LAUSR

Introduction: Neoantigen-based biomarkers have improved predictions of response to immune checkpoint blockade (ICB) therapy, highlighting the importance of accurate prediction of immunogenic neoantigen candidates. A challenge in developing robust immunogenicity prediction models is limited availability of sequencing-associated immunogenicity data for evaluating methods due to the complexity of generating such datasets. We propose a novel approach to optimize prediction models of immunogenic neoantigens using a meta-analysis framework based on multiple ICB cohorts. Methods: To build on the 110 mono-allelic immunopeptidomics-derived SHERPA® MHC binding prediction framework, we engineered T-cell recognition features on two datasets: peptide-centric data aggregated by Schmidt et al. and patient-specific exome and transcriptome sequencing data from the TESLA consortium. We developed two-tiered models based on the feature landscapes of both datasets to predict peptide-MHC (pMHC) immunogenicity, incorporating features with significant performance gains. We systematically re-processed publicly available DNA and RNA sequencing data from over 500 ICB treated patients spanning 12 different cohorts across five different cancer types with a harmonized bioinformatics pipeline. We then evaluated the performance of each model consisting of a unique combination of immunogenic features across the ICB training (N=7) and validation (N=5) cohorts using a meta-analysis framework. Results: We evaluated iterations of SHERPA-Immunogenicity (SI) models using the Schmidt et al. and TESLA datasets, resulting in a range of performance metrics (area under the precision recall curves of 0.74-0.84 and positive predictive values of 0.32-0.54). After aggregating pMHC predictions into patient-specific scores based on the most immunogenic peptide present (SHERPA-Immunogenicity Maximum - SIM) or the quantity of immunogenic peptides identified (SHERPA-Immunogenicity Burden - SIB), we observed that responders had higher SIM and SIB scores compared to non-responders across the melanoma training cohorts. We found SIM scores outperformed SIB scores, suggesting the degree of epitope immunogenicity may be a critical factor in predicting response. The model with the most significant meta p-value for ICB response in melanoma cohorts (OR=2.43, p=0.006) also predicted overall survival in 3/5 melanoma cohorts (p<0.05). Conclusions: We developed a novel framework to predict neoantigen immunogenicity utilizing meta-analysis of ICB cohorts to overcome dataset limitations and gain prediction performance confidence. We look forward to supporting personalized cancer vaccine development with our pMHC immunogenicity predictions and applying our predictive biomarker on additional ICB cohorts. Citation Format: Hima Anbunathan, Neeraja Ravi, Rachel Marty Pyke, Steven Dea, Richard O. Chen, Sean Michael Boyle. Utilizing response in immune checkpoint inhibitor treated cohorts improves clinical applicability of neoantigen immunogenicity predictions. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5472.