Ranking neoantigens for escape-resilience rather than predicted immunogenicity: Associated changes in vaccine design.

2654 Background: Personalised neoantigen vaccines induce detectable CD8 + T cell responses for fewer than one-third of selected peptides. Current pipelines prioritise candidates by predicted immunogenicity and select peptides independently, overlooking two constraints: efficacy depends on the peptide set as a whole, and tumours adapt antigen processing under immune pressure. We developed a framework that optimises peptide combinations for resilience to tumour escape and tested whether escape-resilience predicts clinical immunogenicity. Methods: We modelled peptide susceptibility to five antigen-processing escape mechanisms: TAP downregulation, immunoproteasome-to-constitutive proteasome switching, aminopeptidase upregulation, tapasin loss, and HLA loss of heterozygosity. Selection was formulated as a minimax optimisation, maximising predicted efficacy under worst-case tumour adaptation. We analysed five neoantigen vaccine trials with per-epitope CD8 + T cell response data, consisting of 571 neoantigens, 3,806 peptides, and 174 patients. Mixed-effects models tested associations between escape-resilience and immunogenicity, adjusting for binding affinity (NetMHCpan-4.1 %rank), pMHC stability (NetMHCstabpan), mutation type, and clonality. Six hypotheses were pre-registered with Bonferroni correction. Results: Escape-resilience predicted immunogenicity independently of established features. After adjustment, vulnerability to TAP loss (OR 0.42 per SD, 95% CI 0.24–0.71, p=0.0018) and proteasome switching (OR 0.54 per SD, 95% CI 0.34–0.86, p=0.0089) were associated with failure to elicit CD8 + responses. Among peptides with comparable predicted binding affinity, escape-resilient peptides were significantly more likely to be immunogenic. Composite escape-resilience scores discriminated immunogenic from non-immunogenic peptides (AUC 0.71, 95% CI 0.66–0.76), outperforming binding affinity alone (AUC 0.58) and an affinity-stability model (AUC 0.64). Adding escape-resilience improved discrimination (ΔAUC 0.07, p=0.003). Retrospective re-ranking altered 38% (95% CI 31–45%) of vaccine compositions, replacing high-affinity but escape-vulnerable peptides with lower-affinity, processing-robust alternatives. Associations were stronger for truncal mutations (OR 2.8, 95% CI 1.6–4.9) than subclonal mutations (OR 1.4, 95% CI 0.8–2.4; interaction p=0.041), indicating that processing robustness is most consequential for clonally dominant neoantigens. Conclusions: Optimising neoantigen selection against tumour escape identifies peptides more likely to elicit CD8 + T cell responses, independent of binding affinity and stability. The stronger effects in truncal mutations suggest escape-aware ranking may be particularly valuable for durable, clone-targeted vaccination strategies. Prospective trials are needed to assess clinical impact.