Off-target HLA matches for prediction of response in personalised cancer vaccines: A multi-trial retrospective analysis.

Abstract

2653 Background: Personalized neoantigen vaccines aim to match epitopes to a patient's HLA genotype, yet many peptides bind promiscuously across multiple HLA alleles. Promiscuous binding may reflect intrinsic peptide properties, including enhanced processing efficiency and structural stability, that drive immunogenicity independent of any single HLA match. We hypothesized that peptides with broader HLA binding profiles would show higher clinical immunogenicity, even when restricted to the patient's own HLA alleles. Methods: We curated 17 neoantigen vaccine trials (174 patients) and selected five with per-epitope CD8⁺ T cell immunogenicity data for primary analysis. Across 571 neoantigen sequences (3,806 derived peptides), we predicted binding and presentation using NetMHCpan-4.2, MHCflurry 2.0, and PRIME-2.0, and estimated peptide–MHC stability with NetMHCstabpan. A reference panel of 62 common HLA class I alleles (>95% global coverage) was used to quantify incidental coverage, defined as predicted binding (IC50 <500 nM) to non-patient HLA alleles. Mixed-effects models adjusted for patient-specific binding affinity, predicted stability, mutation type, neoantigen length, and trial structure, with Bonferroni correction for 5 pre-specified hypotheses. Results: Off-target HLA binding independently predicted clinical immunogenicity. Among peptides selected for vaccination, 35.2% bound at least one non-patient HLA allele, with the most promiscuous peptides binding up to 37 alleles. After adjusting for patient-specific binding affinity and stability, peptides from immunogenic neoantigens showed greater incidental coverage than non-immunogenic peptides (mean 2.46 vs 1.80 additional HLAs; p=0.0034). This association was consistent across tumour types and prediction methods. Among immunogenic peptides, the breadth of off-target binding correlated with response magnitude (Spearman ρ=0.46, p=0.034), with the strongest effect observed in glioblastoma. High-affinity off-target matches (IC50 <50 nM) and high-stability interactions showed the most robust associations with immunogenicity. Conclusions: Promiscuous HLA binding independently predicts neoantigen immunogenicity in clinical trials, beyond binding affinity and stability to the patient’s own HLA alleles. These results point to peptide-intrinsic properties linked to MHC stability and processing that are overlooked by current selection pipelines. Explicit modelling of HLA promiscuity and stability may improve neoantigen prioritisation, particularly in low–mutation burden tumours and patients with rare HLA genotypes.