Leveraging Contrastive Learning for Conditional and Robust TCR Sequence Generation

Abstract T cell receptor (TCR) initiates T cell activation programs that are central to antiviral and antitumor immunity, and TCR-engineered T cell is a promising therapeutic approach in treating solid tumors. However, the design of TCR is time-consuming and labor-intensive, and computational TCR sequence generation remains comparatively underdeveloped with fewer dedicated models. We propose an encoder-decoder generation framework (i.e., Pep2Tcr) warm-started with pretrained peptide and TCR language models to conditionally generate TCR CDR3β sequences targeting a specific peptide. It also incorporates contrastive learning over batches to pull each TCR toward the peptide it pairs with. Through a series of progressive evaluations, Pep2Tcr outperforms other state-of-the-art methods under generation distribution, hit (independent testing sets), authenticity (third-party prediction), and generation probability (OLGA). Grafting and docking experiments of some predicted positive TCR sequences to the HLA-A*02:01-AAGIGILTV (melanoma antigen) complex showed plausible scores and potential interaction patterns for downstream experimental validation.