Background:Neoantigens are increasingly critical in immuno-oncology as a therapeutic target for neoantigen-based personalized cancer vaccines and as a potential biomarker for immunotherapy response. However, the methods for identifying which neoepitopes are more likely to provoke an immune response remains an important challenge for improving both the effectiveness of neoantigen-based vaccines and enabling the potential use of neoantigens as a biomarker in immunotherapy.

Methods:We sought to improve overall neoantigen identification performance by systematically improving critical components of our ACE ImmunoID assays and neoantigen pipeline. Personalis' Accuracy and Content Enhanced (ACE) technology was developed to fill critical gaps in conventional exome and transcriptome sequencing that can lead to missed neoantigens. To improve MHC-epitope binding prediction, we trained neural networks on mass spectrometry derived MHC-epitope binding data. This is in contrast to other MHC binding algorithms that have been primarily trained using in vitro competitive binding data, which suffer from having not been processed, loaded, nor shuttled natively into the HLA binding domain. HLA typing, a key input into the neoantigen prediction algorithms, was improved through exome augmentation of the HLA region with an optimized HLA typing algorithm. Other enhancements include RNA based somatic variant calling, peptide phasing, transcript isoform estimation, and identification of indel and fusion derived neoepitopes.

Results:Our ACE augmented exome demonstrates high sensitivity and specificity for SNVs, indels, and fusions at MAF >=10%. These are all variant types that result in putative neoantigens. Further, we show that our augmented ACE transcriptome can achieve high sensitivity for RNA derived variants and can be an important filter for putative neoantigens. When compared with commercially available MHC binding algorithms for specific HLA alleles, our MHC binding prediction algorithm consistently achieves a higher overall sensitivity and specificity than other tools. For example, our MHC class I-epitope binding prediction algorithm demonstrated an aggregative precision value of 0.88 across HLA alleles, as opposed to 0.50 for other widely used tools. To assess overall HLA-typing performance, we performed a blinded clinical HLA typing validation demonstrating 98% and 95% concordance with Class I and II HLA results (respectively) from clinical testing. We also show instances where peptide phasing, SNP, indel and fusion-derived neoepitopes are important for more accurate and comprehensive neoantigen identification.

Citation Format: Sean Michael Boyle, Ravi Alla, Ryan Wang, Eric Levy, Gabor Bartha, Jason Harris, Robert McCord, Rena McClory, John West, Richard Chen. Methods of improving accuracy of neoantigen identification for therapeutic and diagnostic use in immuno-oncology [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 1292.