Epitope prediction and identification- adaptive T cell responses in humans

Abstract

Epitopes, in the context of T cell recognition, are short peptides typically derived by antigen processing, and presented on the cell surface bound to MHC molecules (HLA molecules in humans) for TCR scrutiny. The identification of epitopes is a context-dependent process, with consideration given to, for example, the source pathogen and protein, the host organism, and state of the immune reaction (e.g., following natural infection, vaccination, etc.). In the following review, we consider the various approaches used to define T cell epitopes, including both bioinformatic and experimental approaches, and discuss the concepts of immunodominance and immunoprevalence. We also discuss HLA polymorphism and epitope restriction, and the resulting impact on the identification of, and potential population coverage afforded by, epitopes or epitope-based vaccines. Finally, some examples of the practical application of T cell epitope identification are provided, showing how epitopes have been valuable for deriving novel immunological insights in the context of the immune response to various pathogens and allergens.

1. Introduction

In this review, we present our viewpoint on epitope identification. We stress that this is a rather personal and self-centered account. In particular, this review is focused on bacterial and viral microbes and human T cell epitope identification. Our approach hinges on first defining what is an epitope, and the type of data and metadata usually associated with epitopes. In the following sections we discuss epitope prediction, and epitope immunodominance and prevalence. Additional sections discuss HLA¹ polymorphism and HLA restriction and their impact on population coverage. After discussing various technical approaches to epitope identification, we close by mentioning specific instances of how epitope identification studies have generated novel insights in microbial diseases.

2. What is a T cell epitope?

An epitope can be defined as the molecular structure recognized by adaptive immune responses, namely by T cell and B cell receptors and soluble antibodies [1]. In this review, we will focus on T cell epitopes, as this is our principal area of expertise. We will further focus on the epitopes recognized by classical alpha-beta TCR receptors, and restricted by conventional HLA molecules. We will not discuss the epitope structures recognized by gamma delta, MAIT and other non-conventional T cells, which have been described and reviewed elsewhere [2–8].

In general, the epitopes recognized by T cells are short peptides derived from processing of larger antigens [9], even though epitopes that are partly or completely of non-peptidic nature have been described [10]. In particular, MHC-II-bound glycan-peptide epitopes, which could be relevant for both infectious disease and cancer, have been described [11, 12]. As a general rule, CD4 “helper” T cells recognize peptides of mostly 13–17 residues in size bound to HLA class II, while CD8 “cytotoxic” T cells recognize peptides of mostly 8–10 residues in size bound to HLA class I [9, 13] (Figure 1A).

Figure 1. Epitope definition.

Presentation of peptide ligands by HLA class I and class II molecules is depicted. A) HLA class I and II ligands are generated by proteolytic processing of endogenously expressed proteins (antigen) (class I) or from proteins degraded in endocytic compartments (class II). Longer class II peptides typically overhang the open ends of the HLA class II binding groove, while shorter class I ligands are size constrained due to the closed end of the class I binding groove. With binding, and subsequent presentation on the cell surface, HLA-ligand complexes are available for scrutiny by CD4+ (class II) or CD8+ (class I) T cells. B) Ligands processed from longer protein antigens bind HLA using, in general, a nine-mer core region, where the main energy of binding is provided by interaction of some, but not all, peptide residues with residues forming the main pockets of the HLA binding groove. Definition of partial epitopes reflects that not all residues within an epitope region are necessarily important for HLA binding or T cell recognition, while mutation of other residues may ameliorate or abrogate specific immunity.

Even within the confines of the definition above, it should be pointed out that in practical terms T cell epitopes are defined at different level of granularity. In theory, an unambiguous crystallographic structure of an epitope bound to HLA and its TCR can define all relevant interactions, and unequivocally map the exact structure of the epitope [14]. However, in practice, this exact determination is most often missing, and different experimental epitope definitions are utilized. In some cases, the experimental data points to an epitope “region.” Examples of this situation are the mapping of a CD4 reactivity to a 40–50 residue protein domain, or the identification of a synthetic 15-mer recognized by CD8 T cell responses. In those cases, it is understood that the epitope is contained within the region, but there is no expectation that each and every residue in the region is actually playing a role in HLA binding and/or TCR recognition. In the case of a “partial epitope”, certain peptide residues might be shown to be crucial, for example, because their mutation might abolish or hinder T cell recognition. Here it is clearly understood that those residues are part of the epitope, but there is an expectation that other neighboring residues are actually required for HLA binding and/or TCR recognition (Figure 1B).

