Deep mapping of structural perturbations to energetics enables precise TCR design

Affinity optimization and cross-reactivity profiling are pivotal for T-cell receptor (TCR) engineering and immunotherapy, yet remain hindered by the vast diversity of the TCR repertoire and limited structural insights. Here, we present mpTCRai, a deep learning framework that captures the sequential characteristics of antigen presentation and recognition with high precision. Leveraging these precise structural predictions, we established a contact hotspot-based scoring mechanism that explicitly maps structural perturbations to energetic changes, yielding a strong correlation with experimental affinity (r = −0.88). In application, the model identified critical molecular switches, such as specific G-to-A substitutions, and revealed their dependence on structural context. Furthermore, it effectively distinguished functional variants from dangerous cross-reactive mutations like Y5W, thereby mitigating the risk of off-target toxicity. Guided by these insights, we computationally designed four novel A6-TCR variants, demonstrating a rational strategy for candidate selection in adult T-cell leukemia. This work establishes a unified platform integrating structural and energetic constraints to advance precise, safe TCR therapeutic design.