TransfIGN: A Structure-Based Deep Learning Method for Modeling the Interaction between HLA-A*02:01 and Antigen...

Area of Science:

• Immunology and computational biology
• Structural bioinformatics
• Machine learning in drug discovery

Background:

• Major histocompatibility complexes (MHCs) and their interactions with antigen peptides are crucial for T cell-mediated immunity.
• Deep learning (DL) models accelerate antigen peptide screening, but often neglect 3D structural information.
• Existing sequence-based DL models have limitations in predicting peptide-MHC binding accurately.

Purpose of the Study:

• To develop a novel structure-based DL model, TransfIGN, for predicting antigen peptide interactions with HLA-A*02:01.
• To integrate sequence information from transformers with structural data for enhanced prediction accuracy.
• To provide physically interpretable insights into peptide-MHC binding mechanisms.

Main Methods:

• Developed TransfIGN, a DL model inspired by Interaction Graph Network (IGN) and incorporating transformer-based sequence features.
• Trained the model on a large dataset of 61,816 sequences with binding affinity and eluted ligand labels.
• Evaluated model performance against state-of-the-art sequence-based models and benchmark datasets.

Main Results:

• TransfIGN achieved an AUC of 0.893 on a binary dataset, outperforming NetMHCpan4.1, ANN, and TransPHLA.
• On IEDB benchmark datasets, TransfIGN predictions (AUC = 0.816) surpassed the IEDB consensus (AUC = 0.795).
• Generated interaction weight matrices revealed specific peptide positions with strong interactions, offering physical interpretability.

Conclusions:

• TransfIGN demonstrates superior performance in predicting antigen peptide-MHC interactions compared to existing methods.
• The model's structure-based approach and interpretability offer valuable insights into binding mechanisms.
• This work holds promise for advancing immunotherapies by improving the understanding of T cell epitope identification.