Computational identification of antigen-specific T cell groups through generative epitope modeling

Single-cell T cell receptor (TCR) sequencing enables high-resolution analysis of TCR diversity and clonal expansion. However, inferring antigen specificity for individual TCRs remains challenging and typically requires costly functional assays. We introduce EpitopeGen, a transformer-based model that generates candidate epitope sequences from TCR sequences, enabling population-level identification of T cell groups with shared predicted antigen specificity. EpitopeGen employs a semi-supervised strategy that searches over 70 billion candidate TCR-epitope pairs and incorporates high-confidence binding predictions as pseudo-labels under biologically inspired quality control. Generated epitopes exhibit high predicted binding probabilities, sequence diversity, realistic biochemical properties, and biophysical stability. In cancer datasets, EpitopeGen identifies clonally expanded tumor-infiltrating CD8⁺ T cells enriched for tumor-associated antigen recognition and cytotoxic signatures. In patients with COVID-19, it reveals distinct transcriptomic profiles of T cells targeting spike and non-structural proteins across disease severities. EpitopeGen provides a scalable framework for studying disease-associated T cell populations, complementing experimental antigen validation.