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

Summary

The entirety of human leukocyte antigen (HLA)‐presented peptides is referred to as the HLA ligandome of a cell or tissue, in tumours often termed immunopeptidome. Mapping the tumour immunopeptidome by mass spectrometry (MS) comprehensively views the pathophysiologically relevant antigenic signature of human malignancies. MS is an unbiased approach stringently filtering the candidates to be tested as opposed to epitope prediction algorithms. In the setting of peptide‐specific immunotherapies, MS‐based strategies significantly diminish the risk of lacking clinical benefit, as they yield highly enriched amounts of truly presented peptides. Early immunopeptidomic efforts were severely limited by technical sensitivity and manual spectra interpretation. The technological progress with development of orbitrap mass analysers and enhanced chromatographic performance led to vast improvements in mass accuracy, sensitivity, resolution, and speed. Concomitantly, bioinformatic tools were developed to process MS data, integrate sequencing results, and deconvolute multi‐allelic datasets. This enabled the immense advancement of tumour immunopeptidomics. Studying the HLA‐presented peptide repertoire bears high potential for both answering basic scientific questions and translational application. Mapping the tumour HLA ligandome has started to significantly contribute to target identification for the design of peptide‐specific cancer immunotherapies in clinical trials and compassionate need treatments. In contrast to prediction algorithms, rare HLA allotypes and HLA class II can be adequately addressed when choosing MS‐guided target identification platforms. Herein, we review the identification of tumour HLA ligands focusing on sources, methods, bioinformatic data analysis, translational application, and provide an outlook on future developments.

Abbreviations

AML

acute myeloid leukaemia

CID

collision‐induced dissociation

CLL

chronic lymphocytic leukaemia

DC

dendritic cell

DDA

data‐dependent acquisition

DIA

data‐independent acquisition

EOC

epithelial ovarian cancer

EVs

extracellular vesicles

FDR

false discovery rate

GBM

glioblastoma

GTEx

Genotype Tissue Expression

HLA

human leukocyte antigen

HUPO‐PSI

Humane Proteome Organization Proteomics Standards Initiative

IMAC

immobilised metal affinity chromatography

IP

immunoprecipitation

LC‐MS/MS

liquid chromatography‐coupled tandem mass spectrometry

LFQ

label‐free quantitation

MAE

mild acid elution

MCL

mantle cell lymphoma

MHC

major histocompatibility complex

MM

multiple myeloma

MMM

metastatic malignant melanoma

MOAC

metal oxide affinity chromatography

MS

mass spectrometry

PRIDE

PRoteomics IDentifications

PTM

post‐translational modification

RCC

renal cell carcinoma

SCX

strong cation exchange

sHLA

soluble HLA‐peptide complexes

SRM/MRM

selected/multiple reaction monitoring

TAA

tumour‐associated antigen.

Sources of tumour human leukocyte antigen (HLA) ligands

Primary human tumour samples

HLA ligands have often been isolated from bulk tumour tissue.1, 2, 3, 4, 5, 6, 7, 8 A more sophisticated approach has been proposed by Schuster et al.; enzymatic dissociation with collagenase and subsequent sorting of cell populations allows for fractional HLA ligandome profiling with verification of (target) antigens to be presented on cancer cells.1 Immunopeptidomes of most haematological malignancies can be directly analysed from cell suspensions with optional enrichment of specific cell types.9, 10, 11, 12, 13, 14 In vitro expansion of primary cells prior to HLA immunoprecipitation (IP) is possible in the case of leukaemia samples or isolated tumour‐infiltrating lymphocytes.15, 16 Alternatively, cell lines can be established from primary (solid) tumour tissue and metastases (Fig. 1a).17, 18

Figure 1.

Sources of tumour HLA ligands. (a) Primary human cancer samples. Solid tumours can be analysed in multiple ways: by enzymatic dissociation and sorting of cell populations, from bulk tissue, after establishment of cell lines, and indirectly by studying released soluble HLA‐peptide complexes (sHLA) and extracellular vesicles (EVs). Peripheral blood is a source of cell suspensions from haematological malignancies, which also release sHLA and EVs. Expansion in vitro increases the amount of cells and offers the opportunity to investigate treatment effects. (b) Cell lines as artificial models of human tumours. Permanent cell lines are the most common model system. They offer a variety of options such as treatment, infection with oncogenic viruses, production of sHLA molecules, and transfection with single HLA alleles or tumour‐associated antigens (TAAs). (c) Tumour samples obtained from model organisms. Immunodeficient model organisms are available for xenografting human tissues or cell lines, whereas tumours can be induced in major histocompatibility complex (MHC)‐humanized models.

