the advantage of bistability: i. in “bet hedging”, the optimally adapted phenotype will prevail and ensure the survival of the shared genotype in a changing environment [12]. ii. in “division of labor”, both phenotypes cooperate to ensure survival of the shared genotype [4]. In either way, the bistable expression of certain genes is thought to promote the survival of the genotype. However, it has remained poorly understood whether/how bistability may affect the lifestyle of pathogenic bacteria.
Salmonella enterica subspecies 1 serovar Typhimurium (S. Tm) is a pathogenic Gram-negative bacterium causing numerous cases of diarrhea, worldwide. Its' type III secretion system 1 (TTSS-1) was recently identified as an example for bistable gene expression [1], [3], [5], [13]. TTSS-1 is a well-known virulence determinant of S. Tm required for eliciting diarrheal disease [14], [15], [16]. The needle like TTSS-1 apparatus injects effector proteins into host epithelial cells, thus triggering host cell invasion and pro-inflammatory responses [17], [18], [19]. TTSS-1 is encoded on a genomic island (Salmonella pathogenicity island 1 (SPI-1)), which also harbors genes for effector proteins and for several regulators of ttss-1 expression, e.g. hilA, hilC and hilD [20], [21].
The bistable ttss-1 expression is controlled by a complex regulatory network, which includes coupled positive feedback loops, controls the threshold for ttss-1 induction and amplifies ttss-1 expression [5], [22]. Bistable ttss-1 expression is observed in “ttss-1 inducing” environments, i.e. the gut lumen of infected mice or in non-host environments, e.g. when S. Tm is grown to late logarithmic phase in LB [1], [2], [4], [5]. This yields mixed populations of isogenic S. Tm cells that express ttss-1 (TTSS-1+ phenotype), or do not (TTSS-1− phenotype), in a bimodal fashion. In the mouse gut, only the TTSS-1+ cells can actively invade the mucosal tissue and efficiently trigger inflammation [4], [18]. This inflammatory response may help to overcome the commensal microflora, thus enhancing Salmonella growth and transmission [23], [24], [25], [26], [27], [28], [29]. Experimental data indicate that bistable ttss-1 expression might represent an example of “division of labor” [4], but further data is required to settle this point. At any rate, ttss-1 expression seems to be instrumental for eliciting diarrheal disease and enhancing pathogen transmission. But the functional properties of the TTSS-1+ phenotype are not well understood.
The complex setting of the infected animal gut has hampered the analysis of the TTSS-1+ phenotype. In vitro experiments are essential for gaining detailed mechanistic insights. Here, we have analyzed the induction of ttss-1 expression and its effects on the growth rate of the TTSS-1+ phenotype by single cell reporter assays, competitive growth experiments and mathematical modeling. In such non-host environments, expression of the ttss-1 virulence system expression imposed a growth penalty on the TTSS-1+ cells. This may have important implications with respect to compensatory mechanisms during the infection of animal hosts.
We started our analysis of the TTSS-1+ phenotype by probing ttss-1 expression at the single cell level. For this purpose, we chose the sicA promoter (PsicA), which controls expression of the chromosomal sicAsipBCDA operon (Fig. S1C). This operon encodes key parts of the TTSS-1 virulence system. On the one hand, we employed a transcriptional sipA-tsrvenus reporter gene cassette placing the reporter downstream of the sicAsipBCDA operon (Fig. S1; [2], [3]). Due to its localization at the bacterial poles, the tsrvenus reporter allows detecting <10 proteins per cell [30]. Thus, sipA-tsrvenus provides a highly sensitive reporter for the TTSS-1+ phenotype.
Next, we verified the performance of the sipA-tsrvenus reporter. sipA-tsrvenus expression was bistable and TTSS-1− and TTSS-1+ individuals were distinguishable by the presence/absence of Tsrvenus spots at the bacterial poles ([30]; Fig. 1A; Fig. S1D). TTSS-1 expression and virulence were not compromised (Fig. 1B). The accurate response of sipA-tsrvenus to Salmonella signaling cascades was established by disturbing known elements of the TTSS-1 gene regulation network and FACS analysis of sipA-tsrvenus expression (Fig. 1C, D). In line with the published work on ttss-1 regulation (Fig. 1D): i. Over-expression of positive TTSS-1 regulators increased the abundance of tsrvenus-expressing individuals (Fig. 1C; Fig. S1D). In particular, hilA, hilC and hilD over-expression increased the fraction of sipA-tsrvenus expressing individuals from ∼20% to 80–100%. ii. The median signal intensity per sipA-tsrvenus expressing cell increased when positive regulators were over-expressed (philA: 3.8±0.3-fold; philC: 4.0±0.1-fold; philD: 4±0.1-fold; median ± s.d.). iii. Control experiments in a ΔhilA mutant verified that expression of the TTSS-1+ phenotype depended on the ttss-1 master-regulator, HilA (Fig. 1C; open bars) and iv. The average HilA protein levels of the analyzed strains correlated positively with the fraction of tsrvenus-expressing individuals (r2 = 0.78; quantitative Western blot; Fig. 1E). These data verified the accurate performance of the sipA-tsrvenus reporter and demonstrated that hilA-dependent regulation affects both, the fraction of TTSS-1+ individuals and the level of ttss-1 expression per cell.
In addition, we employed psicA-gfp, a reporter plasmid expressing gfp under control of the sicA promoter. This construct yielded brighter fluorescence than the chromosomal sipA-tsrvenus and was better suited for FACS analysis. Again, this reporter yielded a bistable expression pattern (Fig. 1F). Using wt S. Tm psicA-gfp we separated TTSS-1+ and TTSS-1- subpopulations by FACS. Western blot analysis of the FACS-sorted subpopulations verified coincident expression of psicA-gfp and the TTSS-1 protein SipC (Fig. 1F, G). This indicated that our fluorescent reporter constructs are faithful reporters of the bistable expression of the TTSS-1+ phenotype.
During our experiments, we observed that hilA, hilC and hilD over-expression led to reduced culture densities (e.g. OD600 for wt sipA-tsrvenus: 3.4±0.3 vs. wt sipA-tsrvenus philA: 2.0±0.3; mean ± s.d.). This was a first hint suggesting that retarded growth might be a general feature of the TTSS-1+ phenotype. However, it remained to be shown whether growth retardation occurs in wild type cells expressing normal levels of hilA, hilC and hilD.
The growth rate of the TTSS-1+ individuals was analyzed by time-lapse microscopy. Wild type S. Tm harboring gfp- or tsrvenus-reporters for ttss-1 expression were placed on an agar pad (LB, 1.5% agarose), the TTSS-1+ individuals were identified by fluorescence microscopy and growth was analyzed by time-lapse phase contrast microscopy (1 frame/30 min; Fig. 2A). Imaging did not impose detectable photo damage to the bacteria, as indicated by the unaltered growth rate (Fig. S2). Strikingly, TTSS-1+ individuals grew slower than TTSS-1− individuals (wt S. Tm sipA-tsrvenus (M2001); µT1+ = 0.90 h−1 vs. µT1− = 1.30 h−1; p = 0.027 for the factor ‘phenotype’ in a two-way ANOVA; Fig. 2B). The negative control strain ΔhilA sipA-tsrvenus yielded only TTSS-1− individuals, which grew at the “fast” rate (µT1− = 1.16 h−1; Fig. 2B). Thus, TTSS-1+ individuals seemed to grow at a reduced rate.
To exclude potential artifacts attributable to the sipA-tsrvenus reporter, we analyzed unmodified wild type S. Tm not harboring any reporter (Fig. 2C; Fig. S2). Using a maximum likelihood approach, we identified two populations with distinct growth rates (likelihood ratio test for two populations versus one population, p<0.001, µslow = 0.66 h−1 vs. µfast = 1.27 h−1; Fig. 2C), very similar to the ones described above (Fig. 2B). Furthermore, unmarked mutants lacking the entire SPI-1 region (Δspi-1) or the positive ttss-1 regulator hilD yielded exclusively fast growing cells, while deletion of the negative ttss-1 regulator hilE yielded only slow growing cells (Fig. 2C). Finally, wild type S. Tm harboring psicA-gfp or a chromosomal gfp-reporter for the TTSS-1 gene prgH [1] yielded slow growing TTSS-1+ and fast growing TTSS-1− cells (µT1+ = 0.51 h−1 vs. µT1− = 1.2 h−1; p = 0.006 for the factor ‘phenotype’ in a two-way ANOVA; Fig. 2D). Bacteria expressing the psicA-gfp or prgH-gfp reporters grew even slower than the TTSS-1− sipA-tsrvenus bacteria or the slow-growing wt S. Tm subpopulation (Fig. 2BC). Presumably, this was attributable to the additional “burden” conferred by the GFP expression, as described, before [31].
Thus, the time-lapse microscopy experiments verified bistable ttss-1 expression and revealed that the TTSS-1− phenotype has a reduced growth rate, even at wild type HilA and TTSS-1 levels (µT1+ in the range of 0.7 h−1 vs. µT1− in the range of 1.3 h−1). This was confirmed in a dye dilution assay (Fig. S3).
Our data suggested that ttss-1 expression represents a “cost” to the bacterial cell. However the mechanism explaining this growth retardation had remained unclear. We speculated that expression of the TTS apparatus itself or the sheer load of the proteins transported by the TTSS-1 (effectors, translocon proteins) might play a role. To test these hypotheses, we analyzed two additional S. Tm mutants. In the first mutant, termed Δprg-orgΔinv-spa, we deleted most apparatus-encoding genes (Table S1). This mutant formed two populations with distinct growth rates (likelihood ratio test for two populations versus one population, p<0.001, µslow = 0.72 h−1 vs. µfast = 1.36 h−1; Fig. 2E), very similar to those described for wild type S. Tm (Fig. 2C). The second mutant, termed Δ8Δsip, was lacking the genes for most TTSS-1 effector proteins and the secreted translocon components including sipB, sipC, sipD, sipA, sptP, sopE, sopE2, sopB and sopA (Tab. S1). In contrast to wild type S. Tm, we could not distinguish two subpopulations in this mutant (likelihood ratio test for two populations versus one population, p = 0.73; Fig. 2E). Instead, this mutant displayed a median growth rate of µ = 1.10 h−1, similar to the fast growing subpopulation of S. Tm wt and the mutants Δspi-1 and hilD (Fig. 2C). This data suggests, that expression of the effector proteins and translocon components is “costly” and provides at least in part a mechanistic explanation for the growth retardation of wild type S. Tm cells of the TTSS-1+ phenotype.
When monitoring growth and bistable ttss-1 expression in a wt S. Tm (psicA-gfp) culture, the fraction of TTSS-1+ individuals began to rise after 2.5 h as soon as the culture entered the late logarithmic phase, increased in a linear fashion, and reached approx. 60% after 7 h once the culture entered the stationary phase (Fig. 3A).
Our results implied that two different parameters affect the fraction of TTSS-1+ individuals and the overall growth progression in the late logarithmic phase: i. Competitive growth. TTSS-1+ individuals are steadily outgrown by the fast-growing TTSS-1− individuals (µT1+<µT1−; Fig. 2); this constantly reduces the size of the TTSS-1+ subpopulation. ii. ttss-1 induction. Presumably, initiation of ttss-1 expression in TTSS-1− individuals compensates the “TTSS-1+ losses” attributable to competitive growth and explains the increasing fractions of TTSS-1+ individuals during the late logarithmic phase.
To infer the dynamic initiation rate ri of ttss-1 expression in the late logarithmic phase from our experimental data, we devised a mathematical model describing the growth of the TTSS-1+ (NT1+; growth rate µT1+) and the TTSS-1− population (NT1−; growth rate µT1−) as a function of time (t): (1)(2)
It should be noted that the model does not include a term for “switching off” ttss-1 expression. This was justified by our failure to observe “off switching” in the experiments shown in Fig. 2 and further supported by other data (Fig. S2 and data shown below). During the late logarithmic phase, the relative abundance of the TTSS-1+ individuals increased, and the fraction α of TTSS-1− individuals (NT1-) decreased in a linear fashion (Fig. 3A):(3)
Equation (2) can be rearranged to calculate ri(t) (see Text S1 for details):(4)
With the data from Fig. 3A and by using equation (3) we could determine NT1− (t) and, after fitting an empirical function to NT1− (t), also dNT1−/dt. Using equation (4), this allowed calculating ri(t) during the late logarithmic phase (see Text S1 for details). We found that the mean initiation rate (ri) of ttss-1 expression increased continuously during the late logarithmic phase, e.g. from 0.28 h−1 at 3.5 h to 0.54 h−1 at 5.5 h (SEM = 0.03 h−1; Fig. 3B).
The initiation rate of ttss-1 expression seemed to increase upon entry into the late logarithmic growth phase (Fig. 3A). Therefore, it might be induced by growth-related environmental signals (e.g. oxygen depletion, quorum signals, nutrient depletion, metabolite accumulation). To address this, we analyzed the partial oxygen pressure (pO2) during growth. As expected, pO2 declined to <30% relative aeration during the first three hours (Fig. 3C). After approximately 3.5 h, we detected a transient rebound of the oxygen pressure followed by a steady decline to <3% relative aeration during the next hour. This undulation of oxygen pressure is indicative of a change in the growth physiology at 3.5 h and was in line with the reduced growth rate (Fig. 3A, shaded area).
The data suggested that altered metabolism, nutrient availability, waste product accumulation, the reduced growth rate or the low oxygen pressure might represent cues inducing ttss-1 expression. As a first approach to test the role of pO2, we performed batch culture growth experiments in identical 250 ml culture flasks filled with the indicated volumes of media (wt S. Tm psicA gfp grown in 5, 10, 25, 50 or 100 ml LB; Fig. 3D). This setup allowed analyzing the effect of reduced pO2 (i.e. in larger, poorly aerated culture volumes) at equivalent growth rates. We observed that the fraction of ttss-1 expressing cells increased in larger culture volumes. Therefore, low oxygen tension might represent one environmental cue directly or indirectly inducing bistable ttss-1 expression. However, the evidence is merely circumstantial at this moment and other cues might well be involved. Identification of these cues will benefit from the strategies for determining ri as described above.
In liquid culture, the initiation of ttss-1 expression occurred in the late logarithmic phase. However, our initial time lapse microscopy data for bacteria sampled from this growth phase did not show initiation of ttss-1 expression (Fig. 2). We reasoned that this might be attributable to the lack of inducing environmental signals, as these experiments had been performed on agar pads soaked with fresh LB medium. To test this hypothesis, we modified the time lapse microscopy experiment and imaged bacteria (S. Tm psicA-gfp) placed on agar pads soaked with filter-sterilized spent medium taken from a culture at the same growth phase (OD600 = 0.9, see Materials and Methods). We analyzed growth of 191 micro colonies. At the beginning, 135 did not express ttss-1. But remarkably, we observed 15 of 135 initially TTSS-1− micro colonies, in which individual bacteria induced ttss-1 expression during the course of our imaging experiment (e.g. Fig. 4A, Fig. S4; Video S1). After induction, the TTSS-1+ cells grew at a slower rate than their TTSS-1− siblings. In addition, we observed numerous TTSS-1+ bacteria (56 micro colonies) and TTSS-1− bacteria (120 micro colonies) which did not “switch” their ttss-1 expression status. In line with the results above, ttss-1 expression and the interval between two cell divisions was negatively correlated (Fig. 4A,B,C, Spearman's rho = −0.747, p<0.0001, N = 29).
These experiments support the stochastic initiation of ttss-1 expression. But the initiation rate of ttss-1 expression (<0.04 h−1) was lower than that predicted from the batch culture experiment shown in Fig. 3 (ri = 0.18−0.45 h−1). This might be attributable to the lack of some environmental cue, e.g. low oxygen pressure, as time lapse microscopy was performed at ambient atmosphere. Only two micro colonies showed a decrease in fluorescence as expected for “off-switching”. Hence, the rate of off-switching is not substantial. This indicated that our mathematical model, which assumed that “switching off” the ttss-1 expression would be negligible, was justified (equation (1) did not include ri(t)NT1-(t)). These experiments verified that ttss-1 expression is initiated in a stochastic fashion under “inducing” environmental conditions and that the TTSS-1+ phenotype exhibits a growth defect.
Finally, we confirmed the growth penalty attributable to ttss-1 expression in the late logarithmic phase in competition experiments. Wt S. Tm expresses ttss-1 in a bistable fashion and forms a significant fraction of slow-growing TTSS-1+ cells during the late logarithmic phase (Fig. 3). This slows down the apparent growth of the total wild type population (see above). In contrast, hilA or hilD mutants, which do not express ttss-1, yield a pure population of fast-growing TTSS-1− cells (Figs. 1 and 2). Thus, in a mixed culture, hilA or hilD mutants should outgrow wt S. Tm. Indeed, both mutants out-competed the wt strain during the late logarithmic phase of the mixed culture (ΔhilA, ΔhilD; Fig. 5A,B). In contrast, a hilE mutant, which forms a larger fraction of TTSS-1+ cells than wt S. Tm (Fig. 2), was outcompeted by wt S. Tm in this type of assay (ΔhilE, Fig. 5C). This verified the growth penalty of TTSS-1+ cells in LB batch cultures.
The effect of virulence factor expression on the fitness of an individual pathogen cell has remained unclear. We have analyzed the fitness costs associated with the expression of ttss-1, which encodes a key virulence function of S. Tm. An in vitro system was chosen for a detailed analysis of the growth phenotype of TTSS-1+ cells. We found that these cells have a reduced growth rate. This established that ttss-1 expression represents a burden (and not an advantage) at the level of the individual cell, at least in the non-host environment of our assay system. The growth penalty affects the fraction of TTSS-1+ individuals and the overall growth progression in a S. Tm culture. Mathematical modeling and experimental data demonstrated that this growth penalty and an increasing initiation rate of ttss-1 expression during the late logarithmic growth phase were sufficient to explain the dynamic abundance of TTSS-1+ and TTSS-1− individuals in a clonal S. Tm batch culture.
Evidence for bistability of ttss-1 expression has only recently been accumulated. Under inducing conditions, single cell reporters for expression of ttss-1 or effector proteins yielded cells in the “on” and cells in the “off” state [1], [2], [3], [5], [32]. The regulatory network controlling ttss-1 expression includes at least three positive feedback loops and this architecture is thought to set the threshold for initiating ttss-1 expression and to amplify the level of expression [5], [32], [33]. The TTSS-1+ phenotype can persist for several hours, even if the bacteria are shifted into environments normally not inducing ttss-1 expression (histeresis; shift to fresh LB, Fig. 2; Fig. S2). However, it should also be noted that it has not been possible to define unequivocally where stochasticity is introduced. In fact, stochastic initiation of ttss-1 expression might hinge on different regulators in different environments.
TTSS-1+ cells have at least two important characteristics. First, they express the virulence factors enabling host manipulation and elicitation of disease [13], [17], [18]. Second, as we have found here, they grow at a reduced rate. ttss-1 expression may represent a “burden” in itself. The mechanism explaining the growth defect of TTSS-1+ cells is of significant interest. A partial disruption of the proton gradient by “leaky” TTSS assembly-intermediates and/or the metabolic energy required for biosynthesis of the TTSS may offer plausible explanations. Typical TTSS-1+ cells are estimated to express 20–200 TTS apparatuses and approx. 3−10×104 effector proteins, amounting to a significant fraction of the total cellular protein [2], [3]. Indeed, deleting the translocon and most effector proteins significantly increased the growth rate of the TTSS-1+ cells (Δ8Δsip; Fig. 2E), indicating that these proteins account at least in part for the cost of ttss-1 expression. However, the growth rate of Δ8Δsip (µ = 1.10 h−1) was still lower than that of the TTSS-1− subpopulation of wt S. Tm (µfast = 1.27 h−1), suggesting that other factors do also contribute to growth retardation.
An alternative explanation for the reduced growth rate of TTSS-1+ cells might reside in coordinated expression of a complex regulon. This might be reminiscent of the prf virulence regulon of Listeria monocytogenes, which coordinates metabolism and virulence gene expression thus controlling environment-specific fitness phenotypes in vitro and in vivo [34]. Several global regulators (e.g. crp, mlc, fur; [7], [35], [36]) and silencing proteins (hns, hha; [37], [38]) can control ttss-1 expression. Moreover, HilA may control multiple loci apart from ttss-1 (25). And we have observed co-expression of ttss-1 and of fliC, which encodes a key structural component of the flagella, in the late logarithmic phase (Fig. S5). Accordingly, ttss-1 expression might be one feature of a “differentiated” state which also includes adaptations reducing the growth rate. It is tempting to speculate that this state might be particularly adapted for mucosal tissue invasion. This would be an important topic for future research.
Interestingly, similar phenomena have been observed in other ttss-expressing pathogens. In Pseudomonas aeruginosa, growth in suboptimal media was shown to result in bistable ttss expression [39]. But it remained unclear whether growth might be affected. In contrast, the plasmid-encoded TTSS of Yersinia spp. is well known to cause growth retardation in response to host cell contact or low calcium environments [40], [41]. However, in this case, ttss induction seems to be uniform even in suboptimal media [42]. Thus, bistability and growth retardation do occur in other ttss expressing bacteria, but specific adaptations may exist for each pathogen.
Which environmental cues induce ttss-1 expression in S. Tm? ttss-1 is expressed in the lumen of the host's intestine and in the late logarithmic phase in LB-batch culture. Low oxygen pressure is common to both environments and may represent an inducing signal (see Fig. 3C). In line with this hypothesis, Shigella flexneri, a closely related gut pathogen, can modulate the activity of its TTSS in response to low oxygen pressures typically observed at the gut wall [43]. Similarly, HilA-mediated ttss-1 expression is known to respond to oxygen pressure [21], [44]. In addition, numerous other internal and external cues are known to affect ttss-1 expression, including osmolarity, pH, growth rate, or the presence of short chain fatty acids like acetate [45], [46], [47], [48], [49], [50], [51]. The sum of these environmental cues seems to determine the level of ttss-1 induction. This might explain our observation of a low, but detectable initiation rate of ttss-1 expression on agar pads soaked with spent medium (Fig. 4). This environment should harbor most cues present in the late log culture medium, but lacks low oxygen pressure, which could not be established in the real time microscopy setup.
In summary, our findings indicate that the TTSS-1+ phenotype is more complex than previously anticipated. Currently, we can only speculate how this affects the real infection and transmission in vivo. Our results suggest that the TTSS-1+ subpopulation is constantly drained by the burdens inflicted by immune defenses within the infected gut mucosa [4] and by the reduced growth rate (this work). The latter should represent a competitive disadvantage against all other bacteria (commensals and TTSS-1− S. Tm cells) present in the gut lumen. Moreover, this burden should materialize even before invading the gut tissue and may explain why ttss-1 defective mutants are sometimes (though rarely) found in infected animal flocks and isolated in one case of a human outbreak [52], [53]. In order to explain the evolution and mainentance of bistable ttss-1 expression and the successful propagation of the ttss-1 genotype, one has to predict that the TTSS-1+ phenotype must confer some type of advantage. According to the “division of labor” model, the advantage might emanate from a “public good”, i.e. the TTSS-1 induced gut inflammation fostering Salmonella growth in the gut lumen and enhancing transmission. Alternatively, the TTSS-1+ phenotype might include (unidentified) features enhancing the survival and growth of the ttss-1 expressing bacteria themselves, e.g. in permissive niches of the host's intestine or by enhancing the chances of chronic infection and long-term shedding. Identifying these mechanisms will represent an important step for understanding the evolution of bistable ttss-1 expression.
All strains were derivatives of Salmonella Typhimurium SL1344 or ATCC14028 (see Tab. S1 and Text S2 for references). All plasmids and primers are shown in Tab. S2 and S3. Bacteria were inoculated (1∶100 in LB) from 12 h overnight cultures (LB, supplemented with the appropriate antibiotics) and grown under mild aeration for 4 h at 37°C, if not stated otherwise. In Fig. 1C,E, the medium included 0.01% arabinose.
The mutants were constructed using the lambda red recombination system [54]. The chloramphenicol or kanamycin resistance cassette of pKD3 (cat) resp. pKD4 (aphT) were amplified by PCR using the primer pairs ÄhilA::kan-fw and ÄhilA::kan-rev, ÄhilD::kan-fw and ÄhilD::kan-rev, ÄhilE::cat-fw and ÄhilE::cat-rev and electroporated into SL1344 harboring pKD46 to generate the regulator mutants M2005 (ÄhilA::aphT), M2007 (ÄhilD::aphT) and M2008 (ÄhilE::cat). Mutants were selected by plating on LB-Agar (50 µg/ml kanamycin or 30 µg/ml chloramphenicol). M2072 (termed Δprg-orgΔinv-spa in this paper) was also generated using the lambda red system using the primers invG-fw and spaS-rev as well as prgH-fw and orgC–rev and the plasmids pKD3 and pKD4 to generate prgHIJKorgABC::aphT, invGEABCIJspaOPQRS::cat, a mutant lacking most genes of the TTS apparatus. For construction of strain M2532 (termed Δ8Δsip in this paper), we transduced the ÄsipBCDA-sptP::aphT allele from SB245 (SL1344, ÄsipBCDA-sptP::aphT fliGHI::Tn10; K. Kaniga and J. E. Galan, unpublished data) via P22 into M2400 (SL1344, ÄsopE, ÄsopE2, ÄsopB, ÄsipA, ÄsptP, ÄsopA, ÄspvB, ÄspvC), which has been previously described [55]. M2532 fails to express most TTSS-1 effector proteins and the translocon components.
To create the suicide plasmid pM2002, pVS152Tsr [30] was digested with the restriction endonucleases Eco47III and XmaI. The tsrvenus encoding fragment was ligated into pM1300 (digested with MslI and XmaI, [56]) downstream of a truncated sipA fragment (nt 1156–2058 of the orf), to finally create pM2002 and introduced by homologous recombination into the genome of ATCC14028 to generate the reporter strain M2001. To obtain the tsrvenus reporter for hilA (M2076), the c-terminal region of hilA (nt 114 to 1661 of the orf) was amplified using the primer pair hilA-fw-XmaI-NcoI and hilA-rev-NheI-XbaI and cloned into pBluescriptII (Invitrogen) using the restriction endonucleases XmaI and XbaI, yielding pM2090. This plasmid was digested with NheI and NotI to introduce the tsrvenus encoding PCR fragment (template pM2002, primers: venus-NheI-fw and venus-NotI-rev, digested with NheI and NotI) to obtain pM2095. The entire region ranging from hilA to tsrvenus was cloned into pSB377 using the restriction enzymes NotI and XmaI yielding the suicide plasmid pM2080. This plasmid was used to generate the hilA reporter strain M2076 by homologous recombination into the genome of ATCC14028. To obtain the tsrvenus reporter for fliC, tsrvenus was amplified by PCR (primers: tsr-XmaI-fw and venus-XbaI-rev) and cloned into pBluescriptII using XmaI and XbaI thus yielding pM2533. After amplification of fliC by PCR using SL1344 chromosomal DNA as template and primers fliC-XhoI-fw and fliC-HindIII-rev, the fliC encoding fragment was cloned via XhoI and HindIII upstream of the tsrvenus gene into pM2533, thus yielding pM2539. Subsequently, the construct was moved via XhoI and XbaI into the suicide plasmid pGP704, thus yielding pM2819. This plasmid was used to create the fliC-tsrvenus reporter strain M2821 by homologous recombination into the genome of SL1344.
All over-expression plasmids from pM2010 to pM2042 were obtained by digesting the indicated PCR fragments (Table S2 and S3 for plasmids and primers) with EcoRI and XbaI into pBAD24.
All mutations were verified by PCR or DNA sequencing.
HilA expression was analyzed by quantitative Western blot using an affinity-purified rabbit α-HilA antiserum (Fig. 1E). Recombinant HilA was used for normalization. SipC was detected using an α-SipC serum (Fig. 1G).
For invasion, MDCK cells were grown in MEM (Invitrogen), infected for 30 min (MOI = 5; [57], washed and incubated in MEM (400 µg/ml gentamicin; 1 h). Intracellular bacteria were enumerated by plating.
Prior to analysis, fluorophore formation was ensured (2 h, RT, 30 µg/ml chloramphenicol). Tsrvenus and Gfp emission was analyzed at 530 nm (supplement; FACSCalibur 4-color, Becton Dickinson). Bacteria were identified by side scatter (SSC). Data were analyzed with FlowJo software (Tree Star, Inc.). For Tsrvenus (Fig. 1), ln-transformed fluorescence values for 40000 events were median-normalized (subtraction) and compared to the similarly normalized data from the reporterless control strain, thus yielding the fraction of TTSS-1+ individuals. For sorting bacterial cells, S. Tm (psicA-gfp) cells were sorted by FACS (Aria Becton Dickinson, FACSDiva Software).
Bacteria were placed on a 1.5% agarose pad equilibrated with LB, sealed under a glass coverslip and mounted (37°C temp. control; Axioplan2; Plan-APOCHROMAT 63x/1.4 oil; Zeiss or IX81, UPlanFLN 100x/1.3 Oil, Olympus). Reporter fluorescence (Exc. 470/20 nm; BP 495 nm; Em. 505–530 nm) and micro colony growth (phase contrast) were monitored and evaluated using Axiovision software (Zeiss). The slope of the ln-tranformed bacterial numbers (t), as determined from the logarithmic growth phase, yielded the growth rate µ. For sipA-tsrvenus and prgH-gfp, the micro colonies were scored visually as TTSS-1+ or TTSS-1−. To analyze differences in growth rates between TTSS-1+ and TTSS-1− micro colonies, we performed a full-factorial analysis of variance with the two factors phenotype (fixed) and experiment (random). Variance was analyzed in SPSS 17.0 (SPSS Inc. - Chicago, IL).
Growth rates w/o reporter were analyzed via a maximum likelihood approach to test for two subpopulations with different growth rates. The growth rate measurements from five independent experiments (87 micro colonies) were combined. Using maximum likelihood, we fitted a bi-modal distribution (the sum of two normal probability density functions) and a unimodal (normal) distribution, and compared the two fits with a likelihood ratio test using R software [58].
