Learning residue-level context for modeling protein-protein interactions

Protein language models (PLMs) enable prediction of protein properties by learning residue-level features from sequence, yet most PLM-based approaches to protein-protein interactions aggregate information across entire proteins, limiting resolution and interpretability. Here we present ReCLIP, a transformer-based framework that learns interaction-specific representations at the level of individual residues by combining intra-protein residue neighborhoods with residue-conditioned representations of interaction partners. We show that residue-centered context provides a general framework for modeling protein interactions across diverse biological settings. ReCLIP accurately predicts mutation-induced perturbations (AUROC = 0.973), generalizes to post-translational modifications that do not alter sequence (AUROC = 0.822), and enables zero-shot prediction of peptide-MHC binding across unseen alleles (AUROC up to 0.972). Analysis of learned residue neighborhoods reveals structurally and functionally coherent patterns aligned with known determinants of binding. Applied to clinically annotated genetic variants, ReCLIP identifies disease-associated interaction perturbations that link pathogenic variants to specific molecular interaction contexts. Our results establish a generalizable and interpretable framework for modeling protein interactions and provide insights into how residue-level context shapes interaction specificity and its perturbation.