Class I MHC molecules are polypeptides composed of a polymorphic heavy chain that associates with a constant β2-microglobulin (β2m) subunit, whereas class II MHC molecules are composed of an α and β polypeptide chain. Class I MHC molecules are generally recognized by CD8 cytotoxic T cells, whereas class II MHC molecules are recognized by CD4 T cells. For the purposes of this review, we will only discuss class I MHC molecules and their interactions with CD8 T cells. The mechanisms underlying epitope prediction and discovery for class II MHC molecules, while important, are beyond the scope of this article. The peptide binding groove of an MHC I molecule preferentially binds 8–11mer peptides. The top surface of the groove, where both the MHC protein and bound peptide are exposed, is the portion of the peptide:MHC complex detected by a TCR.10

Humans possess three MHC I genes: human leucocyte antigen A (HLA-A), HLA-B, and HLA-C. These genes encode highly polymorphic proteins, particularly in the peptide interacting region, allowing different HLA molecules to bind different sets of peptides.11 This presents a challenge to the prediction of MHC binding peptides as prediction algorithms must take into account HLA-specific binding preferences.

For a peptide to be presented on an MHC I molecule, it must be processed by the cellular antigen processing machinery. This begins with proteolysis of parental protein precursors. Exogenous proteins that are internalized by the cell through processes such as phagocytosis and endocytosis are processed in the late endosome and primarily presented on MHC II. Endogenous proteins are processed by the proteasome and are presented mainly on MHC I.12

Proteasomes are proteases that break down misfolded, damaged, aberrant, or ubiquitinylated proteins. The core particle of a proteasome is a barrel-shaped structure composed of two outer α-rings and two inner β-rings.13 Variants of the β-subunits are interferon gamma (IFN-γ) inducible and can replace their constitutive counterparts.14 When these are expressed and active, they form a complex known as the immunoproteasome. The constitutive proteasome is capped at each end by a 19S complex that helps proteins into the proteasome lumen by deubiquitinating and unfolding them. The immunoproteasomes are capped by an alternative 11S complex that is also IFN-γ inducible.14 As a result of the different proteolytic subunits and capping structures, the cleavage site specificity differs between constitutive and immunoproteasomes.15

The characterization of proteasomal activity includes two basic approaches. In vitro study involves co-incubation of parental proteins with proteasomes. The resulting peptides are analyzed using mass spectrometry.16 Several proteolysis prediction techniques are trained on in vitro datasets.17 In vivo proteolysis study involves the measurement of peptides that have been eluted from an peptide:MHC complexes and subjected to mass spectrometry.18 A limitation of this strategy is that only a small proportion of cleavage sites can be identified in this manner as most cleaved peptides do not end up being presented on MHC complexes. Furthermore, the influence of other cellular proteases cannot be deconvoluted from pure proteasome function. Several prediction techniques are trained on in vivo datasets and are generally superior at predicting in vivo proteolysis as compared with prediction techniques trained on in vitro datasets.19

The peptide loading complex (PLC) is a multi-subunit, endoplasmic reticulum (ER) membrane complex which coordinates peptide translocation into the ER, editing, and loading onto MHC I molecules. One of the core components is the transporter associated with antigen processing (TAP), which is made up of the TAP1 and TAP2 subunits. Each subunit contains a transmembrane domain and a nucleotide binding domain which interact to mediate a channel for peptide movement between the cytosol and ER lumen.12 TAP preferentially translocates peptides with 9–16 residues, although longer peptides up to 25–30 amino acids in length can also be translocated with lower efficiency.20

The specificity of TAP has been well studied. On the C-terminal end of the peptide, TAP selects for the presence of an aromatic, hydrophobic, or positively charged terminal amino acid such as phenylalanine, tyrosine, arginine, or leucine.21 On the N-terminal end, the first three residues have significant effects on peptide binding. Aromatic and hydrophobic side chains are favored and an N-terminal arginine is optimal.22 Proteasomal specificity at the C-terminal residue is relatively non-specific and aligns with the restraints of both TAP and MHC I molecules.23 However, the N-terminal ends of TAP-translocated peptides require further processing within the ER to conform to MHC I binding requirements.24 ER aminopeptidase 1 (ERAP1) and ERAP2 are ER luminal components tasked with the intraluminal processing of peptides.25 ERAP1 trims N-terminal amino acids to create 8–11mer peptides capable of fitting within the MHC I binding groove.26 Peptides undergo structural changes when reaching this size, preventing ERAP1 from further trimming.27