Downregulation or even loss of HLA class I expression was – and is sometimes still – designated a hallmark of tumour immune evasion. Studies with primary human tumour samples have, however, clearly refuted this for various cancer entities. Using quantitative Edman degradation and immunohistochemistry, both primary tumours and metastases of renal cell carcinoma (RCC) were shown to have significantly elevated HLA class I levels as compared with (autologous) benign kidney tissue.7 HLA class II expression on human malignancies such as RCC, colorectal and urothelial carcinoma was substantiated by immunohistochemistry and oligonucleotide microarray.19 In epithelial ovarian cancer (EOC), both RNA and protein level indicated significantly increased HLA class I and II expression, assigned to EpCAM⁺ CD45⁻ cancer cells, on tumours versus benign fallopian tubes or ovaries.1 Surface HLA expression of most haematological malignancies can be easily determined by flow cytometry.10 Acute myeloid leukaemia (AML) blasts, chronic lymphocytic leukaemia (CLL), and multiple myeloma (MM) cells carry the same or significantly increased numbers of HLA class I and II molecules, when compared with autologous benign or healthy donor leukocytes of the lymphoid or myeloid lineage.9, 10, 14 In heterogeneous bulk tumour tissue, stromal cells contribute considerably to the immunopeptidome. Nevertheless, the isolation of high amounts of both HLA class I‐ and II‐associated peptides from so far all investigated tumour entities does also plead for unimpaired HLA expression on human tumours.2, 3, 4, 9, 10, 13, 14 Fractionated HLA class I IP of cell populations from enzymatically dissociated EOC with high peptide yields from EpCAM⁺ CD45⁻ tumour cells further underscores this statement.1

So far, in‐depth investigations of the non‐mutated immunopeptidome presented on native clinical samples have been published for AML,9 CLL,14 MM,10 mantle cell lymphoma (MCL),13 EOC,1 glioblastoma (GBM),2 and metastatic malignant melanoma (MMM).4 From six of these cancer entities both HLA class I and II data have been acquired and analysed (Fig. 2).

Figure 2.

In‐depth immunopeptidomic datasets generated from primary human tumours. So far, exploratory HLA class I ligandome analyses (black/grey colour coding) have been performed in glioblastoma (GBM), acute myeloid leukaemia (AML), chronic lymphocytic leukaemia (CLL), multiple myeloma (MM), metastatic malignant melanoma (MMM), epithelial ovarian cancer (EOC), and mantle cell lymphoma (MCL). HLA class II data (blue colour coding) are available for all entities except GBM. Peptide numbers are represented as published or as counted from supplementary data (without additionally considering modified variants). Note that the total number of unique class I and II peptides of MCL includes data of two cell lines in addition to 17 patients.

The release of soluble HLA‐peptide complexes (sHLA) by tumours has been suspected to support them in modulating and evading immune responses.20 However, mapping the natural sHLA class I and II (tumour) ligandome from plasma, serum or bone marrow samples as well as cell culture supernatants has not been frequently performed.12, 21, 22 sHLA peptides isolated from clinical samples are often contaminated with blood clotting and plasma proteins.12 Likewise, extracellular vesicles (EVs) containing membrane‐bound HLA class I and II molecules have recently been applied to class I immunopeptidomics for the first time. Their tumour‐association comprises not only oncogenic signalling and immune modulation, but also antigen presentation. EVs such as exosomes are present in most body fluids and could serve as direct source of tumour HLA ligands (Fig. 1a).23, 24, 25 In both tumour‐derived sHLA and EVs, an immunosuppressive function is assigned to the non‐classical HLA‐G.26

Artificial models of human tumours

Permanent cell lines offering high degrees of reproducibility, availability, flexibility, and druggability are the leading artificial models of human tumours used for mapping the tumour HLA ligandome.27 However, monoclonal cultures are not a perfect substitute for primary human tumour samples, as they neither describe a genetically diverse population of patients nor do they correctly reflect human tumour biology. Significant differences in the antigenic landscape of cell lines compared with primary tissues are caused by growing independent of immune selection pressure, hypoxia, starvation, and interaction with other cell types. A comprehensive study in MM found most established tumour‐associated antigens (TAAs) to be presented by cell lines but not by patient samples, most likely leading to lack of clinical benefit, if targeted by immunotherapies.10 In addition, misidentification and cross‐contamination of cell lines resulting in association with incorrect HLA typings are not just rarely observed phenomena with misleading results and conclusions as consequences.28

Nevertheless, cell lines are prominent tools to establish novel methods and standards for tumour HLA ligand identification,29, 30 to examine technical aspects such as sensitivity of neoantigen identification, to study uncommon categories of HLA ligands,18, 31, 32 as well as to investigate the relationship between transcriptome, proteome, and immunopeptidome.13, 33 Addressing the immunopeptidome of cancer stem cells constituting minimal amounts of bulk tumour tissue has so far only been possible based on cell lines but not after direct isolation from patient material.8, 34 Moreover, the impact of therapeutic treatments such as radiation35 or drugs11, 33, 36 on the qualitative and quantitative composition of the presented peptide repertoire can be studied, if this is – as for solid tumours – not possible with primary cancer cells expanded in vitro.15 This allows for selection of robustly presented target antigens for peptide‐specific immunotherapy.11, 15 Apart from that, a specific TAA of interest can be introduced, for example by transfection, and enables intensive studies of its HLA ligands.37, 38 Infected cell cultures may also be used to identify HLA ligands derived from oncogenic viruses (Fig. 1b).39

In cell line‐based approaches, sHLA molecules can be produced allowing direct isolation and subsequent identification of high peptide numbers without major differences concerning motifs, binding affinities, length, and the repertoire of HLA class I‐presented peptides.37, 38, 40, 41 Working at mono‐allelic resolution via transfection or transduction has proven to not only contribute significantly to unveiling HLA ligand motifs,41, 42 but also to improve (tumour) epitope prediction (Fig. 1b).41, 43 Neoantigen predictions based on sequencing data can, in turn, be enhanced by multi‐allelic datasets.16, 41 Xenografted human tumours or cell lines expanded in immunodeficient animals as well as tissue samples obtained from major histocompatibility complex (MHC)‐humanized model organisms – as recently exemplified for spondyloarthritis44 – may furthermore serve as sources of tumour HLA ligands, but have so far not been reported as such (Fig. 1c).

