Abstract 4536: Applying immunopeptidomics and machine learning to improve neoantigen prediction for therapeutic and diagnostic use

Abstract

Background: Neoantigens are increasingly critical in immuno-oncology as therapeutic targets for neoantigen-based personalized cancer vaccines (PCVs) and as potential biomarkers for immunotherapy response. However, identifying which neoepitopes are more likely to provoke an immune response remains an important challenge for improving the effectiveness of PCVs and enabling neoantigens as a biomarker in immunotherapy. In recent years, Immuno-peptidomics has greatly improved in sensitivity and specificity, providing large number of peptides bound to MHC class I alleles in vivo . These advances make it possible to identify processed cell surface MHC bound peptides in an in vivo setting, providing accurate and representative presented peptide data for development of an improved neoantigen prediction pipeline. Methods: We generated high quality allele-specific training data for development of an accurate predictive algorithm. Mono-allelic HLA class I cell lines were generated by transfecting individual class I HLA alleles into the HLA class I null cell line K562, prioritizing alleles which will ultimately allow for development of a pan-class-I-allele prediction algorithm. Cell surface bound MHC class I peptides were identified for each transfected allele using immuno-peptidomics. We then developed and trained neural networks to predict MHC class I presentation for each assayed HLA allele. The predictive accuracy for each allele was comprehensively validated using immuno-peptidomic results derived from three sources: mono-allelic cell lines, deconvoluted cell lines, and patient derived tumor samples. Results: We applied immuno-peptidomics to develop a large and highly representative profile of MHC class I peptidomics across 30 HLA class I alleles. We then utilized this dataset to develop a highly accurate HLA class I presentation neural network. Through our work, we have identified thousands of HLA class I peptides bound to each of 30 unique HLA class I alleles, greatly expanding the known mono-allelic space. Our neoantigen prediction algorithm has been extensively validated, consistently achieving a higher overall accuracy across alleles (precision 0.88) than other publicly available tools (precision less than 0.7) based on both in vitro binding data and immuno-peptidomics, when tested on a broad set of peptide sources: mono-allelic cell lines, deconvoluted cell lines, and patient-derived tumor samples. Conclusions: Effective neoantigen identification can be greatly improved through application of immuno-peptidomics. We have generated extensive mono-allelic HLA class I cell lines and extensively characterized their class I ligandomes. We have used this data to develop and train a novel presentation neural network. Finally, we have extensively validated this tool using multiple newly-derived in vitro and in vivo sources, demonstrating very strong accuracy. Citation Format: Datta Mellacheruvu, Nick Phillips, Gabor Bartha, Jason Harris, Robert Power, Rena McClory, John West, Richard Chen, Sean Michael Boyle. Applying immunopeptidomics and machine learning to improve neoantigen prediction for therapeutic and diagnostic use [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 4536.