In Fig. 4, cell growth and ttss-1 expression were analyzed using a modified version of the cell tracking software described in [9]. The first cell in each micro colony that could be observed over a whole division was used to analyze the statistical association between ttss-1 expression and the interval between two divisions (by non-parametric correlation analysis using PASW Statistics 18.0.0). 157 micro colonies were analyzed to estimate the fraction of micro colonies in which all cells, none of the cells, and a fraction of the cells expressed ttss-1. These groupings were based on visual inspection of each micro colony.
|
10.1371/journal.pgen.1002142 | A Two-Stage Meta-Analysis Identifies Several New Loci for Parkinson's Disease | A previous genome-wide association (GWA) meta-analysis of 12,386 PD cases and 21,026 controls conducted by the International Parkinson's Disease Genomics Consortium (IPDGC) discovered or confirmed 11 Parkinson's disease (PD) loci. This first analysis of the two-stage IPDGC study focused on the set of loci that passed genome-wide significance in the first stage GWA scan. However, the second stage genotyping array, the ImmunoChip, included a larger set of 1,920 SNPs selected on the basis of the GWA analysis. Here, we analyzed this set of 1,920 SNPs, and we identified five additional PD risk loci (combined p<5×10−10, PARK16/1q32, STX1B/16p11, FGF20/8p22, STBD1/4q21, and GPNMB/7p15). Two of these five loci have been suggested by previous association studies (PARK16/1q32, FGF20/8p22), and this study provides further support for these findings. Using a dataset of post-mortem brain samples assayed for gene expression (n = 399) and methylation (n = 292), we identified methylation and expression changes associated with PD risk variants in PARK16/1q32, GPNMB/7p15, and STX1B/16p11 loci, hence suggesting potential molecular mechanisms and candidate genes at these risk loci.
| This paper describes the largest case-control analysis of Parkinson's disease to date, with a combined sample set of over 12,000 cases and 21,000 controls. After combining our findings with an independent replication dataset of more than 3,000 cases and 29,000 controls, we found five additional PD risk loci in addition to the 11 loci previously identified in earlier consortium efforts. This successful study further demonstrates the power of the GWA scan experimental design to find new loci contributing to disease risk, even in the context of complex disorders like Parkinson's disease. These new findings provide insights into the etiology of PD and will promote a better understanding of its pathogenesis.
| Until the recent developments of high throughput genotyping and genome-wide association (GWA) studies, little was known of the genetics of typical Parkinson's disease (PD). Studies of the genetic basis of familial forms of PD first identified rare highly penetrant mutations in LRKK2 [1], [2], PINK1 [3], SNCA [4], PARK2 [5] and PARK7 [6]. Following these findings, GWA scans for idiopathic PD identified SNCA and MAPT as unequivocal risk loci [7], [8], [9], [10], [11] as well as implicated BST1 [8], GAK [12], and HLA-DR [13]. Using sequence based imputation methods [14], the meta-analysis of several GWA scans [7], [9], [10], [11] conducted by the International Parkinson's Disease Genomics Consortium (IPDGC) identified and replicated five new loci: ACMSD, STK39, MCCC1/LAMP3, SYT11, and CCDC62/HIP1R [15] and confirmed association at SNCA, LRRK2, MAPT, BST1, GAK and HLA-DR [15].
We conducted a two-stage association study. Combining stage 1 and stage 2, the data consist of 12,386 PD cases and 21,026 controls genotyped using a variety of platforms (Table 1). Stage 1 used genome-wide genotyping arrays and our initial analysis [15] focused on the subset of SNPs that passed genome-wide significance in stage 1. For stage 2 genotyping, we used a custom content Illumina iSelect array, the ImmunoChip and additional GWAS typing as previously described [15]. The primary content of the ImmunoChip data focuses on autoimmune disorders but, as part of a collaborative agreement with the Wellcome Trust Case Control Consortium 2, we included 1,920 ImmunoChip SNPs on the basis of the stage 1 GWA PD results.
Here, we report the combined analysis for this full set of 1,920 SNPs. This step1+2 analysis identified seven new loci that passed genome-wide significance in the meta-analysis. During the process of analyzing these data and preparing for publication, we became aware that another group was also preparing a large independent GWA scan in PD for publication (Do et al, submitted). Following discussion with this group we agreed to cross validate the top hits from each study by exchanging summary statistics for this small number of loci.
To provide further insights into the molecular function of these associated variants, we tested risk alleles at these loci for correlation with the expression of physically close gene (expression quantitative trait locus, eQTL) and the methylation status (methQTL) of proximal DNA CpG sites in a dataset of 399 control frontal cortex and cerebellar tissue samples extracted post-mortem from individuals without a history of neurological disorders.
In addition to eleven loci that passed genome-wide significance in stage 1 [15], we identified over 100 regions of interest defined as 10 kb windows containing at least one SNP associated at p<10−3. We submitted the most associated SNP in each region for probe design and follow-up genotyping using the ImmunoChip platform. For each region of interest, we also added four SNPs in high level of linkage disequilibrium (LD) to provide redundancy where the most associated SNP would not pass the Illumina probe design step or the assay for that SNP would fail. To complete the array design we also added all non-synonymous dbSNPs located in known PD associated regions [1], [2], [3], [4], [5], [6]. Out of these 2,400 submitted SNPs, 1,920 passed QC and were included in the final array design. For these 1,920 SNPs we combined stage 1 and stage 2 associated data in a meta-analysis of 12,386 cases and 21,026 controls (Table 1) from the IPDGC. We exchanged summary statistics for these most significant hits with an additional large, case-control replication dataset (3,426 PD cases and 29,624 controls) in an attempt to demonstrate independent replication.
On the basis of stage 1+2 results, seven new SNPs passed our defined genome-wide significance threshold (p<5×10−8, Table 2 and Figure 1). These loci are either novel or the previous evidence of association was not entirely convincing in individuals of European descent. We combined these results with the independent replication. Five of these seven loci replicated and showed strong combined evidence of PD association (p<10−10 overall). Taking either the nearest gene (or the strongest candidate when available) to designate these regions, these five loci are 1q32/PARK16 [7], 4q21/STBD1, 7p15/GPNMB, 8p22/FGF20 [16] and 16p11/STX1B.
rs708723/1q32 has been previously reported as PD associated (PARK16, [7], [8]) but this SNP lacked the unequivocal evidence of association in European samples (p = 9.47×10−10 in stage 2 only). To understand the potential biological consequences of risk variation at this locus we tested whether rs708723 was correlated with either gene expression or DNA methylation status of proximal transcripts or CpG sites respectively (Table 3). We found correlations with the expression of NUCKS1 (p = 1.8×10−7) and RAB7L1 (p = 7.2×10−4). We also found correlations with the methylation state of CpG sites located in the FLJ3269 gene (p = 3.9×10−22).
In the case of 16p11/STX1B, the proximal gene to the most associated SNP rs4889603 is SETD1A. However, STX1B is located 18 kb upstream of rs4889603 and is a more plausible PD candidate gene [17] owing to its synaptic receptor function. We therefore used this gene to designate this region. Our methQTL/eQTL dataset identified a correlation between the rs4889603 risk allele and increased methylation of a CpG dinucleotide in STX1B (Table 3).
The SNP rs591323 in the 8p22 region is located ∼150 kb downstream of the FGF20 gene (NCBI build 36.3), for which association with PD has been suggested previously in familial PD samples [16], [18] but which remained controversial [19]. Our findings provide further support for a PD association at this locus, but again, whether the functionally affected transcript is FGF20 or not remains unclear.
The regions 4q21/STBD1 and 7p15/GPNMD have not been previously implicated in PD etiology. We found that the risk allele of rs156429, the most associated SNP in the 7p15 region, is associated in our eQTL dataset with decreased expression of the proximal transcript encoded by NUPL2 (Table 3). The same risk allele is also associated with increased methylation of multiple CpG sites proximal to GPNMB itself (Table 3). Neither of these regions contains an obvious candidate gene.
Two additional loci (3q26/NMD3 and 8q21/MMP16) showed strong evidence of association in stage 1 and 2 but were not disease associated in the Do et al dataset. Further replication is required to clarify the role of variation at these loci in risk for PD.
The strongly associated G2019S variant in the LRRK2 gene [20] was included in the Immunochip design and we replicated the published association: control frequency: 0.045% case frequency 0.61%, estimated odds ratio: 13.5 with 95% confidence interval: 5.5–43. However, the case collections have been partially screened for this variant therefore its frequency in cases and the odds ratio is likely to be underestimated.
The ImmunoChip array design provides some power to detect whether multiple distinct association signals exist at individual loci. Indeed, if a SNP showed an independent and sufficiently strong association in stage 1, it would have been included in stage 2 provided that it was not located in the same 10 kb window as the primary SNP in the region. There is precedent for this in PD, with the previous identification of independent risk signals at the SNCA locus [11]. We therefore used the Immunochip data to test whether any of the seven loci in Table 2 showed some evidence of more than one independent signal. None of these seven loci showed any association (p>0.01) after conditioning on the main SNP in the region. In contrast, after conditioning on the most associated SNPs rs356182 in the SNCA region, several SNPs remained convincingly associated (p = 9.7×10−8 for rs2245801 being the most significant).
Lastly, we performed a risk profile analysis to investigate the power to discriminate cases and controls on the basis of the 16 confirmed common associated variants (Table 4). For each locus, we estimated the odds ratio on the basis of stage 1 data and we applied these estimates to compute for each individual in the ImmunoChip cohort a combined risk score. Solely based on these 16 common variants, and therefore not considering rare highly penetrant variants such as G2019S in LRKK2 [20], we found that individuals in the top quintile of the risk score have an estimated three-fold increase in PD risk compared to individuals in the bottom quintile (Table 4). We note however that the effect size of several of these associated variants could be over-estimated (an effect known as winner's curse, see [21]) but given the consistent estimates of odds ratio across studies (Table 4) we expect this bias to be minimal.
The combination of GWA scans and imputation methods in large cohorts of PD cases and controls has enabled us to identify five PD associated loci in addition to the 11 previously reported by us. Two of these loci (1q32/PARK16, 8p22/FGF20) implicate regions that had been previously associated with PD risk [8], [16]. The 1q32/PARK16 showed convincing evidence of association in the Japanese population [8] but until now the association P-value had not passed a stringent genome-wide significance threshold in samples of European descent [7]. The 8p22/FGF20 locus had been previously reported in a study of familial PD [16] and we provide the first evidence of association in a case-control study. The remaining three loci (STX1B/16p11, STBD1/4q21 and GPNMB/7p15) are new.
Adding the eleven previously reported common variants [15] to the five convincingly associated loci identified in this study, common variants at 16 loci have now been associated with PD. Controlling for the risk score based on the 11 SNPs previously identified [15] in the risk profile analysis (Table 4), the addition of these five new loci provides a modest but significant (p = 2.2×10−3) improvement of our ability to discriminate PD cases from controls.
Combining eQTL/methylation and case-control data implicates potential mechanisms which could explain the increased PD risk associated some of these variants. In particular, the strong eQTL in the 1q32/PARK16 region with the RAB7L1 and NUCKS1 genes (Table 3) suggests that either one of these genes could be the biological effector of this risk locus. However, existing data show that eQTLs are widespread and this co-localization could be the result of chance alone [22]. Additional fine-mapping work will be required to assess whether the expression and case-control data are indeed fully consistent.
While we are unable to unequivocally pinpoint the causative genes underlying these associations, their known biological function can suggest likely candidates. At the 1q32/PARK16 loci our association and eQTL data indicate that RAB7L1 and NUCKS1 are the best candidates. The former is a GTP-binding protein that plays an important role in the regulation of exocytotic and endocytotic pathways [23]. Exocytosis is relevant for PD for two main reasons: firstly, since dopaminergic neurotransmission is mediated by the vesicular release of dopamine, i.e. dopamine exocytosis [24], and secondly because it has been shown that alpha-synuclein knock-out mice develop vesicle abnormalities [25], thus providing a potential direct link between genetic variability in the gene and a biological pathway involved in the disease. Less is known regarding NUCKS1; it has been described to be a nuclear protein, containing casein kinase II and cyclin-dependant kinases phosphorylation sites and to be highly expressed in the cardiac muscle [26]; but an involvement in PD pathogenesis has yet to be suggested.
At the 16p11/STX1B locus, notwithstanding the fact that other genes are in the associated region, STX1B is the most plausible candidate. It has been previously shown to be directly implicated in the process of calcium-dependent synaptic transmission in rat brain [17], having been suggested to play a role in the excitatory pathway of synaptic transmission. Since parkin, encoded by PARK2, negatively regulates the number and strength of excitatory synapses [27] , it makes STX1B a very interesting candidate from a biologic perspective.
FGF20 at 8p22 has been suggested to be involved in PD [16], albeit negative results in smaller cohorts have followed the original finding [28]. FGF20 is a neurotrophic factor that exerts strong neurotrophic properties within brain tissue, and regulates central nervous development and function [29]. It is preferentially expressed in the substantia nigra [30], and it has been reported to be involved in dopaminergic neurons survival [30].
The ImmunoChip data provide limited resolution for the detection of multiple independent association signals in these regions. A previous study [31] reported some evidence of allelic heterogeneity at the 1q32/PARK16 locus but the ImmunoChip data do not support this result. A previous study [11] also reported two independent associations at the 4q22/SNCA locus and our data are consistent with this scenario. However, the newly reported secondary association (rs2245801) is in low LD (r2 = 0.21) with rs2301134, the SNP reported in [11] as an independent association. Taken together, these findings suggest that at least three independent associations exist at SNCA/4q22. A more exhaustive fine-mapping analysis using either sequencing of large cohorts or targeted genotyping arrays will also be required to fully explore this locus.
As yet, we do not know which of the variants and which genes within each region are exerting the pathogenic effect. We cannot exclude that some of the currently reported variants are in fact tagging high penetrance, but rare, mutations [32]. Nevertheless, the successful identification of these 16 risk loci further demonstrates the power of the GWA study design, even in the context of disorders like PD that have a complex genetic component. We therefore expect that further and larger association analyses, perhaps using dedicated high-throughput genotyping arrays like the ImmunoChip, will continue to yield new insights into PD etiology.
Participating studies were either genotyped using the ImmunoChip as part of a collaborative agreement with the ImmunoChip Consortium, or as part of previous GWA studies provided by members of the IPDGC or freely available from dbGaP [7], [9], [10], [11]. Genotyping of the UK cases using the Immunochip was undertaken by the WTCCC2 at the Wellcome Trust Sanger Institute which also genotyped the UK control samples. The constituent studies comprising the IPDGC have been described in detail elsewhere [15], although a summary of individual study quality control is available as part of Table S1. In brief all studies followed relatively uniform quality control procedures such as: minimum call rate per sample of 95%, mandatory concordance between self-reported and X-chromosome-heterogeneity estimated sex, exclusion of SNPs with greater than 5% missingness, Hardy Weinberg equilibrium p-values at a minimum of 10−7, minor allele frequencies at a minimum of 1%, exclusion of first degree relatives, and the exclusion of ancestry outliers based on either principal components or multidimensional scaling analyses using either PLINK [33] or EIGENSTRAT [34] to remove non-European ancestry samples. All GWAS studies utilized in this analysis (and in the QTL analyses) were imputed using MACHv1.0.16 [14] to conduct a two-stage imputation based on the August 2009 haplotypes from initial low coverage sequencing of 112 European ancestry samples in the 1000 Genomes Project [35], filtering the data for a minimum imputation quality of (RSQR>0.3) [14]. Logistic regression models were utilized to quantify associations with PD incorporating allele dosages as the primary predictor of disease. Imputed data was analyzed using MACH2DAT, and genotyped SNPs were analyzed using PLINK. All models were adjusted for covariates of components 1 and 2 from either principal components or multidimensional scaling analyses to account for population substructure and stochastic genotypic variation (except in the UK-GWAS data which were not adjusted for population substructure).
Single SNP test statistics were combined across datasets using a score test methodology, essentially assuming equal odds ratio across cohorts. In addition, fixed and random effects meta-analyses were implemented in R (version 2.11) to confirm that the score test approximation does not affect the interpretation of the results. We also tested the relevant SNPs heterogeneity across cohorts and no significant heterogeneity was detected (Table S2).
We communicated to our colleagues in charge of the independent study (Do et al) the seven SNPs listed in Table 2. For this subset of SNPs they selected the marker with the highest r2 value on their genotyping platform and provided us with the following summary statistics: odds ratio, direction of effect, standard error for the estimated odds ratio and one degree-of-freedom trend test P-value.
Quantitative trait analyses were conducted to infer effects of risk SNPs on proximal CpG methylation and gene expression. For the five replicated SNP associations (Table 2), all available CpG probes and expression probes within +/−1 MB of the target SNP were investigated as candidate QTL associations in frontal cortex and cerebellar tissue samples. 399 samples were assayed for genome-wide gene expression on Illumina HumanHT-12 v3 Expression Beadchips and 292 samples were assayed using Infinium HumanMethylation27 Beadchips, both per manufacturer's protocols in each brain region. A more in depth description of the sample series comprising the QTL analyses, relevant laboratory procedures and quality requirements may be found in [15]. The QTL analysis utilized multivariate linear regression models to estimate effects of allele dosages per SNP on expression and methylation levels adjusted for covariates of age at death, gender, the first 2 component vectors from multi-dimensional scaling, post mortem interval (PMI), brain bank from where the samples were provided and in which preparation/hybridization batch the samples were processed. A total of 670 candidate QTL associations were tested: 87 expression QTLs in the cerebellum samples, 85 expression QTLs in the frontal cortex samples, 249 methylation QTLs in the cerebellum samples and 249 methylation QTLs in the frontal cortex samples. Multiple test correction was undertaken using false discovery rate adjusted p-values<0.05 to dictate significance, with the p-value adjustment undertaken in each series separately, stratified by brain region and assay. A complete list of all QTL associations tested is included in Table S3.
|
10.1371/journal.pgen.1004729 | Identification of Rare Causal Variants in Sequence-Based Studies: Methods and Applications to VPS13B, a Gene Involved in Cohen Syndrome and Autism | Pinpointing the small number of causal variants among the abundant naturally occurring genetic variation is a difficult challenge, but a crucial one for understanding precise molecular mechanisms of disease and follow-up functional studies. We propose and investigate two complementary statistical approaches for identification of rare causal variants in sequencing studies: a backward elimination procedure based on groupwise association tests, and a hierarchical approach that can integrate sequencing data with diverse functional and evolutionary conservation annotations for individual variants. Using simulations, we show that incorporation of multiple bioinformatic predictors of deleteriousness, such as PolyPhen-2, SIFT and GERP++ scores, can improve the power to discover truly causal variants. As proof of principle, we apply the proposed methods to VPS13B, a gene mutated in the rare neurodevelopmental disorder called Cohen syndrome, and recently reported with recessive variants in autism. We identify a small set of promising candidates for causal variants, including two loss-of-function variants and a rare, homozygous probably-damaging variant that could contribute to autism risk.
| Sequencing technologies allow identification of genetic variants down to single base resolution for a whole human genome. The vast majority of these variants (over 90%) are rare, with population frequencies less than 1%. Furthermore, in a specific study, many of the variants identified are not associated with the disease of interest, and identification of the small proportion of truly causal variants is a difficult task. Clearly, for causal variants that are rare enough to only appear a few times in a study, observed frequencies in cases and controls are not enough to distinguish them from the vast majority of random variation, and rich functional annotations can help identify the causal variants. Here we propose to develop a set of statistical methods that leverage diverse functional genomics annotations with sequencing data to identify a small set of potentially causal variants and estimate their effects. Pinpointing a subset of potentially causal variants is crucial for understanding precise biological mechanisms, and for further experimental functional studies.
| The tremendous progress in massively parallel sequencing (aka ‘next generation’ sequencing) technologies enables investigators to obtain genetic information down to single base resolution on a genome-wide scale in a rapid and cost efficient manner [1], [2], [3]. The resulting datasets are high dimensional and very sparse, with millions of genetic variants, the vast majority of which are rare in the population. For example, in any genetic region, it is expected that over 90% of genetic variants have a frequency in the population of less than 1% [4]. Therefore in any given study, most variants are only observed a small number of times (e.g. many of them are singletons or doubletons). This sparse nature of the data poses nontrivial statistical difficulties, and traditional statistical methods employed for association testing with common variants are not powerful in this context [5].
Both empirical and theoretical studies suggest that rare genetic variants are an important contributor to disease risk [6], [7], [8], [9]. Over the past few years several statistical tests have been proposed to test for association with rare variants in a small genetic region, such as a gene [10], [11], [12], [13], [14], [15], [16], [17]. The proposed association tests are based on the idea of grouping together variants in the gene, and testing for association at the gene rather than variant level. While these methods attempt to increase power by cumulating the signal across a larger region, they compromise precision and, in particular, it is not possible to pinpoint individual causal variants and estimate their effects on disease. Prioritizing a small number of plausible causal variant candidates is very important for further follow-up functional studies, since experimental analyses are difficult to implement and expensive for large number of variants [18], [19]. Furthermore, identification of causal variants is essential for understanding the precise molecular mechanisms of disease. Despite its importance, this problem is only now possible to address due to the increasing availability of large-scale sequencing data and the advances in computational methods for predicting the functional effects of genetic variation [19].
The fundamental challenge in pinpointing rare causal variants is that these variants are observed very infrequently in any given dataset and these sparse frequencies on their own are insufficient to provide meaningful risk predictions. In particular, for singletons or doubletons it will be necessary to incorporate prior functional and evolutionary conservation information in order to prioritize them as likely causal. Hierarchical modeling offers a natural strategy to leverage collective evidence from rare variants with sparse data. This can be accomplished in the presence of hierarchical covariates that are associated with disease risk and which can be used for implicitly aggregating the rare variants to permit stronger inferences about individual variants. These hierarchical covariates are characteristics of the variants themselves, such as the degree of conservation across species, the position in the gene, and other features that can be represented using bioinformatic measures. Indeed, many annotation tools (such as ANNOVAR [20], PolyPhen-2 [21], SIFT [22], GERP++ [23]) exist to predict the possible impact of a variant on the function of a human protein, or the level of evolutionary constraint for a variant. Even though such bioinformatic predictions of deleteriousness are not extremely accurate and are continuously being improved, they can provide useful information on the prior likelihood that a variant is causal, especially when multiple such predictors are used, as we show in this work. In earlier work we have developed a hierarchical modeling approach that is capable of estimating odds ratios for variants that occur infrequently in the dataset [24], [25]. Hierarchical regression techniques have also been adopted in a Bayesian framework with the goal of detecting rare causal variants [26], [27], however they can be computationally intensive and can be dependent on the choice of the prior weights [27]. More recently, Pickrell [28] has used hierarchical models to combine rich functional genomics annotations (as generated by the ENCODE project [29]) and summary statistics from GWAS to identify types of genomic elements enriched among disease susceptibility loci.
Here we propose and investigate the performance of two complementary statistical methods that are able to incorporate prior information on the putative function of individual variants in a gene in order to (1) identify a list of likely causal variants, and (2) estimate the effects of these variants on disease. The first approach is a backward elimination procedure based on groupwise association tests that leads to the identification of a small set of “interesting” variants in the gene, which are enriched in causal variants. The second approach complements the first by employing hierarchical models [24],[25] that can incorporate diverse functional and evolutionary conservation annotations, and in turn provides effect size estimates and confidence intervals for individual variants.
First, we review the basics of groupwise association tests, and then we describe in detail the two complementary methods we propose for prioritizing variants for follow-up functional studies.
We assume that n subjects have been sequenced in a region of interest (e.g., a gene), that contains m variants. Let be the genotype matrix. We consider the regression model(1)where is a link function, and can be set to be the identity function when traits are continuous, or the logistic function when traits are dichotomous; are regression coefficients for the covariates that we want to adjust for. is the vector of genotypes for the ith individual, and is its trait value. are regression coefficients for the m genetic variants.
We are interested in testing the null hypothesis of no genetic effects: Testing each individual or using multiple df tests can lack power because of the sparsity of the data and the many variants in a gene. Therefore, we need to impose certain assumptions on 's to make the test more powerful. For example, one of the most widely used tests, the Burden test, assumes that all β's have essentially the same value, say , and the regression model in (1) amounts to . More generally, Lee et al. [15] assume that is a random variable with , and for different j and k. To test the null hypothesis of no genetic effects the variance-component score statistic has been proposed [15]:(2)where and specifies an exchangeable correlation matrix, and is a diagonal weight matrix, where each weight can be related, for example, to the predicted functional effect of a variant (e.g. PolyPhen-2 or SIFT score); for a dichotomous trait, is a vector of estimated probabilities of under the null model. Although this class of tests is more general, the two commonly used tests are the Burden test () and the SKAT test (). These score statistics are easy to compute and can be written simply as
(3)(4)
The null distribution of is approximated by a mixture of distributions. Davies' method [30] or moment matching can be employed to calculate the p value. The relative performance of the two tests will depend on the true underlying disease model. The Burden test tends to be more powerful when disease associated variants are all of the same type (risk or protective) and with effects of similar magnitude. The SKAT test tends to be more powerful when there is a mixture of risk and protective variants, and also when only a small percentage of variants in a region are causal. A parallel framework for family-based designs has also been proposed [16].
The groupwise association tests described above test for association at a gene level, but are not able to pinpoint individual causal variants in the gene. However, once a gene has been shown to contain variants associated with disease (e.g. using the Burden or SKAT tests), identifying the individual causal variants among the many variants in a gene is of considerable interest as it can lead to a better understanding of the molecular mechanisms underlying a complex trait, and is essential for further experimental validation work.
Starting with a groupwise association test, one natural way to identify causal variants that are individually of weak effect is to evaluate their contribution to a given set of variants by removing the variant from the set, and assessing the resulting effect, e.g. the p value for the reduced set. The following iterative algorithm (essentially a backward elimination procedure) is designed for this purpose.
Backward Elimination Algorithm: Step 1. Start with a set of variants . The current set is . Compute the score statistic in eq. (2) (either or ) for this current set , and compute the p value: .
Step 2. Remove each of the variants one at a time from , i.e. consider the sets with , and then compute the corresponding score statistic and p value for each of these reduced sets: .
Step 3. If then remove the variant that leads to the smallest p value:
The current set becomes and repeat steps 2 & 3. If the current p value cannot be improved, then go to step 4.
Step 4. Return the current set of variants.
The results on a typical simulated example are shown in Figure S1. We show there the effect of removing a causal variant on the p value of the reduced set (i.e. in Step 2 above), compared to removing a non-causal variant. As shown, the removal of causal variants will tend to result in an increase in the p value for the reduced set, as desired. There is a highly significant difference in the p values for the reduced sets when removing causal vs. non-causal variants (bootstrap Kolmogorov-Smirnov test p value ).
This algorithm is applicable when the number of variants we start with in Step 1 is not too large (otherwise, the contribution of a weak variant to a large set is difficult to evaluate). However, sequencing a gene in thousands of individuals can lead to the detection of potentially hundreds of variants, or more. Therefore, we use a resampling procedure, whereby each time a small number of variants is chosen (say ) from the large number of variants identified in a gene, and then the above algorithm is applied to such small sets a large number of times (in our examples we use 2000 such re-samplings, although this number can be increased in the case of a large number of variants in the gene). At the end, for each variant in the gene we calculate the number of times it was returned in Step 4; we call this number the return count for a variant. A similar resampling procedure has been applied before in the context of gene-by-gene interaction [31]. Our goal is to use the sample of return counts to partition variants into two groups: “interesting” (higher return counts) and “non-interesting” (lower return counts), with the “interesting” category expected to be enriched in disease causing variants. We use nonparametric EM-like methods [32] to identify the two subgroups (see Text S2 for more details).
A complementary approach to the backward elimination procedure described above is a hierarchical model. Hierarchical modeling has several important advantages in the analysis of rare variant data, because it can naturally integrate various functional prediction scores for individual variants. Such prior knowledge will be essential in pinpointing the likely causal variants in a gene, especially for causal variants that are rare enough to only appear a few times in a study (e.g. singletons and doubletons). For such variants, observed frequencies in cases and controls are clearly not enough to distinguish them from the vast majority of random variation (in the Nelson et al. study [4], more than 74% of variants were singletons or doubletons). Information on the putative functional effect of a variant on the protein or the degree of evolutionary conservation can be an important indicator on the likelihood of a variant being causal.
Such functional information can be incorporated through a hierarchical model [24], [25]. In the first stage, the trait value is related to the genotypes and possible confounders via the following model:(5)with notations similar to those in model (1) above.
A second stage model relates the individual variant risks to prior (e.g. functional annotation) information known about the variants:(6)where Z is an matrix for the k variant covariates (e.g. functional information); is a vector of regression parameters for the second stage covariates, and is a vector of normally distributed residual effects, assumed (for convenience) to be statistically independent. A principal advantage of the hierarchical modeling framework is that it can easily incorporate multiple functional annotations.
Combining the two models, one obtains the following generalized linear mixed effects model:(7)
The parameters of this model can be estimated using a hybrid Bayesian pseudo-likelihood approach which performs Bayesian estimation of the variance component of the model and then conducts pseudo-likelihood estimation of the fixed and random effects using this estimated random effects variance [24], [25]. We can use the resulting estimates for the odds ratios and their standard errors to rank variants in a gene. Naturally the most difficult to identify are causal variants that occur only a few times. The odds ratio estimates for such variants will heavily depend on the higher level covariates, such as information on the predicted functional effect for a variant. For example, for a variant that occurs infrequently in a dataset (e.g. 2 times in cases and 0 times in controls), knowing that it is a LoF variant increases its likelihood to be a causal variant compared with a synonymous variant with the same frequency.
We evaluate the performance of the proposed methods using simulated data and then apply them to two sequencing studies, the Dallas Heart Study and a study on Autism Spectrum Disorders.
We simulated sequence data on 10,000 haplotypes in one genomic region of length 1 Mb under a coalescent model using the software package COSI [34]. The model used in the simulations was the calibrated model for the European population. For our purposes, we randomly sampled small subregions of size 10 kb, and simulated datasets with individuals (equal number of cases and controls). The number of variants and the minor allele frequency (MAF) distribution varies depending on the subregion sampled.
We considered two disease models (Table 1). In these two models, the odds ratio (OR) is a decreasing function of the MAF. For both models, we assume that of the variants with MAF ≤0.05 in the 10 kb region under investigation are causal variants.
For a dichotomous trait, we assumed the logistic model:with chosen such that the disease prevalence was 0.05.
We have also simulated bioinformatic covariates for variants to be used in the backward elimination algorithm, as well as to be incorporated in the hierarchical model. A first bioinformatic covariate we simulate is a binary variable, such as whether a variant is non-synonymous or not. Based on empirical studies [4], we consider the non-synonymous to synonymous ratio (NS:S) for the rare variants in the region to be between (depending on the strength of the purifying selection in the region). We assume that 80% of causal variants are non-synonymous, and then using Bayes' rule we calculate the proportion of non-causal variants that are non-synonymous (see Table 2). Given these settings, the proportions of causal variants among non-synonymous and synonymous variants can be easily derived and are reported for completeness in Table 2.
