itgap 0.1.0

Integrated TCR-Gene-Antigen Prediction: dataset tooling and models for TCR-antigen recognition.

ITGAP — Integrated TCR-Gene-Antigen Prediction

itgap is a Python package for building TCR-peptide datasets from the 10x Genomics CD8+ T-cell multi-omics benchmark and training TCR-antigen recognition models that integrate gene expression (GEX), V/J gene usage, and CDR3 sequence information.

It bundles:

• NegativeSamplingTool — two-stage synthetic negative TCR-peptide sampling.
• Sequence encoding utilities (Atchley factors + positional encoding) with the Atchley table shipped as a package resource.
• Autoencoder + encoder–decoder integration models for combining sequence, V/J, and GEX modalities.
• Residual-MLP binary classifiers, sklearn baselines (logistic regression, random forest), and standard evaluation/plotting helpers.

Install

Core install (small footprint, only numpy, pandas, scikit-learn, matplotlib):

pip install itgap

Optional extras:

pip install 'itgap[gex]'   # adds scanpy + anndata for h5ad loading
pip install 'itgap[tf]'    # adds tensorflow (or tensorflow-macos on Apple Silicon)
pip install 'itgap[all]'   # everything

itgap[tf] resolves to tensorflow-macos>=2.9 on macOS arm64 and to tensorflow>=2.9 elsewhere.

Data

The package ships only the small atchley.txt reference table. The large benchmark file merge_gex_all_donors_all_peptides_meta.h5ad (~200 MB) is not included; download it from the 10x Genomics CD8+ T-cell multi-omics dataset and pass its path to NegativeSamplingTool(data_dir=...) or to load_dataset(h5ad_path=...). The pre-computed CSV embeddings used in the example notebooks live in the project repository under examples/data/.

Quickstart

Generate a negative-sampled training set:

from itgap import NegativeSamplingTool

tool = NegativeSamplingTool(
    data_dir="path/to/10x",   # contains merge_gex_all_donors_all_peptides_meta.h5ad
    negative_ratio=3.0,
    random_seed=42,
)
result = tool.create_combined_dataset(negative_ratio=3.0)
print(result["dataset"].shape, result["statistics"])

Train a residual-MLP classifier on assembled features (requires itgap[tf]):

from itgap import (
    load_atchley, build_residual_mlp, compile_and_train, evaluate_classifier,
)

word_vectors, aa_idx = load_atchley()   # uses the packaged atchley.txt
model = build_residual_mlp(input_dim=X_train.shape[1])
history = compile_and_train(model, X_train, y_train, X_val, y_val, epochs=50)
metrics = evaluate_classifier(model, X_test, y_test)

Command-line

Installing the package exposes a console script:

itgap-negative-sampling   # runs NegativeSamplingTool with default settings

Examples

End-to-end Jupyter notebooks live in examples/ of the repository:

• data_preparation_notebook.ipynb — build the labeled dataset.
• tcr_beta_prediction_notebook.ipynb — beta-chain only model.
• tcr_alpha_beta_prediction_notebook.ipynb — alpha + beta + GEX + VJ.

Development

pip install -e '.[dev,all]'
pytest
python -m build

License

MIT. See LICENSE.

Citation

If you use ITGAP in a publication, please cite the project repository: https://github.com/mlizhangx/ITGAP.