Dataset Card for Bacterial Flagellar Motor Tomograms (Kaggle)

Dataset Description

Homepage: Link to Kaggle Competition: BYU - Locating Bacterial Flagellar Motors 2025

Dataset Summary

This dataset originates from the Kaggle competition "BYU - Locating Bacterial Flagellar Motors 2025". The goal is to identify the presence and location of flagellar motors within 3D cryogenic electron tomography (cryo-ET) reconstructions (tomograms) of bacteria. Flagellar motors are molecular machines essential for motility in many microorganisms.

The data consists of 3D tomograms, presented as stacks of 2D image slices. The task is challenging due to low signal-to-noise ratios, variable motor orientations, and complex intracellular environments within the tomograms. Automating this detection task aims to accelerate research in molecular biology, drug development, and synthetic biology by overcoming the bottleneck of manual identification.

The dataset includes training data with labeled motor locations and test data (originally used in the competition) for evaluation.

Dataset Structure

Data Instances

Each instance corresponds to a single tomogram. A tomogram was represented in the original dataset as a directory containing multiple 2D JPEG image slices along the z-axis, and is now composed of numpy file corresponding to the aggragated files jpg images.

Data Fields

Tomogram Data: Stored as Numpy files named by tomo_id. Each Numpy file represents the 3D Volume.

Training Labels (train_labels.csv):

row_id: Index of the row in the CSV file.
tomo_id: Unique identifier for the tomogram. Name of the numpy tomogram file. Note: Some tomograms in the training set contain multiple motors and thus have multiple rows in this file.
Motor axis 0: The z-coordinate (slice index) of the motor center (float).
Motor axis 1: The y-coordinate of the motor center (float).
Motor axis 2: The x-coordinate of the motor center (float).
Array shape axis 0: The number of slices (z-dimension) in the corresponding tomogram.
Array shape axis 1: The height (y-dimension) of each slice in the tomogram.
Array shape axis 2: The width (x-dimension) of each slice in the tomogram.
Voxel spacing: The scaling factor of the tomogram in Angstroms per voxel (float).
Number of motors: The total number of motors present in the tomogram.

Test Data: The structure follows the training data, but without labels. The competition rules state the hidden test set contains approximately 900 tomograms, each with either zero or one motor.

Okay, let's craft a Hugging Face Dataset Card (README.md) based on the information you provided from the Kaggle competition.

Remember to:

Replace placeholders like [Link to your dataset on Hugging Face Hub] and [Your Hugging Face username/dataset name] with your actual information.
Fill in the size_categories based on the actual size of the dataset you uploaded.
Add any specific preprocessing steps you might have performed before uploading, if applicable.

Markdown

Dataset Card for Bacterial Flagellar Motor Tomograms (Kaggle)

Dataset Description

Homepage: Link to Kaggle Competition: BYU - Locating Bacterial Flagellar Motors 2025
Repository: [Link to your dataset on Hugging Face Hub]
Point of Contact: [Your Name or Contact Info, or link back to Kaggle]

Dataset Summary

The dataset includes training data with labeled motor locations and test data (originally used in the competition) for evaluation.

Supported Tasks and Leaderboards

Object Detection / Localization: The primary task is to detect if a flagellar motor is present in a tomogram and, if so, predict its 3D coordinates (x, y, z).
Leaderboard: The original competition leaderboard can be found on the Kaggle competition page.

Languages

The text descriptions and documentation are in English (en). The data itself consists of volumetric image data (tomograms).

Dataset Structure

Data Instances

Each instance typically corresponds to a single tomogram. A tomogram is represented as a directory containing multiple 2D JPEG image slices along the z-axis.

Example Structure:

./ ├── train/ │ ├── tomo_id_001/ │ │ ├── slice_000.jpeg │ │ ├── slice_001.jpeg │ │ └── ... │ ├── tomo_id_002/ │ │ └── ... │ └── ... ├── test/ │ ├── tomo_id_abc/ │ │ └── ... │ └── ... └── train_labels.csv

Data Fields

Tomogram Data: Stored as JPEG files within directories named by tomo_id. Each JPEG represents a 2D slice of the 3D volume.

