‘Hotspots’ of Antigen Presentation Revealed by Human Leukocyte Antigen Ligandomics for Neoantigen Prioritization

Abstract

The remarkable clinical efficacy of the immune checkpoint blockade therapies has motivated researchers to discover immunogenic epitopes and exploit them for personalized vaccines. Human leukocyte antigen (HLA)-binding peptides derived from processing and presentation of mutated proteins are one of the leading targets for T-cell recognition of cancer cells. Currently, most studies attempt to identify neoantigens based on predicted affinity to HLA molecules, but the performance of such prediction algorithms is rather poor for rare HLA class I alleles and for HLA class II. Direct identification of neoantigens by mass spectrometry (MS) is becoming feasible; however, it is not yet applicable to most patients and lacks sensitivity. In an attempt to capitalize on existing immunopeptidomics data and extract information that could complement HLA-binding prediction, we first compiled a large HLA class I and class II immunopeptidomics database across dozens of cell types and HLA allotypes and detected hotspots that are subsequences of proteins frequently presented. About 3% of the peptidome was detected in both class I and class II. Based on the gene ontology of their source proteins and the peptide’s length, we propose that their processing may partake by the cellular class II presentation machinery. Our database captures the global nature of the in vivo peptidome averaged over many HLA alleles, and therefore, reflects the propensity of peptides to be presented on HLA complexes, which is complementary to the existing neoantigen prediction features such as binding affinity and stability or RNA abundance. We further introduce two immunopeptidomics MS-based features to guide prioritization of neoantigens: the number of peptides matching a protein in our database and the overlap of the predicted wild-type peptide with other peptides in our database. We show as a proof of concept that our immunopeptidomics MS-based features improved neoantigen prioritization by up to 50%. Overall, our work shows that, in addition to providing huge training data to improve the HLA binding prediction, immunopeptidomics also captures other aspects of the natural in vivo presentation that significantly improve prediction of clinically relevant neoantigens.

Pre‐existing heterologous T‐cell immunity and neoantigen immunogenicity

“…These neoantigen prediction algorithms are effective for the majority of HLA class I alleles but perform rather poorly for HLA class II alleles. [8][9][10] Although novel algorithms based on mass spectrometry data have been developed to improve prediction of neoantigens presented by HLA class II alleles, [11][12][13] it still remains unclear whether predicted neoantigen candidates are immunogenic. In fact, the frequency of immunogenic neoantigens among candidates is very low.…”

Mapping the tumour human leukocyte antigen (HLA) ligandome by mass spectrometry

“…Non‐mutated canonical proteins are considered to constitute the majority of HLA‐presented antigens. Presentation hotspots within proteins have been proposed to shape the repertoire of HLA ligands . Differential protein signalling and pathway regulation are a hallmark of cancer.…”

“…Presentation hotspots within proteins have been proposed to shape the repertoire of HLA ligands. 4,[85][86][87] Differential protein signalling and pathway regulation are a hallmark of cancer. Therefore, peptides harbouring PTMs have been increasingly addressed in immunopeptidomic studies.…”

A systematic, unbiased mapping of CD8+and CD4+T cell epitopes in Yellow Fever vaccinees

“…Employing recent advances in mass spectrometry to perform large-scale identification of peptides eluted of HLA molecules, these efforts promise to identify natural ligands thereby capturing information on both antigen processing and HLA binding (6). Over the past decades, substantial progress has been made on predicting peptide-HLA interactions, particularly for HLA class I (HLA-I), which restricts CD8 + cytotoxic T cells (CTL's), and to a lesser degree on predictions for HLA class II (HLA-II), which restricts CD4 + helper T cells (Th) (7)(8)(9)(10)(11)(12)(13)(14)(15). State-of-the-art predictors such as NetMHCpan, an artificial neural network method based on a large collection of experimental peptide-HLA-I binding data, can successfully identify 96.5% of CD8 + T cell epitopes, while rejecting 98.5% of non-epitopes (16).…”

“…State-of-the-art predictors such as NetMHCpan, an artificial neural network method based on a large collection of experimental peptide-HLA-I binding data, can successfully identify 96.5% of CD8 + T cell epitopes, while rejecting 98.5% of non-epitopes (16). However, considering that only 1 of 2000 (2) to 8000 (17) random peptides is a T cell immunogen in the context of a given HLA molecule, even a rejection rate as high as 98.5% translates into a high false discovery rate (FDR) (8, 10,11,18). This is a general problem of current peptide-HLA binding predictors (10,11), and it is particularly problematic when trying to develop a neoepitope-specific, personalized cancer immunotherapy where timely delivery of a few unique cancer neoepitopes is of paramount importance; something that potentially could be achieved with even better predictors (8, [19][20][21].…”