Interpretable TCR-Epitope Prediction

April 1, 2026

Interpretable TCR-Epitope Prediction

Interpretable models for predicting and interpreting how T-cell receptors recognize peptide-MHC complexes.

T cells orchestrate the adaptive immune response by recognizing antigens presented as
peptide-MHC (pMHC) complexes through the T-cell receptor (TCR). Accurately
predicting, and just as importantly explaining, TCR-pMHC binding is fundamental to
understanding immune recognition and to designing immunotherapies. This project develops a
family of transformer-based models that are both accurate and interpretable.

Motivation

State-of-the-art predictors such as transformer encoder/decoder models achieve strong
performance, but their black-box nature limits mechanistic insight. Our goal is to make
TCR-pMHC models explainable, recovering the structural basis of binding without
sacrificing predictive accuracy.

Interpreting attention

Our flagship post-hoc method quantifies cross-attention interaction to reveal which
residues drive binding, benchmarked against experimentally determined structures:

Quantifying Cross-Attention Interaction in Transformers for Interpreting TCR-pMHC Binding

(ICLR 2026) The Fourteenth International Conference on Learning Representations

Jiarui Li, Zixiang Yin, Haley Smith, Zhengming Ding, Samuel J. Landry, Ramgopal R. Mettu

Paper arXiv GitHub

Building explainability into the model

We also designed ante-hoc explainable model layers, so interpretability is intrinsic to
the architecture rather than added after the fact:

TCR-EML: Explainable Model Layers for TCR-pMHC Prediction

(LMRL 2026) Learning Meaningful Representations of Life (LMRL) Workshop at ICLR 2026

Jiarui Li, Zixiang Yin, Zhengming Ding, Samuel J. Landry, Ramgopal R. Mettu

Paper arXiv GitHub HuggingFace

Beyond interpretability, we improved prediction itself with rational multi-modal
transformers that fuse sequence and structural signals. Together with our work on
accelerating the underlying biophysical computations, these efforts form a complete
pipeline from epitope processing to TCR recognition.

Rational Multi-Modal Transformers for TCR-pMHC Prediction

(ACM BCB 2025) The ACM Conference on Bioinformatics, Computational Biology, and Health Informatics

Jiarui Li, Zixiang Yin, Zhengming Ding, Samuel J. Landry, Ramgopal R. Mettu

Paper PubMed GitHub HuggingFace arXiv