Training Labels (train_labels.csv):

row_id: Index of the row in the CSV file.
tomo_id: Unique identifier for the tomogram. Links to the directory containing the tomogram slices. Note: Some tomograms in the training set contain multiple motors and thus have multiple rows in this file.
Motor axis 0: The z-coordinate (slice index) of the motor center (float).
Motor axis 1: The y-coordinate of the motor center (float).
Motor axis 2: The x-coordinate of the motor center (float).
Array shape axis 0: The number of slices (z-dimension) in the corresponding tomogram.
Array shape axis 1: The height (y-dimension) of each slice in the tomogram.
Array shape axis 2: The width (x-dimension) of each slice in the tomogram.
Voxel spacing: The scaling factor of the tomogram in Angstroms per voxel (float).
Number of motors: The total number of motors present in the tomogram.

Test Data: The structure follows the training data, but without labels. The competition rules state the hidden test set contains approximately 900 tomograms, each with either zero or one motor.

Data Splits

Train: Contains tomograms and corresponding labels (train_labels.csv) for model training.
Test: Contains tomograms intended for model evaluation (originally labels were hidden for the competition).

Usage Considerations

Intended Use

This dataset is intended for training and evaluating machine learning models, particularly computer vision algorithms, for the task of object detection and localization in 3D volumetric biological imagery. Specifically, it targets the identification of bacterial flagellar motors in cryo-ET tomograms.

Limitations and Bias

Low Signal-to-Noise Ratio: Cryo-ET images inherently have low SNR, making detection difficult.
Variability: Motors can appear in various orientations and potentially different conformational states.
Complexity: The intracellular environment is crowded, potentially obscuring motors or containing similar-looking structures (decoys).
Data Representativeness: The dataset represents bacteria prepared and imaged under specific cryo-ET conditions. Model performance may vary on data from different sources or imaging modalities.
Train/Test Differences: The training set includes tomograms with multiple motors, while the competition's test set was specified to contain only tomograms with zero or one motor. This difference should be considered during evaluation.

Evaluation

The original Kaggle competition used a metric combining the $F_{\beta}$-score (with $\beta=2$, emphasizing recall) and Euclidean distance. A prediction was considered a True Positive (TP) if its Euclidean distance $|y - \bar{y}|_2$ to the ground truth $y$ was less than or equal to a threshold $\tau = 1000$ Angstroms. The $F_2$-score is calculated as:

$F_2 = (1+2^2) \cdot \frac{\text{precision} \cdot \text{recall}}{(2^2 \cdot \text{precision}) + \text{recall}} = \frac{5 \cdot \text{TP}}{5 \cdot \text{TP} + 4 \cdot \text{FN} + \text{FP}}$

Predictions require submitting the tomo_id and the predicted Motor axis 0, Motor axis 1, Motor axis 2 coordinates. If no motor is predicted for a tomogram, coordinates should be set to -1.

Additional Information

Dataset Curators

The dataset was provided for the Kaggle competition by the individuals listed in the citation, likely affiliated with Brigham Young University (BYU) and collaborating institutions.

Licensing Information

The dataset is provided under the MIT License.

Citation Information

If you use this dataset in your work, please cite the original Kaggle competition:

@misc{byu-locating-bacterial-flagellar-motors-2025,
    author = {Andrew Darley and Braxton Owens and Bryan Morse and Eben Lonsdale and Gus Hart and Jackson Pond and Joshua Blaser and Matias Gomez Paz and Matthew Ward and Rachel Webb and Andrew Crowther and Nathan Smith and Grant J. Jensen and TJ Hart and Maggie Demkin and Walter Reade and Elizabeth Park},
    title = {BYU - Locating Bacterial Flagellar Motors 2025},
    year = {2025},
    howpublished = {\url{https://kaggle.com/competitions/byu-locating-bacterial-flagellar-motors-2025}},
    note = {Kaggle}
}