neoMS: attention-based prediction of MHC-I epitope presentation

Abstract

Personalised immunotherapy aims to (re-)activate the immune system of a given patient against its tumour. It relies extensively on the ability of tumour-derived neoantigens to trigger a T-cell immune reaction able to recognise and kill the tumour cells expressing them. Since only peptides presented on the cell surface can be immunogenic, the prediction of neoantigen presentation is a crucial step of any discovery pipeline. Here, we present neoMS, a MHC-I presentation prediction algorithm leveraging mass spectrometry-derived MHC ligandomic data to better isolate presented antigens from potentially very large sets. The neoMS model is a transformer-based, peptide-sequence-to-HLA-sequence neural network algorithm, trained on 386,647 epitopes detected in the ligandomes of 92 HLA-monoallelic datasets and 66 patient-derived HLA-multiallelic datasets. It leverages attention mechanisms in which the most relevant parts of both putative epitope and HLA alleles are isolated, resulting in a positive predictive value of 0.61 at a recall of 40% on its patient-derived test dataset, considerably outperforming current alternatives.