Related to these issues, are the concept of defining “optimal” versus “minimal” epitopes. In the case of class I, because of the closed nature of the HLA class I binding groove, an exact size is usually required. While longer sizes may also be recognized, this typically requires their processing in vivo or in vitro to the “right” size, which is remarkably more potent than larger or smaller fragments [15]. Thus, in this case, the optimal size coincides with the minimal size, and it can be unequivocally defined [16]. A rather different situation is observed in the case of HLA class II, where the binding groove is open ended, and an exact size is not required. Natural endogenously bound ligands usually “overhang” by a couple of residues, both at the N- and C- termini, with “ragged ends” families of ligands usually found in ligand elution experiments [17]. Furthermore, beyond the 9-residue core sequence directly involved in class II binding, these “overhang” residues have been shown to influence HLA binding and TCR recognition [18–20]. Because of these structural constraints, an “optimal” CD4 epitope size is not usually unequivocally defined, with several alternative frames equipotent, and also the definition of a “minimal” epitope is not usually biologically relevant (Figure 1A).

3. What metadata is associated with epitope identification?

Epitope identification is a context-dependent process. To simply state that “such and such sequence is an epitope” in not very informative. Whether a given sequence is an epitope, or in other words is recognized by T cells, is highly contingent on several context-dependent variables. To name a few: what is the host, where does the epitope come from (e.g. what is the source organism), how did the immune reaction come about (e.g. natural infection, vaccination, in vitro stimulation), and how did the investigator measure and quantified the observed response. For example, it is intuitive that the fact a given HCV sequence induces cell proliferation in vitro following peptide immunization of C57BL6 mice with adjuvant, does not speak to whether the same peptide induces TNF-alpha production from T cells in PBMC derived from naturally infected humans.

The IEDB resource was created [21, 22] to host epitope data and metadata, as published in the scientific literature, or generated by several NIAID sponsored initiatives, such as the large-scale epitope identification contracts. As of the end of July 2020, it hosts data related to over 50,000 different epitopes, derived from approximately 10,000 different journal articles [23]. Each epitope must be identified by an unequivocal and well-defined molecular structure. For each epitope, a rich contextual set of metadata is also captured (Figure 2).

Figure 2. Metadata associated with epitope identification capture by the IEDB.

The IEDB homepage (www.iedb.org), and initial search fields are shown. The IEDB is an NIH-NIAID funded publicly available database of T and B cell epitopes curated from the published literature or by direct submission from NIH-NIAID funded large scale epitope discovery contracts. From the homepage, epitopes can be search using selected criteria, and subsequent results can be further filtered with additional criteria, to include specific assays or receptor(s).

In terms of the source of the epitope, key information is captured relating to the organism and strain (following the NCBI taxonomy nomenclature) of provenance, and the specific antigen/protein and its GeneBank accession number. In terms of the host, species information (e.g. mice, humans, NHPs or any other species for which responses are described) and associated features, such the particular mouse strains used in the investigations or, in the case of human data, ethnicity are collected. Age, gender, HLA type of the responding organism, and HLA restriction of the responding T cells are also captured [24]. In particular, the specific features of the responding T cell population are also captured, also including the specific associated TCR sequences, if known [25].

Additional key metadata captured in the IEDB related to any disease process associated with the response (e.g., were the responses associated with donors in an acute or convalescent phase of a viral infection, and so on), and the assay(s) used to detect the response. Examples of assay metadata are whether the response was induced in vivo (e.g. natural infection or exposure, vaccination) or in vitro. Whether in vitro culture or expansion was utilized, and what type of assay was used (proliferation, tetramer staining, ELISPOT and so on) are also noted. All data and metadata hosted in the IEDB are freely available and searchable by selecting and filtering according to the metadata described above [23].

