Abstract 6255: Patient response to immune checkpoint inhibitors (ICI) informs neoantigen prioritization: Applications to personalized cancer vaccines (PCVs) and ICI response prediction

Abstract Accurate identification of neoantigens that can stimulate a cytotoxic immune response is essential in designing PCVs. Here, we describe the design and performance of a neoantigen prioritization model for PCV design and ICI response prediction.The machine learning (ML)-based neoantigen prioritization model was implemented using tumor and matched normal whole exome sequencing data and clinical information to identify candidate neoantigens. Neoantigens were grouped across combinations of three characteristics: MHC-I binding IC50, T-cell immunogenicity predictions, and inferred gene expression. For each neoantigen combination, ICI response, as measured by progression-free survival (PFS), was compared using Cox proportional hazards tests. Statistically significant hazard ratios were used as a prioritization metric for each neoantigen combination. Model performance for neoantigen prioritization was evaluated using three publicly available datasets (TESLA N=8, NCI N=97, HiTIDE N=10) with known immunogenic, true-positive (TP) neoantigens. The percent of TP neoantigens prioritized within the top 100, 50, and 20 per patient was computed. Furthermore, a real-world clinical cohort (N=543; training 80%, test 20%) and an additional 66 patients with solid tumors from the Bespoke IO clinical study were utilized to validate our model for ICI response prediction as compared to TMB.In this cohort, hazard ratios of candidate neoantigens correlated with higher inferred gene expression and inversely correlated with IC50 and T-cell activation rank. In a separate analysis, we demonstrated that TCGA-based inferred gene expression, in the absence of patient-matched RNAseq, was sufficient to identify immunogenic neoantigens. Across all three test sets, the known TPneoantigens were consistently prioritized in the top 100 (NCI 89.7%, TESLA 91.4%, HiTIDE 96.4%) and top 50 (NCI 77.2%, TESLA 80.0%, HiTIDE 78.6%). Approximately half of TP neoantigens were prioritized in the top 20 (NCI 54.5%, TESLA 54.3%, HiTIDE 46.4%). These results outperformed other methods that have used TESLA data. The classification of patients into non-responders and responders was significantly better by our model for ICI response prediction [melanoma (HR: 13.69, p=0.0005), NSCLC (HR: 5.79, p=0.036) and MSI-H CRC (HR: 23.1, p=0.0001)] compared to TMB [melanoma (HR: 2.56, p=0.07), NSCLC (HR: 1.04, p=0.96) and MSI-H CRC (HR: 3.31, p=0.09)]. Similar trends were observed in the blinded validation cohort [(model: HR=3.16, p=0.006) vs. TMB (HR=1.47, p=0.33)].Our ML-based neoantigen prioritization model is the first to our knowledge to be informed by in vivo ICI response that effectively identifies and prioritizes immunogenic neoantigens for PCV design and ICI response prediction. Citation Format: Robert Burns, Sara L. Bristow, J. Bryce Ortiz, Meenakshi Malhotra, Jasreet Hundal, Aaron M. Rosenfeld, Helio A. Costa, Alexey Aleshin, Matthew Rabinowitz. Patient response to immune checkpoint inhibitors (ICI) informs neoantigen prioritization: Applications to personalized cancer vaccines (PCVs) and ICI response prediction [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 6255.