Biophysical Modeling for Predicting Human T Cell Repertoire Specificity

Accurate prediction of T cell receptor (TCR) recognition specificity remains a major goal in immunology, with broad implications for vaccine development, targeted immunotherapy, and the management of autoimmune diseases. The extraordinary diversity of TCR and peptide–MHC (pMHC) repertoires across individuals presents a formidable challenge for computational modeling. Here, we present a structure-informed framework that integrates experimentally characterized and AlphaFold3-predicted TCR–pMHC complexes to capture the biophysical determinants of antigen recognition. This approach reveals how structural diversity and residue-level interactions govern specificity across tumor-associated and viral epitopes, and demonstrates predictive accuracy on clinically derived donor and patient TCRs in the context of hematopoietic stem cell transplantation (HSCT). Building upon these structural insights, we incorporate a deep learning framework that efficiently captures long-range dependencies and interaction patterns across TCR and antigen sequences. Trained on both experimental and structure-augmented datasets, the model generalizes effectively to unseen epitopes and TCRs, achieving high predictive accuracy and interpretability. Together, these results highlight the power of integrative modeling—linking structural biophysics with advanced computational methods—to elucidate the principles of TCR–antigen recognition and enable predictive immunology at scale.