Within the PLC, TAP is associated with a number of other proteins. Tapasin interacts with TAP and serves to recruit MHC I–β2m complexes to the PLC.28 Tapasin is further linked to ERp57 and calreticulin, which mediate the folding of newly synthesized MHC molecules.12 Following translocation of peptides by TAP, peptides are released into a molecular basket, confined by two editing complexes composed of tapasin, CRT, ERp57, and MHC I. Peptides are then further edited by ERAP1 and loaded onto MHC I molecules.29 Once loaded onto MHC I, peptides undergo additional processing and quality control mediated by tapasin and TAP binding protein-related protein (TAPBPR).30 TAPBPR associates with empty MHC I molecules, stabilizing their structure. The choice of peptide is based on the stability of the final MHC I–peptide complex, and TAPBPR is released once a high-affinity MHC I–peptide complex is stably achieved.31 A similar mechanism is thought to underlie the function of tapasin in the PLC.29 31 If the affinity for the bound peptide is high and the resulting complex is stable, the MHC I–peptide complex will be transported through the Golgi apparatus and to the cell surface.12 The peptide specificity associated with each step of this process beginning from proteasome proteolysis to MHC I binding is what defines natural processing and presentation of epitopes. For a peptide to be clinically relevant, experimental validation should strive to mimic this process.

Endogenous proteins are processed through the proteasome pathway and loaded onto MHC I molecules for detection by CD8 T cells, whereas exogenous proteins are processed through the endosomal pathway and presented on MHC II molecules for detection by CD4 T cells. Tumor cells alone are unable to prime an effective, antigen-specific, CD8 cytotoxic T cell response.32 How do CD8 T cells targeting MHC I-restricted antigens become primed and activated against antigens that originated as exogenous proteins?

Cross presentation refers to the presentation of exogenous proteins on MHC I.33 In 1990, Rock and colleagues found that soluble antigens in the extracellular fluid could be internalized, processed, and presented on MHC I molecules in specialized antigen presenting cells (APC).34 One crucial APC is the dendritic cell (DC). Mechanistic studies in mice have demonstrated that the development of antigen-specific T cell responses depends on functional APC.32

DCs can acquire antigens from multiple sources. Dying cells release antigens and they are an important source of antigenic materials.35 Antigens ingested by DC through micropinocytosis, endocytosis, or phagocytosis are efficiently cross-presented.36 While the exact mechanisms are unclear, it is hypothesized that the corresponding downstream intracellular compartments, such as the phagosome or endosome, contain cellular proteins, which are part of the MHC I presentation pathway.33 Additionally, DCs can internalize cytosolic and membrane materials from live cells to prime T cell responses.37 Possible mechanisms for this transfer of antigens includes ingestion of secreted antigens and trogocytosis, where one cell ‘nibbles’ bits off another.38 39

Once a DC has acquired an antigen, processing and presentation occur through either a proteasome-independent vacuolar pathway or a proteasome-dependent cytosolic pathway. To establish a diverse repertoire of T cells that can recognize tumor-presented antigens, an ideal cross-presentation pathway would mimic the natural processing and presentation machinery of a tumor cell. In the vacuolar pathway, proteins are lysed into peptides within the endocytic compartments and then loaded onto MHC I molecules without ever encountering the cytosol.33 40 Therefore, the proteasome and TAP are never encountered, meaning that the repertoire of selected epitopes may differ from those presented on a tumor.41 In the proteasome-dependent pathway, proteins progress through the cytoplasm and ER.42 Several studies have confirmed that exogenous antigens appear in the cytosol.43 Once in the cytosol, the proteins undergo the same process described earlier. Studies have verified that peptides generated by the immunoproteasome from conventional endogenous antigens are identical to immunoproteasome-generated peptides from exogenously delivered proteins.44 Presentation of tumor antigens by APC to naïve T cells in lymphoid tissues then primes T cells for clonal expansion and cytotoxic effector function.

Many sources of cancer antigens exist. Epitopes derived from single nucleotide variations (SNVs), insertions/deletions (INDELs), transcript splice variants, gene fusions, and endogenous retroviral elements can all serve as tumor-specific antigens.45 46 Some immunotherapies do not rely on the identification of tumor antigens for use, while other forms of immunotherapy do require knowledge of antigen and possibly HLA type. Immune checkpoint blockade immunotherapy is widely used but induces durable tumor control in only one subset of patients and does not require a priori knowledge of antigens.47 Adoptive transfer of bulk cultured tumor infiltrating lymphocytes (TIL) as an immunotherapeutic approach has produced durable regression of malignancy in a limited subset of patients, and this approach does not rely on the identification of specific tumor antigens.48 Conversely, adoptive cell transfer of TCR-engineered T cells does rely on the identification of a specific antigen target, the HLA restriction, and the TCR specific for this antigen:MHC complex. Autologous T cells are engineered to express TCRs that target antigen:MHC complexes presenting tumor-specific neoantigens.49 Personalized cancer vaccine therapies similarly rely on the identification of tumor-specific neoantigens.50 51 However, the utility of these potential therapies is constrained by the difficulties underlying the discovery of tumor-specific epitopes that elicit T cell responses.52 Here, we review the commonly utilized epitope discovery techniques and key considerations underlying these methods.

