Detailed information about the biological samples that were included in this database is provided in Table S1 in Supplementary Material. Informed consent of the participants was obtained following requirements of the institutional review board [Ethics Commission, University Hospital of Lausanne (CHUV)]. W6/32 (anti-pan-HLA-I) and IVA12 (anti-pan-HLA-II) monoclonal antibodies were purified from the supernatant of HB95 and HB145 cells, respectively, as previously described (17). We cross-linked the antibodies to Protein-A Sepharose beads (Invitrogen, CA, USA) with 20 mM dimethyl pimelimidate in 0.2 M sodium borate buffer pH9.

We included here also unpublished immunopeptidomics data of HLA-I and HLA-II peptides extracted from several biological replicates per cell line or patient material. The cell counts ranged from 1 × 10⁸ to 5 × 10⁸ cells or up to 2 g of tissue per replicate. Purification from these additional samples was performed as previously described (17, 33). Shortly, snap-frozen tissue samples were homogenized for 10 s on ice using ULTRA-TURRAX (IKA, Staufen, Germany) in a tube containing 5–10 ml of lysis buffer and incubated at 4°C for 1 h. Cell pellets were resuspended in 5 ml lysis buffer and incubated similarly. Lysis buffer contained 0.25% sodium deoxycholate (Sigma-Aldrich), 0.2 mM iodoacetamide (Sigma-Aldrich), 1 mM EDTA, 1:200 Protease Inhibitors Cocktail (Sigma, MO, USA), 1 mM Phenylmethylsulfonylfluoride (Roche, Mannheim, Germany), 1% octyl-beta-D glucopyranoside (Sigma). The lysates were cleared by centrifugation with a table-top centrifuge (Eppendorf Centrifuge 5430R, Schönenbuch, Switzerland) at 4°C at 14,200 rpm for 20 min. Immuno-affinity purification from tissues was performed by passing the cleared lysates through Protein-A Sepharose beads, then through Protein-A Sepharose beads covalently bound to W6-32 antibodies, and finally through beads covalently bound to IVA12 antibodies. Purification from cell line lysates required only the two last affinity columns. Affinity columns were then washed with at least 6 column volumes of 150 mM NaCl and 20 mM Tris–HCl (buffer A), 6 column volumes of 400 mM NaCl and 20 mM Tris–HCl and lastly with another 6 column washes of buffer A. Finally, affinity columns were washed with at least 2 column volumes of 20 mM Tris HCl, pH8. HLA complexes were eluted by addition of 1% trifluoroacetic acid (TFA, Merck, Darmstadt, Switzerland) for each sample. To further purify the peptides, the elution samples were loaded separately on Sep-Pak tC18 (Waters, MA, USA) cartridges, which were pre-washed with 80% acetonitrile (ACN, Merck) in 0.1% TFA and 0.1% TFA only. After loading, cartridges were washed twice with 0.1% TFA before separation peptides were eluted with 30% ACN in 0.1% TFA. The peptide samples were dried using vacuum centrifugation (Eppendorf Concentrator Plus, Schönenbuch, Switzerland) and re-suspended in a final volume of 12 μL 0.1% TFA. For MS analysis, we injected 5 μL of these peptides per run.

Measurements of HLA-I and HLA-II peptidomics samples were acquired using the nanoflow UHPLC Easy nLC 1200 (Thermo Fisher Scientific, Germering, Germany) coupled online to a Q Exactive or Q Exactive HF Orbitrap mass spectrometers (Thermo Fischer Scientific, Bremen, Germany) with a nanoelectrospray ion source as previously described (33). We packed an uncoated PicoTip with diameter of 50 cm × 75 µm and 8 µm tip opening with a ReproSil-Pur C18 1.9 µm particles and 120 Å pore size resin (Dr. Maisch GmbH, Ammerbuch-Entringen, Germany) resuspended in Methanol. The analytical column was heated to 50°C using a column oven. Peptides were eluted with a linear gradient of 2–30% buffer B (80% ACN and 0.1% formic acid) at a flow rate of 250 nL/min over 90 min.