4. Epitope predictions

Several different, yet complimentary approaches are routinely used to identify T cell epitopes (Figure 3). One is based on bioinformatics prediction of peptides likely to bind or to occur as natural ligands of different HLA molecules [26]. Parallel to these approaches are the actual determination of which peptides bind purified HLA in vitro [27], or the experimental determination of peptides occurring as natural HLA ligands [28–30]. All four of these approaches rely on the fact that generation of the peptide by natural processing and subsequent HLA binding are key necessary (but not sufficient) steps for T cell immunogenicity. It is generally true that all epitopes must bind HLA and be generated by processing, but not all of these peptides are immunogenic. In essence, these methodologies seek to apply a powerful filter to reduce the number of candidates to be considered, and are best used in concert with experimental assays and validations, described in more detail in the following section.

Figure 3. Epitope predictions based on HLA interactions.

The development of bioinformatic tools to aid in identification of T cell epitopes is based on data generated from, for example, HLA-ligand assays or by mass spectrometry analysis of eluted ligands. This data is then utilized to develop machine learning tools to predict potential HLA binding peptides, and in turn candidate T cell epitopes.

In terms of HLA binding, the most widely used assay is based on a classic receptor-ligand inhibition, utilizing purified HLA molecules and synthetic peptides which, if run in the appropriate stoichiometric conditions, produces IC50 values approximating the true binding affinity (KD) [27]. Alternative assay formats have been proposed which measure dissociation rates, and provide limited increase in predictive power [31, 32]. This methodology was particularly effective in early days, serving to establish the basic parameters of HLA binding as related to T cell recognition [33, 34], including the affinity threshold associated with immunogenicity [35], and has produced a large set of experimental data on the basis of which bioinformatics predictive algorithms were trained and refined [26]. As the performance of predictions has increased over time, the usefulness of measuring binding as a pre-screening assay has decreased, as the cost and labor involved in measuring a large number of affinities is easily offset by the less stringent, but more cost efficient, use of sets of predicted peptides in T cell assays [36].

In terms of analyzing the actual pathogen derived natural ligands, the majority of the data available to date relate to general sets of eluted ligands, mostly of self-origin. These data sets, often referred as “immunopeptidomes” [28–30], have recently outnumbered in terms of data points what is available from measured binding data, and have proven invaluable for training bioinformatics-based predictive algorithms [37, 38]. However, relatively few studies [36, 39, 40] have thus far been able to define actual pathogen derived epitopes, as the majority of the natural ligands are of self-origin and the few pathogen-derived epitopes are relatively difficult to detect [28].

There are many different tools to predict HLA I- and II-restricted T-cell epitopes available online. While the current review is focused on the IEDB analysis resource, a more in-depth analysis of the different algorithms, and a review of their strengths and limitations, has been presented elsewhere [26]. The IEDB analysis resource hosts a large panel of different predictive tools [41]. In terms of bioinformatics predictive algorithms, in general class I predictions are more accurate than their class II counterparts, and methods trained on eluted ligands perform slightly better than the ones trained on binding data [26]. Interestingly, the best performance is obtained when both binding and eluted ligand data are used to train the algorithms [26]. Unbiased systematic benchmarking studies have started to shed an objective light on the matter of the performance of different methodologies [36], showing that predictive methods are remarkably accurate, essentially predicting all epitopes (high true positive rates and low false negative rates), while also showing that not all predicted binders or eluted ligands are recognized by T cell responses (relatively high false positive rates), reflecting that the basic biological fact that HLA binding is a necessary but not sufficient requisite for T cell immunogenicity. An additional methodology that was recently described [42] involves directly using epitope data to train predictive networks to predict epitopes (as opposed to binders or eluted ligands). While conceptually interesting, this method has not yielded substantial improvements over the other methodologies.

5. Immunodominance in T cell immunogenicity

To further discuss epitope identification, it is convenient to introduce some additional terminology. We define as immunodominant (Figure 4A) those epitopes that are recognized within a particular HLA haplotype, and a given protein sequence, with an immune response of relatively greater magnitude than other epitope-specific responses. According to this definition, it is intuitive that high HLA binding and/or efficient processing can contribute to immunodominance [43, 44]. The peptide sequence is determinative for proper binding to specific HLA molecules, and it is equally important to assess whether a particular epitope will be processed from an intact protein. However, the performance of tools that predict processing has been disappointing, in general, as discussed in detail elsewhere [26]. The general current trend is to incorporate the influence of processing by utilizing eluted ligands (peptides that have been generated by processing) in the training of predictive algorithms. Another variable that can contribute to immunodominance is the availability of a repertoire of TCR capable of recognizing the particular epitope/HLA combination [45, 46]. Several groups are interested in systematically studying, and one day predicting, which TCR correspond to particular HLA/epitope combinations [47, 48]. This is a very promising area, but at this point in time is not yet ready for implementation for high throughput epitope identification in large donor cohorts.