Biochemical isolation and analysis of tumour HLA ligands

Isolation of HLA ligands by IP

Commonly, HLA class I and class II peptide isolation is achieved from cell lysates by IP. Cell suspensions or solid tissues are mechanically homogenised and lysed, employing non‐denaturing detergents, such as NP‐40,45, 46, 47 Triton X‐100,43 CHAPS,1, 48 sodium deoxycholate49, 50 or IGEPAL CA‐630.51, 52 Lysis buffers contain protease inhibitors to block degradation of HLA‐peptide complexes. Optionally, phosphatase inhibitors can be added to prevent dephosphorylation of phosphopeptides.51, 53 In some cases, iodacetamide is added as well to alkylate cysteins,45, 49, 50, 52 thereby inhibiting disulphide bond formation.

The cleared lysate is subjected to immunoaffinity chromatography employing either the well‐described pan‐HLA class I antibody W6/3243, 49, 50, 51, 52, 54 or antibodies specific for distinct HLA allotypes, such as BB7·2 for A*02,45 GAP‐A3 for A*03,55 ME1‐1·2 for B*07,53 etc. The lysate may also pass a second in‐line column containing antibodies specific for class II HLA molecules, such as L243 for HLA‐DR,47, 48 SPVL3 for HLA‐DQ52 or Tü39 for HLA‐DP, ‐DQ, ‐DR1 (Fig. 3). It appears possible that some antibodies currently employed may be biased in ways not yet understood, either enriching for or excluding a certain subset of HLA class I‐ or II‐presented peptides. Pan‐HLA class I antibodies are widely employed, as prediction algorithms for HLA class I are quite accurate in assigning peptides to the corresponding HLA allotypes of the sample. However, HLA class II peptide annotation is currently not as specific and sensitive as for class I HLA molecules, in part due to promiscuous binding motifs of class II HLA molecules.

Figure 3.

Biochemical isolation of HLA‐presented peptides. HLA ligand isolation is performed by immunoprecipitation (IP) or mild acid elution (MAE). The main steps are described in the illustrated workflow.

Antibodies can be either covalently coupled to sepharose or agarose resins or non‐covalently attached to Protein A or Protein G. Different commercial cross‐linking technologies are available, such as CNBr‐activated sepharose or AminoLink™ coupling resin, which employs aldehyde‐activated 4% beaded agarose.43, 50, 53, 54 Sometimes the lysate is precleared from native antibodies before immunoaffinity chromatography with Protein A or Protein G.48, 49, 50

Elution of HLA complexes can be achieved either through treatment with a strong acid, such as 0·1–0·2% TFA,46, 51, 54 10% acetic acid47 or with 0·1–0·2 N acetic acid followed by heat denaturation.12, 45, 48, 56 Peptides are frequently separated from HLA fragments through ultrafiltration, employing a molecular weight filter. Other separation methods employ C₁₈ fractionation through spin columns, cartridges, tips or in‐house‐generated columns.43, 49, 50, 54, 57 Sample concentration and purification is essential before mass spectrometry (MS), and it is usually achieved via C₁₈ 1, 15, 49, 56 or strong cation exchange (SCX) prefractionation.47

The recurrent question of HLA peptide recovery during classical IP has been previously addressed by spiking heavy isotope‐labelled HLA‐peptide monomers directly into the cell lysate, and medium isotope‐labelled peptides, just before MS analysis.57 Thus, absolute quantification of yields during purification steps indicated HLA peptide recovery is of about 0·5–3%, with most peptides lost during IP.57 The uncertainty of peptide loss during the IP isolation procedure is a technical challenge that has been mentioned in the Human Immuno‐Peptidome Project (HIPP) meeting report. It has been proposed to generate a library of isotope‐labelled HLA‐peptide monomers that should be distributed to multiple laboratories to assess the peptide recovery rate in a systematic decentralised way. If the low recovery rate is reproduced, the method needs to be further developed to improve this impediment.58 The IP method has been improved to allow for miniaturised, parallelised sample preparation. This enables higher sample throughput that is compatible with clinical settings.49

Isolation of HLA‐presented peptides by mild acid elution (MAE)