A general approach to identify neoepitopes and accompany reactive T cells from humans is a reverse immunology approach based on tumor sequencing. A section of surgically resected tumor or metastasis is cultured in media containing cytokines selective for lymphocyte growth, notably IL-2, to expand TIL in vitro. In addition, another portion of the tumor is sent for genome sequencing to identify genetic mutations and the corresponding repertoire of potential neoepitopes. The resulting neoepitopes can be further analyzed in silico to predict the best candidates for MHC I cell surface presentation. Next, T cells are cocultured with APC expressing the antigens of interest and analyzed for activation signals. T cells reactive against antigens can be assessed further via TCR sequencing or expanded as primary effector cells.

Tran and colleagues have successfully applied this strategy to identify T cells reactive against mutant KRAS G12D in a patient with metastatic colorectal cancer.48 Three metastatic lung lesions were resected and used to generate TIL cultures and whole exome sequencing (WES) data. In total, 61 mutations were identified. Cultures of TIL were screened against all identified mutant neoepitopes expressed by the tumor. For each mutation, a minigene encoding the mutation flanked by 12 amino acids on either side was generated and synthesized in tandem to create tandem minigene (TMG) constructs. Five total TMG constructs were made, in vitro transcribed into RNA, and electroporated into autologous DCs to undergo natural processing and presentation of epitopes. TIL cultures were reactive against TMG-1, so the corresponding peptides were synthesized and pulsed onto DCs, followed by coincubation with TIL. Adoptive transfer of four T cell clonotypes that demonstrated antigen-specific reactivity resulted in objective regression of all lung metastases.48 The TMG approach to identify tumor-specific T cell antigens is highly comprehensive as it interrogates responses from all possible neoantigens.

The advent of in silico epitope prediction algorithms began with the hypothesis that certain peptide sequence motifs were more likely to bind to an MHC molecule compared with others.73 These motifs were generally based on primary anchor residues that highly influenced binding and secondary anchor positions that were less important for binding.74 For example, positions 2 and 9 serve as anchor residues in HLA-A*02:01 and are most often occupied by small hydrophobic amino acids such as valine, leucine, and isoleucine.75 From these motifs, matrices were developed which assigned a numerical score corresponding to the contribution of each amino acid at each position; values for each amino acid within a peptide were then combined to generate a score for each peptide:MHC combination.76 By design, these models assumed that the contribution of an amino acid at a particular position was independent of its neighboring positions. This may not be true.77

To handle these interactions and to improve predictions over motif-based models, researchers turned to machine learning algorithms. The machine learning algorithms used for T cell epitope prediction are trained on large-scale datasets, resulting in a model that deciphers patterns from a known set of input data and applies them to new data to generate predictions. Datasets used for T cell epitope prediction can be classified as datasets of peptides with known binding affinity to MHC molecules and datasets containing peptides that either bind or do not bind to MHC molecules.78 Models trained on the combination of MHC binding and ligand elution data show superior performance compared with models trained on binding datasets alone.79 An additional benefit of machine learning is that methods can be developed to predict binding for rare HLA alleles using data from well-known HLA alleles (pan-specific predictors).79 This is important because rare HLA alleles often do not have sufficient data to train their own prediction algorithms. The quality of the available datasets determines the performance of the machine learning algorithms. Several databases compile epitope data in a consistent manner to enable proper training of machine learning algorithms.80

Most in silico epitope prediction algorithms aim to predict binding affinities of peptides to MHC I molecules. This is based on the assumption that peptide binding to MHC is the single most selective step in the antigen presentation pathway.77 79 Comparison of a combined proteasome/TAP/MHC affinity predictor with a predictor based on MHC affinity alone determined that the MHC affinity predictor offers higher sensitivity whereas combined predictor offers higher specificity. For a hypothetical protein of 300 amino acids, one would need to functionally test 11 peptides to identify a new T cell-reactive epitope with 90% likelihood when using a combined predictor versus 13 peptides when using a MHC affinity predictor.81 Furthermore, widely used prediction algorithms do not incorporate peptide:MHC complex stability, which is a more important determinant of immunogenicity than binding affinity.82 Incorporation of peptide dissociation data into current algorithms may improve prediction performance.83 Current immune epitope database (IEDB) guidelines recommend using MHC binding predictions alone by default and using prediction scores that incorporate antigen processing as an additional filter if resources require limiting.84