Data were acquired with data-dependent “top10” method, which isolates the ten most intense ions and fragments them by higher-energy collisional dissociation with a normalized collision energy of 27%. The MS scan range was set to 300–1,650 m/z with a 60,000 (200 m/z) resolution and a target value of 3e6 ions. The ten most intense ions were sequentially isolated and accumulated to an AGC target value of 1e5 with a maximum injection time of 120 ms and MS/MS resolution was 15,000 (200 m/z). The peptide match option was disabled. Dynamic exclusion was set for 20 s.

We employed the MaxQuant computational proteomics platform (36) version 1.5.3.2 to search the peak lists against the UniProt database (Human 85,919 entries, May 2014) and a file containing 247 frequently observed contaminants. All MS/MS datasets were processed in one batch using a global spectrum level false discovery rate (37) cutoff of 1%. Protein N-terminal acetylation (42.010565 Da) and methionine oxidation (15.994915 Da) were set as variable modifications. The second peptide identification option in Andromeda was enabled. The enzyme specificity was set as unspecific. The initial allowed mass deviation of the precursor ion was set to 6 ppm and the maximum fragment mass deviation was set to 20 ppm.

An in-house Java program¹ based on the MzJava class library (38) was used to parse the MaxQuant results and organize them in a database (ipMSDB). The matching and scoring between ipMSDB and query peptides was done by another in-house Java program and all further data analysis and visualization was performed in R² if not otherwise indicated.

The GO enrichment analysis of the source proteins of the presented HLA peptides was performed on the Panther webpage³ (39). All human proteins were taken as a background and compared to the different protein lists based on the biological process classification. Proteins were quantified by their type I, II, or I/II peptide counts, respectively. A statistical overrepresentation test was performed and the resulting p-values were corrected for multiple testing and a p-value threshold of 0.05 was applied.

For visualizing of the protein lists, we used the Proteomaps tool⁴ (40), which is based on the KEGG protein annotation. We used the same protein lists and the same protein peptide counts as for the GO analysis. The resulting Veronoi-tree-map images were slightly edited for better visibility of the text.

HLA-I, HLA-II, and HLA-I/II density profiles were calculated from ipMSDB by summing up for each amino acid in a protein sequence the number of HLA-I, HLA-II, and HLA-I/II peptides covering this amino acid. In order to display profiles of HLA-I peptides of typical length, HLA-I peptides shorter than 15 amino acids were also considered separately. The correlation between profiles was calculated in a way that reflects the overlap of the main peaks in a profile ignoring smaller peaks.

Profiles for NetMHCpan version 3.0 (25) HLA-I-binding affinity prediction for the GILT, SEM4D, and MITF proteins were obtained from the web page.⁵ All HLA supertype representatives and peptides of length 9–11 were selected. All predicted strongly or weakly binding peptides were retained. The profiles are the NetMHCpan scores of the representative binders summed up over all strong or weak binding peptides for each amino acids of the protein.

In order to test whether our ipMSDB based features are able to improve the prioritization of predicted immunogenic peptides, we used data from a recent publication by Stronen et al. (41): 1,034 HLA-I peptides of length 9–11 carrying non-synonymous somatic mutations were obtained by genome sequencing from 3 melanoma patients and were screened with T-cell assays for immune recognition. 16 out of 1,034 neoantigens turned out to be immunogenic. Stronen et al. calculated several features for both mutated and wild-type (wt) peptides to enable prioritization of neoantigens: best predicted binding affinity to one of the patients HLA-I alleles (mutAffinity, wtAffinity), predicted HLA-I-peptide complex stability (mutPeptideStability, wtPeptideStability), proteasomal cleavage probability (mutCleavProb, wtCleavProb), number of mutated and wt reads (mutReads, wtReads), and RNA abundance (rnaExpr). Here, we added three MS-based prediction scores, which evaluate how well the wt counterpart of the predicted mutated peptide is represented in ipMSDB. The first score (nrMatchingPeptides_I) operates on the protein level and counts the number of all wt HLA-I peptides per protein in ipMSDB. The other two scores (matchScore_I and exactMatchScore_I) operate on the peptide level. In order to calculate matchScore_I for a peptide, we sum up the HLA-I density profile height over the position of the peptide. If the peptide is found on multiple proteins and/or several times on the same protein in ipMSDB, we take the highest of all the matchScore_I values. exactMatchScore_I is equal to matchScore_I if there is an exact wt peptide match in ipMSDB and 0 otherwise.

