Biophysical modeling for accurate T cell specificity prediction of viral and tumor antigens.

Accurate predictions of T cell receptor (TCR) specificity remain an important open problem in immunology, with broad implications for vaccine design, optimal immunotherapy, and improved management of autoimmune diseases. However, diversity in peptide antigens and TCR sequences at the level of individual patient repertoires remains a formidable computational challenge. Here, we develop a joint experimental and computational approach for predicting the antigen specificity of clinically-derived TCR sequences. Our model is trained on a combination of experimentally pre-identified and -predicted TCR-pMHC structures using AlphaFold3. We apply our structural model in the clinical setting of hematopoietic stem cell transplant (HSCT) and demonstrate that our model is able to effectively discern the specificity of previously unseen donor and patient-derived TCR sequences against tumor associated and viral antigens. Model performance was further enhanced through the integration of sequence-based clustering and structurally diverse training templates. Our results highlight the predictive capabilities of structurally guided machine learning frameworks, trained on a minority test dataset, for antigen specificity prediction on unseen TCR sequences and their potential impact on a wide range of immunological applications.