Figure 4. Immunodominance and immunoprevalence.

A) Epitopes effectively generated by processing and for which a TCR repertoire is available (light blue bars), and capable of eliciting the (relatively) strongest T cell responses, are termed immunodominant. B) Epitopes, antigens or ORFs that elicit responses with high frequency in an out-bred population are termed immunoprevalent. Typically, only a few epitopes/ORFS/antigens are found to be immunoprevalent, and may be associated with high levels of surface or RNA expression or secretion (class II); kinetics (time) of expression is often associated with dominance of class I or class II responses.

Perhaps the most important variable influencing immunodominance in terms of shaping the TCR repertoire is the relationship of the epitope sequence to other sequences related to the potential epitope. This is because the TCR repertoire is shaped by both positive and negative selection [49], and by constant tuning and expansion based on the actual sequence space encountered by an individual during their lifetime. This influence can be either positive or negative. For example, TCRs recognizing sequences that are self or highly homologous to self might be eliminated or silenced by central [50] or peripheral tolerance [51]. Conversely, cross-reactive TCRs recognizing sequences homologous to other antigens might be preferentially expanded and thus become relatively immunodominant because of repeated exposure to the homologous antigens. Examples include pollen allergen epitopes conserved across different grass species [52, 53], mycobacteria epitopes conserved across multiple bacteria [54–56], Dengue and Zika virus epitopes conserved across different serotypes [57–60], epitopes conserved in multiple herpesvirus and pox species [61–63], influenza [64, 65] and epitopes conserved across multiple coronaviruses [66]. Sequence conservation across the individual microbiome has also been shown to influence immunodominance [67, 68].

6. Immunoprevalence and T cell responsiveness at the antigen level

The previous section reviewed some of the variables that determine, within a given antigen, which epitopes would be recognized as more dominant. Additional variables dictate, when the genome of a pathogen or allergen is considered, which particular antigens would be recognized most prominently. In this context, we also consider the concept of immunoprevalence, (Figure 4B) defined on the basis of how frequently responses to a given epitope or antigen are detected when an out-bread population, such as a human patient cohort, is considered. This is complementary to the concept of immunodominance, defined on the basis of magnitude of responses [69, 70].

At the level of antibody responses, it is routinely observed that certain antigens are recognized with both dominance and prevalence. This is usually directly correlated to protein abundance, expression on the microbe surface, and antigen/epitope surface accessibility [71]. But what determines dominance and prevalence at the level of T cell responses? It is possible different antigens might be recognized as dominant in different individuals, simply because they contain peptide sequences that bind differentially to the specific HLA types expressed by the different individuals. However, for example, in the case of pox viruses it was observed that certain ORFs are immunoprevalent, e.g. recognized in multiple individuals with different HLA types, thus ruling out that prevalent recognition was reflective of HLA type [70]. Further inspection of the data revealed that the targets more prevalently recognized for CD4 responses corresponded to viral products produced abundantly and late in the infection, possibly reflecting uptake and processing of virions. Conversely, targets of CD8 responses tended to map to ORFs that were expressed early in the infection [72], potentially reflecting endogenous generation and presentation before host viral synthesis shut down.

To be able to predict which ORFs might be preferentially recognized is of limited value in the case of viruses with a genome of limited size where all ORFs can be addressed without preselection, but could be of large impact by restricting the potential targets in the case of larger viruses (such as, for example, pox and herpes viruses), bacteria and other pathogens with a complex genome (e. g. plasmodium). Direct determination of mRNA expression levels [73] in combination with HLA binding predictions is a promising approach, that has not been extensively utilized thus far. Gene annotation and expression profiles derived for each target or tools predicting the surface proteome [74–76] or secretome [77] might be of use.

