An experimental benchmark for physics-based scoring of chemically modified peptide-HLA complexes across 45 HLA class I alleles

Peptide binding to major histocompatibility complex (MHC) class I molecules, known in humans as the human leukocyte antigens (HLA), is predicted almost entirely by sequence models trained on the twenty canonical amino acids, which cannot represent chemically modified peptides. Structure-based methods accept such ligands without retraining but have not been evaluated against cellular measurements. We present a paired benchmark for isolevuglandin (IsoLG) adduction, a lysine modification generated by lipid peroxidation and elevated in hypertension and other immune-mediated conditions. We docked 46,000 peptide-HLA complexes with Rosetta FlexPepDock in matched unadducted and IsoLG-adducted states across 45 class I alleles, and independently measured IsoLG-HLA surface complexes by flow cytometry-based FRET on 45 single-HLA-expressing K562 lines, normalized to surface HLA. Across the 44 alleles with both measurements, the mean IsoLG-adducted interface score is weakly and negatively associated with surface presentation (ρ = -0.34 [-0.58, -0.06], p = 0.023), consistent in direction at both loci. The paired ΔScore, the quantity a simple accommodation hypothesis would nominate, shows no association at all (ρ = +0.01, 95% CI [-0.33, +0.33], p = 0.97). We further find that the interface score is confounded by ligand atom count, since the adduct improves van der Waals contacts in 92% of complexes regardless of fit, so in these ensembles the interface score favors the adducted state. We offer this paired resource as an oracle for methods that claim to handle modified ligands.