In order to compare our ipMSDB predictors to the features described in Stronen et al. individually, all the 1,018 control and 16 immunogenic peptides were used. We applied a support vector machine (SVM) regression (42) with Gaussian kernel to compare the predictive power of feature sets. The R package e1071, which is an interface to the LIBSVM SVM implementation (43), was used for this purpose. Other than kernel selection no optimization was performed and all SVM parameters were kept at their default values. Combining several features means that each peptide is represented by a feature vector in a N-dimensional space, where N is the number of features. The task of the SVM regressor is to grade this feature space with values between −1 and 1, where values close to one represent “immunogenic” regions of the feature space. To calculate this peptide immunogenicity grading, the SVM needs to learn from training data, and in order to evaluate the quality of this learning, the grading is compared to independent test data. If the immunogenic test peptides lie in regions in the feature space with grade close to 1 and the control test peptides in regions with grade close to −1, the learning has worked well. To perform the learning and independent testing, the control and immunogenic peptide lists were both randomly split into equally sized training and test parts. One part of the control and one part of the immunogenic peptide list go to the training set (half of all peptides), the other parts to the test set (the other half of all peptides). The SVM was trained on the training set and the trained SVM regression was used to rank the peptides in the test set. The number of immunogenic peptides in the 20 top ranked test peptides was calculated as the prediction performance (value between 0 and 8). The process was repeated 2,000 times in order to calculate average performance values and their standard deviations.

The experimental extraction procedure of HLA peptides highly enriches for the true HLA ligands. More than 95% of the HLA-I peptides we identified by MS matched the typical properties of sequence length and binding motifs that are necessary for binding to the different HLA-I allotypes (31). In order to build the ipMSDB database of HLA peptides, we compiled data from our recent published immunopeptidomics experiments (17, 31, 33, 44) and we added unpublished data (Table S1 in Supplementary Material). Altogether, ipMSDB represents an in-depth repertoire of HLA-I and HLA-II peptides purified separately from dozens of different human cell lines and tissues covering many HLA allotypes. Currently, our ipMSDB includes 15,422 protein groups with at least one valid peptide match (only MaxQuant leading razor proteins were considered) identified from 67 different biological samples, mainly B-cells (13 samples), T-cells (4 samples), and melanoma tissues (35 samples). At the peptide level, this corresponds to 131,402 unique peptides detected in HLA-I peptide samples and 66,420 unique peptides detected in HLA-II peptide samples. The length distribution (mostly 9 to 11 -mer peptides) of the identified HLA-I peptides highlights the purity of the peptidome (Figure 1A). Unlike HLA-I complexes, HLA-II complexes presented families of longer peptides (mainly 13–17 amino acids) (Figure 1A), sharing the core binding region of typically 9 amino acids. The binding restrictions of HLA-II peptides are still rather poorly understood and technically it is more challenging to retrieve them directly from immunopeptidomics data as a way to estimate the purity of HLA-II peptidome samples. We expect that the HLA-II peptidomes have similar high purity level because we purified them similarly to the HLA-I. Also, when cells lack HLA-II expression no peptides are detected (17). Interestingly, 6,819 unique HLA-I/II peptide sequences (3.4%) in ipMSDB were detected in both HLA-I and HLA-II samples. Figure 1B reveals that their length distribution is a mixture of the class I and class II modes.

Remarkably, the broad distribution of the number of peptides presented as HLA-I and HLA-II peptides per source protein implies that the proteins are not randomly selected for presentation (Figure 1C; Figure S1A in Supplementary Material). As we have shown previously (31), the number of HLA-I peptides per protein depends on the protein length (Figure S1B in Supplementary Material), but the HLA presentation of proteins also depends on many other factors. The assembly of the above database has allowed preliminary observations that provide important hints on the biogenesis of the immunopeptidome. These could be exploited in the development of algorithms for optimizing the prediction of neoantigens. Our main working hypotheses are presented below.