Furthermore, additional variant annotation tools for non-synonymous variants exist, and are able, for example, to predict the damaging effect of an amino acid substitution (PolyPhen-2 and SIFT), and to assess the extent of evolutionary conservation at a position (GERP++). Therefore, for non-synonymous variants we simulate two additional predictors, as follows. The first bioinformatic predictor (B1) for non-synonymous variants is defined as a binary indicator whether a variant is predicted to be damaging or not. Following the empirical results in Cooper et al. [18] we assume that 30% of non-synonymous, non-causal variants are damaging (possibly or probably), and that 80% of non-synonymous, causal variants are damaging (Table 2). A second bioinformatic predictor (B2) is also defined as a binary indicator whether a variant is predicted to be probably damaging or not. Again, as in Cooper et al. [18] we assume that 10% of non-synonymous, non-causal variants are probably damaging, and that 80% of non-synonymous, causal variants are probably damaging (Table 2). To assess the effect of using a non-informative predictor, we also simulate a binary predictor with 50% non-synonymous causal and 50% non-synonymous non-causal variants having a value of 1 for this non-informative predictor.
The main goal of the proposed methods is to combine sequencing data with functional predictions about the deleteriousness of variants to identify a set of promising variants, enriched in causal variants. Furthermore, the selected variants can be ranked according to their return counts from the backward elimination procedure, or the estimated effects from the hierarchical model (ranking based on Z scores gave similar results). We use several measures to assess the performance of the methods. The main measures are: (1) the overall ranking of the true causal variants among the variants in the gene, and (2) the bias and coverage accuracy in the estimation of effect sizes for the variants from the hierarchical model.
We employ the backward elimination procedure as well as fit a hierarchical model including the full set of variants and assuming a single functional predictor in the second stage, namely whether a variant is non-synonymous or synonymous. Due to the expected difference in enrichment of causal variants among non-synonymous versus synonymous variants we evaluate the overall ranking of the causal variants separately among non-synonymous and synonymous variants. More explicitly, among the non-synonymous variants selected as “interesting” by the backward elimination procedure we rank the causal variants based on their return counts (this approach is denoted as BE in the figures below). Furthermore, we also use the estimates obtained from the hierarchical model to rank the causal variants among all non-synonymous variants (HM) as well as among the non-synonymous variants selected as “interesting” by the backward elimination procedure (HMS). For each simulation we take the median of the ranks of the causal variants involved and then compute the median of these estimates across simulations. The ranking for the synonymous variants is done similarly.
Figures 1(a) and S3(a) present the median ranks of the causal non-synonymous variants based on the different ranking procedures (HM, BE, and HMS). The hierarchical model (titled “HM” in the figure) results in higher median rank (worse performance) than that of the backward elimination procedure (titled “BE” in the figure). This is of course expected due to the smaller number of variants that the backward elimination procedure returns as “interesting” (Figures 1(a) and S3(a)). However, despite excluding a substantial proportion of variants in the backward elimination process (the “non-interesting” category), we show that the top ranked causal variants in the hierarchical model are kept in the selected list (Figures 1(b) and S3(b)). For the scenarios investigated, the number of causal, non-synonymous variants in the top 10 ranked variants varies between 5 and 8 for the case when the percentage of causal variants in a region is 20% (Figure 1(b)), and 3–6 for the case with only 10% causal variants (Figure S3(b)). When looking only among the “interesting” variants from the backward elimination procedure, the overall ranking of causal variants based on the hierarchical model estimates (titled “HMS”) is similar to the one based on return counts in the backward elimination procedure (Figures 1 and S3). Therefore the backward elimination method can be used as an effective tool to select and rank a set of promising variants, and reduce the overall list of variants to a smaller, more manageable list, followed by further characterization of these variants’ effects within the framework of the hierarchical model. For all methods, and regardless of disease model, the performance tends to decrease as the non-synonymous to synonymous ratio increases from 0.6 to 1.4 (as the effect of purifying selection becomes weaker).
When looking at the synonymous variants separately, the results are qualitatively similar to the non-synonymous case. However, because only 20% of the causal variants are assumed synonymous, the overall ranking of the few causal variants among the synonymous variants is noticeably worse compared with non-synonymous variants, as expected (Figures S4 and S5). For example, in a region with 20% causal variants, overall we detect between 2 and 3 causal synonymous variants among the top 10 ranked variants, and only 1–2 in a region with 10% causal variants. Due to these high false-discovery rates for synonymous variants, it may be more effective to focus initial efforts for causal variant identification among the functional (non-synonymous and LoF) variants. As genomic annotations become richer for synonymous variants, we expect the discovery of causal variants among synonymous variants to become more accurate.
We evaluate the effect on ranking the causal variants among non-synonymous variants when additional bioinformatic predictors are added to the hierarchical model (in addition to the indicator whether the variant is non-synonymous vs. synonymous). Note that synonymous variants were assigned a bioinformatic predictor of 0.
We restrict attention to ranking only among the “interesting” variants, as selected by the backward elimination procedure. As shown in Figure 2, when we add one bioinformatic predictor (B1 or B2; see Table 2), the ranking of causal variants improves significantly compared to the original hierarchical model that only uses a binary predictor (whether a variant is non-synonymous or not). The improvement is more pronounced with predictor B2, due to the higher specificity of this predictor. For example, for model M1, a non-synonymous to synonymous ratio of 1.4 and 20% causal variants in a region, the median number of causal variants among the top 10 ranked non-synonymous variants increases from 5 (in the original hierarchical model) to 8 when using bioinformatic predictor B2. Since we do not always know which of several available bioinformatic predictors may have higher accuracy, the hierarchical model allows us to combine multiple bioinformatic predictors. When combining three bioinformatic predictors (two predictors with the same sensitivity and specificity as B1 and one predictor B2, all independent), we find that the ranking of causal variants is now similar or superior to the ranking obtained when using only the better of the two bioinformatic predictors (i.e. B2). Similarly, when using a combination of four bioinformatic predictors (four predictors with the same sensitivity and specificity as B1), the ranking of causal variants is better than using just a single predictor B1, and similar to using the more accurate predictor B2. These results suggest that using multiple bioinformatic predictors with different accuracies (even multiple weak predictors) can help detection of the causal variants. Similar results are obtained when the proportion of causal variants in a region is 10% (Figure S6).
We have also evaluated the effect of including a non-informative predictor in the analysis, although in practice we expect that functional annotations are correlated with the causal status of a variant. The results are reported in Figures S11 and S12. As shown, including a random (non-informative) bioinformatic predictor does lead to worse performance compared to when such a predictor is not included, although combining an informative predictor (B1) together with a non-informative one does help improve the performance. Again, the ability of the hierarchical model to incorporate multiple functional predictors of varying accuracy is an important feature when the best predictors are not known a priori.
It is possible to incorporate one bioinformatic predictor, such as B1 or B2, in the backward elimination procedure directly (as a weight in the Burden test statistic). We found that for the case of only one bioinformatic predictor, the backward elimination procedure performed similarly with (or slightly worse than) the hierarchical model (Figures S7). However, in general, it is not clear how to choose one single functional annotation from several annotations available. Therefore, the hierarchical model has the important advantage that multiple bioinformatic predictors can be included, and, as shown above, the ranking of the causal variants improves with the addition of several predictors of varying accuracy.
As already mentioned, the hierarchical model has distinct advantages when multiple functional predictions are available for variants. In particular, it is possible to provide effect size estimates and standard errors for individual variants, that take into account such diverse functional predictions. As seen in Table S1, for disease model M1 (Table 1), absolute biases of the log odds ratio estimates from the hierarchical modeling approach are similar among the different scenarios while coverages are close to the nominal level of 95%. In comparison, bias is further increased and coverages are under the nominal level of 95% for disease model M2 (which assumes higher odds ratios than model M1, Table 1), though there is a trend towards reduced bias and improved coverage with the addition of stronger bioinformatic predictor(s).
The biases observed here are due to several causes. One main source of bias is the shrinkage phenomenon that occurs with hierarchical models: in this setting of sparse data the model relies heavily on the higher level covariates and as a result the estimated risks of the non-causal variants with high bioinformatic predictor scores will be biased upwards, while the risks of the causal variants with low bioinformatic predictor values will be shrunk down, resulting in increased bias and loss of power, respectively. As the frequency of carriers increases, the model overrides the misclassifications of the higher level covariate, yielding less biased estimates (data not shown). This shrinkage is even more pronounced for model M2, which assumes higher odds ratios for the causal variants (compared to M1), resulting in the poorer performance noted with model M2. An additional source of bias comes from our analyses being conditional on the groupwise (gene-based) test being significant.
We first show an application of the proposed methods to a well studied re-sequencing dataset for ANGPTL4 for 3,551 individuals of varied ethnicity from the Dallas Heart Study. Rare and low-frequency variants in this gene have been previously associated with low serum triglyceride levels [35]. We consider log-transformed triglyceride level as our phenotype, and adjust for gender and ethnicity. As in the original study [35], we dichotomize the phenotype by considering the individuals in the lowest quartile as cases, and the individuals in the highest quartile as controls, for a total of 898 individuals with variation in this gene. We identify 20 functional variants (missense, nonsense, and frameshift).
In Table 3 we report the functional variants, ranked by the estimated effects from the hierarchical model taking into account their PolyPhen-2 and GERP_RS scores. Also reported is the return count from the backward elimination procedure (due to the small number of variants in this gene we do not fit the nonparametric mixture model in this example; instead we simply rank all variants). All the top ranked variants that appear only in cases (i.e. the lowest quartile) have been shown by Romeo et al. [36] to severely compromise the function of the protein. In particular, the top ranked variant, p.Lys217Ter, is a nonsense variant that appears only once in an affected individual, and is assumed to interfere with protein synthesis. The second, fourth, seventh and eighth variants have been shown using functional studies to lead to impaired protein secretion. The fifth variant showed reduced ability to inhibit LPL (lipoprotein lipase) activity in vitro, while the sixth variant introduced a premature termination codon [36]. The third variant in the list, p.Glu40Lys, is a missense variant (classified as probably damaging by PolyPhen-2 and as evolutionarily conserved site by GERP_RS), with a frequency of 1.3% in this dataset, and has been shown to be significantly associated with plasma triglyceride levels [35]. However, due to its presence even among controls (i.e. the highest quartile), this variant was not investigated in the functional studies in Romeo et al. [36].
We next show an application to a gene with a larger number of functional variants, and for which not much is known on the likely causal variants. Hence the next application is a more difficult example for the proposed methods.
The Vacuolar Protein Sorting 13 homolog B (VPS13B, also known as COH1, MIM #607817) is a gene associated with Cohen syndrome (CS, OMIM #216550), a rare autosomal recessive neurodevelopmental disorder overrepresented in Finland and common in Amish, Irish travelers and Greek/Mediterranean founder populations [37], [38]. At least 200 affected individuals of diverse ethnic background have been reported so far with diverse VPS13B mutations, including nonsense, missense, splicing, indels, microdeletions and microduplications [38]. Despite clinical heterogeneity in part related to ethnic background, the disorder has core features, including non-progressive intellectual disability, motor clumsiness, postnatal microcephaly, a typical facial gestalt, hypotonia, intermittent neutropenia, and chorioretinal dystrophy [39]. Behavioral disturbances are common among CS individuals, and autistic traits have been reported in cases of greek/mediterranean descent [40]. Furthermore, VPS13B mutations have been found in individuals with autism [41] and non-syndromic intellectual disability [42]. It is worth noting that mutations in another member of the VPS13 gene family (VPS13A or CHAC, MIM #605978, encoding for a protein known as Chorein), cause chorea-acanthocytosis [43] (MIM #200150), a recessive disorder of acanthocytosis and adult-onset choreic involuntary movements with significant co-morbidity with psychiatric illness [44].
VPS13B is also an intolerant gene with a Residual Variation Intolerance Score [45] of −2.44 (top 0.55% most intolerant genes) in Europeans and a similar score for African Americans. We applied the proposed methods to the 166 VPS13B variants identified in a whole-exome sequencing autism spectrum disorders (ASD) case/control dataset (; more details on this dataset can be found in Text S1). We tested for association with functional (non-synonymous, nonsense and splice-sites) rare variants in this gene and the Burden test p value was 0.01. We then used the backward elimination algorithm to identify a set of “interesting” (i.e. potentially causal) variants, and for each of these variants we report effect size estimates and standard errors from the hierarchical model. Note that the ratio of non-synonymous to synonymous variants in this gene is 0.84, hence towards the lower end of values in our simulated scenarios.
Of the 166 variants in this gene, we focus on 74 that are non-synonymous, nonsense or splice-sites (two variants affecting the invariant splice acceptor site of the intron between exons 51 and 52 have been excluded from further analyses because they did not validate by Sanger sequencing). Of these, the backward elimination procedure selects 42: 2 of them are LoF (one nonsense and one variant affecting an essential splice site), and of the missense PolyPhen-2 predicts that 14 are probably damaging, 1 possibly damaging, and 25 benign (Figure 3). In Figure 4(a) we show the drop in p value each time a variant is being removed in step 3 of the backward elimination procedure; the process stops when the p value starts to increase as one tries to remove any of the remaining variants. Also shown in Figure 4(b) is the distribution of return counts (from the re-sampling procedure), and overlaid is the fitted mixture with two distinct components. The 42 selected variants belong to the second component of the fitted mixture (these are the “interesting” variants). As a comparison, applying the backward elimination algorithm to the remaining 90 synonymous variants results in no distinguishable “interesting” component (and markedly smaller average return counts compared to the non-synonymous case; Figure S8).
In Table 4 we report the top 20 variants among the selected functional variants (ranked by the estimated effects from the hierarchical model), along with their PolyPhen-2 and GERP_RS scores. Noticeably, among the top ranked variants there is a probably damaging variant (, p.Arg3198Trp, annotated on NM_152564.4 and Q7Z768-2, respectively) with 5 variant copies in cases and 1 in controls, with one case being homozygous at the position. Furthermore, the top two variants in the list have both very high C-scores [46] (36 and 35, top 0.1%), based on the recently introduced measure of deleteriousness Combined Annotation-Dependent Depletion (CADD) that integrates diverse genome annotations. The two LoF variants (one nonsense (, p.Ser3383Ter) and one splice site ()) have been seen only once in cases (i.e. singletons). Notably, the splice variant affects the splice donor site of the intron between exons 18 and 19, and a homozygous mutation in the splice acceptor site of the same intron has been identified in an individual suffering from CS [47]. As a first step toward the characterization of the variants, we used Sanger sequencing to validate two cases with the variant and the cases with and and study their inheritance pattern. This analysis is of particular relevance for singletons, considering that the false discovery rate among those can be high. All variants were validated and found to be inherited (Figure S9). In one family with the , both affected children are homozygous and inherit the variant from their parents (father homozygous and mother heterozygous - Figure S9A). In a second family with , the variant is transmitted from heterozygous parents to one affected child, and untransmitted to the unaffected child (Figure S9B). The variant is inherited from the mother (Figure S9C), and the variant is paternally transmitted to both affected children (Figure S9D).
To understand the impact of the variants on the molecular functions of VPS13B, all 42 variants deemed “interesting” by the backward elimination procedure were projected on the protein topology, reconstructed with the Pfam domains (N-terminal region of Chorein, DUF1162, ATG C-terminal domain), the experimentally ascertained Golgi targeting domain [47], and 11 transmembrane domains predicted with TMPred [48] (Figure S10). Benign variants appear scattered along the protein topology, while some of the predicted damaging variants map to known domains, including a missense in the DUF1162 domain and two missense in the Golgi targeting domain. Prediction of the structural changes that can result from the variants using MutPred [49] further revealed two top deleterious missense variants (p.Tyr1428His, predicted to cause gain of disorder (p = 0.006), loss of beta-sheet (p = 0.008) and gain of alpha-helix (p = 0.049); and p.Asp1475Gly, predicted to cause gain of alpha-helix (p = 0.049)). The LoF variants are upstream the Golgi domain, thus they are likely to cause premature insertion of a stop codon, activating nonsense-mediated mRNA decay or producing protein isoforms lacking the Golgi targeting domain. Although the pathological mechanisms caused by VPS13B insufficiency or mutations are still unknown, fibroblasts isolated from individuals with CS show severe fragmentation of the Golgi apparatus into ministacks [47], a defect observed in neurodegenerative disorders [50] and hypothesized to precede neuronal cell death [51]. Therefore, the LoF variants might prevent proper localization of VPS13B and disruption of its molecular functions on Golgi assembly or maintenance, triggering the pathological cascades underlying Cohen syndrome and/or autism.
Most of the variants selected are singletons (34 out of 42). As previously mentioned, for singletons accurate bioinformatic predictors about their likely functional effects are essential in order to identify such variants as promising, and hierarchical modeling is a natural framework to incorporate such information. Naturally the false discovery rate among these singletons can be high, and dependent on the sensitivity and specificity of the bioinformatic predictors used to characterize the variants in the hierarchical model. For comparison, in Table S2 we show the top 20 variants among the functional variants selected by the backward elimination procedure with ranking based on return count. No functional prediction score was used in this analysis. Although the more common variants still occur among the top variants in this analysis, for singletons, the hierarchical model gives higher priority to variants with high scores for both PolyPhen-2 and GERP_RS (Table 4). This ability of the hierarchical model to prioritize low frequency variants by taking into account multiple functional predictions is a distinct advantage over ranking based on return count alone (with no consideration of the PolyPhen-2 and GERP_RS scores for the variants).
Pinpointing the rare causal variants among a large number of variants that occur in a genetic region is a difficult challenge, but crucial for follow-up functional studies, and for a better understanding of the molecular mechanisms that lead to disease. For many causal variants that occur only a few times in a dataset, incorporation of external information characterizing the variants (such as bioinformatic predictions on the deleteriousness of a variant) is essential to help prioritize these rare variants. We have described here two complementary statistical methods, that are able to integrate diverse functional annotations on individual variants in a region, and produce a selected list of candidates for causal variants, ranked according to their estimated effect sizes. The backward elimination procedure offers a natural way to select a set of promising variants, while using multiple functional predictors in the hierarchical modeling approach provides more in depth characterization of variants’ effects on disease, and can help boost the power to identify causal variants. We have focused attention here on some of the commonly used annotations for coding regions; however we acknowledge that there are other possible functional genomics annotations available both for the coding and non-coding regions [28] and with continued efforts to improve these functional predictions this list will further expand.
We illustrate the proposed methods through an application to a gene implicated in Cohen syndrome and autism, VPS13B. For this gene, we show that among the top selected variants are two LoF variants, and one rare, probably damaging variant that is homozygous in one affected individual. Autosomal recessive mutations associated with autism have been recognized for decades [52]. Recently, whole-exome sequencing has provided strong evidence that rare, recessive LoF variation is a major contributor to risk [53]. It is likely that many recessive missense variants contribute to ASD as well, although there has been insufficient power in whole-exome studies carried out to date to fully explore such variation. VPS13B is indispensable for the Golgi apparatus, and genes important for Golgi morphology and function have been linked to autism disorders, including RAB39B, mutated in a X-linked intellectual disability associated with autism, epilepsy and macrocephaly [54], and UBE3A, responsible of Angelman syndrome [55]. In addition, disturbances in pathways linked to Golgi, e.g. autophagy [56] and protein glycosylation [57], have been associated with autism etiology. Our findings extend the mutational landscape of VPS13B in Cohen syndrome and autism and further strengthen the connection between Golgi homeostasis and autism.
A rather large number of selected variants in the backward elimination procedure are singletons. Causal variants that appear as singletons in a dataset are difficult to distinguish from random genetic variation, and accurate functional predictions on such variants are crucial and will help in identifying those singletons more likely to be causal. Currently, it is not uncommon for predictions on the deleteriousness of a variant to be discordant (e.g. predictions from PolyPhen-2 and SIFT), and combining such multiple predictors can be difficult, although aggregate deleteriousness scores, such as Condel [58] and C-score [46], are available. Since the hierarchical model can easily incorporate multiple functional predictions, it has a distinct advantage over methods that cannot consider multiple predictions at once. Indeed, most of the existing groupwise association tests (including the Burden and SKAT tests discussed here) can only use one functional score at a time, and therefore it is not clear how multiple scores (that are sometimes discordant) can be taken into account. It is also worth noting that in the case of existing Burden (or SKAT) tests, a variant with a low functional score (e.g. PolyPhen-2 score close to 0) will be excluded from analysis regardless of the evidence of association that the data suggests (frequency in cases vs. controls). In contrast, in the presence of sufficient case-control frequencies, the hierarchical model places more weight on the larger case-control frequencies overriding the information from the bioinformatic predictors when that information does not support the likelihood of an increased risk. Therefore, the hierarchical model has an advantage also over simple methods to prioritize based on one functional score (e.g. PolyPhen-2) alone.
The use of next-generation sequencing technologies may lead to higher error rates compared to a traditional Sanger sequencing platform. Sequencing errors may be disproportionately present among singletons or very rare variants, especially for larger sample sizes, although for a single gene the number of errors is expected to be relatively small. Therefore, as a first step toward the characterization of the top ranked variants, Sanger sequencing can be used to validate the variants.
Classical variable selection methods (such as ridge regression [59] and LASSO [60]) are natural tools to employ in this setting in which the causal variants are expected to be just a small subset of all sequenced variants. Such methods have recently been applied to sequencing data [61], [62]. However, because these methods have not been developed to handle such sparse data, they have difficulty in selecting very rare variants (such as singletons). Furthermore, it is not clear how one can take into account multiple functional predictions for variants. Further work in this area needs to be done to assess the ability of these classical variable selection approaches to identify rare causal variants. Other existing methods one could use for causal variant prioritization fall into two extremes. Some methods use only data on observed case-control frequencies. For example, in KBAC [11], variants are weighted using data-adaptive weights, reflecting the estimated effect of a variant on the phenotype, and these weights can be potentially used to rank variants. However, as explained above, for rare and low frequency variants it is essential to make use of rich functional genomics annotations. At the other extreme, one can rank variants based on a functional score alone. This latter class of methods has the important drawback, especially in the case of complex traits, that it ignores case-control frequencies, and relies heavily on the accuracy of the bioinformatics predictor. Based on simulation studies, we have shown that hierarchical modeling that takes into account both association evidence coming from the sequencing data, and available functional genomics predictions, have better performance compared with ranking based on a single bioinformatic predictor alone (Text S3 and Figure S13).
We have focused here on the selection of variants that increase risk to disease. However, one can in principle use the proposed methods to identify protective variants. Instead of a Burden type statistic, one can use a SKAT statistic in the backward elimination procedure (also implemented in our software). Similarly, for the hierarchical model, variants with low functional score and higher frequency in controls compared with cases will be candidates for protective variants.
Further improvements to the backward elimination procedure are possible. For example, instead of partitioning variants into two groups based on a binary bioinformatic predictor, such as non-synonymous and synonymous, an alternative would be to calculate stratified false discovery rates [63] or possibly covariate-modulated local false discovery rates [64]. The advantage of such an approach would be that more than one covariate can be added to the backward-elimination procedure, although this point requires further work. Furthermore, information on the location of a variant within a gene or region of interest, e.g. what functional domain it affects, can be important, especially for missense variants. We have previously described scan statistic approaches to identify clusters of rare disease associated variants, and have shown applications to both autism and schizophrenia studies [65], [66], suggesting that incorporating such location information into the backward elimination procedure could improve the identification of causal variants.
The estimates of a rare variant's effect on disease from the hierarchical model can have substantial bias. This happens because in any particular gene or region only a small proportion of variants are expected to be disease causing, and most of the variants represent random genetic variation. Therefore, when estimating odds ratios of causal variants, there is a strong shrinkage effect toward the overall estimate. Incorporation of accurate functional predictors in the hierarchical model is one possible way to help attenuate this bias; further work is needed on finding better ways to reduce the bias.
In addition to their ability to pinpoint likely causal variants, the proposed methods can be used to prioritize variants for genotyping in independent datasets for the purpose of replication or validation. This is relevant when re-sequencing the gene or region in additional datasets is too expensive, and one chooses instead to genotype variants discovered in the original study [67]. Moreover, the proposed approaches can be used at a genome-wide scale, by first selecting the promising genes based on the overall gene-based test or other criteria (e.g. good biological candidate) followed by the backward elimination and hierarchical model approach to prioritize the variants within the genes identified as promising. Such a genome-wide analysis can, for example, identify classes of functional elements or domains enriched among the top variants in the selected genes.
The proposed methods are applicable to case-control or population-based designs. However, family-based designs represent a natural way to identify causal variants. For example, in multiplex families, significant sharing of a non-synonymous mutation among multiple affected relatives can be an important indication of causality. Bayesian approaches in this context have been developed before [68], and further work in this area is worth pursuing.
In summary, we have proposed and investigated two complementary statistical methods to identify causal variants among the naturally occurring genetic variation at a locus. They are able to incorporate sequencing data with various functional predictors on variants, and select a small number of variants that are enriched in causal variants. In the current study, we applied the proposed methods to a gene known to contain risk variants for ASD as proof-of-principle, and identify several interesting variants, including two LoF variants and a homozygous probably damaging variant likely important to autism risk.
|
10.1371/journal.ppat.1004870 | Sigma Factor SigB Is Crucial to Mediate Staphylococcus aureus Adaptation during Chronic Infections | Staphylococcus aureus is a major human pathogen that causes a range of infections from acute invasive to chronic and difficult-to-treat. Infection strategies associated with persisting S. aureus infections are bacterial host cell invasion and the bacterial ability to dynamically change phenotypes from the aggressive wild-type to small colony variants (SCVs), which are adapted for intracellular long-term persistence. The underlying mechanisms of the bacterial switching and adaptation mechanisms appear to be very dynamic, but are largely unknown. Here, we analyzed the role and the crosstalk of the global S. aureus regulators agr, sarA and SigB by generating single, double and triple mutants, and testing them with proteome analysis and in different in vitro and in vivo infection models. We were able to demonstrate that SigB is the crucial factor for adaptation in chronic infections. During acute infection, the bacteria require the simultaneous action of the agr and sarA loci to defend against invading immune cells by causing inflammation and cytotoxicity and to escape from phagosomes in their host cells that enable them to settle an infection at high bacterial density. To persist intracellularly the bacteria subsequently need to silence agr and sarA. Indeed agr and sarA deletion mutants expressed a much lower number of virulence factors and could persist at high numbers intracellularly. SigB plays a crucial function to promote bacterial intracellular persistence. In fact, ΔsigB-mutants did not generate SCVs and were completely cleared by the host cells within a few days. In this study we identified SigB as an essential factor that enables the bacteria to switch from the highly aggressive phenotype that settles an acute infection to a silent SCV-phenotype that allows for long-term intracellular persistence. Consequently, the SigB-operon represents a possible target to develop preventive and therapeutic strategies against chronic and therapy-refractory infections.
| Staphylococcus aureus is a frequent pathogen of severe invasive infections that can develop into chronicity and become extremely difficult to eradicate. Chronic infections have been highly associated with altered bacterial phenotypes, i.e., the small colony variants (SCVs) that dynamically appear after bacterial host cell invasion and are highly adapted for intracellular long-term persistence. In this study, we analyzed the underlying mechanisms of the bacterial switching and adaptation process by investigating the functions of the global S. aureus regulators agr, sarA and SigB. We demonstrate that a tight crosstalk between these factors supports the bacteria at any stage of the infection and that SigB is the crucial factor for bacterial adaptation during long-term persistence. In the acute phase, the bacteria require active agr and sarA systems to induce inflammation and cytotoxicity, and to establish an infection at high bacterial numbers. In the chronic stage of infection, SigB downregulates the aggressive bacterial phenotype and mediates the formation of dynamic SCV-phenotypes. Consequently, we describe SigB as a crucial factor for bacterial adaptation and persistence, which represents a possible target for therapeutic interventions against chronic infections.
| S. aureus is a major human pathogen that can infect almost every organ in the body and cause destructive infections [1]. Besides tissue damage the ability to develop persisting and therapy-refractory infections poses a major problem in clinical practice, such as endovascular and bone infections [2,3]. Chronic infections require prolonged antimicrobial treatments and can have a dramatic impact on the patients`quality of life, as they often afford repeated surgical interventions with the risk of amputation or loss of function [2,4].
To induce an infection S. aureus expresses a multitude of virulence factors, including surface proteins and secreted components, like toxins and peptides [1]. Toxins and other secreted factors are mainly directed against invading immune cells, but can also cause tissue damage that enables the bacteria to enter deep tissue structures [5,6]. Yet, S. aureus is not only an extracellular pathogen, but can also invade a wide variety of mammalian cells, such as osteoblasts, epithelial- and endothelial cells [7–9]. All cells discussed in the literature as “non-professional phagocytes” possess mechanisms that nevertheless permit endocytotic uptake and degradation of microorganisms [10–12]. To evade the intracellular degradation machineries the bacteria have evolved different strategies, such as the killing of host cells or the escape from the lysosomal compartments and silent persistence within the intracellular location [8,13,14]. Only recently, we demonstrated that S. aureus can dynamically switch phenotypes from a highly aggressive and cytotoxic wild-type form to a metabolically inactive phenotype (small colony variants, SCVs) that is able to persist for long time periods within host cells without provoking a response from the host immune system [9]. In their intracellular location the bacteria are most likely very well protected from antimicrobial treatments and the host´s defense system. This is a possible reservoir for chronic and recurrent infections. Yet, the environmental changes encountered by invading bacteria during the passage from an extracellular to the intracellular milieu and during long term persistence within the intracellular shelter most likely cause diverse stress conditions. The adaptation mechanisms involved and how bacteria cope with this stress, are largely unknown, but probably involve global changes in gene expression to promote survival.
It is well known that S. aureus possess a large set of regulatory factors that control the expression of virulence determinants [15–18]. A very important and well-studied system is the accessory gene regulator (agr)-locus with the effector molecule RNA III [19,20]. Many S. aureus derived factors that induce inflammation or cell death are under the control of this system. Important cytotoxic components regulated by the agr-system are the pore-forming α-hemolysin (α-toxin, Hla) and the phenol-soluble modulins (PSMs), which are strong cytotoxic and pro-inflammatory factors in different host cell types [6,21–23]. SarA, which is the major protein encoded by the sar-locus, is believed to contribute to the activation of agr expression [24,25]. This is supported by findings of many S. aureus infection models demonstrating that mutations of either loci result in attenuation of virulence [26–28]. Beyond that, it has been shown that SarA influences the regulation of several virulence factors independently of agr, e.g. expression of adhesins [25,29,30]. The alternative sigma factor B (SigB; σB) modulates the stress response of several Gram-positive bacteria, including S. aureus [31–33]. SigB is responsible for the transcription of genes that can confer resistance to heat, oxidative and antibiotic stresses [16,31,34,35]. The sigB system is linked to the complex S. aureus regulatory network, as it increases sarA expression, but decreases RNA III production [36].
