STRUCUTRE-GUIDED MACHINE LEARNING MODELS FOR PREDICTING TCR:pMHC SPECIFICITY FROM SEQUENCE DATA

T cells play a pivotal role in the adaptive immune response through the recognition of antigenic epitopes presented as peptides bound to major histocompatibility complex molecules (pMHC), a process mediated by T cell receptors (TCRs). Computational methods have recently emerged to predict TCR:pMHC specificity, which means whether a given TCR-pMHC pair forms a true interaction complex. These predictive models are essential for deciphering TCR antigen specificity, enabling insights into the functional role of TCR repertoires in immune responses and disease diagnostics. In this study, we implemented an automated pipeline for training machine learning models using sequence data enriched with structural information derived from TCR:pMHC complexes available in the Protein Data Bank (PDB). Structural features were incorporated by selecting high-contact interface residues observed across known complexes. Training data were sourced from the public VDJdb database, and models were evaluated on an external test set from the IMMREP23 competition, filtered to include only peptides unseen during training. Model performance was assessed using the Macro AUC0.1 metric. Our results indicate that sequence data alone, even when guided by structural heuristics, are insufficient for robust generalization, with ensemble models achieving a maximum performance below 0.6. However, evaluations involving similar peptides revealed that improvements in data quality significantly enhanced performance, reaching a Macro AUC0.1 of 0.84. These findings highlight the need to develop models that explicitly incorporate structural features during training, to better capture the leverage of physicochemical and biophysical determinants of TCR:pMHC interactions.