TCRLens: structure-aware equivariant graph learning for TCR-pMHC-I recognition and immunogenic epitope discovery

Accurate prediction of T-cell receptor (TCR) recognition of peptide-MHC class I (pMHC-I) complexes is a key challenge due to structural diversity and data sparsity. We introduce TCRLens, a structure-aware deep learning framework that models residue-level interactions across five critical interface zones using multi-scale graph representations and an equivariant graph neural network (EGNN). To mitigate data sparsity and severe class imbalance arising from limited negative samples, TCRLens incorporates a variational autoencoder-generative adversarial network (VAE-GAN) to generate structurally plausible weak-affinity interaction samples. We evaluated TCRlens across three prediction tasks including peptide-MHC binding, peptide-TCR recognition, and full-complex TCR-pMHC-I interaction and observed consistently strong performance.

Using curated dataset of human TCR-pMHC-I structural complexes from ATLAS and TCR3d, TCRLens outperforms state-of-the-art sequence-based, motif-based, and structure-aware methods, including NetMHCpan 4.2, CapsNet-MHC, RPEMHC, NetTCR-2.0, TITAN, PanPep, pMTnet, ERGO II, and STAG. Furthermore, TCRLens demonstrates robust cross-species generalization, achieving high predictive performance in swine and chicken MHC-I systems. These findings highlight the value of geometry-aware representation learning and generative data augmentation for capturing immunological specificity. TCRLens provides a unified and extensible platform for TCR-pMHC-I interaction modeling, with potential applications in epitope discovery and structure-guided vaccine design across both human and veterinary immunology.

The code used in this study is publicly available at https://github.com/paopitsiri/TCRLens.