During the infection process the bacteria encounter different stressful conditions that they need to overcome in order to settle and maintain an infection. In the acute infection they have to fight against invading immune cells and destroy tissue cells to enter deep tissue structures, whereas during the chronic phase a major challenge is most likely the hostile intracellular location deprived of nutrients and lysosomal degradation. Consequently, the host-pathogen interaction needs to be very dynamic. Yet, to date no studies have followed the adaptation mechanisms and the impact of regulators during the whole infection process. In this work we focused on the interaction of the global regulatory systems agr, SarA and SigB and we demonstrate a crucial function for SigB in bacterial adaptation mechanisms and the formation of SCV-phenotypes.
For our study we generated single mutants for the functions of SigB (ΔsigB), agr (Δagr), SarA (ΔsarA) and the complemented mutant for sigB (ΔsigB compl.), three double mutants for SigB, agr, SarA (Δagr/ΔsarA, ΔsigB/Δagr, ΔsigB/ΔsarA) and a triple mutant (ΔsigB/Δagr/ΔsarA) in S. aureus LS1, a strain derived from mouse osteomyelitis [37]. Several mutants were also generated in the rsbU complemented derivative of the laboratory strain 8325–4, SH1000 [33] (Supp. S1 Table). All strains and mutants were characterized by growth curves, which did not reveal any substantial differences between mutants and parent strains (S1 Fig). Yet, differences in hemolysis were present in many mutants, particularly in the agr-, sarA-, double and triple mutants, which exhibited low hemolysis (S1C Fig). Furthermore, the strain LS1 and the corresponding mutants were analyzed by LC-MS/MS mass spectrometry, which was focused onto the culture supernatants to provide an overview on the levels of virulence factors in each strain (Fig 1A and S3 Table). These data show that particularly the sarA-mutant and even more the double and the triple-mutants released a much reduced number of virulence factors associated with disease development compared with the wild-type strain LS1 (reduced levels of virulence factors associated with disease; deep green areas, Fig 1A); in particular different adhesive proteins and toxins were present in reduced levels in culture supernatants (Fig 1B). Yet, for the FnBPs that are important for host cell invasion we detected similar levels compared with the wild-type LS1 in most mutants that can account for the invasive capacity of the strains in host cells (S5 Fig). Most importantly, the sigB-mutant showed only modest alterations in protein levels (Fig 1A), but increased levels of α-toxin that is known as a strong proinflammatory and cytotoxic factor after host cell invasion (hla, Fig 1B). The other listed toxins, e.g. lukD and lukE, are reported as non-hemolytic and only poorly leucotoxic toxins [37].
As secreted virulence factors are particularly directed against professional phagocytes, we tested the effect of bacterial supernatants on neutrophils (PMNs) isolated from humans and mice. In line with the proteomic data, only the strains with high levels of virulence factors (LS1, ΔsigB, ΔsigB compl.) caused cell death, whereas all other mutants with reduced levels of toxins induced significantly less cytotoxicity (Figs 2A and 2B and S2A–S2C). This effect was concentration dependent and revealed highest levels of cell death in response to supernatants of the sigB-mutant (S2C and S2D Fig). Next we analyzed levels of chemokine expression in cultured tissue cells, such as osteoblasts and endothelial cells, 48 h after infection by real time PCR (Fig 2C and 2D and S4C and S4D Fig) and 24 h after infection by ELISA-measurements (S3C and S3D Fig). In contrast to the wild-type strain, all double-and triple-mutants (including mutations in SigB) exhibited reduced cell activating activity, whereas the sigB-single-mutant often caused even more cell activation than the wild-type strain. Furthermore, these effects were independent of the bacterial background and of the infected host cell types, as they were reproduced with selected mutants generated in strain SH1000 (S3D Fig). All effects were equally present in bone and endothelial cells (S4C and S4D Fig) regardless of the observation that endothelial cells took up higher amounts of bacteria than osteoblasts (S4A Fig). Nevertheless, endothelial cells expressed in general lower levels of chemokines than osteoblasts (S4B Fig).
Taken together, these results suggest that the agr- and SarA-systems are required to mount an aggressive and cytotoxic phenotype during acute infection, while SigB appears to have a restraining function on virulence. Nevertheless, since double and triple mutants are weak in virulence, the interaction of SigB with the agr- and SarA-systems is not clear during acute infection.
To test the function of the global regulatory systems in the course from acute to chronic infection, we infected osteoblast and endothelial cell cultures with wild-type and mutant strains and analyzed their ability to persist intracellularly for 9 days. All strains were invasive in osteoblasts to a similar extent (S5 Fig) and induced cell death ranging around 50% immediately after infection (S6A and S6B Fig). Yet, in the following 2–3 days the integrity of the infected cell monolayers were fully recovered and the rate of cell death was reduced to control levels (S6C Fig). In general the numbers of intracellular bacteria were decreased during the whole infection course (Figs 3A and S7A), but considerable differences between the strains appeared after several days (9 days, Figs 3B and S7B). The agr/sarA- and the sigB/sarA-double mutants as well as the triple mutant were able to persist within the intracellular location at significantly higher numbers (up to 100-fold) than the corresponding wild-type strain (Fig 3A and 3B). By contrast, the sigB-mutants were completely cleared from the host cells within 7–9 days, whereas this effect could be fully reversed by the complementation of sigB (Figs 3A and S7A). To test whether this effect is specific for sigB-mutants, we further tested mutants for other virulence or regulatory factors such as for sae and hla which did not reveal any differences in the numbers of intracellular persisting bacteria compared with the wild-type strain (S8 Fig). From our previous work we know that bacterial persistence is associated with dynamic SCV-formation. As recently described [9] we discovered an increased rate of SCV-formation after several days of intracellular bacterial persistence, whereby the recovered SCV were not stable, but the majority reverted back to the wild-type phenotype upon 2 to 5 subcultivating steps on agar plates. Interestingly, in the present study we found that all sigB-mutants completely failed to develop SCV phenotypes after 7 days of intracellular persistence (Fig 3C and 3D and S7C and S7D Fig). By analyzing the recovered colonies from sigB-mutants, we observed much less phenotypic diversity than in the wild-types and other mutants, as the plates revealed only uniform large white colonies. Again these effects could be reversed by complementation of the sigB-mutations with an intact sigB-operon, thus proving a clear and specific connection between the bacterial ability to form dynamic SCVs and the SigB-system.
To further explain the differential ability of the mutants to persist, we evaluated the expression of the global regulatory systems during the long course of the infection. To accomplish this, we extracted RNA from infected host cells (HUVEC; as they can host higher numbers of bacteria, S4A Fig) at day 2 and day 7 p.i. and measured the expression of agr, sarA and sigB and the related factors hla (regulated by agrA), aur (repressed by sarA [38]) and asp23 (regulated by sigB [39]) in the wild-type strain and corresponding mutants by quantitative real-time PCR (Fig 4). As expected high levels of both agrA and sarA were only expressed by the strains LS1, ΔsigB and ΔsigB compl. that resulted in high levels of hla expression in the acute phase of the infection. By contrast, the agr/sarA-double mutant expressed sigB and asp23 at significant higher levels than the wild-type strain during chronic infection (Fig 4) and was able to form higher numbers of SCV phenotypes (Figs 3C and S7C and S7D). Taken together, our results show that a concomitant downregulation of agrA and sarA promotes long-term intracellular persistence of S. aureus. SigB promotes chronic infections and is highly associated with the bacterial ability to form SCVs.
Only recently, phagosomal escape to the cytoplasm was reported for different S. aureus strains early after host cell invasion [13,40]. In a further approach we analyzed whether an early phagosomal escape is a prerequisite for persistence. Therefore we used a reporter recruitment technique based on the host cell line A549 genetically engineered to produce a phagosomal escape marker [13]. Within the first 2 h after host cell infection we detected phagosomal escape for the wild-type strain LS1, the sigB-, the sigB compl.- and the sigB/agr-mutants (Fig 5A–5D). These strains showed only weak changes and down-regulation of virulence factor expression by proteomic analysis compared with the wild-type (Fig 1A) and were not able to persist at high bacterial numbers (Fig 3B). As the single agr and sarA mutants readily lost their ability to translocate to the cytoplasm, apparently both agr- and sarA-regulated factors are required for the escape mechanism. The activity of SarA alone is sufficient only in case of a non-functional SigB-system, indicating a modulating role of SigB in virulence factor expression. Our results show that an early phagosomal escape is not required for persistence. Further on, mutants that persisted at high bacterial numbers did not escape to the cytoplasm. As phagosomal escape could not be detected at later stages of infection (up to 24 h; as shown in for the triple mutant; Fig 5E), it must be assumed that these mutants are not degraded within phagolysosomes and thus persist in phagosomes at high bacterial numbers.
To test the ability of the wild-type and the mutant strains to establish an infection in vivo, we next performed experiments using a rat localized osteomyelitis model [41], where defined numbers of bacteria (1x106 CFU; strain SH1000 and mutants) were directly injected into the bones (Fig 6). Using this model we aimed to study the complex interaction of the bacterial strains with the immune response of the host organism. After 4 days (acute stage) and 14 weeks (chronic stage) groups of rats were sacrificed and the tibial bones were used for histology or morphometric analysis (osteomyelitis index; OI, Fig 6F) to assess for infection severity. After 4 days and 14 weeks the bacterial loads were determined by quantitative culture of bone homogenates. The data clearly demonstrated that the single mutants ΔsigB, Δagr and ΔsarA were found in lower numbers within the bones and caused less inflammation than the wild-type strain (Fig 6B–6E). We further tested as a selective double mutant the agr/sarA double mutant that persisted in high numbers within cells in culture experiments (Figs 3 and S7). By contrast, in the animal model we detected drastically reduced CFU and a lower osteomyelitis index in rats challenged with the agr/sarA double mutant when compared with rats infected with the parental wild-type strain. The agr/sarA-double mutant was found almost as avirulent as the low-pathogenic Staphylococcus carnosus strain TM300 [42], which lacks most S. aureus virulence factors (Fig 6B and 6C). Nevertheless, histological analysis 4 days p.i. revealed that in all experimental groups and control rats the bones were densely infiltrated with immune cells after S. aureus challenge (Fig 6A), indicating that both the wild-type and the mutant strains attract immune cells and sustain the infection. Only the wild-type strain, however, was able to settle and replicate at high bacterial numbers, to induce severe bone destruction and to develop into chronicity. These findings supports the hypothesis that regulators agr, SarA and SigB need to be functional to enable S. aureus to successfully survive during the whole infection process.
S. aureus is one of the most frequent causes of osteomyelitis and endovascular infections that often take a therapy-refractory or chronic course. Many clinical studies show that persistent infections are highly associated with the SCV phenotype [4,43]. Only recently, we were able to demonstrate that the formation of dynamic SCVs is an integral part of the long-term infection process that enables the bacteria to hide inside host cells, but also to rapidly revert to the fully-aggressive wild-type phenotype, when leaving the intracellular location and causing a new episode of infection [9]. Therefore, the pathogen-host interaction must be very dynamic and most likely requires global transcriptional changes on the bacterial side to promote survival of the pathogen. This study was aimed at detecting regulatory factors that mediate this dynamic infection and adaptation strategies. To this purpose, we focused on a set of important S. aureus regulators/regulatory loci agr, SarA, and SigB that are linked together in a global regulatory network. Each one of these regulators/regulatory loci is involved in the control of the expression of many virulence factors such as adhesive and cytotoxic components. For our experiments we used various in vitro and in vivo infection models to analyze the impact and dynamics of the regulators from acute to chronic infection.
In models of acute infection we demonstrated that active agr- and SarA-systems are required to cause inflammation and cytotoxicity. Our results are in line with many published infection models, showing that both factors contribute to disease development [27,44–46]. We further confirmed that strains become almost avirulent, when both factors are inactive. In case of a single mutation, agr and sarA may partly compensate for each other [47], but the compensation is not sufficient in all functions, e.g., in phagosomal escape (Fig 5) and in in vivo infections (Fig 6). In the acute stage of infection the bacteria need to express a set of virulence factors, including toxins and exoenzymes, to fight against recruited immune cells [48] and to destroy and invade deep tissue structures at high bacterial numbers. Particularly toxins and cytotoxic factors are under the tight control of agr and sarA [19,47,49]. The role of SigB in acute infection is less clear. Inactivation of sigB in S. aureus has been reported to decrease infectivity in some murine infection models [50–52], but was ineffective in others [53,54]. In line with these conflicting findings, we observed on the one hand that single mutants in sigB express higher levels of toxins, e.g., α-toxin, and become even more virulent (Fig 2C and 2D and S2C and S2D Fig) suggesting a moderating role of sigB on the expression of secreted proinflammatory factors. On the other hand double mutants lacking sigB were almost avirulent (Fig 2C) indicating a further function of SigB within the regulatory network during acute inflammation. This is supported by recent work showing a fast and transient upregulation of sigB in the first hours following host cell invasion and the requirement of SigB for early intracellular growth [55].
In the longer course of the infection bacteria can be situated within host cells, like in the cell culture models. As almost all types of host cells contain killing and clearing machineries [10], the persisting bacteria need to develop mechanisms to resist degradation that can be achieved by two different pathways according to the results from our cell culture experiments:
Certain mutants that reveal significant metabolic changes including downregulation of virulence factors (Fig 1), and do not escape from the phagosomes after host cell invasion (Fig 5) can persist within their host cells partly at higher numbers than the wild-type phenotype (Fig 3). An example of this is the triple mutant ∆sigB∆agr∆sarA that fails to form SCVs, but is largely avirulent and can persist at high bacterial numbers. Recent work demonstrated that phagosomal escape is largely dependent on the agr-regulated phenol-soluble modulins (PSMs) [6,13], but further agr or sarA regulated factors could also be involved, as single mutants in agr or sarA were already compromised in their escape [56–58]. This mechanism of persistence is restricted to strains that lack expression of agr and/or sarA-regulated virulence factors. Our results indicate that these strains are less prone to degradation and can “passively” persist inside host cells, possibly within their initial phagosomes after host cell invasion even without forming SCVs.
Further on, persistence is also possible when bacteria express virulence factors and escape from their phagosomes to the cytoplasm. Persistence obviously requires an adaptation to the intracellular environment that could be attributed to the function of SigB during the long course of the infection. SigB is an important staphylococcal transcription factor that is associated with various types of stress-responses [31,33,35,59], was shown to be upregulated in stable clinical SCVs and was associated with increased intracellular persistence [60]. According to our results SigB is also involved in stress-resistance that promotes “active” intracellular survival during long-term persistence: The first important finding is that ΔsigB-mutants were not able to persist, as they were completely cleared by their host cells within 9 days (Fig 3B). Additionally, the agr/sarA-double mutants that were found at the highest numbers during long-term persistence (Figs 3B and S7B) displayed the highest levels of sigB expression (Fig 4). Finally, a result of major significance is that SigB is required for dynamic SCV formation, as all mutants deficient in sigB were not able to form SCV-phenotypes (Figs 3C and S7C) and agr/sarA-double mutants that highly expressed sigB (Fig 4) developed the highest levels of SCVs. Only recently SigB was described as an important virulence factor in stable SCVs that mediates biofilm formation and promotes intracellular bacterial growth [61]. Yet, in our work the recovered SCVs were not stable, but rapidly reverted back to the wild-type phenotype upon subcultivation. Consequently, we describe the formation of dynamic SCVs for persistence as an additional central function that is dependent on an intact SigB-system.
Taken together, our results demonstrated that intracellular bacterial persistence is promoted by the silencing of agr- and sarA-regulated factors and/or requires an intact SigB-system. Although strains with deletions in the agr and/or sarA-system were able to persist at high bacterial numbers in cell culture systems that lack most components of a functioning immune system, they showed severe disadvantages in the in vivo model, as they were unable to defend themselves from invading immune cells (Fig 2A and 2B) and were rapidly cleared from the infection focus (Fig 6A and 6B). Consequently, SigB represents a crucial factor to dynamically adapt fully virulent wild-type strains to switch to long-term persistent phenotypes. SigB was described to turn down the agr system [36], which is most likely responsible for the enhanced inflammatory activity. The downregulated agr-system helps the bacteria to form biofilm [62] and silences aggressiveness for persistence within the host cell [63]. This expression pattern (high sigB and low agr) of global regulators appears to be characteristic for SCVs and seems to represent a general adaptation response, as it had been described for stable SCVs generated by aminoglycoside treatment as well [35]. Yet, the varying stress conditions encountered by S. aureus on its way to the intracellular location are less defined and many questions remain to be answered to fully elucidate the complete dynamic bacterial adaption strategies: e.g., (i) which intracellular conditions affect the staphylococcal regulatory factors? (ii) which staphylococcal regulatory factor(s) is/are directly influenced by the intracellular milieu (agr or sigB or further systems, such as the mazEF toxin-antitoxin module [64])? (iii) How are the changes of regulatory factors transferred to an increased SCV formation? (iv) which factors of the SigB regulon are required for persistence [65]? (v) Does SigB increase bacterial resistance against antibiotics? All questions require extensive additional laboratory work to decipher the bacterial adaptation mechanisms in more detail.
In our study we used different bacterial backgrounds and mutants, as well as different in vitro and in vivo infection models to demonstrate that bacteria apply general adaption strategies via the crosstalk of regulatory factors with a central function for SigB. By this means, bacteria can rapidly react to changing environmental conditions and dynamically adjust their virulence factor expression at any time of the infection. As the regulatory network involving SigB appears to be the central factor that enables the bacteria to persist and cause chronic infections, it represents a novel therapeutic target for prevention and treatment of chronic and recurrent infection.
The bacterial strains and mutants used in this study are listed in Supp. S1 Table. All the experiments were performed with wild type and mutants in the background of S. aureus LS1 and S. aureus SH1000. LS1 is a murine arthritis isolate that has been used in infection models before [66]. The strain SH1000 has a complementation of the rbsU gene in the strain 8325–4 which is deficient in SigB activity (a stress-induced activity) due to a mutation in the rsbU gene. This gene encodes for a phosphatase required for the release of the sigma factor SigB from inhibition by its anti-sigma factor RsbW. To create a strain with an intact SigB-dependent stress response, the rsbU gene was restored in S. aureus 8325–4, with the resulting strain called SH1000 [66].The antibiotic resistance cassette-tagged agr, rsbUVWsigB, and sarA mutations of RN6911 [67], IK181 [68] and ALC136 [69] were transduced into LS1 using phages 11, 80α and 85, respectively. ΔrsbUVWsigB derivatives were cis-complemented by phage transduction with a resistance cassette-tagged intact sigB operon obtained from strain GP268 [70] using phage 80α. The sarA mutant of SH1000 was constructed by replacing the sarA gene with an erythromycin resistance cassette. The sigB mutant of SH1000 was constructed by transducing the sigB mutation (sigB::Tn551) from RUSA168 into SH1000.
The strains were cultivated in tryptic soy broth (TSB) at 37°C with linear shaking at 100 rpm in a water bath (OLS200, Grant Instruments, England). Strains were grown in two sets. In the first set the wild type and the ∆sigB, ∆agr, ∆sarA and ∆agr/∆sarA mutants were cultured and in a second set the wild type and the ∆sigB∆agr, ∆sigB∆sarA and ∆sigB∆agr∆sarA mutants were cultured. The wild type always served as control. During exponential growth phase bacteria were pelleted by centrifugation and culture supernatants were mixed with 10% final concentration of TCA and proteins precipitated at 4°C overnight. Pelleted proteins were washed five times with 70% ethanol and then incubated for 30 min at 21°C and mixed at 600 rpm in a thermomixer (Eppendorf, Germany). Afterwards, pellets were washed once with 100% ethanol and dried in a speed vacuum centrifuge (Concentrator 5301, Eppendorf, Germany). Subsequently, protein pellets were dissolved in a suitable volume of 1x UT buffer (8 M urea and 2 M thiourea) and incubated for 1h at 21°C with shaking at 600 rpm in a thermomixer (Eppendorf, Germany). No soluble components were pelleted via centrifugation. Protein concentration was determined according to Bradford [71]. 4 μg of protein were reduced and alkylated with Dithiothreitol and Iodoacetamid prior to digestion with Trypsin.
Peptides were purified and desalted using μC18 ZipTip columns and dried in a speed vacuum centrifuge (Concentrator plus, Eppendorf, Germany). Dried peptides were dissolved in LC buffer A (2% ACN, in water with 0.1% Acetic acid) and subsequently analysis by mass spectrometry was performed on a Proxeon nLC system (Proxeon, Denmark) connected to a LTQ-Orbitrap Velos- mass spectrometer (ThermoElectron, Germany). For LC separation the peptides were enriched on a BioSphere C18 pre-column (NanoSeparations, Netherlands) and separated using an Acclaim PepMap 100 C18 column (Dionex, USA). For separation a 86-minute gradient was used with a solvent mixture of buffer A (2% Acetonitrile in water with 0.1% Acetic acid) and B (ACN with 0.1% acetic acid): 0–2% for 1min, 2–5% for 1 min, 5–25% for 59 min, 25–40% for 10 min, 40–100% for 8 min for buffer B. The peptides were eluted with a flow rate of 300 nL/min. The full scan MS spectra were carried out using a FTMS analyzer with a mass range of m/z 300 to 1700. Data were acquired in profile mode with positive polarity. The method used allowed sequential isolation of the top 20 most intense ions for fragmentation using collision induced dissociation (CID). A minimum of 1000 counts were activated for 10 ms with an activation of q = 0.25, isolation width 2 Da and a normalized collision energy of 35%. The charge state screening and monoisotopic precursor selection was rejecting +1 and +4 charged ions. Target ions already selected for MS/MS were excluded for next 60 seconds.
Analysis of MS data was performed with Rosetta Elucidator version 3.3.0.1 (Rosetta Biosoftware, MA, USA). An experimental definition for differential label free quantification was created with default settings: instrument mass accuracy of 5 ppm, spectral alignment search distance of 4 minutes, peak time width minimum of 0.1. For identification the S. aureus NCTC 8325 FASTA sequence in combination with SEQUEST/ Sorcerer was used. Oxidation of methionine, carbamidomethylation and zero missed cleavages were specified as variable modifications. After identification an automatic Peptide/Protein Tellers annotation was performed and only Peptide Teller results greater than 0.8 were used. At least two peptides per protein or one peptide with protein sequence coverage of at least 10% were necessary for reliable protein identification.
Protein intensities received with the Rosetta Elucidator software were further processed using Genedata Analyst version 7.6 (Genedata AG, Basel, Switzerland). Protein intensities were normalized using the central tendency normalization with a dynamic target and the median as central tendency. For relative quantification the ratio of the protein intensities between the wild type and the respective mutants from set one or two were calculated. A protein with a ratio of ≥2 was assigned to be upregulated in the wild type and with a ratio of ≤0.5 was assigned to be upregulated in the mutant strain. For prediction of protein localization the PSORT database was used (PSORTdb 3.0, http://db.psort.org/browse/genome?id=9009). Furthermore, for categorizing the proteins into subgroups The SEED (Overbeek et al., Nucleic Acids Res 33(17)) annotation for S. aureus NCTC8325 was used. A heat map with log2 ratios of the wild type versus the respective mutants was created using the package heatmap.plus version 1.3 in R version 2.15.1 (http://www.R-project.org).
For cell culture experiments, bacteria were grown overnight in 15 ml of brain-heart infusion (BHI) with shaking (160 rpm) at 37°C. The following day, bacteria were two times washed with PBS and were adjusted to OD = 1 (578nm). Bacterial supernatants were prepared as described [21]. Briefly, bacteria were grown in 5 ml of brain-heart-infusion broth (Merck, Germany) in a rotator shaker (200 rpm) at 37°C for 17 h and pelleted for 5 min at 3350 g. Supernatants were sterile-filtered through a Millex-GP filter unit (0.22 μm; Millipore, Bedford, MA) and added to the cell culture medium in the indicated concentrations.
For growth curves, Staphylococcus aureus wild-type LS1, SH1000 and their respective mutants were grown overnight in 15ml of Mueller Hinton infusion (MH) with shaking (160rpm) at 37°C. The following day, 100ml of MH were inoculated with each strain in order to obtain the starting optical density 0,05 (578nm). The growth of each strain was monitored spectrophotometrically (578nm) and by plating on blood agar for counting of CFU every hour during 8h. The growth rate (μ, growth speed) and the generation time (g) for each strain used in this study were calculated according the followings formulas [72].
µ=LnN−LnN0t−t0g=Ln2/µ
μ = growth rate
g = generation time
N = final population
N0 = initial population
t = final time
t0 = initial time
Various types of host cells, namely primary isolated human umbilical venous endothelial cells (HUVECs) and primary isolated human osteoblasts were cultivated as outlined before [73–75] and were infected with different S. aureus strains as previously described [14]. Briefly, primary cells were infected with an MOI (multiplicity of infection) of 50. After 3 h cells were washed and lysostaphin (20 μg/ml) was added for 30 min to lyse all extracellular or adherent staphylococci, then fresh culture medium was added to the cells. The washing, the lysostaphin step and medium exchange was repeated daily to remove all extracellular staphylococci, which might have been released from the infected cells. To detect live intracellular bacteria at different time points post infection (p.i.) host cells were lysed in 3 ml H2O (for 25cm2 bottles) or 20 ml H2O (for 175cm2 bottles). To determine the number of colony forming units (CFU), serial dilutions of the cell lysates were plated on blood agar plates and incubated overnight at 37°C. The colony phenotypes were determined on blood agar plates by a Colony Counter Biocount 5000 (Biosys, Karben, Germany).
A549 cells stably expressing the escape marker YFP-CWT in the cytoplasm were generated by transduction with lentiviral particles as described [76]. Transductants were passaged three times and subsequently selected by FACS. The phagosomal escape signal is based on the recruitment of YFP-CWT to the staphylococcal cell wall upon rupture of the phagosomal membrane barrier. Recruitment thereby is mediated by the cell wall targeting domain of the S. simulans protease lysostaphin [77]. Accumulation of the fusion protein YFP-CWT was recorded by fluorescence microscopy. For this the A549 YFP-CWT cells were infected in imaging dishes with coverglass bottom (MoBiTec, Göttingen, Germany) with the S. aureus LS1 wild type strain or the corresponding mutants for 1 h, followed by lysostaphin treatment to remove all extracellular bacteria. 2 h post infection the YFP-signal was observed with a Zeiss Axiovert 135 TV microscope (Carl Zeiss, Jena, Germany) equipped with a 50 W HBO mercury short-arc lamp and a Zeiss filter set (excitation BP 450–490 nm, beam splitter FT 510 nm, emission LP 515 nm). For quantitative assays a 100×/NA 1.3 plan-neofluar objective (field of view 25mm) was used. Data were acquired with an AxioCam MRm camera and processed using Zeiss AxioVision software. 10 fields of view were recorded per experiment and phagosomal escape was enumerated for 5 independent experiments per strain.
Human polymorphonuclear cells (PMNs) were freshly isolated from Na-citrate-treated blood of healthy donors. For neutrophil-isolation, dextran-sedimentation and density gradient centrifugation using Ficoll-Paque Plus (Amersham Bioscience) was used according to the manufacturer's instruction. Cell purity was determined by Giemsa staining and was always above 99%. PMNs were suspended to a final density of 1×106 cells/0.5 ml in RPMI 1640 culture medium (PAA Laboratories GmbH) supplemented with 10% heat-inactivated FCS (PAA Laboratories GmbH) and immediately used for the experiments. All incubations were performed at 37°C in humidified air with 5% CO2.
To measure cell death induction 1x106 cells/0.5 ml PMNs were cultured at 37°C in a 5% CO2 atmosphere in 24-well plates and bacterial supernatants were added as indicated. After 1 h cell death induction was analyzed as described previously [78,79].
For the flow cytometric invasion assay, primary human osteoblasts or endothelial cells were plated at 2x105 cells in 12-well plates the day before the assay. Cells were washed with PBS, then cells were incubated with 1 ml of 1% HSA, 25 mM HEPES (pH 7.4) in F-12 medium (Invitrogen). The bacteria were grown in BHI, washed and the bacterial suspension (OD 1, 540 nm) was prepared and was added to cells. After the lysostaphin step the infected cells were incubated for 24 h and cell death assays were performed by measuring the proportion of hypodiploid nuclei as described [8].
To measure cell death during the bacterial persistence, this protocol was performed every two days.
Hemolysis analysis was performed as described previously with some modifications [80]. The lysis efficacies of human red blood cells were measured using whole culture supernatants of S. aureus LS1, SH1000 and their respective mutants. Briefly, S. aureus cells were cultured in BHI for 17 h at 250rpm. Staphylococcal cells were centrifuged and the supernatants were used for measuring hemolytic activity. Supernatants (100μL) were added to 100μl human red blood cells (previously prepared, see preparation of erythrocytes). To determine hemolytic activities, the mixtures of blood and S. aureus were incubated at 37°C for 30 minutes. Supernatants were collected by centrifugation at 1000g for 5 min and optical densities were measured at 570nm in an ELISA plate reader. The strain S. aureus Wood46 was used as a positive control.
For measurement of cytokine release, osteoblasts were seeded in 12-well plates and stimulated with live staphylococci as described above. After the lysostaphin step, cells were washed and incubated with new culture medium for 24 h. Conditioned media were centrifuged to remove cells and cellular debris, and samples were frozen at 20°C until levels of cytokines and chemokines were measured. The levels of RANTES [regulated on activation of normal T cell expressed and secreted] was measured using immunoassays from RayBiotech according to the manufacturer`s description. Results are given as absolute values (ng/ml) of two independent experiments performed in duplicates. Statistical analysis was performed using the ANOVA test.
