Abstract 2085: Precision neoantigen discovery using a pan-allelic machine learning model for enabling the development of composite biomarkers and personalized immunotherapy

Abstract

Abstract Background: Technologies for neoantigen discovery are critical for developing more advanced, composite biomarkers for immunotherapy as well as personalized cancer therapies. Precision neoantigen discovery entails comprehensive detection of tumor specific genomic variants and accurate prediction of MHC presentation of epitopes originating from such variants. Here we present our pan-allelic machine learning model for predicting MHC class I presentation and identifying potentially immunogenic patient-specific neoantigens. We then apply our predictions to develop composite biomarkers that can stratify patients by response to immunotherapy. Methods: Mono-allelic cell lines were generated by transfecting a single allele of interest into HLA-null K562 cell lines. Immunoprecipitation mass spectrometry (IPMS) was performed as follows: W6/32 antibody was used for immunoprecipitation of HLA complexes, followed by elution of bound peptides and identification using liquid chromatography-mass spectrometry. Machine learning models were implemented to predict neoantigen presentation. Our prediction model, integrated into the ImmunoID NeXT platform, was used with other genomic features to generate composite neoantigen-based biomarkers. Results: We generated a high quality and unambiguous immunopeptidomics training dataset by performing IPMS on ~60 mono-allelic cell lines, with ongoing efforts to expand it to ~100 alleles. These alleles, selected to optimize both allelic diversity and population coverage, enable accurate and comprehensive modeling of MHC ligand processing and presentation. Our advanced prediction model combines multiple modelling strategies, including deep neural networks, convolutional neural networks and gradient boosted decision trees. New features that model antigen-processing were implemented using large scale public and private datasets to improve presentation specificity (60% PPV in dataset containing 999-fold decoys). As a result, our pan-allelic model has significantly higher specificity across a range of sensitivity values in comparison to NetMHCPan 4.0 and generalizes to both trained and untrained alleles. Our comprehensive validation strategy includes: evaluation of overall performance of the model on an independent, multi-allelic immunopeptidomics dataset generated from tumor samples; validation of top ranking tumor specific neoantigens nominated using an integrated patient-centric model that incorporates HLA loss of heterozygosity using targeted proteomics (Parallel Reaction Monitoring); and evaluation of our model's utility to predict neoantigens that drive immunogenic responses to tumors. Finally, a composite neoantigen-based biomarker score calculated using our model stratifies patients by response to immunotherapy. Conclusions: In summary, we present here a pan-allelic MHC presentation prediction model trained on a large mono-allelic data set and evaluated using tumor samples and known immunogenic peptides. Through integration with ImmunoID NeXT, our platform enables precision neoantigen discovery by comprehensively surveying neoantigens and accurately predicting MHC presentation. These methods can significantly enhance the creation of composite biomarkers and applications in personalized immunotherapy. Citation Format: DATTATREYA MELLACHERUVU, Rachel Marty Pyke, Charles Abbott, Nick Phillips, Rena McClory, John West, Richard Chen, Sean Michael Boyle. Precision neoantigen discovery using a pan-allelic machine learning model for enabling the development of composite biomarkers and personalized immunotherapy [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 2085.