Machine Learning for TCR Repertoire Epitope Annotation and Pattern Discovery

T cells are central to adaptive immunity, recognizing antigenic peptides, called epitopes, via the T cell receptor (TCR). The immense diversity and cross-reactivity of the TCR repertoire makes direct interpretation of antigen specificity from repertoire sequencing challenging. High-throughput sequencing enables large-scale profiling of TCRs but does not directly reveal their target epitopes, requiring computational approaches to bridge this gap. This review outlines two complementary strategies, bottom-up and top-down approaches, to annotate TCR specificity. Bottom-up methods predict TCR-epitope specificity from curated TCR-epitope databases, identifying recurring patterns through distance-based, feature-based, or deep learning models. While effective for well-characterized epitopes, they are limited by biased training data, absence of negative data, and weak generalization to unseen epitopes. Top-down approaches instead infer antigen-driven responses from repertoire-level signals such as sequence similarity, enrichment, and TCR convergence. These methods enable discovery of disease- or exposure-associated TCR signatures without prior epitope knowledge but are sensitive to technical noise and biological confounding. Both approaches are complementary as bottom-up provides mechanistic specificity, while top-down enables discovery in complex datasets. Their integration, alongside multimodal modeling and improved benchmarking, is key to advancing TCR-epitope annotation and understanding adaptive immune responses.