Gritstone presents improvements to EDGE Platform at AACR 2024

Gritstone bio presented an update on its T cell epitope discovery platform, EDGE, at the 2024 American Association for Cancer Research Annual Meeting in San Diego, CA. The presentation details improvements Gritstone has made in prediction of peptide presentation by HLA Class I, associated with CD8+ T cells, since the publication of the initial results in 2018. It also reviews how Gritstone has expanded EDGE’s application to predict peptide presentation by HLA Class II. EDGE for Oncology: Class I antigens – predicted using allele-specific and pan-specific models Allele-specific model is an improved version of published 2018 EDGE model that predicts for 116 HLA alleles and achieves an Average Precision of 0.63 and Positive Predictive Value at 40% Recall of 0.79 Pan-specific model trains using the HLA allele sequence and is applicable to any Class I allele with known sequence, achieving an AP of 0.65 and PPV40 of 0.81 2-fold better performance vs MHCFlurry 2.0 when ranking mutations from 80 cancer patients based on immunogenicity Detectable CD8 responses to over half of the 20 administered candidate neoantigens per patient after treatment with Gritstone’s personalized cancer vaccines Class II antigens – predicted using EDGE-II model Uses a pretrained protein language model, a novel learned HLA allele-deconvolution strategy, and in-house immunopeptidomics training data Achieves a test set AP of 0.92 and outperforms NetMHCIIpan and BERTMHC on an externally curated validation set with an AP = 0.71 CD4 immunogenicity in a personalized cancer vaccine context is better predicted by EDGE-II than NetMHCIIpan and MARIA EDGE for Infectious Disease: Class I antigens – predicted using EDGE-ID model Optimizing EDGE for use on infectious diseases results in improved performance Trained using both human immunopeptidomics and infectious disease binding affinity datasets and tested on publicly available infectious disease datasets Better performance on HIV and Influenza A datasets vs. MHCFlurry 2.0; comparable SARS-CoV-2 performance