Alternatively, HLA ligands can be isolated by MAE from whole cells to induce dissociation of the non‐covalently bound β ₂‐microglobulin and the peptide from HLA complexes on the cell surface.59, 60, 61 Typically, citrate phosphate buffer at pH 3·3 is used for about 1 min. MAE is supposed to isolate HLA ligands with fewer purification steps, detergent‐free, and without the bias linked to preferential loss of low‐affinity peptides. However, contaminating peptides interacting with the cell membrane via hydrostatic forces may also be eluted by mild acid treatment.62 These could be discriminated from HLA ligands by analysing an equivalent negative control as well, possibly a β ₂‐microglobulin‐deficient cell line.61 However, a negative control is not always feasible especially when considering patient‐derived samples. An advantage of MAE over IP methods is when studying kinetics of immune responses or when input material is limited, as the cells recuperate from the acid wash (Fig. 3).62 A recent comparative study showed up to sixfold increased HLA class I peptide recovery when employing IP compared with MAE.63

Enrichment of post‐translational modifications (PTMs)

Peptides harbouring PTMs, such as phosphorylation or glycosylation, are currently in the scientific focus. Detection of phosphopeptides, as being the most abundant PTM, is technically challenging due to their substoichiometric abundance compared with non‐modified peptides, their low ionisation efficacy and lability when fragmented with collision‐induced dissociation (CID). Therefore, enrichment prior to MS increases identification rates. Enrichment strategies are frequently variations of the immobilised metal affinity chromatography (IMAC), in which metal ions such as Ti⁴⁺ or Fe³⁺ chelate phosphonate groups.51, 64, 65 Derivatisation of the carboxy‐termini to methyl esters has been employed to decrease unspecific interaction of acidic peptides with metal ions by some groups.53 Metal oxide affinity chromatography (MOAC) follows a similar enrichment mechanism, but uses metal oxides such as Ti₂O as solid phase.51 SCX chromatography is often used as a primary51 or secondary prefractionation strategy in combination with either IMAC or MOAC enrichment.66

Furthermore, the identification of HLA class I peptides with an O‐linked β‐N‐acetylglucosamine (O‐GlcNAc) from leukaemia cell lines was achieved by applying an enrichment method based on esterification of alcohol groups with boronic acid coupled to POROS20 AL beads. The MS detection method was tailored to fragment ions with electron‐transfer dissociation (ETD) when CID spectra contained doubly charged ions corresponding to loss of dehydro‐N‐acetylglucosamine (Δm = 203 Th).67

Sequence identification by tandem mass spectrometry

Peptide sequencing by liquid chromatography‐coupled tandem mass spectrometry (LC‐MS/MS) is achieved by prefractionation of complex peptide solutions, followed by MS.68 Typically, MS1 survey spectra are acquired and abundant peptides are selected for fragmentation yielding MS2 spectra. Prefractionation is often performed by reversed‐phase or SCX chromatographic separation (Fig. 3). Sometimes #bib2D prefractionation strategies can be employed, such as a first fractionation performed on a SCX column, from which salt fractions can be eluted with pulsed concentrations of ammonium acetate to a C₁₂ precolumn.61 Better fractionation is beneficial for detection of low‐abundant peptide species; however #bib2D fractionation is very cumbersome and currently less widely used.

MS sequencing is frequently accomplished by using CID or beam‐type higher‐energy CID (HCD). These methods produce peptide fragment ions that can be used in automated database search strategies or de novo analysis to identify peptide sequences. However, these methods have been optimised for tryptic peptides and are not ideal for HLA ligands, as generated spectra often lack sufficient fragment information for confident identification. Therefore, the use of hybrid fragmentation methods such as electron‐transfer/higher‐energy‐induced dissociation (EThCD) has been proposed, and is especially well suited for analysis of labile PTMs.29, 51 For a comprehensive review about MS acquisition strategies, visit Refs 27, 62.

MS analyses are generally performed in data‐dependent acquisition mode (DDA), which is well suited for discovery proteomics but offers limited analytical reproducibility, which is required for quantitation of analytes between different conditions.30 These limitations can be overcome by data‐independent acquisition (DIA), where spectra are acquired following a predefined scheme, independent of analyte abundance and distribution.52 Besides improved technical reproducibility, input material can be reduced without a negative effect on identifications and intensities.52 DIA spectra can be annotated to corresponding peptide sequences via the use of spectral libraries. Spectral libraries can be generated by pre‐analysing samples in DDA mode, and using SpectraST to concatenate database search results. There are community efforts such as the SWATH atlas (http://www.swathatlas.org) to overcome this impediment; however, this strategy requires retention time stability, which is not always given between laboratories. Approaches independent of spectral libraries have also been developed, such as DIA‐Umpire69 and PULSAR from Biognosys (https://biognosys.com/shop/spectronaut-pulsar).

