EPIC-TRACE: predicting TCR binding to unseen epitopes using attention and contextualized embeddings.

Area of Science:

• Immunology
• Computational Biology
• Machine Learning

Background:

• T cells are crucial for adaptive immunity against pathogens and cancer, but aberrant T cell responses can cause autoimmune diseases.
• T cell receptor (TCR) recognition of peptide-MHC (pMHC) complexes is fundamental to initiating immune responses.
• Predicting TCR-pMHC interactions is vital for understanding immune function and disease, yet current models face generalization challenges.

Purpose of the Study:

• To develop an advanced machine learning model for predicting TCR-pMHC interactions.
• To enhance the generalization capabilities of predictive models for TCR-pMHC binding.
• To investigate the contribution of individual features and data augmentation strategies in TCR-pMHC prediction.

Main Methods:

• Utilized ProtBERT embeddings for amino acid sequences of TCR alpha and beta chains, and epitopes.
• Employed convolution and multi-head attention architectures to model complex interactions.
• Incorporated epitope data with limited TCRs into training to improve model robustness.

Main Results:

• The developed model demonstrates strong performance in predicting TCR-pMHC interactions.
• Feature importance analysis highlights the significance of TCR chains, epitope, and MHC information.
• Including sparse epitope data improved model generalization compared to existing state-of-the-art methods.

Conclusions:

• The novel machine learning approach effectively predicts TCR-pMHC interactions, outperforming current models.
• The model's ability to generalize to unseen pMHCs offers significant potential for immunological research.
• This work provides a valuable tool for studying adaptive immunity, cancer immunology, and autoimmune diseases.