Artificial intelligence predicts the immunogenic landscape of SARS-CoV-2: toward universal blueprints for vaccine designs

Brandon Malone; Boris Simovski; Clement Moline; Jun Cheng; Marius Gheorghe; Hugues Fontenelle; Ioannis Vardaxis; Simen Tennoe; Jenny-Ann Malmberg; Richard Stratford; Trevor Clancy

We carried out an epitope mapping of the entire SARS-CoV-2 virus proteome using cell-surface antigen presentation and immunogenicity predictors from the NEC Immune Profiler suite of tools. Antigen presentation (AP) was predicted from a machine-learning model that integrates in an ensemble machine learning layer information from several HLA binding predictors (in the case three distinct HLA binding predictors trained on ic50nm binding affinity data) and 13 different predictors of antigen processing (all trained on mass spectrometry data). The outputted AP score ranges from 0 to 1, and was used as input to compute immune presentation (IP) across the epitope map. The IP score penalizes those presented peptides that have degrees of "similarity to human" when compared against the human proteome, and awards peptides that are less similar. The resulting IP score represents those HLA presented peptides that are likely to be recognized by circulating T-cells in the periphery i.e. T-cells that have not been deleted or tolerized, and therefore most likely to be immunogenic. Both the AP and the IP epitope predictions are "pan" HLA or HLA-agnostic and can be carried out for any allele in the human population, however for the purpose of this study we limited the analysis to the 100 of the most frequent HLA-A, HLA-B and HLA-DR alleles in the human population. Class II HLA binding predictions were also incorporated into the large scale epitope screen from the IEDB consensus of tools [33] , and B cell epitope predictions were performed using BepiPred [33] .