Unsupervised HLA Peptidome Deconvolution Improves Ligand Prediction Accuracy and Predicts Cooperative Effects in Peptide–HLA Interactions

Abstract

Ag presentation on HLA molecules plays a central role in infectious diseases and tumor immunology. To date, large-scale identification of (neo-)Ags from DNA sequencing data has mainly relied on predictions. In parallel, mass spectrometry analysis of HLA peptidome is increasingly performed to directly detect peptides presented on HLA molecules. In this study, we use a novel unsupervised approach to assign mass spectrometry–based HLA peptidomics data to their cognate HLA molecules. We show that incorporation of deconvoluted HLA peptidomics data in ligand prediction algorithms can improve their accuracy for HLA alleles with few ligands in existing databases. The results of our computational analysis of large datasets of naturally processed HLA peptides, together with experimental validation and protein structure analysis, further reveal how HLA-binding motifs change with peptide length and predict new cooperative effects between distant residues in HLA-B07:02 ligands.

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

“…The only critical limitation for such data integrations is the criteria that each data point must be associated with a specific MHC element. This information is not always readily available, but can in most cases be inferred by unsupervised clustering of the available data (using GibbsCluster (29), position weight matrix mixture models (16), or similar approaches), and association of each cluster to an MHC molecule of the given host.…”

“…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.…”

Mass Spectrometry Profiling of HLA-Associated Peptidomes in Mono-allelic Cells Enables More Accurate Epitope Prediction

“…The majority of LC-MS/MS studies of the HLA peptidome have used cells expressing multiple HLA molecules, which requires peptides to be assigned to one of up to six class I alleles through the use of pre-existing bioinformatics predictors, or “deconvolution” (Bassani-Sternberg and Gfeller, 2016). Thus, peptides that do not closely match known motifs cannot confidently be reported as binders to a given HLA allele.…”

“…Currently, the only single-allele approach available is based on the isolation of soluble HLA from cell lines grown in bioreactors, a setup that is not straightforward to implement and requires several orders of magnitude more input material (Hawkins et al, 2008; Trolle et al, 2016). Our approach, which isolates peptides from cells engineered to express a single HLA allele, provides a scalable means to improve the predictive power of algorithms for class I HLA-presented peptides and avoid in silico allelic deconvolution (Bassani-Stern-berg and Gfeller, 2016). Meanwhile, it leverages advances in instrumentation for rapid collection of high-resolution data and database search strategies that dynamically learn and leverage HLA peptide-binding motifs.…”

Tools to define the melanoma‐associated immunopeptidome

“…This is especially important as prediction analyses are highly limited to frequent HLA types for which a large data set is already available. Using a computational approach for the assignment of a data set of mass spectrometry‐identified epitopes to its respective HLA restriction element, existing in silico prediction tools can be further improved in their accuracy if they are trained with data sets derived from mass spectrometry analyses . In addition, measurement of an immunopeptidome that can be clearly assigned to one single HLA molecule leads to more input and a high‐quality data set for adjustment of prediction algorithms, such as the detection of novel anchor residues .…”