Mitigating negative data bias to enhance TCR–epitope binding and residue interaction prediction

Accurate prediction of the binding specificity between T-cell receptors (TCRs) and epitopes, along with the elucidation of their molecular interaction mechanisms, is pivotal for advancing immunotherapy and vaccine development. In this study, we propose a negative dataset construction strategy based on region-directed random mutations as an effective complement to traditional negative sampling methods. This strategy preserves the conserved amino acid motifs encoded by the V and J gene segments of the CDR3$\beta$ sequence while introducing key residue mutations within the central junctional region. By constructing hard negatives, this approach encourages the model to capture more discriminative TCR-epitope binding features. Based on this optimized dataset, we developed TranTCR, a computational framework comprising two models: TranTCR-bind, which focuses on global sequence-level binding probability prediction, and TranTCR-map, which leverages transfer learning to translate global binding knowledge into fine-grained characterizations of residue-level interactions, such as inter-residue distances and contact scores. Experimental results demonstrate that TranTCR-bind exhibits superior predictive performance and generalization robustness across various negative sampling protocols. Furthermore, TranTCR-map utilizes attention mechanisms to deeply resolve complex inter-amino acid associations, enabling the identification of latent binding patterns and the revelation of TCR cross-reactivity characteristics. This study provides an efficient computational tool for the high-throughput screening of TCR repertoires and the digital characterization of immune recognition mechanisms.