Deciphering HLA-I motifs across HLA peptidomes improves neo-antigen predictions and identifies allostery regulating HLA specificity

Abstract

The precise identification of Human Leukocyte Antigen class I (HLA-I) binding motifs plays a central role in our ability to understand and predict (neo-)antigen presentation in infectious diseases and cancer. Here, by exploiting co-occurrence of HLA-I alleles across ten newly generated as well as forty public HLA peptidomics datasets comprising more than 115,000 unique peptides, we show that we can rapidly and accurately identify many HLA-I binding motifs and map them to their corresponding alleles without any a priori knowledge of HLA-I binding specificity. Our approach recapitulates and refines known motifs for 43 of the most frequent alleles, uncovers new motifs for 9 alleles that up to now had less than five known ligands and provides a scalable framework to incorporate additional HLA peptidomics studies in the future. The refined motifs improve neo-antigen and cancer testis antigen predictions, indicating that unbiased HLA peptidomics data are ideal for in silico predictions of neo-antigens from tumor exome sequencing data. The new motifs further reveal distant modulation of the binding specificity at P2 for some HLA-I alleles by residues in the HLA-I binding site but outside of the B-pocket and we unravel the underlying mechanisms by protein structure analysis, mutagenesis and in vitro binding assays.

NetMHCpan-4.0: Improved Peptide–MHC Class I Interaction Predictions Integrating Eluted Ligand and Peptide Binding Affinity Data

“…Here, we repeated this benchmark analysis using NetMHCpan-4.0. The results are shown in figure 10 and confirm the finding by Bassani-Sternberg et al (14), that predictors trained on MS eluted ligand data information in most cases show very high predictive power for the identification of cancer neoantigens. Both the NetMHCpan-4.0 and MixMHCpred method proposed by Bassani-Sternberg et al (14) identify the known neoantigens within the top 25 peptides in 6 out out 10 cases.…”

“…This large number of potential peptide candidates clearly underlines the need for tools to rationally downsize the peptide space in the search for cancer neoepitopes. A recent study by Bassani-Sternberg et al (14) demonstrated how this downsizing could be effectively achieved by a prediction method trained on a large set of MS eluted ligands. Here, we repeated this benchmark analysis using NetMHCpan-4.0.…”

“…A rank value of 1 corresponds to the ligand obtaining the highest score (lowest percentile rank) of all peptides from the given sample. Data and performance values for MixMHCFpred are from (31). NetMHCpan-4.0 and NetMHCpan-3.0 are performance values obtained by assigning to each peptide in the given data set the lowest percentile rank score to each of the HLA-A and B molecules expressed by the given cell line.…”

“…Advances in mass spectrometry (MS) have allowed the field of MS peptidomics to move forward. In this context, recent studies (14, 15, 16) have suggested that training prediction methods on such data rather than binding affinity data could improve the ability to accurately identify MHC ligands. As such, MS peptidome data would contain the comprehensive signal of antigen processing and presentation rather than just MHC binding affinity.…”

“…While recent works have demonstrated the improved ability to identify MHC ligands using methods trained on MS peptidome data (14, 15), limited data is available on their impact for the identification of T cell epitopes. In this work, we focus on demonstrating the improved prediction performance not only on large sets of MS peptidome data but also on T cell epitope data independent from the data used to train the new predictor.…”