otk: ecDNA Analysis Toolkit

otk (ecDNA Analysis Toolkit) is a deep learning-based tool for analyzing extrachromosomal DNA (ecDNA), predicting whether genes are detected as ecDNA cargo genes at the gene level, and classifying focal amplification types at the sample level.

Core Features

• Deep learning-based ecDNA cargo gene prediction
• Sample-level focal amplification type classification
• Support for analysis from BAM files or processed copy number data
• Efficient command-line interface
• GPU acceleration support

Technology Stack

• Python 3.8+
• PyTorch 2.0+
• NumPy
• Pandas
• scikit-learn
• Click (command-line interface)

Installation Guide

From Source

1. Clone the repository:

git clone https://github.com/WangLabCSU/otk.git
cd otk

2. Install with pip:

pip install -e .

Dependencies

The following dependencies will be installed automatically:

• pandas>=2.0
• numpy>=1.24
• torch>=2.0
• scikit-learn>=1.3
• tqdm>=4.65
• click>=8.1
• matplotlib>=3.7
• seaborn>=0.12
• pyyaml>=6.0

Usage

otk provides two main command-line subcommands: train and predict.

Model Training

Use the otk train command to train the model:

otk train --config configs/model_config.yml --output models/ --gpu 0

Parameters:

• --config, -c: Path to configuration file (default: configs/model_config.yml)
• --output, -o: Output directory for trained models (default: models/)
• --gpu, -g: GPU device ID to use (default: 0)

Model Prediction

Use the otk predict command for predictions:

otk predict --model models/best_model.pth --input data/test_data.csv --output predictions/ --gpu -1

Parameters:

• --model, -m: Path to trained model (required)
• --input, -i: Path to input data file (required)
• --output, -o: Output directory for predictions (default: predictions/)
• --gpu, -g: GPU device ID to use (default: -1, i.e., use CPU)

Data Format

Input Data Format

Input data should be in CSV format with the following columns:

Required identifier columns:

• sample: Tumor sample ID
• gene_id: Gene ID

Copy number features:

• segVal: Total gene copy number
• minor_cn: Minor gene copy number
• intersect_ratio: Proportion of overlap between copy number detection segment and gene region

Sample-level genomic features (same value for all genes in a sample):

• purity: Tumor purity estimate
• ploidy: Tumor genome ploidy estimate
• AScore: Aneuploidy score
• pLOH: Proportion of genome with loss of heterozygosity (LOH)
• cna_burden: Proportion of genome with copy number alterations

Copy number signature features:

• CN1 to CN19: 19 copy number signature activity estimates

Clinical features:

• age: Patient age
• gender: Patient gender (0/1 encoded)

Tumor type features (one-hot encoded, 24 cancer types):

• type_BLCA, type_BRCA, type_CESC, type_COAD, type_DLBC, type_ESCA, type_GBM, type_HNSC
• type_KICH, type_KIRC, type_KIRP, type_LGG, type_LIHC, type_LUAD, type_LUSC, type_OV
• type_PRAD, type_READ, type_SARC, type_SKCM, type_STAD, type_THCA, type_UCEC, type_UVM

Gene frequency features:

• freq_Linear: Prior estimated frequency of gene in linear focal amplifications
• freq_BFB: Prior estimated frequency of gene in breakage-fusion-bridge (BFB) events
• freq_Circular: Prior estimated frequency of gene in circular focal amplifications (ecDNA)
• freq_HR: Prior estimated frequency of gene in homologous recombination events

Target column (for training data):

• y: Binary label indicating whether the gene is detected as an ecDNA cargo gene (1) or not (0)

Output Data Format

Prediction results are saved as a CSV file with the following columns:

Gene-level predictions:

• sample: Tumor sample ID
• gene_id: Gene ID
• prediction_prob: Probability of being an ecDNA cargo gene (0-1)
• prediction: Binary classification result (0 = not ecDNA cargo, 1 = ecDNA cargo)

Sample-level predictions:

• sample_level_prediction_label: Sample-level focal amplification type classification:
  • nofocal: No focal amplification detected
  • noncircular: Non-circular focal amplification detected
  • circular: Circular focal amplification (ecDNA) detected
• sample_level_prediction: Numerical encoding of sample-level classification (0 = nofocal, 1 = noncircular, 2 = circular)

Note: Sample-level classification follows these rules:

1. If any gene in the sample is predicted as ecDNA cargo (prediction = 1), the sample is classified as circular
2. If no ecDNA cargo genes but any gene has segVal > ploidy + 2, the sample is classified as noncircular
3. Otherwise, the sample is classified as nofocal

Model Architecture

otk supports multiple model architectures with unified interface:

Available Models

Model	Type	Description
xgb_new	XGBoost	Optimized with feature engineering
xgb_paper	XGBoost	Paper reproduction (11 features)
baseline_mlp	Neural Network	Simple MLP baseline
transformer	Neural Network	Transformer architecture
deep_residual	Neural Network	Deep residual network
optimized_residual	Neural Network	Optimized residual network
dgit_super	Neural Network	Deep gated interaction transformer
tabpfn	TabPFN	TabPFN ensemble

Unified Interface

All models inherit from BaseEcDNAModel and provide:

• fit(X_train, y_train, X_val, y_val) - Training
• predict_proba(X) - Probability prediction
• predict(X) - Binary prediction
• save(path) / load(path) - Persistence

Data Split

All models use unified data split (80/10/10) with seed=2026 for reproducibility.

Training Script

Use the unified training script:

# Train single model
python train_unified.py --model xgb_new

# Train all models
python train_unified.py --all

Configuration File

Model configuration uses YAML format, with example configuration files located in configs/. You can modify parameters in the configuration files as needed, such as model architecture and training parameters.

Examples

Training Examples

# Train model with default configuration
otk train

# Train model with custom configuration file
otk train --config my_config.yml

Prediction Examples

# Make predictions using a trained model
otk predict --model models/best_model.pth --input test_data.csv

Performance Metrics

The following performance metrics are recorded during model training:

• auPRC (Area under Precision-Recall Curve)
• AUC (Area under ROC Curve)
• F1 Score
• Precision
• Recall

Contribution Guide

We welcome community contributions! If you have any questions or suggestions, please submit them through GitHub Issues.

Development Process

1. Fork the repository
2. Create a feature branch
3. Implement features or fix bugs
4. Run tests
5. Submit a Pull Request

License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

Citation

If you use otk in your research, please cite the following paper:

Wang, S., Wu, C. Y., He, M. M., Yong, J. X., Chen, Y. X., Qian, L. M., ... & Zhao, Q. (2024). Machine learning-based extrachromosomal DNA identification in large-scale cohorts reveals its clinical implications in cancer. Nature Communications, 15(1), 1-17.

Contact

• Project homepage: https://github.com/WangLabCSU/otk
• Email: wangshx@csu.edu.cn