Rational Multi-Modal Transformers for TCR-pMHC Prediction.

Area of Science:

• Immunology
• Computational Biology
• Deep Learning

Background:

• T cell receptor (TCR) recognition of peptide-MHC (pMHC) complexes is crucial for adaptive immunity and T cell-based immunotherapies.
• Existing transformer models for TCR-pMHC prediction often lack systematic and explainable architectural design.

Purpose of the Study:

• To develop a novel, explanation-driven encoder-decoder transformer model for predicting TCR-pMHC interactions.
• To enhance model explainability, robustness, and generalization through a principled, data-informed approach.

Main Methods:

• Utilized a post-hoc explainability method to guide the construction of a novel encoder-decoder transformer architecture.
• Optimized cross-attention strategies and incorporated auxiliary training objectives based on input sequence informativeness.
• Introduced a new early-stopping criterion based on explanation quality.

Main Results:

• Achieved state-of-the-art predictive performance in modeling TCR-pMHC binding.
• Significantly improved model explainability, robustness, and generalization capabilities.
• Provided mechanistic insights into sequence-level binding behavior via deep learning.

Conclusions:

• Established a principled, explanation-driven strategy for modeling TCR-pMHC interactions.
• Demonstrated the utility of explainability methods in enhancing deep learning model design for biological systems.
• Advanced the understanding of TCR-pMHC binding mechanisms through a novel computational framework.