HLA ligand quantitation methods

Label‐free quantitation (LFQ) has been frequently employed for assessing quantitative immunopeptidome changes between conditions, such as autologous tumour and benign samples1 or upon treatment with different drugs.11, 15, 36 LFQ helps visualising peptide intensity modulations between conditions, based on mean intensities of extracted ion chromatograms in technical replicates. Another labelling approach for relative quantitation has been employed in comparative profiling studies of tumour versus benign samples. After classical IP, isolated peptides can be derivatised with ‘light’ or ‘heavy’ nicotinic acid.5, 64, 70 Accurate peptide quantitation is then possible by mixing peptides from both conditions at a ratio of 1 : 1 followed by LC‐MS/MS. An absolute quantitation approach (AQUA) is based on isotope‐labelled peptides spiked into the sample at known concentrations. Comparing signal intensities of the known isotope‐labelled peptide and the native counterpart allows for peptide quantitation.57

Bioinformatic analysis of MS data

Sequence identification by automated database search

Large MS datasets are generated from HLA ligandomics studies, rendering their manual interpretation impossible. Three main strategies have been established for the identification of peptide‐derived tandem mass spectra: sequence database searching, de novo spectrum sequencing, and spectral library searching. In high‐throughput studies, the most efficient method is based on searching experimental MS2 spectra against a reference protein sequence database. There are a multitude of software implementations, called search engines, for database search, reviewed in Ref. 71. Database search was pioneered by the SEQUEST algorithm,72 followed by the commercial alternative Mascot,73 and the number of available search algorithms has steadily increased since, including open source alternatives. Search algorithms frequently employed in HLA ligandomics studies are Mascot,1 COMET and its predecessor SEQUEST,13, 15 and Andromeda integrated into the MaxQuant framework.4 Unbiased de novo spectrum sequencing with PEAKS DB74 to supplement database search increases the identification rate, without enlarging the search space.13, 71

All database search strategies function in a similar manner: they take an MS2 spectrum as input and compare it against theoretical fragmentation patterns constructed from the database queried. The search output is a list of peptide sequences ranked according to the scoring scheme implemented in each algorithm. Special care needs to be taken in choosing and restricting the search space for each experiment, and controlling the global false discovery rate (FDR).75 Strategies to estimate the FDR are based on decoy databases, in which the desired database is either reversed or scrambled. Thus, one can suppose that a similar number of hits obtained from the decoy search are falsely annotated in the target dataset as well. Common laboratory contaminants, such as Staphylococcus aureus Protein A and cell culture supplements can be added to the database to reduce peptide misidentification.13, 16, 74, 76, 77

Additional software for upstream and downstream processing of data is available and search engines can be integrated into larger platforms, such as OpenMS78 and the Trans Peptidomic Pipeline (TPP),76 allowing the design of complex workflows. Due to different results obtained from database search employing various algorithms, MS data can be uploaded directly into the SysteMHC Atlas project79 and the PRoteomics IDentifications (PRIDE) database.80

Generally, proteomic workflows both for MS analysis and automated sequence identification have been adapted to the field of HLA peptidomics, often at the expense of HLA ligands. Discrepancies between shotgun proteomic workflows and HLA peptide discovery have been summarised comprehensively in a recent viewpoint.81

HLA annotation to deconvolute multi‐allelic datasets

Peptides extracted from HLA molecules either after IP or by MAE and identified by MS may not always be HLA ligands, as contaminant peptides often accumulate through the enrichment procedure. Furthermore, by employing pan‐HLA class I antibodies such as W6/32, one cannot easily infer which HLA molecule (‐A, ‐B or ‐C) they originated from. The most common approach is to predict binding affinity of each peptide to each HLA allotype present in the sample with available prediction algorithms.

SYFPEITHI, one of the first epitope prediction methods described, employs position‐specific scoring matrices (PSSM), which are generated based on eluted peptides from mono‐allelic cells.42, 82, 83

HLA motifs can be deduced by unsupervised clustering of HLA ligandomics data allowing peptide grouping according to their sequence similarity.16 This concept was implemented into MixMHCpred and led to the refinement of known HLA motifs. However, a strong bias against infrequent alleles and HLA‐C alleles with less stringent motifs became evident.16, 84 To complement this impediment, HLA‐C motifs from mono‐allelic transfectants were recently published.42

Other prediction algorithms such as netMHC and netMHCpan are based on machine learning algorithms trained with IC₅₀ values from HLA binding assays. A drawback of these training sets is that peptide‐MHC complexes displaying weak binding affinities are also included, but they would not necessarily represent physiological interactions. Furthermore, binding assays show that peptides can bind to HLA molecules in vitro, but do not take into account any intracellular processing preferences. However, netMHCpan‐4·0 integrates both publicly available HLA ligandomics data and binding affinity data, thus increasing the sensitivity and specificity of their binding prediction.84

Gibbs clustering is also a widely adopted unsupervised approach to define HLA motifs.42 Deconvolution of multi‐allelic data can also be performed in a motif‐based way.30, 51 HLA class II prediction is less sensitive and specific, due to degenerate anchor positions in most HLA class II motifs, but it is possible with netMHCpan‐4·0 for a limited set of alleles.84

Origin of HLA ligands – what is recordable?

