Protein language models accurately predict polymorphic peptide-modulated NK cell receptor-HLA class I interaction strengths.

Killer-cell immunoglobulin-like receptors (KIRs) are key determinants of natural killer cell function and are associated with the outcomes of infective, inflammatory, and neoplastic diseases. They form a polymorphic family of activating and inhibitory receptors that interact with polymorphic class I human leukocyte antigen (HLA-I) molecules. This interaction is dependent on the short peptides bound by the HLA-I molecules, including those derived from viruses and cancers. Identifying these peptides among the vast space of possible peptides based on the sequences of the interacting molecules can provide a valuable tool for developing personalized immunotherapy against infection and cancer. To address this challenge, we leveraged foundation protein language models and trained our model on available datasets for KIR-binding peptide-HLA complexes. Our tool generated excellent predictions with an area under receiver operator characteristic (AUROC) >0.8 for the majority of inhibitory KIRs and performed well (AUROC >0.7) for peptides generated during HIV and HCV infections. Our model holds substantial potential for advancing our understanding of immune regulation and the biophysical factors responsible for it, paving the way for KIR-specific therapeutic interventions.