Alternative methods include the experimental determination of surface expressed proteins by proteolysis, and mass spectrometry approaches which have been of value in determining targets of protective antibody immunity in several bacteria [71]. We have utilized an immunoproteomic approach to determine the targets of antibody recognition in complex allergens and shown that this analysis is useful for T cell epitope identification, because of the association between the targets of humoral responses and CD4 T cell immunity [78–80]. This association, however, does not appear to necessarily apply to infectious agents. For example, in the case of Mycobacterium tuberculosis, the targets of antibody and CD4 T cell responses are not correlated [81]. This might reflect that Mycobacterium tuberculosis is an intracellular pathogen and, therefore, recognition of antibody accessible antigens might not be relevant for T cell immunodominance. An additional potential experimental approach involves the elution of peptides bound to a limited set class I or class II HLAs, as a preliminary probe to experimentally determine which ORFs are more represented and therefore most likely to be recognized by CD4 and CD8 responses. This approach has been shown to greatly restrict the number of ORFS to be considered in the search for cancer or autoimmune epitopes [82].

7. HLA polymorphism and epitope recognition

At the genetic level, the HLA is a multi-locus gene complex. Two main loci encode for HLA A and B class I molecules, and additional loci encode for different class II molecules, to include DR (DRB1 and DRB3/4/5), DP and DQ molecules [83]. Thus, each chromosome encodes a total of two main HLA class I and four main HLA class II molecules. Each of the corresponding loci are extremely polymorphic, with several thousand different alleles expressed at different frequencies in different ethnicities [84, 85]. Thus, because of heterozygosity, each individual may express up to four different HLA class I and eight different HLA class II allelic variant molecules (Figure 5).

Figure 5. HLA polymorphism and polygeny.

The HLA region is highly polymorphic, with thousands of different allelic variants expressed at each locus. With heterozygosity, each individual will express up to two different alleles at each locus. Further, with class I and II presentation across several loci (polygeny), up to six different class I (2 HLA-A, 2 HLA-B and 2 HLA-C) and eight different class II (2 HLA-DRB1, 2 HLA-DRB3/4/5, 2 DQ and 2 DP) molecules may be expressed by an individual.

We developed targeted approaches to deal with this apparently daunting degree of multi-genicity and genetic polymorphism. Specifically, we have shown that the large majority of different molecules and allelic variants can be classified in relatively few “supertype” specificities [86–99], which are characterized by similar binding specificity (motifs) and, most importantly, largely overlapping peptide binding repertoires. Since several of the HLA allelic variant molecules are relatively more frequent, we have further shown that focusing on approximately 30–40 HLA class I [87, 88, 92, 100, 101] and 20–30 HLA class II molecules [86, 97, 102, 103] allows to cover the vast majority of HLA molecules expressed in different populations, irrespective of ethnicity.

Human T cells recognize epitopes bound to specific HLA molecules. The specific HLA molecule that binds and presents the epitope is called the epitope’s HLA restriction. Determining HLA restriction is an important component of epitope identification studies, since knowledge of HLA restriction is, for example, necessary to produce multimeric staining reagents [104–107], or to project which fraction of given population can respond to a given epitope or set of epitopes (population coverage) [101, 103, 108, 109]. Certain epitopes can bind and be recognized in the context of multiple HLA, and are termed “promiscuous epitopes” [110–115]. It has been shown that promiscuous epitopes also tend to be immunodominant/immunoprevalent [86, 116–119], accounting for a large fraction of responses to a given antigen.

8. HLA restriction of epitope responses

A variety of different approaches are used to determine or infer HLA restriction. Indeed, the IEDB database lists for each scientific study, in the form of “evidence codes,” which methodology is used to assign HLA restriction [120].

The classic approach originally developed to address this question involved the use of antigen presentation assays, utilizing loci or allele specific antibodies to inhibit presentation [121, 122] or using panels of matched/mismatched Antigen Presenting Cells (APC) [102, 117–119, 123–125]. These methodologies are laborious, and suffer from specific drawbacks. Loci/allele specific antibodies are not always available, and the assays require careful titration, as too little antibody will not inhibit, and high doses will kill APCs, or cause non-specific inhibition. Panels of matched/mismatched APCs are also not easy to assemble, especially at the population level, and because of linkage disequilibrium of alleles expressed at different loci. Furthermore, their use often does not allow to determine HLA restriction of promiscuous epitopes.

