/epiTCR-KDA

epiTCR-KDA: Knowledge Distillation model on Dihedral Angles for TCR-peptide prediction

Primary LanguagePythonMIT LicenseMIT

epiTCR-KDA: Knowledge Distillation model on Dihedral Angles for TCR-peptide prediction

This repository contains the code and the data to train epiTCR-KDA model.

Requirements

  • Python >= 3.6.8
  • Keras 2.6.0
  • TensorFlow 2.6.0

How to run epiTCR-KDA

pipeline

1. Clone the repository

git clone https://github.com/ddiem-ri-4D/epiTCR-KDA
cd epiTCR-KDA/
conda env create -f environment.yml
source activate kda

2. Prepare data

  • Download training and testing data from datasets folder.
  • Download the 3D structure and dihedral angles of TCR and peptide from folders 3DS_PDBFiles and DA_TSVFiles.

3. Check if TCR/pep sequences are present in the DA folders

  • Prepare a list containing unique TCR/peptides from the data for training/testing.
  • Check if the unique TCR/peptides are already present in the DA_TSVFiles folders or not by executing the following command:
cd utils
python3 check3DSDA.py
  • If they are already complete, proceed to step 4.
  • If not, run the 3D structure using OmegaFold and add the structure to the PDB folders, following the steps below:

3.1 Run OmegaFold

  • Prepare a FASTA file containing the TCR/peptide sequences to run OmegaFold, see an example here.
  • Refer to the OmegaFold running steps here, and place the output into the PDB files directory.
  • Double-check for any TCR/peptides that might still lack a structure. If all structures are present, proceed to step 3.2.

3.2 Run Biopython

  • After obtaining the 3D structure, run Biopython to retrieve Dihedral Angles information, resulting in an output *.tsv file, see an example here.
  • The output *.tsv files containing Dihedral Angles information are placed into the DA folders directory.
cd utils
python3 PDB2DA.py

4. Retrain and Predict Model

  • train.parquet/test.parquet: input parquet file with 3 columns named as "CDR3b, epitope, binder (if training)": TCR-beta CDR3 sequence, peptide sequence, and CDR3b and peptide bind together or not.
CDR3b epitope binder
AASSYGQNFV QIKVRVDMV 1
AIRAGGDEQ HSKKKCDEL 1
AISETDKLG LPPIVAKEI 1
SARDRVRTDTQY FVSKLYYFE 0
SARDRVRTDTQY KLSHQPVLL 0
  • An example for training and testing
python3 train.py \
        --trainfile ./datasets/DATA_4MODEL/TRAIN-TEST/train.parquet \
        --testfile ./datasets/DATA_4MODEL/TRAIN-TEST/test.parquet \
        --savemodel ./models/KDA_model.h5 \
        --outfile ./datasets/DATA_4PRED/test_prediction.parquet
python3 test.py \
        --testfile ./datasets/DATA_4MODEL/TRAIN-TEST/test.parquet \
        --savedmodel ./models/KDA_model.h5 \
        --outfile ./datasets/DATA_4PRED/test_prediction.parquet

5. Contact

For more questions or feedback, please post an Issue.

6. Citation

Please cite this paper if it helps your research:

@article {Pham2024.05.18.594806,
	author = {Pham, My-Diem Nguyen and Su, Chinh Tran-To and Nguyen, Thanh-Nhan and Nguyen, Hoai-Nghia and Nguyen, Dinh Duy An and Giang, Hoa and Nguyen, Dinh-Thuc and Phan, Minh-Duy and Nguyen, Vy},
	title = {epiTCR-KDA: Knowledge Distillation model on Dihedral Angles for TCR-peptide prediction},
	elocation-id = {2024.05.18.594806},
	year = {2024},
	doi = {10.1101/2024.05.18.594806},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2024/05/21/2024.05.18.594806},
	eprint = {https://www.biorxiv.org/content/early/2024/05/21/2024.05.18.594806.full.pdf},
	journal = {bioRxiv}}

7. Reference

My-Diem Nguyen Pham, Thanh-Nhan Nguyen, Le Son Tran, Que-Tran Bui Nguyen, Thien-Phuc Hoang Nguyen, Thi Mong Quynh Pham, Hoai-Nghia Nguyen, Hoa Giang, Minh-Duy Phan, Vy Nguyen, epiTCR: a highly sensitive predictor for TCR–peptide binding, Bioinformatics, Volume 39, Issue 5, May 2023, btad284, https://doi.org/10.1093/bioinformatics/btad284