We compared the characteristics of the source proteins presented as HLA-I and HLA-II peptides in terms of the biological process GO annotations (see Materials and Methods for details). We further visualized the data with the Proteomaps tool (see Materials and Methods), which is based on the KEGG protein annotation. The Proteomaps tree-map visualization tool shows quantitative composition of proteomes arranged in multiple levels. As a pseudo quantitative score we used the number of assigned HLA peptides per protein in ipMSDB. On the lowest level, each protein is represented by a polygon, whose area reflects the number of HLA-I, HLA-II or HLA-I/II peptides, times the protein length, respectively. Functionally related proteins according to a KEGG hierarchy tree were arranged in adjacent and similarly colored regions. On higher levels, similar proteins were grouped into regions. We investigated cellular proteins separately for the following groups:

Cellular proteins presented as HLA-I peptides, collectively named “type I.” A large fraction comprising 93.2% of the source proteins in ipMSDB were presented as HLA-I peptides. HLA-I molecules present most of the cellular proteome. GO annotation enrichment analysis (Table S2 in Supplementary Material) revealed that compared to the reference human proteome, “type I” were enriched in the following biological processes: nuclear chromosomes (p-value = 2.93E−02), nucleus (7.02E−10), and nuclear envelope (2.10E−03) and also lysosome (1.01E−02), endosome (4.67E−04) the Golgi apparatus (4.42E−02), vacuole (2.69E−04) and more general annotation like the ribosome (3.54E−23) and the cytoskeleton (6.81E−07). The MHC protein complex was enriched (6.88E−03), while membrane proteins in general were depleted (3.87E−02). The proteomaps were in agreement with the GO annotations enrichment analysis (Figure 2A and in more details in Figure S2 in Supplementary Material). This pattern was independently observed in B-cells, T-cells and in melanoma tissues. Differences were partially related to differences in protein expression between cell lineages and/or between in vivo tissues and cells growing in culture. For example, ribosomal and cytoskeleton proteins were more prominent in melanoma tissues than in B- and T-cells, while proteins related to DNA replication were presented more in rapidly dividing cells growing in culture (Figure S3A in Supplementary Material).

Collectively, these results indicate that the sampling of the self proteome for presentation on HLA-I and on HLA-II complexes is not random and the cellular localization of proteins, possibly also related to the mechanism of their degradation, has an impact. More than that, a subset of the proteome is presented by both machineries and resembles “type II” source proteins.

3.4% of peptide sequences in ipMSDB were detected as HLA-I/II peptides. Such long peptides detected in HLA-I peptidome could also be a technical artifact of contaminating HLA-II peptides that occurs during the purification. However, several main observations argue against this option: first, a significant part of the long HLA-I peptides fit the P2/PΩ-anchor mode of binding to the expressed HLA-I allotypes. We showed this for the UWB289 ovarian cancer cells that do not express HLA-II, and melanoma tissues from Mel15 and Mel16 patients form which both HLA-I and HLA-II peptidomes were obtained (Figure S6 in Supplementary Material). The remaining peptides could still bind with alternative internal anchors leaving the ends of the peptides to protrude beyond the binding groove (45, 46). Second, we calculated the proportion of long peptides (equal or longer than 14-mers) detected in cell lines or tissue samples that express HLA-II, and those that lack HLA-II expression, i.e., in which no or only sparse amounts of less than 100 HLA-II peptides could be detected by MS. The average proportion of longer HLA-I peptides in the group expressing HLA-II peptides was 5.6%, whereas the average proportion in the group lacking HLA-II peptides was 4.8% which was not significantly different (standard deviation is 3%). Therefore, it is very unlikely that the long HLA-I peptides were contaminations from HLA-II peptides. Furthermore, HLA-II peptides were purified from the lysates after the HLA-I had been purified, which minimized the chance that HLA-II peptides would contaminate class I peptidome samples. No HLA-II peptides could be detected in samples from which thousands of HLA-I were identified, which supports the claim that there is no significant cross contamination related to sample handling. We further elaborate on the possible biogenesis of the HLA-I/II peptides below.

