Generative Design of High-affinity T-cell Receptors by Progressive Learning with Structural Confidence

The engineering of T-cell receptors (TCRs) for precision immunotherapy is constrained by the difficulty of modeling the conformational dynamics of hypervariable regions, a challenge exacerbated by the scarcity of experimental structures. In this paper, we constructed ProTSC-TCR, a specialized model for the sequence-structure co-design of TCRs conditioned on specific peptide-major histocompatibility complexes (pMHCs), through a progressive learning framework that synthesizes the sequence-structure distribution of TCR-pMHCs by knowledge transfer from antibody-antigen complexes and structure prediction tools. ProTSC-TCR establishes a new state-of-the-art in TCR design, demonstrating robustness across experimental and computationally predicted TCR-pMHC templates, enabling the construction of a million-scale design database (ProTSC-TCR-DB) covering 957 pMHC targets. In experimental validation targeting SARS-CoV-2 Spike-Y453F and KRAS-G12V antigens, the model achieved a 55.0% success rate across 20 variants, delivering a 10-fold affinity improvement (37.8 μM to 3.5 μM) over wild-type receptors. These results underscore the broad utility of ProTSC-TCR and its learning framework.