Outbred Wistar adult rats (groups of 10–12 rats) weighing 250–350 g were anesthetized with ketamine/xylazine. The left tibia was exposed and a hole made with a high-speed drill using a 0.4 mm diameter bit. Each tibia was injected with a 5 μl suspension containing 1x106 CFU of bacteria suspended in fibrin glue (Tissucol kit 1 ml, Baxter Argentina-AG Viena, Austria). Groups of rats were sacrificed at 4 days or 14 weeks after intratibial challenge by exposure to CO2. Both left and right tibias were removed and morphometrically assessed by measuring the Osteomyelitic Index (OI) as described previously [41]. Both epiphysis were cut off from the infected tibias and 1 cm bone segments involving the infected zone were crushed and homogenized. Homogenates were quantitatively cultured overnight on trypticase soy agar and the number of CFU was determined. The OI and the S. aureus CFU counts from each experimental group (infected and control) were compared.
For RNA extraction, we used the kit RNeasy Mini kit (Qiagen). The RNA extraction was performed following the manufacture instruction and the suggestions of the protocol described by Garzoni et al. [81].
All the primers used in this study are listed in Supp. S2 Table. Real-time PCR was performed by using the RNA isolated from infected cells, infected tissues or different bacterial isolates. The cDNA was obtained using the kit QuantiTect reverse transcription (Qiagen) and iQSYBR Green Supermix (BIO-RAD) was used. The reaction mixtures were incubated for 15 min at 95°C followed by 40 cycles of 15 s at 95°C, 30 s at 55°C and 30 s at 72°C using the iCycler from BIO-RAD. PCR efficiencies, melting-curve analysis and expression rates were calculated with the BIO-RAD iQ5 Software.
In order to analyze the expression of bacterial and host cell genes, respectively, the primers listed in S2 Table were used [14,63]. The expression analysis experiments were performed using the software CFX Manager software which calculates the normalized expression ΔΔCT (relative quantity of genes of interest is normalized to relative quantity of the reference genes across samples). The genes used as housekeeping genes to analyze the chemokine expression were GAPDH and B-actin. As controls we used uninfected host cells. For bacterial factors, we used as housekeeping genes aroE, gyrB and gmk. The results shown in each graph are normalized to that control (control expression = 1).
The isolation of human cells and the infection with clinical strains were approved by the local ethics committee (Ethik-Kommission der Ärztekammer Westfalen-Lippe und der Medizinischen Fakultät der Westfälischen Wilhelms-Universität Münster). For our study, written informed consent was obtained (Az. 2010-155-f-S). Taking of blood samples from humans and animals and cell isolation were conducted with approval of the local ethics committee (2008-034-f-S; Ethik-Komission der Ärztekammer Westfalen-Lippe und der Medizinischen Fakultät der Westfälischen Wilhelms-Universität Münster). Human blood samples were taken from healthy blood donors, who provided written informed consent for the collection of samples and subsequent neutrophil isolation and analysis.
For the local osteomyelitis model care of the rats was in accordance with the guidelines set forth by the National Institutes of Health [82]. The experimental protocol involving rats in the experiments was approved by the Institutional Committee for the Care and Use of Laboratory Animals (CICUAL), School of Medicine, University of Buenos Aires (resolution CD 2361–11).
The relationship between WT and all the different mutants was established by the one way ANOVA test with the Dunnett multi-comparison post-test. Significance was calculated using the GraphPad Prism 6.0 software and results were considered significant at P = 0.05.
|
10.1371/journal.ppat.1000790 | Lethal Antibody Enhancement of Dengue Disease in Mice Is Prevented by Fc Modification | Immunity to one of the four dengue virus (DV) serotypes can increase disease severity in humans upon subsequent infection with another DV serotype. Serotype cross-reactive antibodies facilitate DV infection of myeloid cells in vitro by promoting virus entry via Fcγ receptors (FcγR), a process known as antibody-dependent enhancement (ADE). However, despite decades of investigation, no in vivo model for antibody enhancement of dengue disease severity has been described. Analogous to human infants who receive anti-DV antibodies by transplacental transfer and develop severe dengue disease during primary infection, we show here that passive administration of anti-DV antibodies is sufficient to enhance DV infection and disease in mice using both mouse-adapted and clinical DV isolates. Antibody-enhanced lethal disease featured many of the hallmarks of severe dengue disease in humans, including thrombocytopenia, vascular leakage, elevated serum cytokine levels, and increased systemic viral burden in serum and tissue phagocytes. Passive transfer of a high dose of serotype-specific antibodies eliminated viremia, but lower doses of these antibodies or cross-reactive polyclonal or monoclonal antibodies all enhanced disease in vivo even when antibody levels were neutralizing in vitro. In contrast, a genetically engineered antibody variant (E60-N297Q) that cannot bind FcγR exhibited prophylactic and therapeutic efficacy against ADE-induced lethal challenge. These observations provide insight into the pathogenesis of antibody-enhanced dengue disease and identify a novel strategy for the design of therapeutic antibodies against dengue.
| Dengue is the most common vector-borne viral disease of humans, with over 3 billion people at risk for infection and 50–100 million infections in tropical and subtropical regions each year. Dengue virus (DV) causes a spectrum of clinical disease ranging from an acute debilitating, self-limited febrile illness (DF) to a life-threatening vascular leakage syndrome, referred to as dengue hemorrhagic fever/dengue shock syndrome (DHF/DSS). There are four serotypes of DENV; infection with one serotype is thought to protect against re-infection with the same serotype, but may either protect against or enhance infection with one of the other three serotypes. Epidemiological and in vitro data has implicated anti-DENV antibodies in mediating pathogenesis of a second DENV infection. However, it is unclear which antibody conditions are protective and which exacerbate disease in vivo, in part because no animal model of antibody-enhanced dengue disease has been available. Here, we present the first animal model of antibody-enhanced severe DENV infection. Importantly, this model recapitulates many aspects of human disease, including vascular leakage, elevated serum cytokine levels, reduced platelet count, and disseminated infection of tissue phagocytes. Furthermore, we demonstrate the utility of this model by showing that a genetically modified anti-DENV antibody that fails to bind the Fcγ receptor has prophylactic and therapeutic efficacy against lethal DENV challenge in vivo.
| The four serotypes of dengue virus (DV) are mosquito-borne flaviviruses responsible for 50–100 million human infections annually. Primary infection in individuals over the age of one year with any DV serotype is usually asymptomatic or results in self-limited dengue fever (DF), but secondary infection with another DV serotype carries an increased risk of severe disease, including life-threatening dengue hemorrhagic fever/dengue shock syndrome (DHF/DSS) [1],[2]. Fatal disease is characterized by increased vascular permeability leading to hemoconcentration and hypovolemic shock [3]. The increased severity of secondary infections is believed to result, at least in part, from antibody-dependent enhancement (ADE) of DV infection, in which FcγR engagement by antibody-virus immune complexes facilitates virus entry into susceptible myeloid cell types [4]. A role for ADE in human dengue pathogenesis is supported by observations that maternally-derived anti-DV antibodies increase the risk of DHF in infants during primary infection with DENV2 [5],[6]. Antibody-mediated increases in DV viremia have been demonstrated in macaques, but a limited number of antibody conditions were examined, and exacerbation of dengue disease by passively transferred antibodies was not observed [7],[8]. Consequently, fundamental questions about the immunology and pathogenesis of ADE have remained unanswered, and small animal models for testing antiviral interventions in the context of ADE have not been available.
Recently, we derived a mouse-adapted DV2 strain, D2S10, that produces a TNF-α-dependent fatal vascular permeability syndrome in interferon-α/β and γ-receptor-deficient (AG129) mice 4–5 days after intravenous (iv) infection with 107 plaque forming units (pfu) of virus [9]. DV infection in AG129 mice reproduces important features of human DV infection, including similar tissue and cellular tropism, viremia, vascular leakage, and elevated serum cytokine levels [9]–[12]. Antibodies elicited by DV infection are a mixture of serotype-specific and serotype-cross-reactive antibodies, including long-lasting neutralizing antibodies [13]. Memory immune responses are present after primary DV infection, and serotype cross-protective immunity was observed in three different sequential infection scenarios ([13] and data not shown). Thus, we utilized the AG129 model to examine the effects of serotype cross-reactive antibodies on DV2 infection in vivo. In this report, we demonstrate lethal enhancement of DV infection and disease by both polyclonal and monoclonal antibodies. We also show that ADE functions to increase the viral burden in blood and tissues, resulting in a vascular permeability syndrome that is similar to that seen in mice with a higher inoculum in the absence of immune antibody and that shares clinical features of human dengue disease. Finally, we confirm the critical role of FcγR interaction in ADE in vivo and provide proof-of-principle for a pre- and post-exposure treatment strategy utilizing genetically engineered monoclonal antibodies that can no longer bind FcγR.
Serum containing anti-DV1 antibodies was collected from AG129 mice 8 weeks after subcutaneous inoculation with 105 pfu of DV1 strain 98J. Heat-inactivated anti-DV1 serum exhibited a 50% neutralizing titer (NT50) against DV2 strain D2S10 of 1∶296 and against DV1 98J of 1∶1,069 using a flow-based neutralization assay [14], peak enhancement titers of 1∶75 against DV2 D2S10 (fold-enhancement 14.8%) and 1∶225 against DV1 98J (fold-enhancement 10.7%) in an in vitro ADE assay with FcγR-bearing human K562 cells, and ELISA titers of 400 and 3200 against purified DV2 and DV1, respectively (data not shown). In addition, no residual DV1 could be isolated following inoculation into C6/36 mosquito cells (data not shown). The effects of anti-DV1 serum on DV2 infection were investigated after intraperitoneal (ip) injection of 100 µl of either naïve mouse serum (NMS) or anti-DV1 serum, followed 24 hours later by iv challenge with 104–106 pfu of DV2. Lethal infection controls received 107 pfu of DV2, and all mice were monitored for mortality for 10 days. While no mortality was observed in NMS-recipient mice infected with 106 pfu or less of DV2, 92–100% of anti-DV1 recipients died after inoculation with 105–106 pfu of DV2 (Figure 1A and Table S1) between 4 and 5 days post-infection. In both the 107 pfu infection controls and anti-DV1 recipients infected with 105 or 106 pfu, lethal disease was accompanied by fluid accumulation in visceral organs characteristic of the vascular permeability syndrome induced by DV2 D2S10 [9] (Figure 1B). Mice administered anti-DV1 serum and challenged with DV2 D2S10 also experienced significant increases in serum TNF-α (p<0.01) and IL-10 (p<0.01) and greater platelet depletion (p<0.02), as compared to NMS-recipient controls (Figure 1C–F); each of these disease parameters also correlates with dengue severity in humans [15]–[17].
Viral burden was subsequently compared between anti-DV1 and NMS-recipient mice infected with 105 or 106 pfu of DV2. Viral burden was systemically increased in anti-DV1 versus NMS-recipient mice, with a 20-fold increase (p<0.02) in viremia accompanied by significant 3- to 12-fold increases in viral load in multiple tissues (p≤0.04) including peripheral blood mononuclear cells, liver, small intestine, lymph node, and bone marrow (Figure 2A); non-significant increases in the large intestine and spleen (p>0.08) and lungs (data not shown) were observed. No statistically significant differences were observed in tested disease parameters, viral burden, or tissue tropism between 107 pfu of D2S10 infection in the absence of antibody and antibody-enhanced infection with 105 pfu of D2S10. Notably, anti-DV antibodies also enhanced infection with non-adapted low-passage human DV isolates DV1 Western Pacific-74 (Figure 2B) and DV2 TSV01 (Figure 2C), as determined by significant increases (p≤0.04) in viral burden in the liver and small intestine for both viruses, and in serum for DV2 TSV01. Although mortality was not observed, a subset of animals infected with DV1 Western Pacific-74 under antibody-enhanced conditions displayed fluid accumulation in visceral organs and gross morphology similar to but less pronounced than that observed with enhanced DV2 D2S10 disease.
ADE is predicted to facilitate infection of FcγR-bearing cell types such as tissue macrophages and dendritic cells [4]; therefore, we examined the cellular tropism of DV2 in mice by immunostaining for the viral NS3 protein, which is only present during active replication of the virus. As found in humans [12],[18], infected cells with morphology and location consistent with tissue macrophages or dendritic cells [19],[20] were detected in lymph node, small intestine, large intestine, and bone marrow under all infection conditions, and NS3+ cells with endothelial and/or phagocyte morphology were also observed in liver (Figure 3 and data not shown). Infection in myeloid cells was confirmed by co-staining of serial sections and bone marrow aspirates for NS3 and the myeloid/macrophage marker F4/80 (data not shown). Furthermore, using flow cytometry, DV NS3 and E protein were detected in bone marrow cells expressing myeloid markers CD11b, CD11c, and F4/80, and in liver DV infection was primarily in CD31+CD45− sinusoidal endothelial cells, which also express FcγR (Figure S1). Notably, significantly greater numbers of NS3+ cells (p≤0.05) were present in tissues of anti-DV1 recipient mice compared to naïve serum recipient controls infected with the same dose of DV2 (Figure 3B), supporting the hypothesis that ADE functions to increase the viral burden in cells and tissues.
While serotype cross-reactive immunity is implicated in the pathogenesis of severe dengue, serotype-specific immunity typically protects against re-infection with the same DV serotype [1]. However, in vitro studies suggest that all antibodies that neutralize infection are capable of ADE at some lower concentration [21]; therefore, we examined the effects of anti-DV1 and anti-DV2 sera on DV2 D2S10 infection in mice over a range of doses. While the highest dose (400 µl) of anti-DV1 serum lethally enhanced infection (Figure 4A and Table S2), recipients of 400 µl of anti-DV2 serum developed no signs of illness and lacked detectable viremia (Figure 4B and C; Table S2), confirming that serotype-specific antibodies can provide robust protection in this model. However, lower doses of both anti-DV1 and anti-DV2 serum caused lethal enhancement, showing that serotype-specific as well as serotype-cross-reactive antibodies can also enhance infection in vivo in a dose-dependent manner (Figure 4A and B). To assess the ability of the BHK PRNT50 assay to predict in vivo protection and enhancement in this mouse model, neutralizing activity was measured in the sera of mice immediately prior to infection with D2S10. Serum was collected approximately 18 hours post-transfer of anti-DV antibodies, and 4 hours prior to infection. Similar to human studies [22], lethal enhancement occurred even in mice that had detectable neutralizing antibodies, although no lethality was observed in mice with PRNT50 values of >200.
To further define the characteristics of enhancing antibodies, we examined the ability of monoclonal antibodies (mAbs) to enhance DV disease in mice. Mice were inoculated with DV2 D2S10 24 hours after transfer of increasing amounts of the flavivirus cross-reactive, neutralizing mAb 4G2 (Figure 4D). 4G2 caused lethal enhancement at doses of 0.062–4 mg/kg (1.25–80 µg/mouse), but no mortality occurred in mice receiving 20mg/kg (400 µg/mouse) or in IgG2a isotype control antibody recipients (Figure 4D and Table S2). 4G2, anti-DV1 serum, and anti-DV2 serum all enhanced infection and disease over a ∼60-fold range in concentration.
Since FcγR engagement is required for ADE in vitro [23], we hypothesized that eliminating the ability of antibodies to bind to FcγRs would prevent ADE in vivo. To test this, we first generated F(ab′)2 fragments of 4G2. These fragments were indistinguishable from intact 4G2 in their ability to bind to DV2 E protein by ELISA (Figure S2A), but were unable to enhance DV infection of the human FcγR-bearing cell line K562 (Figure 5A). The lack of the Fc portion in the F(ab′)2 fragments of 4G2 was confirmed by ELISA (Figure S2B). In vivo, F(ab′)2 fragments have a shorter serum half-life than intact IgG, thus it was necessary to identify a dosing regimen that would maintain serum concentration of F(ab′)2 fragments within the known enhancing range for intact IgG antibodies. Serum F(ab′)2 levels were measured one and 24 hours after iv transfer of 20 µg of F(ab′)2 by E protein ELISA; this dose maintains E-reactive antibodies at levels within the range where IgG causes enhancement for 24 hours (Figure S2C). To examine the effects of intact IgG and F(ab′)2 in vivo, we compared the enhancing effects of a single dose of 4G2 mAb with daily 20 µg doses of 4G2 F(ab′)2 (Figure 5B). Whereas significant mortality was observed in 4G2 mAb recipients (p≤0.04), no illness occurred in 4G2 F(ab′)2 or IgG2a isotype control recipients (Figure 5C). Viremia measured at 3.5 days post-infection in F(ab′)2 recipients was significantly reduced (p<0.03) compared to isotype control antibody recipients (Figure 5D), suggesting that loss of FcγR interaction not only diminished enhancement but also promoted neutralization to reduce viral load.
We followed up these studies using a mAb genetically engineered to eliminate FcγR binding. MAb E60 is a flavivirus cross-reactive neutralizing mouse IgG2a antibody that, similar to 4G2, binds to an epitope in the fusion peptide of domain II on the E protein [24],[25]. This mAb was cloned and the constant regions replaced [26] with those from human IgG1 to create an E60-chimeric human IgG1 (E60-hIgG1). In addition, an asparagine to glutamine variant at position 297 in human IgG1 was engineered (E60-N297Q), as this mutation abolishes FcγR binding without altering the half-life of the antibody in mouse serum [27]. Affinity measurements conducted by surface plasmon resonance indicated that E60-mouse IgG2a (E60-mIgG2a), E60-hIgG1, and E60-N297Q all exhibited similar binding to purified E protein (Figure S3A) and DV2-infected cells by flow cytometry (data not shown), as well as similar neutralizing activity against DV2 by both PRNT50 assay (0.23, 0.25, and 0.42 µg/ml, respectively) and a neutralization assay using DC-SIGN-expressing human target cells (Figure S3B). However, as expected, E60-mIgG2a and E60-hIgG1 enhanced DV2 infection of K562 cells in vitro whereas E60-N297Q did not (Figure 6A).
To test the ability of the E60-N297Q variant to enhance DV infection in vivo, mice were administered 20 µg of E60-mIgG2a, E60-hIgG1, and E60-N297Q 24 hours prior to infection with 106 pfu of D2S10. Whereas both E60-mIgG2a and E60-hIgG1 consistently caused antibody-dependent mortality 4 to 5 days post-infection, equivalent doses of E60-N297Q caused neither morbidity nor mortality (Figure 6B). Instead, viremia and tissue viral burden in E60-N297Q recipients were substantially reduced, demonstrating that the N297Q mutation converted the in vivo effect of E60 on viral burden from enhancement to neutralization (Figure 6C, and data not shown).
The N297Q mutation also abolishes binding to complement component C1q [27]. Consequently, we generated a second variant antibody, E60-A330L, to assess whether the loss of C1q binding or the loss of FcγR binding explained the inability of E60-N297Q to mediate ADE. E60-A330L does not bind C1q but retains binding to FcγR [28], and we confirmed this by surface plasmon resonance (data not shown). E60-A330L exhibited similar binding and neutralization activity to E60-hIgG1, enhanced DV infection in vitro in K562 cells, and lethally enhanced a DV2-D2S10 infection in vivo (Figure S3B, C, and D, and data not shown). Thus, C1q interaction was not required for ADE in vitro or in vivo, and a loss of C1q binding does not explain the inability of E60-N297Q to enhance DV infection.
The experiments above suggested that an N297Q variant antibody against DV could have potential as an antiviral intervention. To assess this, 20 µg of E60-hIgG1 or E60-N297Q was administered concurrently with 25 µl of anti-DV1 serum 1 day prior to infection with DV2. E60-N297Q protected mice against any signs of illness, whereas all recipients of anti-DV1/E60-hIgG1 succumbed to infection (Figure 7A). Post-exposure therapeutic application of E60-N297Q was evaluated by administering 25 µl anti-DV1 serum to mice, followed by infection with DV2 the next day, and iv administration of E60-N297Q or E60-hIgG1 on day 1 or 2 post-infection. While all mice treated with E60-hIgG1 succumbed to infection, lethality was completely prevented by a single 20 µg dose of E60-N297Q on day 1 (Figure 7B and data not shown), and E60-N297Q treatment significantly decreased viremia, tissue viral burden, and serum TNF-α levels as measured 3.5 days post-infection (Figure 7C and 7D, p<0.04). Moreover, 20 or 50 µg doses of E60-N297Q administered on day 2 resulted in 40% and 80% survival, respectively, demonstrating therapeutic efficacy for this antibody in mice (Figure 7B).
Understanding the immunopathogenesis of DV infection has been severely hampered by the lack of a small animal disease model. Thus, studies of ADE have been limited to epidemiological observations and in vitro experimentation. Here, we present the first model of antibody-enhanced lethal dengue disease in vivo. This work describes a long-sought mouse model for investigation of dengue pathogenesis, characterizes a clinically important mechanism of immunopathogenesis, has implications for vaccine development, and identifies a possible antibody-based antiviral strategy to treat life-threatening DV infection.
Numerous attempts have been made to establish a mouse model of dengue disease and have been limited by the relatively low susceptibility of mice to DV infection. Previous models have included intracerebral inoculation of DV or injection of very high (>109 PFU) doses of virus into immunocompetent mice [29],[30]; infection of SCID [31]–[34] or NOD/SCID or RAG2γ (c)−/−mice [35],[36] implanted with human cells or cell lines; and use of various immunodeficient strains of mice [37],[38]. The most common outcome is neurovirulent disease, with a few recent exceptions [35],[36]. Of these, the AG129 mouse model has proven both useful and tractable, as it is permissive to infection with all four DV serotypes, displays relevant tissue and cellular tropism, produces long-lasting serotype-specific and serotype-cross-reactive anti-DV antibodies of a balanced isotype ratio, and generates a vascular leakage syndrome that in many respects resembles human dengue disease [9]–[11],[13],[12]. Nonetheless, we acknowledge that the lack of IFN receptors limits reproduction of some facets of human disease, especially in relation to cytokine profiles or infection conditions that are modulated by IFNs. However, the many similarities with specific features of human DV infection and the critical role for FcγR in ADE in vivo that we demonstrate here support the use of the AG129 model for specific avenues of dengue research. Interestingly, IFN-receptor deficient mice (A129) have recently been successfully adapted for other mosquito-borne viruses, including both Chikungunya and Yellow Fever [39],[40].
In vivo ADE models have also been established for other viruses, including Yellow Fever Virus (YFV), Murray Valley Encephalitis Virus (MVEV), Japanese Encephalitis Virus (JEV), and Feline Infectious Peritonitis Virus (FIPV) [41]–[46], in which passive transfer of antibody increases viral titers and/or mortality. These models show several parallels with our model of antibody-enhanced DENV infection: with FIPV, immune sera are able to enhance macrophage infection and disease during subsequent challenge with the same FIPV serotype in kittens [41],[46], and with MVEV, JEV, and YFV, enhanced mortality was observed in mice administered flavivirus cross-reactive polyclonal antibodies or non-neutralizing YFV-specific monoclonal antibodies [42]–[45]. However, none of these pathogens are associated with antibody-enhanced disease in humans. By modelling ADE with a pathogen known to cause antibody-enhanced disease in humans and using a model that displays a variety of relevant disease phenotypes, this report extends previous work on ADE to the ability to model human disease parameters and aid in the development of therapeutics.
In vivo evidence of ADE of DV infection was first described by Halstead et al [7] following the passive transfer of antibodies in the rhesus macaque. Similar data was recently obtained by Gonçalvez et al [8], where passive transfer of the serotype-cross-reactive mAb 1A5 enhanced DV4 viremia over a ∼30-fold concentration range (0.22–6 mg/kg). While both of these studies described elevated viremia, neither resulted in a clinical phenotype with parallels to human disease. Here, we describe enhancement of a mouse-adapted strain of DV2 by serotype-specific and cross-reactive sera as well as by monoclonal antibodies. Importantly, characterization of antibody-dependent dengue disease in the AG129 mouse maintains several parallels with severe disease in humans. Hallmark features of human DHF/DSS are vascular leak, higher viral burden, increased levels of serum cytokines such as TNF-α and IL-10, and platelet depletion [47]. All of these features were observed in our mouse model of ADE. Moreover, the magnitude of DV enhancement also mimics that seen in humans and non-human primates. We observed a 20-fold increase in viremia triggered by ADE; DV viremia in humans is reported to be 10- to 100-fold higher in DHF cases than in DF cases [48],[49], and ADE in macaques increases viremia 5–100 fold [7],[8]. Interestingly, in all of the disease parameters we examined, there was no apparent difference between lethality resulting from antibody-enhanced infection with a sublethal viral dose and lethality resulting from direct inoculation with a 100-fold higher viral dose. Thus, this model did not reveal any fundamental difference in the mechanisms of pathogenesis between antibody-enhanced and non-enhanced infection; rather, lethality here appears to be a result of higher viral burden, regardless of how such a burden was achieved.
To ensure that enhanced disease in the AG129 model was not solely a feature of the mouse-adapted strain, mice were infected with clinical isolates DV1 Western Pacific-74 and DV2 TSV01 in the presence of anti-DV antibodies, and enhanced viremia was observed in both cases. The lack of mortality in infections with these viruses is likely a result of the lower viral burden established by non-adapted strains even in the presence of enhancing antibody. Interestingly, mild fluid accumulation was also observed in the gastrointestinal organs in a subset of mice experiencing enhanced infection of non-adapted DV. As only a small fraction (0.5%) of human secondary DV infections results in severe disease, and some DV strains are more virulent than others based on genetic differences [50], the observed spectrum in disease severity is not surprising, but rather parallels the human condition.
Immunohistochemical (IHC) characterization of the cellular tropism associated with ADE using NS3-specific antibodies indicated infection in cells with morphology consistent with dendritic cells and tissue macrophages in the lymph node, small intestine, large intestine and bone marrow. Further characterization by flow cytometry supported the IHC data and demonstrated infection, as evidenced by both anti-E and anti-NS3 staining, in cells with surface markers of monocytes and macrophages in the bone marrow and sinusoidal endothelial cells in the liver. By both methodologies, the infected cell types identified in the murine model agree with those cells defined as the natural targets of DV in the human host [12],[18]. Interestingly, the infected cell types did not change between an enhanced and non-enhanced DV infection; rather, quantification by both IHC and flow cytometry indicated an increase in the number of infected cells. Taken together, antibody-enhanced disease appears to result in increased infection in the natural targets of DV infection and resulting pathogenesis that does not significantly differ from the disease that results when a 100-fold higher dose of DV is used in the absence of enhancing antibody.
In human infants who have acquired maternal anti-DV antibody, severe dengue can occur even when calculated neutralizing antibody titers against the secondary infecting serotype are >1∶100 [51]. Similarly, children with detectable neutralizing antibody against the infecting virus strain can develop DHF during secondary DV infections [22]. These studies indicate that the in vitro neutralization assay using BHK21 cells is not a consistent correlate of protection in humans. Similarly, our PRNT50 assays performed on serum samples from mice after antibody transfer but before virus challenge demonstrated that despite in vitro neutralizing activity at the time of infection, anti-DV1 sera, anti-DV2 sera, and 4G2 all enhanced infection in vivo. Enhanced disease was consistently observed in antibody-recipient mice with pre-infection neutralizing titers of <1∶200, but not greater. Thus, substantial neutralizing antibody levels appear to be required to prevent severe disease in this model. Of note, the passive transfer and primary infection scheme used does not examine anamnestic B and T cell immune responses, and thus, more accurately models DHF/DSS in infants with primary DV infection rather than secondary DV infections.
In vitro evidence had previously indicated that an interaction between the Fc portion of the antibody and the FcγR was necessary for ADE [8]; however, this hypothesis had never been corroborated in vivo. Using two different reagents – F(ab)′2 fragments of 4G2 and the N297Q variant of hE60-IgG1, we demonstrate that binding of the Fc portion of the antibody to the FcγR is required for ADE-induced disease. Further analysis with F(ab)′2 or the N297Q variant showed a reduction in viral titer below the level in mice receiving PBS in place of mAb. Thus, under conditions where the antibody cannot bind the FcγR, the F(ab) portion of the antibody can neutralize infection. These data also demonstrate that antibodies directed to the fusion loop in E domain II are capable of neutralizing DV infection independently from effector functions mediated by FcγR and C1q. Because the N297Q mutation also ablated the C1q binding site, we tested a second hE60 variant, hE60-A330L, that contained a mutation disrupting the complement C1q receptor binding site, but not the FcγR interaction. Mice receiving the hE60-A330L variant succumbed to an enhanced DV infection. This confirms that interaction of the anti-DENV mAb with the FcγR, and not binding of C1q, is essential for ADE in vivo.
Given the promising data with the hE60-N297Q variant, we tested the prophylactic and therapeutic efficacy of this antibody. When given as prophylaxis together with an enhancing amount of anti-DV1 serum, hE60-N297Q was completely protective. Although interesting, a DENV prophylactic is not likely to be a clinically useful reagent. However, when given 24 hours after an enhanced DENV infection, E60-N297Q completely protected against mortality; likewise, tissue viral load and systemic TNF-α levels in these mice at 3.5 days post-infection were significantly reduced. Two different doses of E60-N297Q, 20 and 50 µg, were administered 48 hours post-infection and resulted in 50% and 80% survival, respectively. Given the condensed timeframe of DENV pathogenesis in the AG129 model, E60-N297Q or similar therapeutic mAbs may have a broader time window for intervention and efficacy in humans or other animal models that display more protracted kinetics of DV infection.
In summary, we report the first animal model of lethal antibody-mediated enhancement of DV infection, describe virologic and pathologic changes induced by ADE, and define antibody conditions for protection and ADE in passive antibody transfer recipients. Furthermore, we show that antibodies engineered to prevent FcγR interaction exhibit prophylactic and therapeutic efficacy against DV infection, and thus have potential as a novel antiviral strategy against DV.
All experimental procedures were pre-approved by the UC Berkeley Animal Care and Use Committee and were performed according to the guidelines of the UC Berkeley Animal Care and Use Committee.
