GRASP-MHC: structure-based representations narrow the unseen-allele gap in pMHC binding affinity prediction

Ranking peptides by MHC class I binding affinity is critical for antigen selection, yet sequence-based predictors fail on unseen alleles under-represented in training data. Here, reading interface representations from a frozen Boltz-2 structure model bridges this generalization gap. A 0.48M-parameter 2D convolutional head regressing affinity directly from the uncompressed 365 × 14 × 128 residue-pair block achieves Spearman ρ = 0.631 on unseen alleles, outperforming engineered inter7 face features (ρ = 0.538) and global confidence scores (ρ = 0.211). On unseen alleles across two independent partitions, this structural head beats a NetMHCpan9 4.1 baseline retrained on identical splits despite using five times less data, while performing comparably on known alleles. Unsupervised latent space organization reveals locus- and species-level clustering, mapping unseen alleles adjacent to structural rather than sequence neighbors. Extracting fine-grained spatial inter face representations enables robust antigen identification across rare patient HLA genotypes without requiring experimental affinity profiling.