Cross-presentation has been investigated for many years and it has been shown to be central for the priming of naïve T-cells against exogenous antigens. These antigens are taken up by professional antigen-presenting cells that process and consequently present them on HLA-I molecules (47). Cross-presentation happens via two orthogonal routes: a proteasome- and TAP-independent route where proteins digested in endosomes are loaded on HLA-I molecules imported into the endosomes, and a proteasome- and TAP-dependent route where endosomal proteins are exported to the cytosol and processed by the HLA class I presentation machinery (Figure 3).

Several lines of evidence, which we discuss below, led us to propose that “type I/II” source proteins are processed partially by the machinery involved in cross-presentation in the endosome-lysosome compartments. We hypothesize that cross-presentation of peptides cleaved in the endosomes consequently leads to the generation and loading of longer HLA-I peptides that are likely to be in common with the peptides generated by the class II processing machinery, and stem from the same source proteins (Figure 3). It is important to note that cross-presented peptides may also be generated after the polypeptides have been transferred from the lysosome–endosome compartments into the cytosol. Following the conventional class I presentation pathway that is proteasome- and TAP-dependent, these peptides will then become indistinguishable from the normal pool of HLA-I peptides characterized with a typical length (9–11 aa) (47).

We observed that HLA-I/II peptides stemmed mainly from self proteins localized within the endosome-lysosome compartments and were enriched with phagosomal structural proteins and phagosomal cargo proteins that are degraded by autophagy (Figure 2). Among them are ribosomal proteins and mitochondrial proteins that may be selectively degraded and eliminated in the processes called ribophagy (48) and mitophagy (49), respectively. Previously, we studied HLA-I presentation in several cell lines and we have shown that ribosomes and mitochondrial proteins are presented to a higher extent than what would be expected from their abundance (31). One example of a protein that belongs to the “type I/II” group is the “probable serine carboxypeptidase” (CPVL) localized in phagosomes. It may be involved in the digestion of phagocytosed particles in the lysosome and in trimming of peptides for antigen presentation (50). Another example is the PMEL protein from which several peptides were detected to be presented on HLA-I and HLA-II complexes. PMEL is involved in melanosome formation and disintegration of melanosomes is assumed to take place in the lysosomes (51). Furthermore, the autophagic pathway has a substantial role in the degradation of melanosomes in keratinocytes (52). The confined space within the endosome-lysosome compartments may indeed favor cross-presentation of this set of proteins and could also explain how low abundant proteins may still be presented with multiple ligands and out-compete very abundant proteins. Furthermore, HLA-I/II peptides seem to be more prominent in B-cells and T-cells compared to melanoma tissues where class II presentation machinery might not be fully functional (Figure S3B in Supplementary Material).

Based on these observations, we hypothesize that in cancer or upon infection, professional antigen presenting cells that take up antigens released by dying cells and degrade them in the endosome-lysosome compartments, would present their longer peptides as either HLA-II or HLA-I peptides generated through the proteasome-independent pathway (Figure 3). Furthermore, in case cells are directly infected with intracellular pathogens or at a steady state condition, autophagy may lead to the presentation of longer HLA-I peptides from the pathogens or from the self-proteome. For example, a recent study investigated the HLA-I peptidome of cells upon infection with the intracellular pathogen Toxoplasma gondii (45), and reported that the T. gondii ligands were significantly longer than host ligands. The average length of T. gondii ligands was 14.6 amino acids compared to 11.4 amino acids of host ligands for infected and 9.8 amino acids for uninfected cells. Furthermore, they observed that the long ligands did not follow the P2/PΩ-anchor binding mode of HLA-I but instead were predicted to bind via a canonical N-terminal binding core preceding the C-terminal extension. Both the length preference and the mode of binding of these peptides may be explained by the alternative processing we describe here. Notably, the 9–11-mers could potentially be mainly driven by ER-resident chaperones and peptidases that are known to play a role in the ER conventional class I presentation pathway.