DV was propagated in the Aedes albopictus cell line C6/36 (American Type Culture Collection [ATCC]) as described elsewhere [52]. DV2 strain D2S10 (passaged 4 times in C6/36 cells) was derived in our laboratory [9] from the parental DV2 PL046 Taiwanese isolate as previously described [9]. The DV1 strain 98J was isolated in our laboratory from a patient from Guyana in 1998 [53] and passaged 7 times in C6/36 cells. The DV1 strain Western Pacific 74, originally isolated in Nauru in 1974, was obtained from the National Institutes for Biological Standards and Control (Hertfordshire, UK) and passaged 3 times in C6/36 cells. The DV2 strain TSV01, isolated in Townsville, Australia, in 1993 was obtained from W. Schul, passaged ∼10 times in C6/36 cells (Novartis Institute for Tropical Diseases, Singapore) [11]. Virus titers were obtained by plaque assay on baby hamster kidney cells (BHK21, clone 15) as described [52]. For mouse experiments, virus was concentrated by centrifugation at 53,000×g for 2 hours at 4°C and resuspended in cold PBS with 20% FBS (HyClone, Thermo Scientific). U937 DC-SIGN cells were obtained from A. de Silva (University of North Carolina, Chapel Hill) and grown in RPMI media (Invitrogen) at 37°C in 5% CO2. K562 cells were used for all enhancement assays and grown in RPMI media (Invitrogen) at 37°C in 5% CO2. The hybridoma of mAb 4G2 was purchased from ATCC, grown in serum-free medium (Invitrogen), and purified using protein G affinity chromatography (Thermo Scientific). Mouse mAb E60 and human E60-IgG1 (hE60), were obtained from M. Diamond, and hE60-N297Q was obtained from S. Johnson (MacroGenics, Inc.). The mouse E60 IgG2a mAb was originally generated against WNV E protein, reacts with an epitope in the fusion peptide in domain II, and cross-reacts with DV E proteins [25]. The generation of a chimeric human-mouse E60 with the human IgG1 constant regions and the mouse VH and VL was performed as described previously [26]. Point mutations in the Fc region that abolish FcγR and C1q binding (N297Q) or C1q binding alone (A330L) were introduced by QuikChange mutagenesis (Stratagene). All recombinant antibodies were produced after transfection of HEK-293T cells, harvesting of supernatant, and purification by protein A affinity chromatography.
AG129 mice [54] were originally obtained from M. Aguet (Swiss Institute for Experimental Cancer Research, Epalinges, Switzerland) and were bred in the University of California (UC) Berkeley Animal Facility. All experimental procedures were pre-approved and were performed according to the guidelines of the UC Berkeley Animal Care and Use Committee.
Cytokines were measured using commercially available ELISA kits (EBioscience). Platelet counts were obtained by diluting 20 µl of anticoagulated blood into Unopette reservoirs (BD) and counting on a hemocytometer.
Viral load was determined in the indicated tissues as previously described [52], and expressed as either pfu/g (all solid tissues) or pfu/109 cells (bone marrow and PBMCs). To obtain PBMCs, 200–300 µl of whole blood was collected into EDTA-coated microtainer tubes (Becton Dickinson) after cardiac puncture. Samples were washed 3 times in red blood cell lysis buffer (eBioscience) and once in cold PBS, and resuspended in 250 µl alpha-MEM with 5% fetal bovine serum (FBS, Hyclone), 10 mM Hepes (Invitrogen) and 100 U penicillin/100 µg streptomycin (P/S; Invitrogen).
Viral RNA was extracted from 60 µl serum aliquots using Qia-Amp Viral RNA recovery kit (Qiagen). Quantitation of viral RNA utilized Taqman reagents (One Step RT-PCR Kit, Applied Biosystems, Foster City, CA) and an ABI PRISM 7700 sequence detection system as described [55]. Viremia is expressed as plaque-forming unit equivalents/ml, which was calculated by dividing the genomic RNA copy number in each sample by the genome:pfu ratio of C6/36-derived virus as determined by plaque assay and qRT-PCR.
Tissues were collected at day 3.5 (n = 3–6 mice per group), formalin-fixed, and processed into paraffin sections. Serial sections from each tissue were stained for NS3 using MAb E1D8 or an isotype control as previously described [12]. For quantification of NS3+ cells, at least ten visual fields were counted for each sample except bone marrow, where four fields from four independent sections were counted due to the small area of mouse bone cross-sections. All pairwise comparisons were performed by two-sided Wilcoxon Rank Sum tests.
Bone marrow aspirates were collected by perfusing two femurs with cold, complete RPMI media (Invitrogen) containing 10% FBS (Hyclone), 10 mM Hepes (Invitrogen) and 100 U penicillin/100 µg streptomycin (P/S; Invitrogen). Resuspended cells were washed once in red cell lysis buffer and once in D-PBS (Invitrogen). The cells were subsequently resuspended in flow cytometry buffer containing D-PBS, 2.0% bovine serum albumin (BSA; Fisher Scientific) and 0.02% sodium azide (Sigma-Aldrich) and plated in a 96-well U-bottom plate (Becton-Dickinson) at 1×106 cells/well. Cells were blocked with 5% normal rat serum (Jackson Laboratories) diluted in flow cytometry buffer. Bone marrow cells were stained extracellularly using CD11b-PeCy7 (eBioscience), CD11c-PE (eBioscience), and F4/80-TC (Caltag) or isotype control, and then fixed in 2% paraformaldehyde (Ted Pella, Inc.), washed and permeabilized with 0.1% saponin (Sigma-Aldrich). Intracellular staining was then performed with either 1) human anti-DV E mAb 87.1 (F. Sallusto and A. Lanzavecchia, Institute for Research in Biomedicine, Bellinzona, Switzerland) or isotype control mAb hIgG1 WNV-E16 (M.S. Diamond) followed by secondary goat anti-human IgG conjugated to Alexa488 (Invitrogen) or 2) mouse anti-NS3 mAb E1D8 conjugated to Alexa488 (Invitrogen) or isotype control (mIgG2a-Alexa488 (Invitrogen)). Livers were harvested into 10 mL cold, complete RPMI media and subsequently digested using 20 mg/mL collagenase VII (Sigma-Aldrich), washed, and the digested tissue passed over a 70-µM cell strainer (Fisher). The resulting cells were centrifuged over an Optiprep gradient (14.7%/22.2%), washed once with D-PBS, and plated in a 96-well plate at 1×106 cells/well. The cells were stained extracellularly using CD31-PE (eBioscience), fixed in 2% PFA, permeabilized with 0.1% saponin, and stained intracellularly with either anti-E or anti-NS3 mAbs or isotype control as above. Data was collected using either an LSR II or FC-500 flow cytometer (Becton-Dickinson) and analyzed using FlowJo v8.8.6 software (TreeStar).
Monoclonal antibodies at a concentration range of 12.5 to 200 nM were injected over the surface of a Biacore 3000 instrument with immobilized E protein (∼300 RU) at a flow rate of 30 µl/min for 120 seconds and a dissociation time of 180 seconds. Binding curves at concentration zero were subtracted as blank. Kinetic parameters were calculated by fitting binding curves to a bivalent analyte binding model. The kinetic parameters were similar for binding of both mAb variants to E protein, as the difference between affinities is less than two-fold.
4G2 F(ab)′2 fragments were generated using the F(ab)′2 Preparation kit (Pierce) according to the manufacturer's instructions. To ensure that the F(ab)′2 fragments did not contain residual Fc portions, the 4G2 F(ab)′2 proteins were diluted in SDS-PAGE loading dye, boiled, and electrophoresed on a 10–20% Tris-glycine gel (Bio Rad) and stained with Colloidal Blue (Invitrogen) overnight. To measure the stability of F(ab)′2 fragments in vivo, sera from mice given different amounts of F(ab)′2 were tested by ELISA for DV2 E protein binding. In brief, ELISA plates (Fisher Scientific) were coated with 2 µg/ml of recombinant DV2 E protein (Hawaii Biotech Inc.) in carbonate coating buffer, pH 9.6 overnight at 4°C. The plate was blocked for 1 hour at room temperature in 5% nonfat dry milk and 5% donkey serum (Jackson Laboratories) in PBS-0.5% Tween 20. After washing, 50 µl of serum containing intact 4G2 or F(ab)′2 4G2 diluted 1∶10 in blocking buffer was added to the plates. After washing, 100 µl of either goat anti-mouse anti-F(ab)′2 (Jackson Laboratories) or goat anti-mouse anti-Fc (Jackson Laboratories) diluted 1∶1000 in PBS-T was added as secondary antibody. Biotinylated mouse anti-goat antibody (Jackson Laboratories) was added as a tertiary antibody, followed by streptavidin-alkaline phosphatase (Zymed). P-Nitrophenyl phosphate (PnPP; Sigma Aldrich) was added as the substrate, and the reaction was stopped with 3M NaOH and read in an ELX-808 ultra microplate reader (Bio-Tek Instruments) at 405 nm.
Serial 3-fold dilutions of antibodies were mixed with DV2 D2S10 virus at a multiplicity of infection (MOI) generating 7–15% infection of U937 DC-SIGN cells in a 96-well U bottom plate as described previously [14]. After infection for 24 hours, the cells were washed once with flow cytometry buffer and fixed in 2% PFA for 10 minutes at room temperature. The cells were then permeabilized in FACS buffer with 0.1% saponin (Sigma Aldrich) and stained with 2.5 µg/mL 4G2-Alexa 488 (Invitrogen). The cells were washed twice, and percent infection determined by flow cytometry on a Beckman Coulter EPICS XL flow cytometer. The resulting raw data was expressed in GraphPad Prism 5.0 software as percent infection versus log10 of the serum dilution, and a sigmoidal dose-response curve with a variable slope was applied to determine the antibody titer coinciding with a 50% reduction in infection as compared to the no-serum control (NT50). The plaque reduction neutralization test (PRNT) was performed in duplicate as described previously [13].
Serial 3-fold dilutions of antibody were mixed with DV2 D2S10 virus in duplicate for 45 min at 37°C, then mixed with K562 cells at MOI of 1 for 48 hours [23] in a 96-well plate. The cells were subsequently washed once with FACS buffer and fixed in 2% PFA for 10 minutes at room temperature. To stain, the cells were permeabilized in FACS buffer with 0.1% saponin (Sigma Aldrich), and then stained with 2.5 µg/mL 4G2-Alexa 488 (Invitrogen). The cells were washed twice, and percent infection was determined by flow cytometry on a Beckman Coulter EPICS XL flow cytometer. The resulting data was expressed as percent cellular infection versus log10 of the serum dilution in Microsoft Windows Excel.
Kaplan-Meier survival curves were used to display mortality data, and log rank analyses were used to determine statistical significance between experimental groups. Non-parametric analyses using the two-sided Wilcoxon rank sum tests were used for pairwise comparisons of viral load, cytokines, and platelet counts. A Fisher's exact test was used to examine survival on day 4 post-infection in F(ab′)2 experiments because the instability of F(ab′)2 fragments necessitated comparison at a single time point. Calculations were performed in GraphPad Prism 5.0 software.
|
10.1371/journal.pntd.0001419 | In Vivo Expression of Salmonella enterica Serotype Typhi Genes in the Blood of Patients with Typhoid Fever in Bangladesh | Salmonella enterica serotype Typhi is the cause of typhoid fever. It is a human-restricted pathogen, and few data exist on S. Typhi gene expression in humans.
We applied an RNA capture and amplification technique, Selective Capture of Transcribed Sequences (SCOTS), and microarray hybridization to identify S. Typhi transcripts expressed in the blood of five humans infected with S. Typhi in Bangladesh. In total, we detected the expression of mRNAs for 2,046 S. Typhi genes (44% of the S. Typhi genome) in human blood; expression of 912 genes was detected in all 5 patients, and expression of 1,100 genes was detected in 4 or more patients. Identified transcripts were associated with the virulence-associated PhoP regulon, Salmonella pathogenicity islands, the use of alternative carbon and energy sources, synthesis and transport of iron, thiamine, and biotin, and resistance to antimicrobial peptides and oxidative stress. The most highly represented group were genes currently annotated as encoding proteins designated as hypothetical, unknown, or unclassified. Of the 2,046 detected transcripts, 1,320 (29% of the S. Typhi genome) had significantly different levels of detection in human blood compared to in vitro cultures; detection of 141 transcripts was significantly different in all 5 patients, and detection of 331 transcripts varied in at least 4 patients. These mRNAs encode proteins of unknown function, those involved in energy metabolism, transport and binding, cell envelope, cellular processes, and pathogenesis. We confirmed increased expression of a subset of identified mRNAs by quantitative-PCR.
We report the first characterization of bacterial transcriptional profiles in the blood of patients with typhoid fever. S. Typhi is an important global pathogen whose restricted host range has greatly inhibited laboratory studies. Our results suggest that S. Typhi uses a largely uncharacterized genetic repertoire to survive within cells and utilize alternate energy sources during infection.
| Salmonella enterica serotype Typhi is the cause of typhoid fever and infects over 21 million cases and causes 200,000 deaths each year. S. Typhi only infects humans and this has greatly limited studies of S. Typhi pathogenesis. To study bacterial gene expression in human hosts, we used Selective Capture of Transcribed Sequences (SCOTS) and array hybridization to identify S. Typhi mRNAs expressed in the blood of 5 patients with S. Typhi infection. In total, we detected the expression of 2,046 S. Typhi genes (44% of the S. Typhi genome) in human blood; of these, 1,320 (29% of the S. Typhi genome) had significantly different levels of detection in human blood compared to in vitro cultures. Our results provide insight into S. Typhi pathogenesis, identifying both previously described and novel interactions occurring between host and microbe during the natural course of human infection. Further study of these genes, especially those of unknown function, may further our understanding of S. Typhi pathogenesis and aid in vaccine, diagnostic, and/or drug target development.
| Salmonella enterica serotype Typhi is a Gram-negative bacterium and the cause of typhoid fever. Typhoid fever affects over 21 million people each year, killing 200,000 [1]. S. Typhi is a human-restricted pathogen and this has greatly limited studies of S. Typhi pathogenesis. Our current understanding of S. Typhi responses during infection is largely based on the study of murine models with the related bacterium S. Typhimurium (i.e., a bacteria that causes a typhoid-like illness in mice) [2], a separate mouse model of S. Typhi infection [3], and ex vivo macrophage and epithelial cell models of S. Typhi and S. Typhimurium [4] However, these studies have limitations, and do not fully replicate human disease. For instance, despite high sequence similarity, 13% of the genes in the S. Typhi genome are absent from S. Typhimurium, and the S. Typhi chromosome contains over 200 pseudogenes that S. Typhimurium does not [5], [6].
Here we report the application of an mRNA/cDNA capture and amplification technology, Selective Capture of Transcribed Sequences (SCOTS), combined with cDNA hybridization technology [7]–[12], to directly assess the gene expression profile of S. Typhi in the blood of humans with typhoid fever in Bangladesh. We previously applied this technology to S. Paratyphi A, the 2nd leading cause of enteric fever, and detected expression of over 1700 bacterial genes during human infection [12]. Here we report the extension of this analysis to S. Typhi.
This study was approved by the Ethical and Research Review Committees of the International Centre for Diarrhoeal Disease Research, Dhaka, Bangladesh (ICDDR,B) and the Human Research Committee of Massachusetts General Hospital; the study was conducted according to the principles expressed in the Declaration of Helsinki/Belmont Report. Written informed consent was obtained from all individuals or their guardians prior to study participation.
Individuals presenting to the International Centre for Diarrhoeal Disease Research, Bangladesh (ICDDR,B) Hospital or the Kamalapur field site of ICDDR,B were eligible for enrollment if they met the following criteria at presentation: age of 1–59 years, fever duration of 3–7 days (≥39°C), no obvious focus of infection, and no alternate diagnosis. We collected 2 ml of venous blood from participants, immediately placed these specimens in TRIzol (Invitrogen Life Technologies, Carlsbad, CA) at a 1 (blood)∶2 (TRIzol) volume ratio, and stored the samples at −70°C for later analysis. We simultaneously obtained 3–5 ml of blood for microbiologic analysis using a BacT/Alert automated system. We sub-cultured positive bottles on MacConkey agar, and identified S. Typhi isolates using standard biochemical tests and reaction with Salmonella-specific antisera [13]. After we collected blood, we treated patients with parenteral ceftriaxone, oral ciprofloxacin, or oral cefixime for up to 14 days at the discretion of the attending physician.
To generate S. Typhi cDNA from blood samples, we used TRIzol-preserved blood of patients whose initial cultures were subsequently confirmed to grow S. Typhi. To create a corresponding in vitro S. Typhi cDNA sample for comparison, we grew each patient's bacterial isolate to mid-log growth phase (OD600 0.45–0.6) in Luria Bertani (LB) broth, and preserved the samples in TRIzol at a 1 (mid-log culture)∶2 (TRIzol) volume ratio. We extracted total RNA from TRIzol preserved samples per the manufacturer's instructions (Invitrogen) and treated recovered RNA with DNase I on RNeasy columns (Qiagen Inc., Valencia, CA). We then converted 5 µg of total RNA into cDNA for each sample, as previously described with a few modifications [12]. Briefly, we used random priming (T-PCR) to obtain a representative amplifiable double-stranded cDNA population by using Superscript III (Invitrogen) with a conserved primer with a defined 5′ end terminal sequence and a random nonamer at the 3′ end [14]. We then synthesized second strands using the same primers and Klenow fragment (Invitrogen) according to the manufacturer's instructions, and then equilibrated samples based on 16S S. Typhi rRNA.
We separated bacterial cDNA from host DNA using SCOTS, as previously described [12]. Briefly, we mixed denatured biotinylated S. Typhi gDNA with blocking ribosomal S. Typhi DNA, and added this denatured mixture to both in vivo and in vitro cDNA samples. After hybridizing samples overnight at 67°C, we captured biotinylated S. Typhi gDNA-cDNA hybrids using streptavidin-coated magnetic beads (Dynabeads M-280 streptavidin, Invitrogen), eluted captured cDNA with NaOH, PCR-amplified cDNA samples with conserved primers, and purified products using Qiagen PCR column purification kits. We performed three rounds of capture and amplification to separate S. Typhi cDNA from host DNA and to generate the cDNA mixture used for microarray hybridization.
We labelled in vivo and in vitro cDNA recovered from SCOTS with Cy3 and Cy5, respectively, and hybridized these preparations to Salmonella ORF microarrays (version STv7S; McClelland Laboratory, Vaccine Research Institute of San Diego, CA, http://www.sdibr.org/Faculty/mcclelland/mcclelland-lab) in duplicate and with two dye reversals as previously described [12]. These microarrays contained gene-specific PCR-products of 4,600 ORFs from Salmonella enterica serotype Typhi CT18 (98.6% genome coverage) and 4,318 ORFs of strain Ty2 (98.0% genome coverage. The arrays also contained 1049 S. enterica ORFs absent from the S. Typhi genome. We used an equal amount of in vivo and in vitro Cy dye-labeled product on all slides for a given patient. We used ScanArray software (ScanArray express, version 3.0.1) to quantify signal intensities.
For each individual patient, we considered a gene to be detected in vivo if at least 2 of the 3 replicate gene spots on each of the four slides for that infected human was at least ten median absolute deviations greater than the median of spots on the microarray corresponding to genes absent from the S. Typhi CT18 or Ty2 genomes. For those genes we detected in vivo, we evaluated whether there was a difference in expression when compared to detection levels for in vitro grown organisms. For this latter statistical analysis, we included genes with a coefficient of variation in signal intensity less than 50% within an array, and employed repeated measures ANOVA (to within slide replicate spots) with type (in vivo versus in vitro) and dye effects to LOESS-normalized, log-transformed data. Those genes with a False Discovery Rate of less than 0.05 computed using Benjamini-Hochberg multiple testing adjustment and a 2-fold variation in signal intensity were considered differentially expressed in vivo versus in vitro. We deposited data in the NCBI Gene Expression Omnibus (GEO, www.ncbi.nlm.nih.gov/geo), accessible through GEO accession number GSE30565. We based functional classification of genes on J. Craig Venter Institute annotations (http://cmr.jcvi.org/tigr-scripts/CMR/CmrHomePage.cgi).
We used quantitative real time PCR (RT-qPCR) to confirm microarray results for a subset of genes. We compared mRNA levels in the peripheral blood (in vivo sample) of infected patients (i.e. the 5 patients included in our SCOTS array analysis and 5 additional patients) to three in vitro culture replicates of a S. Typhi isolate (from Patient 1) grown to mid-logarithmic phase in LB (in vitro sample), as previously described [12]. To maximize the likelihood of detecting differences in gene expression in comparative samples, we selected eight representative genes from operons involved in intra-cellular invasion or survival (STY4609, sopE, invasion-associated secreted protein; STY3639, trxA, thioredoxin); alternate energy usage (STY2244, pduB, putative propranediol utilization protein; STY0417, psiF, phosphate starvation-inducible protein; STY2701, eutN, a putative ethanolamine utilization protein; STY0634, fepC, a ferric enterobactin transport ATP-binding protein); and bacterial adhesion (STY0207, staA, putative fimbrial protein and STY4543, pilO, putative pilus assembly protein), focusing on genes with high baseline signals and fold-increases by SCOTS-cDNA hybridization analysis comparing in vivo (high signal) to in vitro (low signal) samples. We also quantified by RT-qPCR the expression levels of two house-keeping genes that were predicted by SCOTS-cDNA hybridization to be equally expressed in in vivo and in vitro samples (STY0724, encoding a glutaminyl-tRNA synthetase, glnS; and STY3081, encoding an enolase, eno). We were unable to reproducibly assess expression levels of genes predicted by SCOTS-cDNA hybridization to be down-regulated in blood samples compared to in vitro grown organisms. To generate cDNA for quantitative RT-PCR from TRIzol-preserved samples, we used SuperScript II (Invitrogen) with random hexamers (Sigma, St. Louis, MO) according to the manufacturer's instructions, and performed RT-qPCR analysis using iQ SYBR Green Supermix reagent (Bio-Rad; Hercules, CA) and a CFX96 Real-time PCR detection system (Bio-Rad; Hercules, CA) as previously described [12]. Primers are listed in the Supplemental Table S1. We used no-template controls and samples lacking reverse transcriptase as baseline reactions for each sample. After calculating the threshold cycle (CT) in the low/linear portion of product curves, we quantified gene copy numbers using pGEM-T Easy-based plasmids (Promega, Madison, WI) containing the gene of interest. To calculate the control gene copy number, we used plasmid size and A260 readings, and normalized gene copy numbers based on cDNA copies of 16S rRNA. We assessed singularity of product species and size by melting curve analysis, as previously described [15].
Of the 89 patients enrolled for blood sample collection, we identified 10 patients with confirmed S. Typhi bacteremia at the time of TRIzol-preserved blood collection. We performed SCOTS-cDNA hybridization screening analysis using samples from patients 1–5, and performed RT-qPCR on samples from patients 1–10, as sample quantity permitted.
Using SCOTS-cDNA hybridization technology, we detected expression of 2046 S. Typhi genes in the blood of bacteremic patients. This represents approximately 44% of the S. Typhi ORFeome (Figure 1A, Supplemental Table S2). Of these, we detected expression of 912 genes in all 5 patients (45% of detected transcripts), and 1100 in at least 4 of the 5 patients (54% of detected transcripts).
The products encoded by the 1100 genes identified in 4 or more patients fell into a number of functional categories (Figure 1B). The most highly represented group were genes currently annotated to encode hypothetical proteins or proteins designated as unknown or unclassified. The next most highly represented groups were genes that encode products involved with energy metabolism, transport and binding, followed by genes encoding products of the cell envelope or associated with cellular processes and pathogenesis. Ninety-five of the 1100 genes were located within known Salmonella pathogenicity islands (SPI 1–7, 9, 10, 13, and 16), and 29 are known components of the PhoP regulon, a major virulence regulon in Salmonella, involved in intra-macrophage survival.
A total of 31 genes were detected in 4 or more patients in vivo, but not detected in any in vitro sample (Table 1). The majority of these genes are involved with survival in nutrient-limited conditions including psiF, a phosphate starvation-inducible protein; bioF and thiG involved in vitamin biosynthesis; eutD, oadG, and pduB involved in use of alternative carbon sources; and fepD involved in iron acquisition.
Of the 2046 transcripts detected in human blood samples, 1320 (representing 29% of S. Typhi ORFeome) had significantly different levels of detection in in vivo samples compared to bacterial samples grown in vitro (Figure 2A, Table S2). Detection levels for 141 transcripts were significantly different between in vivo and in vitro samples in all 5 patients, and 331 in at least 4 patients. These 331 encode products that fall into a number of functional categories (Figure 2B). The most highly represented group included proteins annotated as hypothetical, unknown, or unclassified. Other highly represented groups included energy metabolism, transport and binding, the cell envelope, and cellular processes and pathogenesis.
To confirm S. Typhi mRNA expression levels in human blood compared to in vitro grown bacteria, we used RT-qPCR to assess the copy number of the following eight genes that had high in vivo baseline reactivity as well as fold-change between in vivo and in vitro samples by SCOTS array analysis: thioredoxin, trxA (STY3639); a putative fimbrial protein, staA (STY0207); an invasion-associated secreted protein, sopE (STY4609); a putative propranediol utilization protein, pduB (STY2244); a putative pilus assembly protein, pilO (STY4543); an phosphate-inducible starvation protein, psiF (STY0417); a putative ethanolamine utilization protein, eutN (STY2701); and a ferric enterobactin transport ATP-binding protein, fepC (STY0634). Compared to expression levels in in vitro grown bacteria, we found increased expression of all 8 genes in the blood of infected humans, including in humans not analyzed by the SCOTS-cDNA hybridization screening protocol (Figure 3, A–H). As predicted by our SCOTS screening, we found no differences by RT-qPCR in the expression of housekeeping genes glnS (STY0724) and eno (STY3081) in blood versus in vitro bacterial samples (Figure 3, I-J).
S. Typhi is a human-restricted pathogen, the cause of typhoid fever, and a significant cause of global morbidity and mortality. Despite this, there are limited data on bacterial events within humans infected with S. Typhi. Here we describe the application of a cDNA capture-amplification approach combined with microarray hybridization technology to assess S. Typhi gene expression directly in the blood of infected humans. In total, we detected 2046 S. Typhi transcripts in human blood (45% of S. Typhi transcriptome); we detected 1100 in at least 4 of 5 patients. Two major virulence determinants of Salmonella are the ability to invade host cells and the ability to survive and replicate within host cells. The PhoPQ-two component regulatory system is involved in intra-macrophage survival and antimicrobial resistance [16], and Salmonella pathogenicity island-1 (SPI-1) and SPI-2 encode type three secretion systems (T3SSs) involved in invasion of host cells and intracellular survival and replication, respectively [17], [18]. In our analysis, we identified 29 genes involved in the PhoP regulon as more highly expressed in human samples, including the two component regulator itself, phoPQ; virk, a virulence protein; mgtBC, involved in magnesium transport; pmrF, a antimicrobial resistance protein; and slyB, an outer membrane lipoprotein [19], [20]. We also identified 95 genes located within previously described SPIs, including SPI-1 and 2, as well as genes within SPI-3–7, 9, 10, 13, and 16.
The role of SPI-1 in invasion of epithelial cells has been well established [21]. We detected a number of transcripts associated with SPI-1 genes, including a number that encode effector proteins injected into eukaryotic cells via the SPI-1 T3SS, such as SipB. We also detected a number of transcripts encoding SPI-1 T3SS effector proteins expressed from other SPIs, including sopE (expressed from SPI-7) and sopB/sigD (expressed from SPI-5); SopB/sigD is involved in creation and maintenance of the Salmonella Containing Vacuole (SCV), crucial to intra-cellular survival of Salmonella in eukaryotic cells [22]. Of note, we similarly identified SPI-1 transcripts in our recent analysis of S. Paratyphi A cDNA in the blood of infected humans in Bangladesh [12]. Our detection of these transcripts in the blood of infected humans builds upon recent suggestions that the SPI-1 T3SS is involved in pathogenic events beyond intestinal epithelial cell invasion during enteric fever [23]–[25]. In addition to sopE, we also detected transcripts from the Type IV pilus operon encoded within SPI-7, including pilL, pilO, pilQ, pilR, pilU, and pilV, which facilitates invasion of Salmonella into epithelial cells and monocytes [26], [27]. Identification of SPI-7 genes in our analysis is of particular interest since SPI-7 is absent from S. Typhimurium and S. Paratyphi A, but present in S. Typhi, S. Paratyphi C, and S. Dublin [28].
In addition to those associated with SPIs and the PhoPQ regulon, we detected transcripts from a number of virulence-associated Salmonella genes in human blood. These include aromatic amino acid biosynthesis pathway genes (aroG, aroD, aroH, aroE, aroB); mutations in this pathway have been the basis of live attenuated S. Typhi vaccines [29]. We also detected transcripts from genes involved in purine biosynthesis (guaB, purG, purA) [30] and divalent cation transport including Mg2+ (corA, mgtBC) [31]–[33], and Fe 2+ and Mn2+ uptake systems (sitBC and mntH) [34] that have all been associated with virulence in Salmonella.
In order to adapt to the intracellular environment, Salmonella must alter its metabolism to available nutrient and energy sources. We detected transcripts of genes involved in the use of alternative carbon sources, the coenzyme B12-dependent 1,2-propranediol utilization pathway (encoded by the pdu operon), and the ethanolamine utilization pathway (encoded by the eut operon). We also found these operons to be up-regulated in our analysis of S. Paratyphi A genes detected in the blood of humans [12], and mutations in these operons result in attenuation of virulence in S. Typhimurium infection models [35]–[37]. We also identified transcripts expressed from genes encoding three NiFe-uptake hydrogenases that have been associated with virulence in S. Typhimurium, including hydrogenase A, B and D [38]. Prior studies have shown that the hya and hyd operons are upregulated in murine and human phagocytes; hya genes are required for survival within macrophages, and both hya and hyd genes were detected in mice using the RIVET (Resolvase In-Vivo Expression Technology) reporter system that identifies genes expressed in vivo [39]. Our analysis shows that these genes are also expressed by S. Typhi during human infection. Other potential virulence-associated genes that we identified included genes involved in thiamine biosynthesis (e.g. thiG, thiJ, abpA), biotin biosynthesis (e.g. bioB, bioF, kbl), iron acquisition via siderophore biosynthesis (e.g. iroA gene cluster, fes, fepECDB), and phosphate transport (ugpBAEC operon), many of which were also detected in our transcriptional analysis of S. Paratyphi A in infected humans [12].
In addition to survival in nutrient-limited conditions, Salmonella must also be able to survive the action of antimicrobial peptides, oxidative killing, and nitric oxide in various ecologic niches within the human body. We detected genes that may be involved in survival of stressful environments, including a number involved in antimicrobial resistance (e.g. pqaB, virK, pmrF, smvA, bacA, emrA, mdtC) [40]–[45], oxidative stress (e.g. trxA) [46], resistance to acid tolerance (e.g. narZYWV operon) [47], and genes involved in DNA recombination and repair (e.g. recA, recBD, recN, recG, xthA) [48]. Of note, the most highly represented group were genes currently annotated to encode hypothetical proteins or proteins designated as unknown or unclassified.