In recent years, HLA restriction is often inferred based on the presence of specific HLA binding motifs, or on predicted binding affinity. This knowledge is often combined with the knowledge of the HLA expressed in a given donor. For example, in one situation HLA-A*02:01 positive donors are tested for recognition of HLA-A*02:01 binding peptides (either predicted or experimentally verified), and positive responses are assumed to be HLA-A*02:01 restricted. In another experimental setting, a CD8 response was observed in a donor heterozygous for HLA-A*02:01, B*07:02, A*11:03 and B*27:03. Upon further experiments the peptide is mapped to a specific nine-mer predicted to bind B*27:03, but not A*02:01, B*07:02 and A*11:03. Based on these results, the HLA restriction is inferred to be B*27:03. In a further variation, HLA restriction can be inferred based on response frequencies. For example, it might be noted that a given peptide is recognized in 60% of the individuals expressing HLA DRB1*04:01, but only in 5% of individuals that do not express that particular allele. The RATE method [126, 127] automates these types of calculations. Taken together, inference of HLA restriction is effective, but also suffers from several drawbacks, particularly in the case of peptides binding multiple HLAs, individuals expressing HLAs with overlapping binding capacity, or inferences related to HLA loci in strong linkage disequilibrium.

Two additional methodologies are available to experimentally determine restrictions. Multimer staining and use of single HLA transfected cell lines. In the case of multimer staining, a reagent is prepared with a specific epitope /HLA pair, and used to bind the TCR of epitope specific T cells [104, 105, 107, 128, 129]. This is remarkably effective, but does often present a “chicken and egg” situation, in that knowledge (or a reasonable guess) of the HLA restriction is required to prepare the reagent in the first place. This strategy is therefore particularly effective when an educated guess or inference of HLA restriction is available. An alternative methodology is the use of single HLA transfected APCs [102, 130–133]. In these experiments, the peptide is tested in antigen presentation assays utilizing panels of cell lines expressing single HLA molecules that match an allele expressed in the donor from which the T cell response is generated. These assays are laborious and require availability of matching cell lines, but represent to date the most unequivocal methodology to establish HLA restriction.

9. Experimental approaches for actual epitope mapping

A variety of different approaches are used for epitope identification and mapping. The two most popular approaches involve testing overlapping peptides spanning the entire sequence of a given antigen, and the use of predicted epitopes. The relative methods of the two approaches have been compared in murine systems [134] and found to be associated with similar efficiency, even though the sets of epitopes identified by the two methods were not completely overlapping (some epitopes were identified by predictions, but were not identified by the overlapping peptides approach and vice versa). Testing overlapping peptide sets is comprehensive, when compared to predictive methods that are by definition associated with variable levels of accuracy. The drawback is that overlapping peptides require synthesis and testing of a larger number of peptides, which in turn requires availability of larger cell numbers, and some epitopes are missed because of fact that the 15-mer is a suboptimal ligand. Conversely, prediction approaches might miss unusual “non-canonical” motifs and ligand sizes, even though algorithm efficacy is constantly improving. In the case of both predicted or overlapping peptides, it is common practice to test peptide pools that are subsequently deconvoluted to identify the actual epitope recognized.

Regardless of the particular set of peptides tested, different assay and read-out strategies are available to investigators. Often, the cell numbers available to be tested is limiting. An in vitro expansion step is a possible avenue to increase signal by expanding rare T cells [117, 119, 135, 136], but this comes at the expense of a more laborious cost and time intensive set up, and alterations in the phenotype of responding T cells. For these reasons, if practically feasible, a direct ex vivo assay strategy is usually preferred.

Whether an in vitro re-stimulation step or a direct ex vivo assay is used, different final readout options need to be considered, each associated with different degrees of robustness, information content, and costs in terms of effort and resources. Popular assay options include cytokine readouts, such as Enzyme-Linked Immunospot (ELISPOT) [137] and Intracellular Cytokine Staining (ICS) [138]. The ELISPOT is a popular robust assay, but the ICS assay offers the potential advantage to simultaneously perform additional phenotypic determinations, such as determining whether the responding T cells are CD4 or CD8, and whether other common cell markers are expressed. Altogether these assays provide information regarding the functionality of the responding T cells in terms of cytokine responses. Conversely, these assays are by definition only detecting the pre-selected cytokines and are “blind” to additional cytokines.