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.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. Cancer‐specific phosphopeptides have been isolated from HLA molecules of primary RCC64 and leukaemia samples, and shown to be targeted by PTM‐specific immune responses in the latter.53 HLA ligands have been also described to be presented through a TAP‐independent processing pathway that competes with the TAP‐dependent conventional processing route. These peptides, designated TEIPPs (T‐cell epitopes associated with impaired peptide processing), can be targeted by CD8⁺ T‐cells.88

HLA‐presented cryptic peptides

HLA ligands have been proposed to originate from various unconventional sources, such as proteasomal and secretory granule splice products,89, 90, 91 or non‐canonical translation products.6, 32 Cryptic peptides are generated by translation of: (i) protein‐coding sequences in non‐canonical reading frames, (ii) allegedly non‐coding sequences (introns #bib5’‐UTRs #bib3’‐UTRs), and (iii) antisense transcripts.32, 92 Cryptic HLA ligands were estimated to represent 6·5%–13% of the total HLA‐presented peptide repertoire.32 A proteogenomic workflow was successfully employed, identifying cryptic peptides by MS. Cryptic peptides were filtered stringently, and 18 could be confirmed via spectral comparison with the synthetic counterpart. Immunogenicity was shown for four of these peptides via interferon (IFN)γ release.³² Cryptic HLA ligands have been described previously, such as the VEGF peptide SRFGGAVVR, presumably derived from a non‐canonical start codon.6 The majority of cancer mutations are located in non‐exomic regions, but the search for tumour‐specific mutations has focused on exomic regions. If non‐exomic regions are translated and peptide sequences are loaded onto HLA molecules, a large number of cryptic HLA ligands would contain tumour‐specific mutations and thus be genuine TAAs.

Search for neoantigenic peptides

The findings of Pierre Coulie in 1995 coined the term neoepitope defined as an MHC‐presented peptide that harbours an amino acid exchange caused by tumour‐specific mutation recognised by T‐cells.93 Their contribution in eliciting clinically relevant immune responses has become widely accepted.4, 13, 17, 38, 94, 95, 96 Neoantigenic peptides can arise from single nucleotide variants, insertions or deletions, as well as more complex variants such as frameshift mutations or fusion proteins. On RNA level, loss or gain of splice sites and erroneous translation products of untranslated regions can lead to mutated HLA ligands. The identification of neoantigenic peptides is of large interest for cancer immunotherapies as they offer maximum tumour‐specificity and have been correlated to the success of checkpoint inhibitor therapies.41, 97

The state‐of‐the‐art proteogenomic approach combines whole‐exome sequencing with variant calling against autologous normal tissue and LC‐MS/MS.4, 13, 17, 50 For cell lines, using somatic variants as deposited in the Cancer Cell Line Encyclopedia (CCLE)98 or the Catalogue of Somatic Mutations in Cancer (COSMIC)99 represents a justifiable substitute for sequencing.100 Customised databases to search LC‐MS/MS data are often concatenated FASTA‐formatted files integrating somatic and germline variants with the human reference proteome.13, 16, 17, 50 Orthogonal evidence for mutated gene products can be obtained by RNA sequencing. Generating personalised databases from expression data does, in addition, reduce the search space and thus the FDR.41, 77 Mutated HLA ligands have also been identified based on sample‐specific in silico predicted neoantigens. Only short sequences harbouring the mutation and predicted to be HLA ligands are incorporated into the database, instead of including the entire translated exome data.16, 97, 100 However, this approach is compromised for not sufficiently studied HLA allotypes, and not all MS‐identified mutated HLA ligands are predictable.4, 16

The direct identification of naturally presented neoantigenic peptides by MS started in murine cell lines77 and tissues97, followed by human cell lines17, 100 and native tumour samples.4, 13 Fragment spectrum confirmation via synthetic peptides – as also done for viral,39 cryptic,32 and non‐mutated TAA candidates14 – is an essential part of the validation pipeline.4, 13, 17, 38, 77, 97 Ideally, co‐elution of heavy isotope‐labelled synthetic peptides spiked into tumour HLA ligand eluate is performed.101 Low‐abundance peptides falling below the detection limit in discovery MS may be identified by applying selected/multiple reaction monitoring (SRM/MRM), a highly sensitive targeted technique for triple quadrupole (QQQ) instruments. In advance, the fragmentation profile of each synthetic peptide is determined to define a set of SRM transitions, representing specific pairs of precursor and fragment ions with corresponding m/z values. If required, transition conditions such as collision energy can be optimised based on these first measurements. The natural peptide and its isotope‐labelled synthetic counterpart co‐elute with precursor ions being selected for fragmentation only from fixed mass windows (Q1). Subsequent to CID (Q2), the signal of selected fragments (Q3) is recorded.97, 101, 102

Neoantigenic peptides constitute only a small proportion of the tumour immunopeptidome. Identification by MS has remained a major challenge, and requires either high mutational loads or an enormous depth of the investigated ligandome (Fig. 4a and b).101 Considering all samples included in Fig. 4a, one mutated peptide is on average identifiable per 1·8 × 10³ non‐synonymous mutations and per 1·1 × 10⁴ unique HLA class I peptides, respectively.

Figure 4.