When comparing expression levels of S. Typhi genes detected in our analysis in humans to expression levels of S. Typhi genes in in vitro grown cultures, equilibrating for S. Typhi 16S rRNA, we noted differing levels of S. Typhi mRNA for 65% of the genes detected in humans. In total, 331 S. Typhi transcripts had significantly different levels of detection in at least 4 patients compared to in vitro cultures, and 141 had significant differences in all 5 patients compared to mRNA detected in in vitro cultures. Identified genes were involved in iron (fepB, fepC, fepD), thiamine (thiG), and biotin (bioF) metabolism; use of alternative carbon sources including ethanolamine (eutB, eutC, eutD, eutA, and eutN), oxacelatate (oadAB and oadG), and propranediol (pduB and pduK); and antimicrobial resistance (bacA, mdtC). We also identified these operons in our analysis of S. Paratyphi A, further supporting a potential role of these operons in the pathogenesis of enteric fever [12]. In addition, we identified 24 genes with significantly different levels of expression in in vivo compared to in vitro samples that are not present in the S. Typhimurium genome and may play an important role in S. Typhi pathogenesis, including genes encoded within the Type IV pilus cluster of SPI-7 (i.e. pilO and pilL), and fimbrial proteins staA and steD. Of note, the largest grouping of S. Typhi genes identified in our comparison encoded proteins of unknown or unclassified function.
Our findings are similar to prior Salmonella transcriptional analyses. We previously applied SCOTS-microarray analysis to S. Paratyphi A in the blood of infected humans, and the homologs of 75% of the bacterial transcripts identified in S. Paratyphi A infected patients were also identified in S. Typhi infected patients [12]. SCOTS analysis has also been previously applied to S. Typhi using an ex vivo macrophage model system by Faucher et al. [9]. Similar to our current analysis using blood of infected patients, the ex vivo analysis also detected transcripts of genes involved in intracellular survival including a number of genes encoded within SPI-2, mgtBC in SPI-3, the SPI-1 effector, sopE, and genes involved in antimicrobial peptide resistance. Both analyses suggested a role of SPI-1 beyond invasion of the intestinal epithelium and the potential role of alternative carbon sources in S. Typhi pathogenesis. In contrast to Faucher's analysis, we found higher levels of transcripts of genes involved in iron acquisition and transport in vivo including fes, fhu, feo, iro, and ent. Our detection of these genes may reflect a greater complexity or degree of iron-limitation in the blood of infected humans versus in a cultured macrophage model system.
To our knowledge there has not been a prior analysis of S. Typhi gene expression across the transcriptome in humans. Our results highlight potential survival adaptations of S. Typhi within the human host, including expression of genes required for utilization of alternative carbon and energy sources, divalent cation transport, antimicrobial resistance, and oxidative stress resistance, as well as many genes whose function is currently unknown. Further study of these genes, especially those of unknown function, may further our understanding of S. Typhi pathogenesis and aid in vaccine, diagnostic, and/or drug target development.
|
10.1371/journal.pcbi.1002094 | A Dynamical Model Reveals Gene Co-Localizations in Nucleus | Co-localization of networks of genes in the nucleus is thought to play an important role in determining gene expression patterns. Based upon experimental data, we built a dynamical model to test whether pure diffusion could account for the observed co-localization of genes within a defined subnuclear region. A simple standard Brownian motion model in two and three dimensions shows that preferential co-localization is possible for co-regulated genes without any direct interaction, and suggests the occurrence may be due to a limitation in the number of available transcription factors. Experimental data of chromatin movements demonstrates that fractional rather than standard Brownian motion is more appropriate to model gene mobilizations, and we tested our dynamical model against recent static experimental data, using a sub-diffusion process by which the genes tend to colocalize more easily. Moreover, in order to compare our model with recently obtained experimental data, we studied the association level between genes and factors, and presented data supporting the validation of this dynamic model. As further applications of our model, we applied it to test against more biological observations. We found that increasing transcription factor number, rather than factory number and nucleus size, might be the reason for decreasing gene co-localization. In the scenario of frequency- or amplitude-modulation of transcription factors, our model predicted that frequency-modulation may increase the co-localization between its targeted genes.
| Transcription is a fundamental step in gene expression, yet it remains poorly understood at cellular level. Textbooks are full of descriptions of promoter-bound transcription factors recruiting RNA polymerase, which initiates transcription before sliding along the transcription unit. However, increasing evidence supports the view that the DNA template bound with transcription factors slides through a relatively immobile RNA polymerase at discrete nuclear sites (known as transcription factories), rather than RNA polymerase sliding along DNA template. Based on this transcription factory model, we build a virtual space in which genes and transcription factors move randomly while transcription factories are immobile. We find that under a large number of parameter ranges, this simple dynamical model is valid for a number of experimental observations. Moreover, we suggest the occurrence of gene co-localization might be mainly due to limited numbers of transcription factors, rather than other factors such as nucleus size or transcription factory number. This work offers insight into the general principles of regulation of transcription and gene expression by simulating the translocation of transcriptional units (genes and transcription factors) using purely random diffusion processes that result in non-random organization of co-regulated genes.
| A central theme in the regulation of transcription is the binding of transcription factor proteins to specific sites along the DNA. Though these sites can be several tens or hundreds of kilobases from a target gene promoter, regulation is achieved by the formation of chromatin loops that bring the sites together to form transcriptional hubs. It is thought that proximity between distal regulatory elements and their target genes increases the local concentration of specific regulatory factors to affect transcriptional control. Recent studies have also shown that active genes co-localize in the nuclear space at focal concentrations of the active form of RNA Polymerase II (RNAPII) called transcription factories [1], [2], [3], . A genome-wide enhanced 4C (e4C) screen demonstrated that specific combinations of genes from different chromosomes share factories with a high frequency, suggesting that active genes have preferred transcription partners. Co-localization of these spatial gene networks at transcription factories was found to be dependent on the transcription factor Klf1, which co-regulates many of the partners [7]. Just as distal regulatory elements are thought to affect gene regulation by spatial clustering, intra- and inter-chromosomal associations between co-regulated genes may affect expression by creating specialized microenvironments that are optimized for their transcription. Thus, the transcriptional program of a cell may be reflected by, or may even be dependent upon, the spatial organization of the genome. The appreciation that a very large proportion of the genome is transcribed with relatively few transcription sites suggests that the organization of the transcriptional machinery plays a major role in shaping the nuclear organization of the genome. The positioning of genes, regulatory sequences and transcription factors in relation to each other and to landmarks in the nucleus, such as nuclear bodies and lamina, are important determinants in gene expression [8].
How specific subgroups of active genes and transcription factors come to be positioned at factories is still unknown. Gaining an understanding of the emergence of complex spatiotemporal patterns of behavior from the interactions between genes in a regulatory network poses a huge scientific challenge with potentially high industrial pay-offs [4], [9], [10], [11], [12]. Experimental techniques to dissect regulatory interactions on the molecular level are critical to this end. In addition to experimental tools, mathematical modeling and computer tools will be indispensable. As most genetic regulatory systems of interest involve many genes connected through interlocking feedback loops, an intuitive understanding of their behavior is hard to obtain. By explicating hypotheses on the topology of a regulatory network in the form of a computer model, the behavior of possibly large and complex regulatory systems can be predicted and explained in a systematic way. One of such recent examples is described in Misteli [13] and Rajapakse et al [14], where the authors developed a model based upon self-organization. It is probably the most successful model in the area, as confirmed in Misteli [13]. However, the model is phenomenological with an oversimplified system of Kuramoto oscillators and the random effect is largely ignored.
Here we have developed a model based upon known experimental data, aiming to account for experimental results and for predicting and guiding further experiments. Co-localization ratio is introduced to characterize the gene co-localizations in transcription factories. Within a wide parameter region, we demonstrate that gene co-localization is plausible for both two and three dimensional cases. Experimental data tells us that sub-diffusion is observed in various cell-cycle phases (S and G phase) in yeast [15], which implies that fractional Brownian might be required to model gene movement, at least locally in time and space. Using fractional Brownian motion, gene and gene pairs co-localized with transcription factories are estimated, and tested against experimental data obtained from RNA-immuno-FISH experiments. We find that the model, albeit simple, can account for observed experimental data.
Previous research [1] showed that mouse embryonic fibroblasts (MEFs) which have flattened nuclei have relatively high numbers of transcription factories (∼2,000) while embryonic, fetal and adult erythroid cells and normal adult spleen, adult thymus, and fetal brain cells with spherical nuclei, have fewer transcription factories (100–300). Therefore, one may wonder what effects varying numbers of transcription factories and nuclear shape have on gene co-localization. Our simulation using both flattened and spherical nucleus, showed that co-localization is not very sensitive to the number of transcription factories, but is sensitive to the number of transcription factors.
Transcription factor entry to the nucleus may occur in two ways: either in a frequency-modulation mode (NF-B for example [16]) or via amplitude modulation (Klf1 might be an example) fashion. Recently Cai et al. observed that Crz1, a stress-response transcription factor, translocates to nucleus in response to extracellular calcium signal, showing short bursts of nuclear localization [17] (frequency modulation). They proposed that frequency-modulation, rather than amplitude-modulation, of localization bursts of transcription factors may be a control strategy to coordinate gene responses to external signals. Interestingly, we found that frequency-modulation, in comparison with amplitude-modulation, facilitates gene co-localization. This might reveal a key advantage of frequency modulation over amplitude-modulation in coordinating gene expression in cell nuclei.
Recent studies show that active genes dynamically co-localize to shared transcription sites and that specific networks of genes share factories at high frequencies [1], [7], [18], [19]. We built a model by randomizing the movement of genes and transcription factors, with a defined number of immobile transcription factories [6], [18]. Live cell studies have shown that chromatin is highly mobile but regionally constrained within eukaryotic nuclei. We therefore created a defined space for random diffusion of genes (genes restricted to a square in the 2D model or a cubic in the 3D model) based on the observed mobility of chromatin in vivo. Each gene regulatory element is regarded as a point for simplicity, rather than as a polymer. The transcription process is simulated as follows: when a gene and its transcription factor come within a defined proximity, they bind and diffuse together for a limited time (which is called binding time). If the bound complex encounters a transcription factory before their binding time elapses, the gene engages with the polyerase and becomes active, remaining associated with the transcription factory until termination (the transcription time); if the gene-factor complex does not encounter a transcription factory during the binding time, the gene and factor dissociate and continue their brownian motion separately. During a productive transcriptional event, the transcription factor may be released from the gene before it finishes transcription and is available to randomly interact and bind another gene of the same family.
We examined the behavior between genes of the same family (X genes) and two different families of genes (X genes and Y genes) (see Fig. 1A). Genes and their corresponding factors (X factors and Y factors) are allowed to move randomly within the restricted region. However, to simulate the constrained diffusion of chromatin, the genes cannot exit the defined space. Transcription factors are allowed to exit the space, but exit of one factor is followed by entry of an identical factor on the opposite/same side of the space (We have simulated both cases and found no significant difference in the simulation). This maintains the concentration of factors within the space and simulates the observed behavior of factors to explore the entire nucleus. The number of genes and factors for each family are equivalent. X factors can only bind to X genes and Y factors to Y genes, and the two families of genes and factors are independent of each other. We also assume that each gene has only two states (Fig. 1B and 1C) — either it is being transcribed (X gene: u(t) = 1; Y gene: v(t) = 1) or not being transcribed (X gene: u(t) = 0; Y gene: v(t) = 0). Fig. 1B and 1C illustrate X-X gene co-localization and X-Y gene co-localization, respectively, where u(t) is the state function of X gene and v(t) the state function of Y gene.
The choices of parameter values are based on the literatures and previous experimental observations, and are presented in Table 1. For example in our 3D model, the number of genes, factors and factories in our restricted region (about 1/50 in terms of volume) is proportional to the total number of active genes in nucleus (10,000–20,000 alleles, around 2000 transcription factors, and 100–300 factories observed per nucleus, reviewed by [20]). The size of genes, transcription factors and factories were also chosen in consistency with biological data. Since transcription factor binding sites are often clustered in regulatory elements in chromatin, we have given genes a binding radius of 25 nm. The diffusion rate of a gene (0.001 µm2/sec) and transcription factor (0.01 µm2/sec) is chosen in consistency with previous work [20], . The volume of the restricted region is based on Chubb et al's results (displacements of genes were in the range between 0.5 µm to 2 µm [22]). The gene-factor binding time is on the timescale of seconds [23], and transcription time is consistent with the speed of the polymerase across a relatively small gene [24].
All results above tell us that there exist co-localization regions between genes, even though the model is set up completely symmetric (equal number of genes and factors for each family). We further compared our simulation results with recently obtained experimental data (Fig. 5A) to validate our model.
Schoenfelder et al [7] reported the intra- and inter-chromosomal co-localization frequencies of 33 mouse genes relative to the Hbb and Hba globin genes in erythroid tissues (Fig. 5A). Gene regulated by the transcription factory Klf1 preferentially clustered in factories containing high levels of Klf1. Fig. 5Aa shows the spatial distribution of transcription factor Klf1 (Kruppel like factor) relative to RNAPII factories by immunofluorescence in mouse erythroid cells. The data exhibits nearly all nuclear Klf1 foci overlapped with RNAPII-S5P foci, indicating 10–20% of transcription factories contain high levels of Klf1. Therefore, we restrict 20% of Klf1 associated level with RNAPII (as the background association level) by selecting the factor number and binding time in our simulation (Fig. 5B). Fig. 5Ab is the double-label RNA immune-FISH of nascent transcripts (Hbb, green) and Klf1 foci (red). This image shows the positions of transcriptionally active, Klf1-regulated gene (e.g., Hbb) relative to Klf1 foci. It is found that majority (59%–72%) of actively transcribed alleles of Hbb, Hba, Hmbs and Epb4.9 (regard as X genes) were preferentially associated with Klf1 transcription factories. Cpox genes (regard as Y gene) associate with Klf1 factories at marginally higher frequencies (26%) than expected by a purely random distribution. For actively transcribed alleles of the Klf1-independent Tubb5 and Hist1 genes (regard as Z genes), they show no preferential localization to Klf1-containing factories (20%). Hence, we regard Klf1 as the X factor, and the X gene - X factor association level is estimated to be around 64% in experiments, while Y gene – X factor association level is around 20% from this experiment, matching the Klf1 background association level (20%). Therefore, we understand that X factors and Y genes are independent to each other. Fig. 5Ac is the triple-label RNA immune-FISH for pairs of nascent transcripts (Hbb and Hist1, blue and green, respectively) and Klf1 foci (red). From experimental observation, this colocalizing pair of genes relative to Klf1 foci reveal that colocalizing pairs of Klf1-regulated genes are associated with Klf1 transcription factories at very high frequencies (63–79%), and the colocalizing Klf1-independent gene pairs show no preferential association with Klf1 transcription factories.
In simulation, we calculated the association level between X gene (Hbb, Hba, Hmbs, and Epb4.9) and X transcription factor (Klf1), the association level between Y gene (Cpox) and X factor (Klf1), when confining the Klf1 background association level as 20%. Using sub-diffusion (H = 0.4) to simulate the translocation of regulatory elements, we calculated the X gene – X factor association level (Eq. (25)) both in 2D and 3D, by fixing gene number but varying factor number and binding time (Fig. 5Ba), or by fixing factor number but varying gene number and binding time (Fig. 5Bb). It is clearly shown that the experimental data can be well matched with our model with one set of parameters (gene number 5, factor number 2, and binding time 130 s) in 3D case.
Next we examine the co-localization between a pair of genes (X genes, Y genes or Z genes) and Klf1 (X factor). Our simulation result shows that the co-localization of paired X genes with Klf1 is 0.8, and paired X-Y genes with Klf1 is 0.6, again in agreement with experiments [7]. All experimental results and our simulation results except for Z genes are summarized in Table 2.
We tried to understand why X-Z gene pairs (Hbb/Tubb5, Hba/Tubb5, Hbb/Hist1, Hba/Hist1) association level with Klf1 factors is low. One method is to introduce interactions (negative correlation) between X gene transcription factors and Z gene transcription factors or X genes and Z genes themselves. In other words, Z genes (or Z factors) might be negatively correlated with X genes (or X factors), while Y genes and X genes are independent. To assess this, we ran simulations with the following exclusive rules: if an X gene (factor) is in a factory, it will prevent the entry of a Z gene (factor) with a probability p. In factor case, it simply implies that Z gene is co-regulated by X and Z factors. The simulation results on the gene-factor association level (Fig. 5Bc) did not show much difference after including the preventing probability among genes or factors, and it is not easy to simultaneously fit the experimental data which implies that Z gene – X factor association level as 0.2, indicating the negative correlation between different families of genes (factors) might not be the primary reason for different values of Klf1 association rate among different families of genes, as observed from experiments [7]. Hence, more sophisticated interactions are required, and we will further investigate this phenomenon in our future work.
Osborne et al. [1] shows approximately 2000 transcription factories in the extended and flattened nuclei of mouse embryo fibroblast. In contrast (Fig. 6A), they found that erythroblast, B cell, T cell and fetal brain cells, which have spherical nuclei with significantly smaller radii and nuclear volumes, have dramatically fewer transcription factories (100–300 per nucleus). It was argued that the large differences in factory numbers seen in nuclei from tissues versus cells grown on a surface appear genuine and may be a consequence of a reduced potential for inter-chromosomal sharing of factories in flattened cell nuclei [28]. To test how the changes in nuclei shape and transcription factory number will effect on gene co-localization, we ran simulations with flattened cells, squashing the original cubic from 2×2×2 to 0.5×4×4 but maintaining its volume (Fig. 6A). We have also tested the situation when the flattened cell is of volume five times bigger than the spherical cell (0.5), according to the experimental observation (data available by request). We partitioned the flattened cell into four subunits (0.5×2×2) and restricted the translocation of genes within each subunit, so all genes are restricted locally for consistency with experiments while transcription factors are free to move within the entire space. We found, consistent with our ‘limited resource’ theory, that no matter if we increase the volume of the flattened cell or not, the colocalized transcription is increased rather than reduced in the flatten cell, and is almost independent of the number of transcription factories (Fig. 6A). Moreover, it is also observed from Fig. 6A that increasing the volume of nucleus might increase the co-localization ratio. We will investigate the reason mathematically in our future research work.
From the analysis above, we propose the possibility to reconcile the facts observed in experiments and our model simulations: increasing the transcription factors might be the only possible mechanism to prevent gene co-localization. To confirm this, we ran the simulations with different number of transcription factors (Fig. 6B), for different volumes of the flattened cell. The left panel of Fig. 6B (unchanged volume) clearly demonstrated that when the number of transcription factors is around 30, the colalization ratio is reduced to around 0.5 (non-co-localization) and is independent of the number of transcription factories. When the nuclear volume is enlarged 5 times bigger that of the original spherical nucleus (Fig. 5B right panel), co-localization is even easier to happen for various cases, but increasing the number of transcription factory can hardly be the only reason for higher chance of co-localization.
In eukaryotic cells, external signals can modulate the expression of target genes by regulating the nuclear versus cytoplasmic localization of transcription factors. Experimentally, we have observed two possible types of modulations: one is amplitude modulation, implying that external signals regulate a static number of transcription factors into the nucleus (Klf1 might be an example [7]); the other is frequency modulation, in which external signals alter the frequency of nuclear bursts (entry/exit cycles) of the transcription factor (for example, p53 [29], NF-κB [16] and Crz1 [17]).
Cai et al observed that the nuclear localization burst frequency of Crzl, a transcription factor that regulates more than 100 target genes, increases in response to the increase in extracellular calcium concentration [17]. In addition, they suggested and experimentally verified that this frequency-modulation mechanism of transcription factor localizationcan coordinate the expression of multiple target genes, whereas amplitude-modulation cannot.
We assessed whether co-localization is affected by these two different modulations, using our model. Our previous model setting is equivalent to a (fixed) amplitude modulation scenario where the number of transcription factors is kept as a constant in the nucleus. We investigated whether frequency-modulation of factors could be involved in the control of multiple target gene co-expression compared to amplitude-modulation. In our simulations, we regard Crzl as X factor (no Y factor is present), and assume that Crzl binds to two families of target genes (X1 gene and X2 gene), which have completely different diffusion rates, radiuses and transcription times (binding time as 10 sec, and other detailed parameters are presented in Table 3). We ran our simulations in 3D cubic with a sub-diffusion (H = 0.4).
Fig. 7A demonstrates the dynamics of transcription factor translocation into and out off the nucleus with various frequencies, and Fig. 7B shows the factor entry profile into nucleus under frequency-modulation. Each period is composed of an active burst part and refractory part. During the refractory time, only very few (residual) factors are in nucleus, and as a result, only very few genes can be transcribed. While in the active burst time (2 min as reported in [17]), many transcription factors swarm into the nucleus, and diffuse as the model described in the amplitude modulation case.
In Fig. 7C and Fig. 7D, we illustrate the evolution of the normalized expression level of two kinds of genes (X1 and X2 gene) as a function of the factor amplitude and burst frequency, respectively. Clearly, X1 gene and X2 gene yield uncoordinated expression patterns under amplitude modulation; while the curves of X1 and X2 gene normalized expression levels almost coincide under frequency modulation, as suggested in [17].
Now we are in the position to assess the impact of frequency and amplitude modulation on gene co-localization. To this end, we have two types of factors and corresponding genes. One type is frequency modulated (X gene and X factor), the other is amplitude modulated (Y gene and Y factor). Like the symmetric model we simulated in above sections, X factor and Y factor bind to X gene and Y gene respectively. Besides, X gene and Y gene have identical properties (parameters are in Table 1). For comparability, the average number of X factors in the whole simulation time should be the same as the static number of Y factor. Unlike the previous symmetric models, here X gene co-localization ratio differs greatly from Y gene co-localization ratio, as indicated in Fig. 7E. For all burst frequencies, X gene's co-localization ratio is larger than Y gene, which implies that genes regulated by frequency modulated factors may colocalize in the nucleus more than genes of amplitude regulated factor. Although in our simulation there is almost no X-X gene co-localization event in refractory time of burst, X-X gene co-localization event in active time happens more than Y-Y genes since the average number of X factor in active time is larger than Y factor number. We conclude that in additional to the coordination of target gene expressions, another functional role of frequency modulation of factor entry may be to facilitate co-localization between target genes. This can be one interesting biological experiment to evidence whether frequency modulation allows higher co-localization and higher levels of coordinate expression of groups of genes.
In the current paper, we have investigated whether a simple diffusion model can account for the co-localization observed in experiments, based upon parameters measured from experiments. We first assess the ratio of gene co-localization. It is found that the co-localization ratio is determined by the inter-co-localization intervals and is biased. We then applied the theory and numerical simulations to two and three dimensional cases with standard and fractional Brownian motion. We have shown that the experimentally observed co-localization is possible in both two and three dimensional cases and conclude that our dynamical model can match many experimental data.
However, a direct comparison with experimental data is still not easy since we do not have data of the dynamics of multi-genes. All experimental results are static results [7]. With the development of new experimental techniques, we expect that the dynamic data should be available soon. Such data would be valuable for us to understand the interactions between genes.
It is clear that our model is a simplified version of gene mobilization in the nucleus: each gene is treated as independent (fractional) Brownian motion which is only true in local loci and small time intervals (Fig. 1A) [30], [31]). The transcription process is also a simplified process. Moreover, we tried to introduce negative correlation between different families of genes and factors with preventing probability p. However, the simulation results on the gene-factor association level (Fig. 5Bc) did not show much difference after including the preventing probability among genes or factors. Hence, more sophisticated interactions are required, and we will further investigate this phenomenon in our future work.
In the models above, all genes are treated as a point (point model). Modelling of genes as segments on 3D chromosomes as polymer chains [32] would be more appropriate. The 3D whole genome conformation will be based on Hi-C data (see Lieberman-Aiden et al [33]). The dynamics of each polymer chain will be modeled according to the well known polymer physics [34], in collaboration with our experimental data. The simulation would be computationally very expensive and would therefore need to be run on state-of-the-art clusters. The interactions in the model between genes (chromosomes) etc. should fit well with the known experimental data accumulated in our experimental teams for the past years.
After having a biophysically realistic model (with some coarse-grain approaches), we would expect to use the model to predict some key stages of hematopoietic differentiation. These predictions will then be tested by our experimental groups. Certainly this would be a very challenging task and it is a multi-scale spatio-temporal dynamics. Ideally we should be able to predict key decision making mechanisms at the molecular and cellular level that control genome function and may lead to the lymphoid versus myeloid differentiation. The transcription factories story might fit well with some general computational principle as reviewed in Oehler et al [35].
In general, we have to take into account the interactions between genes, both in cis and in trans, between genes and transcriptions, and between genes and transcription factories. As mentioned in Methods section (Eq. (1)), we can include the interactions between transcription units in the drift terms [36]:where m is the mass, xiT is the position of transcription factor of X gene, Ψ is the shape parameter of the harmonic potential, ci is the centre of each factory and Nf is the number of transcription factories, kB is the Boltzman constant, ςT is a friction constant, Te is the temperature and KH(t) is a kernel so that the fluctuation-dissipation theorem holds true. For each gene, it obeys similar equationbut with a potential depending on whether its corresponding transcription factor is in a factory or not (the term Ψ(xT(t)). How to find the right parameters of the interactions in the equations above would be an interesting issue. In the past decades, many techniques have been developed, mainly using the idea of Markov chain Monto Carlo and Bayesian approaches (for example, Pavliotis and Stuart [37]). With the drift term introduced here, we could expect that sub-diffusion has a larger co-localization region than super-diffusion.
Assume that we have m X genes and k Y genes, with n transcription factors of X gene and l transcription factors of Y gene. Denote their positions at time t asAll genes and transcription factors move according to diffusion processes, i.e.(1)where b is the drift term depending on the global activity in the nucleus Θ, is the diffusion coefficient of transcription element z (where z can be x, y, xT or yT) and Bz(t) is the independent fractional Brownian motion. The genes and transcription factors move around with a constant diffusion coefficient(2)(3)The drift term summarizes the interactions between association of centromeres, clustering of co-regulated genes, association of a regulatory element and its target genes, interaction of a genome region with the nuclear envelope etc [13]. For a gene h, define a sequence of stopping (binding) times for X genes and Y genes as(4)(5)with τh0 = 0, ωh0 = 0, and j = 0, 1, …. Moreover, Tb is the binding time of a transcription factor, is the distance, ε0 is the minimal distance between gene and factor if they are not bound (in simulation we set ε0 = rg+rf, where rg and rf are the radiuses of gene and factor, respectively), and XT and YT are the sets of all available (unbinded) transcription factors at time t, i.e.(6)Once a transcription factor binds to a gene of the same family, they will move together with gene diffusion coefficient σg (which is much slower than transcription factor diffusion coefficient σT), i.e.,(7)When the bound gene-factor enters a factory, transcription starts. The transcription time for both X and Y gene are given by(8)(9)where j = 0, 1, …, ξ0h,s = 0, η0h,s = 0, Fs is the sth factory, s = 1, 2, …, Nf, and Tt is the transcription time length. We have(10)(11)The physical meaning is clear: when the transcription starts, the gene is frozen and stays in the factory. For a given factory s, we can calculate the co-localization event. Define the co-localization event as the counting process of the inter co-localization interval Tc,xx between one X gene and another X gene (see Fig. 1b) asWe can define the co-localization event between X gene and Y gene Nxy(s,[0 T]) similarly. Let uh,s(t), vh,s(t) (or uh,sT(t), vh,sT(t)) be the indicator function of the gene (or factor) transcription event of the sth factory for the hth gene (X or Y). Note that each process uh,s(t), vh,s(t) (or uh,sT(t), vh,sT(t)) is a dichotomous random process.
The quantity we intend to calculate is(12)However, there is a problem if we calculate the ratio as above. When we count the events of Nxx(s,[0 T]), the population size is m (m−1), but for Nxy(s,[0 T]), it is mk. Hence we define(13)as the co-localization ratio of X gene (co-localization ratio of Y gene can be similarly defined as ryy/yx). When rxx/xy is larger than 0.5, an X gene tends to transcript with another X gene more often in a factory. Let us first confirm that rxx/xy is independent of time and converges to a constant rapidly. From the definition of Nxx(s, [0 T]), it is the counting process of a renewal process with the inter co-localization interval Tc,xx. From the renewal theorem [38], [39] we know that when T→∞,(14)Hence, as T is large enough we should have(15)Therefore whether there is a co-localization event in the nucleus is completely determined by the inter co-localization interval distribution Tc,xx and Tc,xy.
Is the standard Brownian motion good enough to match the experimental data? Chromatin loci are highly mobile but their motion is restricted within confined volumes. Each gene is constrained by interactions with immobile nuclear structures. Although chromosomes are relatively static, individual chromatin domains undergo Brownian motions and can extend far beyond the edges of their chromosome territory.
A normalized fraction Brownian motion BH(t) is a continuous-time Gaussian process starting at zero, with mean zero, and having the following covariance function(16)where H, called the Hurst index or Hurst parameter associated to the fractional Brownian motion, is a real number in [0,1].
The value of H determines what kind of process the fraction Brownian motion is:
According to Eq. (16), when H is greater than 0.5, it moves faster than the normal diffusion (H = 0.5), hence it is called superdiffusion. We also use fractional Brownian motion in the model developed in the previous subsection.