An alternative assay that has become increasingly popular is the Activation Induced Marker (AIM) assay [139, 140]. This is an agnostic assay in that it detects all activated T cells, and is not focused on a particular cytokine. In particular, this assay is suited to detect T cell subsets that secrete small amounts of cytokines, such as Tfh cells [141–144]. Tetramer staining is also not dependent on any functional activity of the antigen specific T cells, and allows to detect specific T cells in absence of antigen stimulation. However, this methodology is not ideally suited to epitope screening, and more suited for in-depth characterization of epitope specific T cells once a given epitope is identified.

10. Epitope identification and population coverage

Particular attention needs to be devoted to how different approaches achieve and address population coverage. One approach relies on testing a given population/patient group and defining the epitopes recognized. This is an unbiased approach, but suffers the drawback of yielding coverage skewed toward the specific HLA most represented in the test population, raising questions in terms of whether the results will be applicable to other populations with potentially different HLA representation.

The alternative strategy is to define epitopes restricted by specific HLA alleles, thus allowing to project coverage in different populations, and ensuring adequate coverage in different ethnicities if the most common alleles [86, 88] expressed in populations worldwide are selected. A variation along the same lines is the algorithm described by Paul et al. [145], which combined different prototype HLA class II specificities to predict the most dominant promiscuous class II epitopes recognized regardless of ethnicity. Achieving good population coverage is perhaps most challenging in the case of HLA multimeric reagents, because a different reagent is required for each epitope/HLA pair. A quantitative analysis in a Mycobacterium tuberculosis field study [132] revealed that to cover approximately 50% of the total response would require synthesis and validation of several hundred different tetramers.

11. Examples of practical application of epitope identification studies

In this section, we provide a few selected examples of epitope identification studies that led to novel immunological or microbiological insights. A first example is provided by the study of Lindestam Arlehamn et al. [81] that provided the first truly unbiased genomic map of human CD4 T cell responses to Mycobacterium tuberculosis. This study was enabled by the combined use of prediction of promiscuous epitopes and an ELISPOT screen of a large library of over 20,000 different peptides. The results revealed that the CD4 T cell response was generally focused on a few genomic islands associated with Type IV secretion. But, the study also mapped a large number of previously undiscovered antigens and epitopes, highlighting the value of a systematic and unbiased approach.

The study by Lindestam Arlehamn et al. also defined hundreds of different epitopes, and subsequent work showed that these epitopes could be arranged in a “megapool” which could in turn be utilized as a reagent to comprehensively quantitate antigens specific responses [132, 146]. The epitopes defined in those studies were further utilized to map relationships with Mycobacterium tuberculosis recognition to non-tubercoloid-complex Mycobacteria [56], and map the evolution of responses as a function of response to therapy and microbiome similarity [68]. Similar genome-wide analyses are underway for Bordetella pertussis [147], another bacterium of high significance for human health. The megapool approach has also been validated for a variety of different targets, including, amongst others, tetanus toxoid [148], Zika virus [57], and various allergens [78, 149, 150]. In the case of Dengue virus, CD4 and CD8 megapools have been produced and utilized to quantitate responses in endemic areas and in response to vaccination [58, 151, 152].

Megapools have also played a key role in extracting and isolating antigen specific T cells for a variety of different indications, and establishing specific immune signatures [153–162]. The purpose of these studies is to identify specific correlates of immune protection versus immunopathology. Epitope identification studies have played a key role in defining the extent to which responses against antigens from different flaviviruses are cross-reactive with each other, which is of relevance for interpreting results from Dengue virus vaccines that utilize other flaviviruses as a delivery backbone [163]. Similarly, in recent studies, the megapool and epitope identification approaches outlined here have played a key role in characterizing T cell responses in SARS-CoV-2 exposed and unexposed subjects alike [66, 164, 165].

12. Conclusions

Identification of T cell epitopes is a crucial step in establishing a capacity to understand the immune response to pathogens and allergens. Various bioinformatic approaches have been developed over the years that have afforded significant advances in the epitope identification process. In addition, various experimental approaches using defined epitopes, particularly with an understanding of the impact of HLA polymorphism, immunodominance and immunoprevalence, have provided important insights into disease mechanisms and correlates of immunity.