TCR-ESM: employing protein language embeddings to predict TCR-peptide-MHC binding

Description

This is the GitHub repository for the project TCR-ESM: employing protein language embeddings to predict TCR-peptide-MHC binding by Shashank Yadav, Dhvani Sandip Vora, Durai Sundar, Jaspreet Kaur Dhanjal

Getting Started

Dependencies / Requirements

• numpy==1.23.5
• pandas==1.5.3
• tqdm==4.66.1
• natsort==8.4.0
• matplotlib==3.7.1
• tensorflow==2.14.0
• torch==2.1.1
• fair-esm==2.0.0
• scikit-learn==1.2.2

Repository directories & files

• 0_ESM_Embeddings_Extractor.ipynb contains code for extraction of embeddings from the ESM1v model using the fair-esm library.
• 1_netTCRdata_trainESMmodel_BP_MIRA.ipynb contains code for model training on NetTCR 2.0 (CDR3beta+peptide) data and evaluation on the MIRA dataset.
• 2_netTCRdata_trainESMmodel_ABP.ipynb contains code for model training on NetTCR 2.0 data, cross validation evaluation and peptide based metrics.
  • Models:
  • CDR3alpha+CDR3beta+peptide
  • CDR3beta+peptide
  • CDR3alpha+peptide
• 3_MCPAS_trainESMmodel.ipynb contains code for model training on McPAS dataset data.
  • Models for McPAS Subset 1 data (without MHC information)
  • CDR3alpha+CDR3beta+peptide
  • CDR3beta+peptide
  • CDR3alpha+peptide
  • Models for McPAS Subset 2 data (with Human MHC information)
  • CDR3alpha+CDR3beta+peptide+MHC
  • CDR3beta+peptide+MHC
  • CDR3alpha+peptide+MHC
  • CDR3alpha+CDR3beta+peptide
  • CDR3beta+peptide
  • CDR3alpha+peptide
  • Also, includes cross validation evaluation, test set evalution, peptide and mhc based metrics.
• 4_VDJDB_trainESMmodel.ipynb contains code for model training on VDJDb dataset data.
  • Models for VDJDb Subset 1 data (without MHC information)
  • CDR3alpha+CDR3beta+peptide
  • CDR3beta+peptide
  • CDR3alpha+peptide
  • Models for VDJDb Subset 2 data (with Human MHC information)
  • CDR3alpha+CDR3beta+peptide+MHC
  • CDR3beta+peptide+MHC
  • CDR3alpha+peptide+MHC
  • CDR3alpha+CDR3beta+peptide
  • CDR3beta+peptide
  • CDR3alpha+peptide
  • Also, includes cross validation evaluation, test set evalution, peptide and mhc based metrics.
• 5_Embeddings_analysis_netTCR_data_ESMmodel_ABP.ipynb contains code for analysis of learned embeddings in the input and intermediate layers of the model trained on NetTCR2.0 dataset.
• 6_Embeddings_analysis_MCPAS_ESMmodel.ipynb contains code for analysis of learned embeddings in the input and intermediate layers of the model trained on McPAS dataset.
• 7_Embeddings_analysis_VDJDB_ESMmodel.ipynb contains code for analysis of learned embeddings in the input and intermediate layers of the model trained on VDJDb dataset.
• 8_MCPAS_VDJDB_ext_validation.ipynb contains code for external evaluation of models (CDR3beta,peptide) trained on McPAS / VDJDb dataset and evaluated on pmtNet dataset with different similarity thresolds (90% and 80%).

Note: Using ESM1v to generate embeddings of 100 randomly generated peptides using ESM1v takes approximately 17 seconds, model training time may vary from 3 minutes to 3 hours, and predictions involve a quick forward pass lasting up to 3 seconds. These computations were performed on a GPU with 16 GB VRAM, demonstrating the computational efficiency of TCR-ESM. This highlights the feasibility of implementing TCR-ESM for scanning numerous peptides swiftly, aligning to facilitate cognate target identification for T-cell receptors in the context of T-cell therapy development.

We have compared our approach to three models from the literature.

• For the netTCR 2.0 model, please visit NetTCR 2.0
• For the ERGO II (AE and LSTM), please visit ERGO-II

Authors

Contributors names and contact info

• @Shashank Yadav
• @Dhvani Vora
• @Durai Sundar
• @Jaspreet Kaur Dhanjal