New tools for MHC research from machine learning and predictive algorithms to the tumour immunopeptidome

Summary

At a time when immunology seeks to progress ever more rapidly from characterization of a microbial or tumour antigen to the immune correlates that may define protective T‐cell immunity, there is a need for robust tools to enable accurate predictions of peptide–major histocompatibility complex (pMHC) and peptide–MHC–T‐cell receptor binding. Improvements in the curation of data sets from high throughput pMHC analysis, such as the NIH Immune Epitope Database (IEDB), and the associated developments of predictive tools rooted in machine‐learning approaches, are having significant impact. When such approaches are linked to the powerful empirical immunopeptidome data sets from peptide MHC elution and mass spectrometry, there is considerable potential for rapid translation to T‐cell therapies and vaccines.

In this 60th anniversary year of the journal Immunology, there are many aspects of the field on which we can look back and assess our substantive contribution with pride. Somewhere near the top of this list is the journal's long history of publishing outstanding research on the major histocompatibility complex (MHC), stretching back to keynote papers by the likes of Gorer, Medawar, Benacerraf and Mitchison.1 Over this period, appreciation of the MHC has progressed from the era of describing the biological phenomena mediated by these polymorphisms, through the gradual application of molecular biology to generate a fully sequenced region of more than 200 genes,2 to the current state of play of more than 18 000 sequenced alleles at the key functional MHC loci.3

Our ability to make use of this enormous body of MHC data has moved forward rapidly. The explosion of sequence‐derived data, both on polymorphisms of the antigen‐presenting molecules and on the genomics of the pathogens, allergens, autoantigens and tumours being presented by them, has necessitated the development of improved tools, including those for curation of vast empirical databases, these data sets in turn feeding in silico predictive approaches. The incentive to develop these tools comes partly from the requirement to be able to take the genomic sequence of a known or emerging pathogen and rapidly deduce with a high degree of confidence which expressed peptides of all the open‐reading frames possible should be screened for T‐cell recognition and the elucidation of immune correlates.4

A major contribution to the database curation and predictive toolkit aspects of this work has come through a long‐term NIH‐NIAID resource termed the Immune Epitope Database, IEDB.5 Recently, there have been a number of exciting new iterations of these tools. This issue of Immunology contains important contributions from Morten Nielsen's laboratory.6, 7 Jensen et al.6 offer an update on the MHC class II–peptide binding affinity prediction tools, NetMHCII and NetMHCIIpan. These programs have become indispensable to the immunologist who observes a T‐cell response to an epitope in a donor of a given human leucocyte antigen (HLA) haplotype and needs a best‐guess of the restricting HLA molecule to, for example, design T‐cell binding tetramers.

The new iterations depend on quantitative MHC–peptide binding affinity data sets from the IEDB resource. Training with the extended IEDB data set is shown to improve performance with respect to HLA–peptide binding predictions. Jappe et al.7 look at the need for an improved in silico toolkit, able to relate those peptides with high HLA binding to those peptides that are actually immunogenic. This study takes the wealth of empirical peptide‐binding data from mass spectrometry to reconsider the hypothesis that there may be ‘immunological hotspots’ associated with class I peptide immunogenicity. The authors define a terminology for these hotspots and conduct a systematic analysis of in silico predicted hotspots and those derived from mass spectrometry, reiterating that hotspots defined from mass spectrometry data are well predicted by the peptide‐binding predictions made in silico.

Although any such predictions start from the notion of defined anchor residues binding to defined MHC pockets, some peptides show unconventional spacing of anchor residues to enable interaction with MHC.8 A recent study by Andreatta et al.8 argues that such examples involve either stretching of the peptide backbone, or bulging out of the peptide from the MHC groove (a 10mer core). This is an important caveat: the frequency of these non‐canonical binding cores can reach 10% for some MHC class II alleles.

Two recent papers have focused specifically on this renaissance of defining peptide presentation in complex antigenic environments by MHC elution and then mass spectrometry: the notion of defining an ‘immunopeptidome’.9, 10 This has recently become a key resource for tumour immunology and the definition of targeted tumour antigens. A recent review article highlighted the progress made in the sphere of melanoma immunology through partnering direct analysis of tumour‐infiltrating lymphocytes, MHC predictive algorithms and the MHC‐bound, mass‐spectrometry‐determined melanoma immunopeptidome.9 From the University of Tubingen, arguably the birthplace of immunopeptidome research, now comes a major, comprehensive review on state‐of‐the‐art use of mass spectrometry to characterize tumour immunopeptidomes.10 For anyone seeking new insights into tumour immunology and new prospects for immunotherapy, Freudenmann et al. offer a masterclass that crosses different tumour types, peptide elutions and bioinformatics strategies.

At a time when the therapeutic dividends from decades of basic tumour immunology are really paying off, these approaches to peptide–MHC prediction and characterization are playing an ever‐larger part.11