TridentTCR: Predicting T Cell Receptor Specificity for Antigens and Autoimmune-Related Targets via a Topology-Aware Graph and Large Language Model
T cells recognize and eliminate diseased cells by binding their T cell receptors (TCRs) to short endogenous peptides presented on the cell surface, commonly referred to as antigens. Accurate prediction of antigen-TCR specificity is essential for advancing T cell-based immunotherapies while minimizing autoimmune toxicity. However, current computational methods struggle to generalize across diverse TCR repertoires and reliably differentiate therapeutic antigens from autoimmune-related antigens (arAgs). Here, we introduce TridentTCR, a computational framework integrating a pretrained large language model (LLM) for high-fidelity sequence embedding with a topology-aware graph neural network to capture structural insights from the antigen-TCR interaction network. TridentTCR significantly advances TCR specificity prediction by enabling robust trinary classification, distinguishing TCR interactions involving general antigens, arAgs, and non-binding. Additionally, TridentTCR outperforms existing state-of-the-art models in binary classification (binding vs. non-binding). Validation on independent clinical datasets demonstrated strong generalization capability for unseen TCRs and antigens. In a disease-specific context, TridentTCR identifies notable cross-reactivity between ZnT8186-194-specific TCRs and a mimotope from Bacteroides stercoris, suggesting that T cell cross-reactivity may contribute to the initiation of type 1 diabetes. Furthermore, we introduced a quantitative metric, antigenic immune response entropy (AIRE), which leverages TridentTCR predictions alongside clonotype frequencies and repertoire diversity to precisely quantify antigen-specific immune responses from single-cell profiling data. Collectively, TridentTCR provides an interpretable and clinically relevant tool, enabling improved understanding of TCR specificity, cross-reactivity, and off-target autoimmune risks in clinical immunotherapy.