Abstract

BackgroundThe development of personalized neoantigen-based therapeutic cancer vaccines relies on computational algorithm-based pipelines. One of the critical issues in the pipeline is obtaining higher positive predictive value (PPV) performance, i.e., how many are immunogenic when selecting the top 5 to 20 candidate neoepitopes for the vaccination. We attempted to test the PPV of a neoepitope prediction algorithm Neopepsee. MethodsSix breast cancer patients and patient-derived xenografts from three lung cancer patients and their paired peripheral blood samples were subjected to whole-exome and RNA sequencing. Neoantigen was predicted using two different algorithms (Neopepsee and pVACseq). Response of induced memory T cells to neopeptide candidates was evaluated by IFN-{gamma} Enzyme-linked immune absorbent spot (ELISpot) assays of peripheral blood mononuclear cell (PBMC) from three HLA-matched donors. Positive ELISpot response to a candidate peptide in at least 2 of 3 donor PBMC was regarded as an immunogenic response. ResultsNeopepsee predicted 159 HLA-A matched neoepitope candidates out of 898 somatic mutations in nine patients (six breast and three lung cancer patients), whereas pVACseq predicted 84 HLA-A matched candidates. A total of 26 neopeptide candidates overlapped between the two predicted candidate pools. Among the candidates, 28 (20%, 28/ 137) and 15 (20%, 15/ 75) were positive by ELISpot assay, respectively. Among 26 overlapped candidates, 20 could be tested, and 7 of them (35%) were validated by ELISpot. Neopepsee identified at least one neoepitope in 7 of 9 patients (range 0-6), compared to 6 by pVACseq (range 0-5). ConclusionAs suggested by Tumor Neoantigen Selection Alliance (TESLA), our results demonstrate low PPV of individual prediction models as well as the complementary nature of the Neopepsee and pVACseq and may help design neoepitope targeted cancer vaccines. Our data contribute a significant addition to the database of tested neoepitope candidates that can be utilized to further train and improve the prediction algorithms.