To compare with experimental data, next we introduce some quantities which are experimentally measurable. The transcription rate of an X gene is(17)where E[⋅] stands for the expectation, and Nf is the number of transcription factories. This means that for an X gene, its transcription rate depends on the number of factories and the probability that this gene is being transcribed over time inside each factory. Similar definition can be given for Y gene transcription rate. The co-localization between two X genes is defined as(18)describing the probability that when an X gene is being transcribed in a factory, another X gene is also being transcribed in the same factory at the same time. Similarly the co-localization ratio for an X and a Y gene is given by(19)When X genes and Y genes are independent, we have(20)The X-X genes co-localization ratio defined before is simply given by(21)This should give us a clear explanation why we call rxx/xy the co-localization ratio. Therefore, when two X and X genes are colocalized, it should have
Defined(22)as the X transcription factor association rate with all factories, where NT is the number of transcription factors and χ(w) is the indicator function, i.e.We assume that all processes are stationary. When uTi,j(t) is sparse, we have(23)Hence the transcription factor association rate r1(uT) is simply Nf NT E[uT1,1(t)]. The advantage of our approach over the experimental is that we have a dynamical model and we can concentrate on each individual transcription factory. To this end, we will concentrate on the dynamic behaviour of a single transcription factory: peer through one single factory. Under the ergodicity assumption, we intend to match the modelling results with experimental results which are obtained with spatio average. Hence we drop the transcription factory subscript j from now on. In our simulations, for a fixed number of transcription factors, we find a binding time so that X factors (as Klf1 in Schoenfelder et al [7]) are colocalized with the factories with a rate(24)The colocalizaton ratio between the first X gene and X factors is given by(25)This gives us the Klf1-associated ratio for the first X gene (say, Hbb, Hba, Hmbs and Epb4.9 in Schoenfelder et al [7]) per factory. Again, when the event is sparse, we have(26)Similarly,(27)is the Klf1-associated rate with Y genes or Z genes. When they are sparse and independent, it equals NT E[uT1,1(t)]. A pair of X genes colocalized ratio per factory with Klf1 is(28)Since X and Y genes are independent in our model, the co-localization ratio of an X and a Y gene pair with Klf1 becomes(29)When they are sparse and independent, we have(30)
|
10.1371/journal.ppat.1005761 | CD4+ T Cells Expressing PD-1, TIGIT and LAG-3 Contribute to HIV Persistence during ART | HIV persists in a small pool of latently infected cells despite antiretroviral therapy (ART). Identifying cellular markers expressed at the surface of these cells may lead to novel therapeutic strategies to reduce the size of the HIV reservoir. We hypothesized that CD4+ T cells expressing immune checkpoint molecules would be enriched in HIV-infected cells in individuals receiving suppressive ART. Expression levels of 7 immune checkpoint molecules (PD-1, CTLA-4, LAG-3, TIGIT, TIM-3, CD160 and 2B4) as well as 4 markers of HIV persistence (integrated and total HIV DNA, 2-LTR circles and cell-associated unspliced HIV RNA) were measured in PBMCs from 48 virally suppressed individuals. Using negative binomial regression models, we identified PD-1, TIGIT and LAG-3 as immune checkpoint molecules positively associated with the frequency of CD4+ T cells harboring integrated HIV DNA. The frequency of CD4+ T cells co-expressing PD-1, TIGIT and LAG-3 independently predicted the frequency of cells harboring integrated HIV DNA. Quantification of HIV genomes in highly purified cell subsets from blood further revealed that expressions of PD-1, TIGIT and LAG-3 were associated with HIV-infected cells in distinct memory CD4+ T cell subsets. CD4+ T cells co-expressing the three markers were highly enriched for integrated viral genomes (median of 8.2 fold compared to total CD4+ T cells). Importantly, most cells carrying inducible HIV genomes expressed at least one of these markers (median contribution of cells expressing LAG-3, PD-1 or TIGIT to the inducible reservoir = 76%). Our data provide evidence that CD4+ T cells expressing PD-1, TIGIT and LAG-3 alone or in combination are enriched for persistent HIV during ART and suggest that immune checkpoint blockers directed against these receptors may represent valuable tools to target latently infected cells in virally suppressed individuals.
| The persistence of HIV in a small pool of long-lived latently infected resting CD4+ T cells is a major barrier to viral eradication. Identifying cellular markers that are preferentially expressed at the surface of latently infected cells may lead to novel therapeutic strategies to cure HIV infection. We identified PD-1, TIGIT and LAG-3 as markers preferentially expressed at the surface of infected cells in individuals receiving ART. CD4+ T cells co-expressing these markers were highly enriched for cells carrying HIV. Our results suggest that PD-1, TIGIT and LAG-3 may represent new molecular targets to interfere with HIV persistence during ART.
| Although antiretroviral therapy (ART) is highly effective at suppressing HIV replication, viral reservoirs persist despite treatment and lead to rapid viral rebound when ART is interrupted [1–4]. A major step to achieve natural control of HIV replication after ART cessation would be to eliminate, or at least reduce, the number of long-lived infected cells from which HIV reignite infection. The characterization of cell surface markers that could identify HIV-infected cells persisting during ART is a research priority towards an HIV cure [5] as it could lead to the development of novel eradication strategies.
Several subsets of CD4+ T cells harbor replication-competent HIV during ART. These CD4+ T cells are usually defined on the basis of their differentiation stage [6–8], functionality or homing potential [9,10]. Central memory (TCM) and transitional memory (TTM) CD4+ T cells were identified as the major cellular reservoirs for HIV during ART [6]. More recently, a less differentiated subset of long-lived cells with high self-renewal capacity, the stem-cell memory CD4+ T cells (TSCM), has been identified as a main contributor to long-term HIV persistence [7,8]. The functional and homing capacities of CD4+ T cells also dictate their capacity to serve as persistent reservoirs for HIV: Th17 and Th1/Th17 CD4+ T cells as well as cells expressing CCR6 and CXCR3 show increasing contribution to the viral reservoir with duration of ART [11,12].
Immune checkpoint molecules (ICs) are co-inhibitory receptors which down-modulate immune responses to prevent hyper-immune activation, minimize collateral damage, and maintain peripheral self-tolerance [13]. ICs are up regulated upon T-cell activation and constrain the effector response through feedback inhibition. Overexpression of these molecules is associated with T-cell exhaustion and dysfunction in cancer and chronic viral infections, including HIV [14–17]. We hypothesized that ICs, through their ability to inhibit T-cell activation, will favour HIV latency during ART, and that CD4+ T cells expressing ICs would be enriched for persistent HIV in individuals receiving ART.
We focused our analysis on 7 ICs, namely PD-1 (programmed cell death-1), CTLA-4 (cytotoxic T-lymphocyte-associated protein 4), LAG-3 (lymphocyte activation gene 3), TIGIT (T-cell immunoglobulin and ITIM domain), TIM-3 (T cell immunoglobulin and mucin 3), CD160 and 2B4 (CD244).
PD-1, a member of the B7-CD28 superfamily, enforces an inhibitory program that blocks further TCR-induced T-cell proliferation and cytokine production [18,19]. In HIV infection, high levels of PD-1 are associated with T cell exhaustion [14–16,20] and incomplete immunological response to ART [21]. CTLA-4, a CD28 homolog, regulates the amplitude of T-cell activation by both outcompeting CD28 in binding CD80 and CD86, as well as actively delivering inhibitory signals to T cells [13]. TIGIT, which also belongs to the B7/CD28 superfamily, acts as a co-inhibitory molecule by directly down regulating proliferation of human T cells [22], but also by modulating cytokine secretion of DCs, decreasing IL-12 and enhancing IL-10 productions [23]. TIGIT has been recently associated with CD8+ T-cell dysfunction during HIV infection [24]. The expression of 2B4 (CD244), a member of the signalling lymphocyte activation molecule (SLAM) is also modulated on T cells during HIV infection [17,25]. LAG-3, a member of the immunoglobulin superfamily, is structurally highly homologous to the CD4 receptor and share MHC-II as a ligand [26]. Its expression on T regulatory cells plays a role in the modulation of T cell homeostasis and effector T cell responses [27,28]. TIM-3 is also an immunoglobulin superfamily member and its expression is increased on HIV-specific CD8+ and CD4+ T cells [29,30]. Finally, CD160, through its binding to its ligand Herpes Virus Entry Mediator (HVEM), an atypical member of TNF-receptor superfamily, delivers a co-inhibitory signalling to CD4+ T cells or CD8+ T cells dampening their activation in HIV-infected individuals [31,32].
To assess the relationship between the expression of these ICs and HIV persistence, we analysed the association between their levels of expression on CD4+ T cells and the size of the HIV reservoir in individuals receiving ART for at least 3 years.
Forty-eight HIV-infected participants receiving suppressive ART were recruited at the University of California San Francisco (UCSF) for this cross-sectional study. Participants were receiving ART for >3 years, had CD4+ T-cell count >350 cells/μl and HIV RNA <40 copies/mL as measured by the Abbott real time HIV-1 PCR for at least 3 years. Whole blood (50mL) was collected by regular blood draw. For cell sorting experiments, 27 HIV-infected individuals were enrolled at UCSF and at VGTIFL and underwent leukapheresis.
All subjects signed informed consent forms approved by the UCSF and Martin Memorial Health Systems review boards (IRB #10–1320, Ref # 068192 and FWA #00004139, respectively).
PBMCs were isolated from peripheral blood and leukapheresis using previously described methods [6,33]. Cryopreserved PBMCs were thawed, washed and stained for phenotyping or cell sorting. Two antibody panels were used to measure the expression of IC in subsets of memory CD4+ T cells. The same antibody backbone was used in the two panels: CD3-Alexa700 (clone UCHT1, BD#557943), CD4-QDot605 (clone S3.5, Invitrogen#Q10008), CD8-PB (clone RPA-T8, BD#558207), CD14-V500 (clone M5E2, BD#561391), CD19-AmCyan (clone SJ25C1, BD#339190), LIVE/DEAD Aqua marker (Invitrogen#L34957), CD45RA-APC-H7 (clone HI100, BD#560674), CD27-BV650 (clone O323, Biolegend#302828) and CCR7-PE-Cy7 (clone 3D12, BD#557648). The following antibodies were added to this backbone: PD-1-AF647 (clone EH12.1, BD#560838), CTLA-4-PE (clone BNI3, BD#555853), LAG-3-FITC (R&D#FAB2319F), TIGIT-PerCP-eF710 (clone MBSA43, eBioscience#46-9500-41), TIM-3-PE (clone F38-2E2, Biolegend#345006), CD160-AF488 (clone By55, eBioscience#53–1609), 2B4-PerCP-Cy5.5 (clone C1.7, Biolegend#329515). For expression of all ICs, gates were defined using fluorescence minus one controls. CD4+ T-cell subsets were identified by CD27, CD45RA, and CCR7 expression on CD4+ T cells after exclusion of dump positive cells (LIVE/DEAD, CD14 and CD19). ICs were measured in gated CD4+ T-cell subsets including naïve CD4+ T cells (CD3+CD8-CD4+CD45RA+CCR7+CD27+), central memory CD4+ T cells (CD3+CD8-CD4+CD45RA-CCR7+CD27+), transitional memory CD4+ T cells (CD3+CD8-CD4+CD45RA-CCR7-CD27+), effector memory CD4+ T cells (CD3+CD8-CD4+CD45RA-CCR7-CD27-) and terminally differentiated CD4+ T cells (CD3+CD8-CD4+CD45RA+CCR7-CD27-). Data was acquired on a BD LSR II flow cytometer using the FACSDiva software (Becton Dickinson) and analysed using Flow Jo version 9 (Treestar).
Central, transitional and effector memory CD4+ T cells were sorted based on their expression of PD-1, TIGIT or LAG-3. The antibodies used for sorting were similar than those used for phenotyping with the exception of CD27-QDot655 (clone CLB-27/1, Invitrogen#Q10066). In a second set of experiments, total memory CD4+ T cells (CD3+CD4+CD45RA-) were sorted based on their expression of PD-1, TIGIT and LAG-3. Sorted cells were collected using an ARIA FACS sorter (Becton Dickinson).
Total CD4+ T cells were isolated from cryopreserved PBMCs using magnetic depletion as per the manufacturer’s protocol (Stem Cell Technologies, Vancouver, Canada).
Total CD4+ T cells or sorted CD4+ T cell subsets were used to measure the frequency of cells harboring HIV DNA (total, integrated and 2-LTR circles) by real time nested PCR as previously described [34] (S1A Text).
The CA-US RNA was measured by real time nested PCR as previously described [35]. The frequency of CD4+ T cells with inducible multiply spliced HIV RNA was determined using Tat/rev inducible limiting dilution assay (TILDA) [36].
Data distributions were assessed through descriptive statistics and scatter plots. Negative binomial regression models were run for each set of comparisons with the percentage of CD4+ T cells expressing ICs being the predictor and the measure of HIV persistence the outcome. We chose this approach for reasons described previously [12, 35], and for consistency with those previous publications (S1B Text). The approach allowed us to fit models adjusting for the effects of absolute current or nadir CD4+ T-cell, which were examined for all combinations of IC predictors and HIV persistence outcome measures. In addition, the negative binomial regression models take into account that copies/input is measured with less precision when the number of copies is lower and when the amount of input is lower. The methods also permit proper quantitative use of instances where zero copies were present in the specimen assayed, without a need for ad hoc modifications to permit taking logarithms. We did not evaluate the results of alternative analysis methods and did not choose the methods post-hoc based on the results that they produced. Analyses were run in Stata version 13.1 (Stata Corp, College Station, TX).
For TILDA results analysis, we estimated the within-person fold difference in TILDA between the 2 cell subsets analyzed (memory CD4+ T cells expressing any versus none of the ICs) by fitting a maximum likelihood model to the raw data on numbers of positive and negative wells at each dilution (S1C Text).
To determine the relationship between ICs and HIV persistence, 48 HIV-infected participants on suppressive ART for a median time (IQR) of 8.5 years (5.0–12.4) and a median CD4+ T-cell count (IQR) of 684 cells/μL (533–858) were recruited (Table 1). The expressions of 7 ICs on CD4+ T cells (PD-1, CTLA-4, LAG-3, TIGIT, TIM-3, CD160 and 2B4) were measured by multiparametric flow cytometry (S1 Fig). The frequencies of CD4+ T cells expressing these ICs were variable (median (IQR) of 16.7% (13.2–22.7), 12.2% (8.8–16.4), 12.0% (8.9–16.1), 9.5% (3.5–18.5), 1.1% (0.8–2.5), 0.8% (0.6–1.5) and 0.7% (0.6–1.0) for TIGIT, PD-1, LAG-3, 2B4, CD160, TIM-3 and CTLA-4 respectively) (Fig 1A).
The size of the HIV reservoir was determined by measuring the frequencies of CD4+ T cells harboring integrated HIV DNA, total HIV DNA and 2-LTR circles as well as cell-associated unspliced (CA-US) HIV RNA (Fig 1B and S1 Table). Total HIV DNA and cell-associated US HIV RNA were detected in all samples tested, whereas integrated HIV DNA and 2-LTR circles were detected in 98%, and 80% of the samples, respectively.
We evaluated the association between markers of HIV persistence and the frequencies of CD4+ T cells expressing ICs using a negative binomial regression model that was adjusted for current and nadir CD4+ T-cell counts when indicated (Table 2 and S2–S4 Tables). Using these tailored analytical methods for HIV reservoir measurements, we identified 3 ICs for which the expression on CD4+ T cells was statistically significantly associated with the frequency of CD4+ T cells harboring integrated HIV DNA, namely PD-1, TIGIT and LAG-3 (Fig 1C–1E and Table 2). These correlations persisted after adjusting for nadir CD4+ T-cell counts but were no longer significant after adjusting for current CD4+ T-cell count, a clinical parameter strongly associated with the size of the reservoir during ART [6,37,38].
The frequency of PD-1 expressing CD4+ T cells was also associated with the frequency of CD4+ T cells harboring total HIV DNA (S3 Table), but only marginally (1.23-fold effect, p = 0.07) when the model was adjusted for current CD4+ T-cell count. CA-US HIV RNA and 2-LTR circles did not show statistically significant correlation with any IC expression levels, with the exception of a negative association between the frequency of CD160+ CD4+ T cells and 2-LTR circles that remained statistically significant after adjusting for current and nadir CD4+ T-cell count (S3 and S4 Tables).
ICs are co-expressed on exhausted CD4+ and CD8+ T cells during untreated HIV infection [39]. Using a Boolean gating strategy, we determined the frequency of CD4+ T cells co-expressing PD-1 and/or TIGIT and/or LAG-3 in our cohort of 48 HIV-infected participants receiving suppressive ART (Fig 2A and 2B). The majority of CD4+ T cells did not express any of these markers (median (IQR) of 65.8% (59.0–72.4)) (S6 Table). Less than 10% (8.5%) of CD4+ T cells expressed more than one of these markers and 0.9% simultaneously expressed PD-1, TIGIT and LAG-3. We further assessed if the frequency of these discrete CD4+ T-cell subsets was associated with markers of HIV persistence. Using the negative binomial regression model, we found that the frequency of CD4+ T cells not expressing PD-1, TIGIT and LAG-3 was strongly and negatively correlated to the frequency of CD4+ T cells harboring integrated HIV DNA (p = 0.002, Table 3 and Fig 2C). Conversely, the frequency of CD4+ T cell co-expressing PD-1, TIGIT and LAG-3 was strongly and positively associated with the frequency of CD4+ T cells harboring integrated HIV DNA (p = 0.001, Table 3 and Fig 2F). Interestingly, the frequencies of CD4+ T cells co-expressing TIGIT with either PD-1 or LAG-3 were also positively associated with the frequency of CD4+ T cells harboring integrated HIV DNA (p = 0.002 and p = 0.029 respectively, Table 3 and Fig 2D and 2E). Although several of these associations were less or no longer statistically significant when the model was adjusted for the current CD4+ T-cell count, the association between the size of the HIV reservoir and the frequency of triple positive cells (PD-1+, LAG-3+ and TIGIT+) remained statistically significant after adjustment (p = 0.038). Adjusting for duration of ART did not produce any substantial changes to the results from the unadjusted analysis (all fold-effects adjusted for ART duration within 7% of those unadjusted). All together, these results indicate that the co-expression of PD-1, TIGIT and LAG-3 identifies a unique subset of CD4+ T cells that strongly predicts the frequency of cells harboring integrated HIV DNA during ART.
HIV persists preferentially in memory CD4+ T-cell subsets [6–8]. To determine the role played by ICs in each individual CD4+ T-cell memory subset, we first analyzed the expression of PD-1, TIGIT and LAG-3, the 3 ICs we identified to be associated with HIV persistence, on naïve (TN), central memory (TCM), transitional memory (TTM), effector memory (TEM) and terminally differentiated (TTD) cells in 48 HIV-infected participants (Cohort 1: clinical characteristics in Table 1) (Fig 3A, 3B and 3C respectively). As expected, TN cells expressed low levels of these ICs. The frequency of CD4+ T cells expressing PD-1 or LAG-3 increased with differentiation, with TEM cells displaying the highest levels of expression of these markers. The highest frequency of TIGIT+ cells was found within the TTM subset. These results demonstrated that the subsets of memory cells that were previously shown to harbor persistent HIV during ART express PD-1, TIGIT and LAG-3.
To determine whether PD-1, TIGIT and LAG-3 identify cells more likely to carry persistent HIV in virally suppressed participants, individual memory CD4+ T-cell subsets were sorted based on their expression of PD-1, TIGIT or LAG-3 in a subset of subjects who underwent leukapheresis (Cohort 2: clinical characteristics in Table 1) and the results were analyzed by negative binomial regression model (S5 Table). The frequency of cells harboring integrated HIV DNA was moderately higher in PD-1 expressing TTM when compared to their PD-1 negative counterparts (p = 0.053, fold-difference = 1.5) (Fig 3D). TEM cells expressing TIGIT were enriched for integrated genomes when compared to their TIGIT- counterparts (p = 0.001, fold-difference = 2.7) (Fig 3E). Finally, all the memory CD4+ T-cell subsets (TCM, TTM and TEM cells) expressing LAG-3 were enriched for integrated HIV DNA when compared to their negative counterparts (p<0.0001, fold-difference = 1.9, p = 0.003, fold-difference = 1.8 and p = 0.030, fold-difference = 2.5 respectively) (Fig 3F). All together these results indicate that PD-1, TIGIT and LAG-3 enrich for infected cells in distinct memory CD4+ T-cell subsets in individuals on ART. We calculated the contribution of cells expressing PD-1, TIGIT or LAG-3 to the total reservoir by taking into account the frequency of these subsets within the CD4 compartment and their relative infection frequencies. The mean contributions of CD4+ T cells expressing PD-1, TIGIT and LAG-3 were 29%, 34% and 31%, respectively (S2 Fig). As a comparator, TCM, TTM and TEM cells contributed 43%, 27% and 24% to the pool of infected cells in these same virally suppressed individuals. These data indicate that a third of the reservoir is encompassed in cells expressing each individual marker.
We then determined if the combination of PD-1, TIGIT and LAG-3 would further enrich memory CD4+ T cells for HIV-infected cells during ART. The average frequency of cells expressing 0, 1, 2 or 3 of these markers in the memory CD4+ T compartment (CD45RA-) from our cohort of 48 individuals (Table 1) indicated that an average of 33% of memory CD4+ T cells expressed one of the 3 IC only, 12% expressed 2 and 2% expressed the 3 markers simultaneously. Large numbers of memory CD4+ T cells were sorted based on their expression of PD-1, TIGIT and LAG-3 from 5 individuals. The combination of these 3 markers allowed us to sort eight subsets of cells to high purity, namely PD-1/TIGIT/LAG-3 triple -, PD-1 single +, TIGIT single +, LAG-3 single +, PD-1/TIGIT double +, TIGIT/LAG-3 double +, PD-1/LAG-3 double + and PD-1/TIGIT/LAG-3 triple + cells. The frequency of cells harboring integrated HIV DNA was measured by qPCR in each sorted subset (S3 Fig) and the mean frequency for each category was calculated relative to total CD4+ T cells (Fig 4B). Memory CD4+ T cells showed a gradual enrichment in HIV-infected cells when expressing an increasing number of ICs. Memory CD4+ T cells expressing simultaneously PD-1, TIGIT and LAG-3 were enriched for HIV-infected cells up to 10 times more when compared to total CD4+ T cells, with a median fold increase (IQR) of 8.15 (4.92–9.59). These results demonstrated that memory CD4+ T cells expressing a combination of ICs were highly enriched for integrated HIV DNA during ART.
As the majority of HIV genomes, even when integrated, are defective [40], we assessed if PD-1, TIGIT and LAG-3 identify cells in which HIV production can be induced. As the frequency of triple positive cells was too low to perform this experiment, we sorted memory CD4+ T cells (CD45RA-) expressing any (i.e. at least one) versus none of PD-1, TIGIT or LAG-3 (mLPT+ and mLPT- respectively). We measured the frequency of cells in each population that transcribe multiply spliced HIV RNA molecules upon induction with PMA/ionomycin using the Tat/rev induced limiting dilution assay (TILDA) [36]. Tat/rev transcripts were detectable by TILDA in both cell subsets from all of the 8 individuals tested. The rate of inducible virus per million cells was estimated in our maximum likelihood model to average 3.0-fold higher in mLPT+ cells than in mLPT- cells from the same participant (95% CI 1.0 to 9.0, p = 0.049, S1C Text) (Fig 4C). Taking into account the frequency of these cell subsets, the contribution of cells expressing at least one of these markers to the total pool of memory CD4+ T cells infected with inducible HIV genomes was calculated. This contribution ranged from 30 to 98% (median of 76%), indicating that the majority of inducible HIV genomes were found in memory CD4+ T cells expressing at least one of these markers. These experiments provide evidence that memory CD4+ T cells expressing PD-1, TIGIT and/or LAG-3 are enriched for HIV-infected CD4+ T cells harboring inducible proviruses during ART.
In this study, we identified PD-1, TIGIT and LAG-3 as novel markers of cells that are more frequently infected in HIV-infected individuals receiving suppressive ART. Co-expression of the 3 ICs identified a unique subset of CD4+ T cells that was strongly associated with the size of the HIV reservoir and that was highly enriched for integrated HIV DNA. Finally, our results provide evidence that memory CD4+ T cells expressing at least one of these markers are the major contributors to the pool of inducible HIV genomes during ART.
The frequencies of CD4+ T cells expressing PD-1, CTLA-4, LAG-3 and TIM-3 were similar to those reported by other groups [41–44], indicating that the cohort of participants used for this study is likely to be representative of the HIV population receiving suppressive ART. In addition, the association between PD-1 and TIGIT expression on CD4+ T cells and the frequency of CD4+ T cells carrying HIV proviruses was in agreement with previously reported findings [6,24,41].
We found positive associations between the expression of PD-1, LAG-3 and TIGIT in CD4+ T cells and the frequency of cells harboring integrated HIV DNA. Of note, these three markers showed the strongest inverse associations with CD4+ T cell counts among the 7 markers we examined, suggesting a link between T cell homeostasis and HIV persistence (S4A, S4B and S4C Fig). The associations between individual IC expression and HIV persistence marker were substantially smaller and no longer statistically significant after adjusting for current CD4+ T-cell count. These findings from the negative binomial regression models suggest that the current CD4+ T-cell count is an important predictor of the size of the HIV reservoir when measured as the frequency of cells harboring proviral genomes [6,37]. Importantly, and in contrast to cells expressing a single marker, the frequency of cells co-expressing simultaneously PD-1, TIGIT and LAG-3 was strongly associated with the size of the reservoir and remained after adjusting for nadir and current CD4+ T-cell counts. This result reinforces the possibility of a direct—and maybe synergistic—role for these molecules in HIV persistence during ART. In addition, the frequency of CD4+ T cells co-expressing simultaneously PD-1, TIGIT and LAG-3 positively correlated with the frequency of CD4+ T cells expressing HLADR/CD38 (p = 0.003, r = 0.42) and Ki67 (p = 0.022, r = 0.33) (S1D and S1E Text and S4D and S4E Fig). These associations suggest that the persistence of the small pool of cells expressing the 3 markers is associated with T cell activation and proliferation.
Interestingly, we observed a strong negative association between CD160 expression and 2-LTR circles. Notably, this correlation remained after adjusting for current and nadir CD4+ T cell counts. A possible explanation for these findings is that CD160+ cells may be preferential targets for infection and depletion during ART, which would explain the strong negative association between the frequency of CD160+ CD4+ T cells and a putative marker of persistent viral replication.
By sorting TCM, TTM and TEM cells expressing PD-1, TIGIT and LAG-3, we observed that cells expressing these markers were enriched for HIV-infected cells in different memory CD4+ T-cells subsets during ART. While LAG-3 enriched for integrated HIV DNA in all memory subsets (TCM, TTM, and TEM), PD-1 and TIGIT enriched for HIV genomes exclusively in TTM and TEM cells, respectively. These observations suggest that ICs may exert different pro-latency effects in subsets endowed with distinct proliferative and activation status. One may hypothesize that different ICs provide infected cells with different selective advantage to persist by counteracting distinct stimuli specific to an individual memory cell subset. Further investigations will be needed to characterize the mechanisms by which these ICs may specifically contribute to HIV persistence within these distinct subsets.
Overall, the majority of inducible HIV genomes were found in memory CD4+ T cells expressing at least one of these markers (median of 76%). Although triple negative cells also contain inducible HIV genomes, our data provide evidence that there is an enrichment for inducible viral genomes in CD4+ T cells expressing these markers.
Importantly, we found a gradual enrichment in integrated HIV DNA in cells that express multiple ICs simultaneously. This observation mirrors the synergistic mechanisms of action of these receptors to dampen T cell functions. Indeed, LAG-3 and PD-1 are commonly co-expressed on exhausted or dysfunctional T cells in models of chronic infections [45], autoimmune diseases [46], and cancers [47,48]. Potential synergistic functions were highlighted in murine models of autoimmune diseases [49]. Anti-LAG-3 blocking mAb has recently entered clinical testing in cancer in monotherapy or in combination therapy with anti-PD-1 (NCT01968109). Additionally, TIGIT is co-expressed with PD-1 on activated CD8+ tumor-infiltrating lymphocytes from patients with melanoma [50]. Blockages of TIGIT and PD-1 synergize to improve T cell proliferation, cytokines production and degranulation in vivo in melanoma treatment and in vitro in HIV infection [24,50]. All together, these studies indicate that ICs can synergize to repress T cell functions and suggest that these synergies may also play a role in HIV persistence during ART.
The expression of PD-1, LAG-3 and TIM-3 on CD4+ and CD8+ T cells prior to ART was recently identified as a strong predictor of time to viral rebound after treatment interruption in the SPARTAC study [43]. It is possible that CD4+ T cells expressing these markers before ART represent a preferential niche for the establishment of a stable reservoir for HIV and that latently infected cells expressing these markers preferentially persist during ART, as suggested by our observations. In our study, we identified a discrete subset of CD4+ T cells co-expressing PD-1, TIGIT and LAG-3 as an important predictor of the frequency of cells harboring integrated HIV DNA during ART. Of note, the expression of TIGIT before ART initiation was not measured in the SPARTAC study and further studies will be needed to determine if this IC could also represent a pre-ART predictor of viral rebound.
Our data provide a rationale for the use of immune checkpoint blockers (ICBs) to target latently infected cells during ART. Targeting ICs by ICBs, a novel class of molecules in development in oncology, may have a double benefit in the context of HIV remission by both targeting latently infected cells and restoring HIV-specific T cell immunity. By enhancing T cell activation and increasing viral transcription, ICBs may facilitate HIV reactivation in latently infected cells when used alone or in combination with latency reversing agents. The anti-CTLA-4 antibody iplimumab was recently shown to significantly increase CA-US HIV RNA in an HIV-infected individual on ART, consistent with latency reversal [51]. An alternative mechanism of action of some ICBs would be to directly deplete cells expressing these markers, as observed with the anti-CTLA-4 ipilumimab, which induces direct elimination of CTLA-4+ regulatory T cells in tumor tissue in patients with melanoma [52]. Our results suggest that the administration of antibodies with effector functions targeting PD-1, LAG-3 and TIGIT may significantly reduce the size of the latent HIV reservoir during ART by targeting cells in which HIV persists.
Several limitations are associated with our study. We have not adjusted p-values for multiple comparisons, because such adjustment would neglect the biological relationships among our positive results and would require that each analysis detract from the others, rather than reinforcing one another when there is biological coherence [53] (S1F Text). Nevertheless, our evidence may be weaker than if it had arisen from a narrower set of analyses, and, in any case, additional studies will be needed to confirm the hypotheses supported by our results. Most of our analyses were performed using integrated HIV DNA as a marker of HIV persistence. We chose this readout as it was applicable to small subsets of CD4+ T cells on which measures of replication competent HIV cannot be performed. The majority of viral genomes persisting during ART are known to be defective [40,54], and although our experiments indicate that cells that express ICs can produce multiply spliced RNA upon activation (TILDA), they do not demonstrate that replication competent virus persists in these cells. In addition, our results are limited to circulating T cells. It is possible and indeed likely that the biology of ICs expression and HIV persistence will differ in tissues, particularly in secondary lymphoid tissues where many of the ligands for these receptors are likely to be expressed.
A better understanding of the nature of the cells that encompass the latent HIV reservoir is a prerequisite to the development of novel curative strategies. Despite similarities in their mechanisms of action, PD-1, TIGIT and LAG-3 are likely to be non-redundant in their functions. Blocking these pathways simultaneously may show synergies in latency reversal, as suggested by their synergistic activities in the restoration of T cell immunity.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.