The ways in which researchers use MHC epitope prediction data are varied. Studies of MHC I binding affinity and T cell epitope recognition have shown that predicted MHC binding IC₅₀ of 500 nM or less results in approximately 90% sensitivity for identification of true T cell reactive epitopes.85 Genetic differences between HLA alleles create disparities; some alleles are better able to bind oncogenic mutations while others are predicted to have very few peptide binders with IC₅₀ <500 nM.86 Therefore, the rank of an epitope’s MHC binding affinity in comparison with all other possible epitopes (percentile rank) has been suggested as an alternative. When generating predictions across multiple HLA alleles, percentile ranks are preferred to avoid bias towards certain HLA alleles.87

In the development of a personalized melanoma vaccine, Ott et al utilized NetMHCpan to identify up to 30 patient-specific MHC I-restricted high-affinity binders derived from mutant peptides to include in each vaccine.88 The final list of epitopes was chosen based on a rank list that prioritized neo-ORFs over SNVs, mutations in anchor residues over secondary residues, and low MHC IC₅₀. For each patient, up to 20 putative neoantigens were ultimately incorporated into a personalized vaccine and delivered. All neoantigens had predicted MHC IC₅₀ <500 nM or rank in the top one percentile. Ex vivo analyses demonstrated that 15 of 91 (16%) predicted immunizing peptides-induced CD8 antigen-specific T cell responses across six patients. The total number of predicted epitopes that can be included in each vaccine is limited, but the use of MHC I binding prediction algorithms with astute prioritization criteria generated sufficient specificity to generate T cell reactive epitopes for each patient. Of six vaccinated patients, four were recurrence-free 25 months after vaccination and the remaining two experienced complete tumor regression after treatment with anti-programmed death receptor-1 (PD-1) therapy.88

We utilized the IEDB proteasomal cleavage/TAP transport/MHC class I combined predictor tool to retrospectively generate total prediction scores and MHC binding affinities of a list of known and validated epitopes derived from the Cancer Antigenic Peptide Database (https://caped.icp.ucl.ac.be/).89 The total score is a combined score incorporating proteasomal cleavage, TAP transport, and MHC binding predictions where higher scores represent higher efficiency for MHC presentation.90

Validated tumor-specific neoantigens are summarized in table 1 and validated tumor-associated antigens derived from germ-line or differentiation antigens are summarized in table 2. Tumor-specific and tumor-associated antigens displayed similar MHC IC₅₀ and total score profiles and were combined for further analyses. A significant proportion of epitopes displayed weak predicted binding (IC₅₀ >500 nM, 30.3%) to their respective MHC alleles, with several epitopes exhibiting very weak predicted binding (IC₅₀ >10,000 nM, 9.7%) (figure 1A). One possible explanation for why low-affinity peptides elicit strong T cell responses is that these weakly immunogenic antigens have not been subjected to immunoediting and therefore represent the immunodominant antigen following elimination of more immunogenic targets. When considering discovery of tumor-specific putative neoantigens, nearly all studies utilized a comprehensive querying technique that did not use, or minimally used, epitope prediction algorithms. Methods that heavily incorporated epitope prediction methods were also successful, but rarer in this database. Robbins et al tested the top 55 predicted 9mer or 10mer MHC I binders from a melanoma cell line and demonstrated that the 5th, 18th, 19th, and 39th ranked peptides elicited T cell responses.7 The hit rate of 4/55 and seemingly random rank of peptides validated as antigens raises the question of whether peptide prediction algorithms perform better than random testing. Unfortunately, it is impossible to know how many similar studies failed to recognize successful epitopes. The only solution to achieve maximum sensitivity for detecting antigens is comprehensive screening of all possible neoantigens, which should be used when resources are available.

There was high correlation between MHC IC₅₀ and total prediction score in each group (figure 1B), indicating that use of MHC binding predictions alone compared with total prediction scores is likely to generate similar results. When comparing validated neoantigens with their wild-type (WT) counterparts, many neoantigens had predicted MHC I binding affinities and total prediction scores that were similar to that of the WT sequence (figure 1C). Since WT sequences are widely presented throughout the body, central T cell tolerance should have protected the body against auto-immunity. Shaping of T cell repertoires through thymic selection is a critical step in the determination of tolerance on exposure to various antigen:HLA complexes.91 The simplest experimental models of thymic negative selection have demonstrated that negative selection is specific enough that single amino acid substitutions in self-peptides can effectively be recognized as foreign by T cells and serve as neoantigens.92

The stability of the peptide:HLA class I complex may determine presentation to and development of an activated T cell clone to a greater degree than simple peptide binding affinity.93 The algorithms detailed earlier do not incorporate stability as a variable, illustrating one deficiency of current, widely utilized modeling approaches. Weighted inclusion of stability in predicting immunogenic epitopes outperformed predicting based on binding affinity alone, and such parameters may be incorporated into future prediction pipelines.83