Abstract Adaptive immunity is governed by both qualitative and quantitative features of antigen presentation. In anti-cancer T cell responses, ‘quality’ reflects peptide-MHC binding, whereas ‘quantity’ reflects how abundantly each epitope is displayed. Although binding has been studied extensively, the contribution of peptide abundance to neoantigen immunogenicity remains poorly understood, mainly because of limited immunopeptidome datasets. To bridge this gap, we developed epiVIP, a novel virtual immunopeptidomics method that predicts HLA-I peptide abundance. We first established that mass spectrometry (MS) intensity provides a robust and scalable approximation of absolute epitope abundance. Leveraging this, we curated and uniformly quantified 1.7 million HLA-I peptides from 254 tumors with paired transcriptomes, creating the largest quantitatively standardized immunopeptidome resource to date. We then developed epiVIP, a deep neural network that integrates peptide and HLA sequences with the expression of the peptide’s source gene and 472 putative regulators of antigen presentation. To minimize batch effects, we used a pairwise ranking loss strategy. epiVIP achieved high prediction accuracy and generalizability for within-sample abundance, with AUC0.8 for 20 held-out samples and 24 independent samples. Applying epiVIP to 33,782 neoantigens from four studies, we observed that higher predicted abundance was strongly associated with increased immunogenicity. Importantly, the effect was conditional on self-discrimination, defined as the sequence similarity between the neoantigen and its wild-type counterpart. Neoantigens with low self-discrimination required high abundance to elicit T cell responses, whereas those with high self-discrimination were immunogenic regardless of abundance. In three immune-checkpoint blockade cohorts, the summed abundance of low self-discrimination neoantigens outperformed tumor mutational burden in predicting response and survival (p = 4.7*10-4 vs 1.3*10-3 in one cohort). To identify regulators of epitope abundance, we first validated that the predicted abundance changes accurately recapitulated antigen-repertoire remodeling after PSME4 knockdown in A549 cells. We then extended predictions to 409 regulatory gene knockdowns in HCT116 and HEK293T using pseudobulked perturb-seq profiles. We observed that perturbation effects clustered by C-terminus amino acid properties and identified 32 regulators with C-terminal-specific effects, including PSME4 and PSMF1. In summary, epiVIP establishes epitope abundance as a key quantitative determinant of neoantigen immunogenicity, provides a model to predict abundance when immunopeptidomics is unavailable, and offers a framework to identify gene perturbations that enhance presentation of desired epitopes for TCR-T and cancer vaccine development. Citation Format: Yuhao Tan, Ziqi Yang, Julia Fleming, Hailong Hu, Bo Li, . AI-empowered virtual immunopeptidomics uncovers novel regulators of neoantigen immunogenicity abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6889.
Tan et al. (Fri,) studied this question.