Design of TCR-mimicking binders for pHLA with high potency

The rational design of high-specificity binders to peptide–HLA (pHLA) complexes remains a major challenge in personalized immunotherapy, particularly for shared neoantigens with single-point mutations. To address this, we have developed an integrated framework that combines knowledge-based deep learning with physics-based simulation for the design of highly specific pHLA binders. Applied to p53 R175H–HLA-A*02:01, molecular-dynamics–guided electrostatic filtering yielded nanomolar binders with >10-fold selectivity over wild type. Motivated by the challenge of KRAS G12V–HLA-A*03:01, which lacks electrostatic cues, we implemented message-passing neural network (MPNN)–based optimization strategies to broaden binder design beyond charge-dependent interfaces. When reapplied to p53 R175H, these strategies improved success rates from 2/5 to 5/5, with most candidates exhibiting minimal or no binding to the wild-type complex. Chimera-based and residue-retention approaches within the MPNN optimization further expanded the sequence–structure search space while preserving critical hotspot interactions. This integrated framework thus enables the design of nanomolar, mutation-selective binders to pHLA complexes, advancing next-generation personalized cancer immunotherapies.