The bottleneck in verifying naturally HLA‐presented neoantigens by mass spectrometry (MS). (a) Funnel chart with five levels of class I neoantigen discovery as inferred from published datasets. All levels of data analysis are traceable for each sample, with few missing values (n.a.). Total numbers of non‐synonymous mutations were either integrated as published, retrieved from supplementary materials or the Cancer Cell Line Encyclopedia (CCLE; excluding gained and retained stops, nonsense, silent and synonymous mutations). From non‐synonymous mutations identified by sequencing (upper panel), peptides spanning the mutated amino acid and possibly binding to HLA molecules are predicted in silico (second panel). The total number of unique HLA‐eluted peptides identified by liquid chromatography‐coupled tandem mass spectrometry (LC‐MS/MS) is depicted in the third panel. Panels 4 and 5 illustrate confident neoantigen identifications and positively tested mutation‐specific immunogenicity. (b) Relation between non‐synonymous mutations, HLA class I‐presented peptides and confirmed neoantigenic peptides. For the data included in (a), the number of non‐synonymous mutations was plotted against the number of unique HLA class I peptide identifications. Samples with spectrum‐identified mutated peptides are depicted by stars. A clear separation becomes evident between the majority of samples without neoantigen identifications and the few with more than one mutated peptide identified.

Translational application

Strategies for target definition

Defining the correct targets is crucial for T‐cell‐mediated cancer immunotherapy, with several different strategies being pursued. Lack of tumour‐specificity can cause severe, even fatal, side‐effects, especially when applying highly active therapies such as adoptive transfer of engineered T‐cells.103 On RNA level, the Genotype‐Tissue Expression (GTEx) Consortium has examined gene expression in various human tissues and made the data publicly available.104 However, it has been shown several times that the immunopeptidome does neither mirror the transcriptome nor the proteome.5, 33, 61, 100 Thus, mapping the benign immunopeptidome across multiple organs and HLA allotypes lays the foundation to achieve sufficient tumour‐specificity.

MS‐guided target definition for cancer immunotherapies is often based on comparative profiling of tumour against benign HLA ligandome data. Potential non‐mutated targets are characterised by frequent representation in malignant but not in benign immunopeptidomes.1, 9, 10, 14 Tumour‐exclusivity can either be on the level of HLA ligands, for example in terms of differential antigen processing in cancer cells,105 or on the level of the entire antigen.1, 8, 9, 10, 14 To identify differentiation antigens for cancer entities such as ovarian or prostate carcinoma, benign counterparts of the respective organs have to be excluded from comparative profiling.1 Similarly, cancer‐testis antigen definition requires exclusion of benign gametogenic tissues.

Further criteria can be the robust presentation across different subtypes, stages, and under (standard) therapy.11, 14, 15 Orthogonal evidence for tumour‐association of identified antigens can be provided by additional immunohistochemistry, transcriptome or proteome analyses, and queries of databases such as The Cancer Genome Atlas (TCGA),106 GTEx104 or The Human Protein Atlas,107 both embedded in HumanMine108 and the Gene Expression Profiling Interactive Analysis (GEPIA).1, 8, 109 XPRESIDENT (Xpression profiling and analysis of peptide PRESentation by HLA molecules for IDEntification of New tumor antigens in combination with T‐cell screening) is a proteogenomic approach computing the overlap of overexpressed genes in tumour versus adjacent benign and HLA ligands eluted from the malignant sample.105 Network‐ or pathway‐based approaches require malignant and benign data, and start with assigning HLA ligand sequences to their source proteins. Assisted by (open source) online tools or software such as GeneMANIA110 and Cytoscape,111 these are allocated to biological processes, molecular functions or interaction networks. Pathways involved in tumourigenesis or metastasis formation, which are highly enriched in malignant samples, serve as basis to select antigens, thus permitting almost tumour‐exclusive HLA presentation patterns.3 When using databases, clustering or annotation tools, it is essential to know the experimental origin of the data input, how up‐to‐date it is, and which curation procedures it underwent. The Proteomics Standards Initiative Common QUery InterfaCe (PSICQUIC), launched by the Humane Proteome Organization Proteomics Standards Initiative (HUPO‐PSI), assists in answering these questions and provides a web interface for simultaneous queries to its resources.112

Independent of which strategy is pursued to define a list of candidate targets, it is mandatory to evaluate these in subsequent immunological assays, for example after in vitro priming of naïve T‐cells of healthy donors or to determine pre‐existing T‐cell responses in patients.1, 3, 8, 9, 14 The search for neoantigenic peptides, per se designated to be tumour‐specific, and HLA ligands carrying cancer‐specific PTMs by MS is described in detail above. Targeting these by cancer immunotherapies inevitably requires mutation‐ and PTM‐specific immunogenicity.4, 13, 17, 53, 67

Databases and compendia of established tumour antigens and peptides

Besides defining novel sets of TAAs, tumour HLA ligandome data can be searched for established and successfully targeted peptides or antigens.1, 4, 8, 9, 105 Table 1 provides an overview of currently accessible databases and compendia. The SEREX113 database is not online any more.

Table 1.

