Abstract
Identification of T cell receptor (TCR) specificities for antigens from large-scale single-cell or bulk TCR repertoire data plays a vital role in disease diagnosis and immunotherapy. In silico prediction models have emerged in recent years. However, the generalizability and transferability of current computational models remain significant hurdles in accurately predicting TCR-pMHC binding specificity, primarily due to the limited availability of experimental data and the vast diversity of TCR sequences. In this paper, we propose a lightweight contrastive TCR-pMHC learning with context-aware prompts, named LightCTL, to infer TCR-pMHC binding specificity. For each TCR and peptide-MHC sequence, we utilize a TCR encoding module and a pMHC encoding module to transform them into latent representations. Specifically, we introduce a contrastive TCR-pMHC learning paradigm to enhance the generalization ability of TCR-pMHC binding specificity prediction by learning the matching relationship between TCR-pMHC and MHC-peptide. We fuse the TCR and pMHC latent representations and employ a novel context-aware prompt module to consider the varying importance of different feature maps. Compared with existing methods, LightCTL substantially improves the accuracy of predicting TCR-pMHC binding specificity. Moreover, comparative experiments across eight independent datasets demonstrate the generalization ability of LightCTL, showing superior performance for predicting unknown TCR-pMHC pairs. Finally, we assess LightCTL's efficacy across different TCR sequence lengths and distinct unseen epitopes, as well as estimate cytomegalovirus-specific TCR diversity and clone frequency from peripheral TCR repertoire data. Overall, our findings highlight LightCTL as a versatile analytical method for advancing novel T-cell therapies and identifying novel biomarkers for disease diagnosis.