ReCLIP advances residue-level protein interaction modeling

Learning residue-level context for modeling protein-protein interactions

ReCLIP utilizes a transformer architecture to learn interaction-specific representations at the individual residue level, addressing limitations in current protein language models that aggregate information across entire proteins. The framework demonstrates strong performance in predicting mutation-induced perturbations (AUROC = 0.973) and enables zero-shot prediction of peptide-MHC binding across unseen alleles with an AUROC up to 0.972. By identifying structurally coherent residue neighborhoods, the method links pathogenic genetic variants to interaction perturbations, offering a precise tool for analyzing neoantigen-MHC interactions.