Currently accessible databases and compendia of established tumour antigens and peptides

Database/compendium	Access	Last update	Ref.
Cancer Immunity Peptide Database	Open access https://www.cancerresearch.org/scientists/events-and-resources/peptide-database	2016	116
Cheever et al. (2009)	Table 3	–	117
CTdatabase	Open access http://www.cta.lncc.br	2009	118
IEDB	Open access http://www.iedb.org/	2018	119
Novellino et al. (2005)	Tables 1–6	–	120
SYFPEITHI	Open access http://www.syfpeithi.de/index.html	2012	82
TANTIGEN	Open access meta‐database http://cvc.dfci.harvard.edu/tadb/index.php	2017	121

Clinical application in personalised cancer immunotherapies

High degrees of inter‐ and even intra‐patient heterogeneity in tumour biology demand tailored therapies. Personalised cancer immunotherapy can be divided into three categories. Stratified approaches include biomarker‐based selection of patients subsequently treated with the same drug. Passively personalised therapies are based on autologous cellular material, whereby tumours did not undergo molecular characterisation. In turn, actively personalised approaches do not only apply molecular markers for patient selection, but also for definition of drug composition. Active personalisation can be categorised into warehouse and fully individualised concepts, explained at the example of peptide vaccination. Warehousing includes selecting off‐the‐shelf peptides for individual vaccine cocktails, whereas fully individualised therapies depend on de novo synthesis of patient‐specific peptides identified by immunopeptidomics, epitope prediction or immunogenicity screening.114 Detailed insights into the design of personalised immunotherapeutic intervention or an overview of ongoing clinical trials are beyond the scope of this review, but we would like to point out three examples including MS‐guided target definition.

1. IMA901, a multi‐epitope vaccine for RCC, was tested in a Phase II clinical trial (NCT01265901) and is an example of stratified cancer immunotherapy. The composition of the invariant peptide cocktail was determined via the aforementioned XPRESIDENT platform. IMA901, tested as monotherapy in a cohort of 28 HLA‐A*02‐positive patients, showed an association of multi‐peptide responses with longer overall survival.115

2. The warehouse of the ongoing Phase II clinical trial iVAC‐L‐CLL01 (NCT02802943) comprising 40 non‐mutated HLA class I and five HLA class II peptides was defined by intensive immunopeptidome analyses of a cohort of 30 CLL patients.14 Individual vaccines consist of five class II and five class I off‐the‐shelf peptides, whereby the latter are selected depending on the individual immunopeptidome landscape.

3. For a fully individualised Phase I dendritic cell (DC) vaccination study (NCT00683670) in three melanoma patients, exome sequencing revealed non‐synonymous mutations with subsequent in silico prediction of mutated HLA‐A*02:01 ligands. Seven neoantigenic peptides with confirmed HLA‐A*02:01 binding and expression of the respective source protein were – along with two gp100 peptides – loaded onto autologous DCs. The vaccine both induced and enhanced pre‐existing neoantigen‐specific T‐cell immunity. By minigene transfection of a sHLA‐producing melanoma cell line, two of seven tested neoantigenic peptides were confirmed to be HLA‐presented by MS.38

Two further Phase I clinical trials targeting the individual mutanome have recently been completed in melanoma patients. Mutated epitopes were targeted by fully individualised RNA (NCT02035956) or long peptide vaccines (NCT01970358) based on up to 20 in silico predicted neoepitopes. Both approaches induced or enhanced neoantigen‐specific T‐cell responses associated with remarkable clinical benefit. This further emphasises the potential of neoantigens as targets for peptide‐specific immunotherapies.95, 96

Future directions

The field of large‐scale immunopeptidomics is still in discovery and far from validation mode. Refined protocols to isolate HLA ligands, higher resolution and sensitivity of LC‐MS/MS devices as well as optimised acquisition methods accompanied by tailored bioinformatic tools will permit deeper insights into immunopeptidomes. This opens new avenues to scarcer clinical samples and sources of tumour HLA ligands will be extended, for example to sHLA and EVs obtained from urine (bladder cancer), cerebrospinal fluid (brain tumours) or bronchoalveolar lavage fluid (lung carcinoma). Other conceivable options comprise more sophisticated human tumour models and circulating tumour cells as less invasive sources of HLA‐presented peptides. Furthermore, mutated, cryptic, post‐translationally modified or even proteasomally spliced HLA ligands, marginally contributing to the naturally presented peptide repertoire, will likely be increasingly addressed. De novo sequencing will support the exploration of the immunopeptidome in its full complexity, including neoepitopes.

The amount of publicly available datasets will substantially rise. Besides encouraging meta‐analyses, these will make a significant contribution to the improvement of search engines and HLA annotation algorithms. Moreover, novel computational tools for immunoinformatics and multiomics approaches will be developed. Increasing numbers of high‐confidence peptide identifications raise the informative value of newly and previously acquired MS datasets (re‐)analysed with novel bioinformatics. Altogether, this will lead to both maximum exploitation of MS data and foster inter‐laboratory standardisation following the example of the HUPO‐PSI guidelines.

Not only standardised workflows, but also reduced sample input and declining costs for exome and RNA sequencing supporting proteogenomic approaches may bring us closer to the implementation of immunopeptidomics in clinical practice. Faster acquisition and interpretation of LC‐MS/MS data as well as growing experience with rare HLA allotypes, HLA class II, and tumour entities are expected to advance immunopeptidomics as antigen discovery platform for personalised cancer immunotherapies.

Disclosure

The authors declare no competing financial interests.