Interestingly, we have noticed that there are “hotspots” of antigen presentation within proteins, and that domains within proteins are presented at a higher extent. We separately aligned HLA-I, HLA-II, and HLA-I/II peptides to the protein sequences. As an example, we show the hotspots we detected for the gamma-interferon-inducible lysosomal thiol reductase (GILT) protein (UniProt P13284), the semaphorin-4D (SEM4D) protein (Uniprot Q92854), and the microphthalmia-associated transcription factor (MITF, Uniprot O75030) (Figures 4A–C, respectively). More examples are provided in Figures S7A–F in Supplementary Material. GILT is the only enzyme known to catalyze disulfide bond reduction in the endocytic pathway. It facilitates presentation of a subset of HLA peptides from disulfide bond-containing antigens (53). GILT is expressed constitutively in antigen-presenting cells and is induced by gamma-interferon in other cell types. It has an important role in HLA-II-restricted antigen processing and was reported to be expressed in most of primary and metastatic melanomas (54). Indeed, we also detected GILT in B- and T-cells and in melanoma tissues (Figure 4A, right inset). SEM4D belongs to the semaphorin family and it regulates the sensitivity of the B-cell antigen receptor that is required for proper B-cell homeostasis (55). We observed that SEM4D was mainly presented in B-cells and also in T-cells (Figure 4B). The length distribution of peptides derived from GILT and SEM4D (left insets in Figures 4A,B) reveals that the longer HLA-I peptides were also detected as HLA-II peptides. Therefore, these peptides were mainly HLA-I/II (gray line). Globally, such HLA-I/II hotspots overlapped significantly more often with HLA-II hotspots than with HLA-I hotspots (Figure S8 in Supplementary Material).

Naturally, tissue specificity will further restrict the presentation of the antigens. For example, MITF is a transcription factor that regulates the expression of genes with essential roles in cell differentiation, proliferation and survival (56). MITF plays an important role in melanocyte development by regulating the expression of tyrosinase (TYR) and tyrosinase-related protein 1 (TYRP1) (56). Indeed, we also identified MITF, TYR, and melanocyte protein PMEL ligands almost exclusively in the melanoma tissues (Figure 4C; Figures S7D,E in Supplementary Material).

It is important to note that HLA-binding affinities cannot accurately predict the hotspots we detected for these proteins (Figures 4D–F). Therefore, the immunopeptidomics data provide critical additional information to capture the true in vivo presented ligandome. The height of the hotspots and the distribution of hotspots along the protein sequence reflect the level of its presentation. Hotspots may be related to sequence and structure dependent proteasomal or endosomal cleavage preferences. Alternatively, hotspots may be merely the outcome of de facto presentation of the rather more stable polypeptides surviving the highly proteolytic cytosolic environment as the expected half life of peptides in the cytosol of living cells is 6–10 s (57). Furthermore, some posttranslational modifications may interfere with protein cleavage and with binding of the modified peptides to the HLA, eliminating them from the presented repertoire. Currently, no prediction algorithms incorporate these factors. Since our database comprises of peptidomes of dozens of different HLA allotypes and binding specificities, these hotspots reflect an average propensity of a protein sub-sequence to be presented on a HLA molecule. Hotspots may encompass HLA peptides with N′ or C′ terminal extensions of several amino acids to accommodate peptides that fit a variety of HLA-binding specificities. Hotspots of “type I” proteins are highly enriched in 9–11 HLA-I peptides (Figures S7A,B in Supplementary Material). Similarly, hotspots of “type II” proteins are longer, and the length distribution of their peptides is centered on 15-mer peptides (Figure S7C in Supplementary Material).

MS-based immunopeptidomics is a powerful approach to shed light on the selective sampling of the proteome. Because it captures the actual presented peptidome, it may bypass the need to computationally predict ligands. Therefore, it is also a promising method to directly identify presented neoantigens. However, enough tumor tissue is available only for the minority of patients and only rarely such neoantigens can be detected due to current limitation in sensitivity. Therefore, the widespread approach for identification of neoantigens for personalized cancer vaccines is still based on in silico predictions of the binding affinity to the respective HLA class I allotypes of 9–11-mer peptide sequences harboring the non-synonymous mutations. RNA expression data are further interrogated to exclude non-expressed genes. This in silico approach has high false positive rate and, consequently, the list of proposed targets must be further filtered to enrich for true positives. In addition, as we have just discussed above, some HLA ligands are longer and do not conform to the classical binding mode, hence are expected to be false negatives. All the hints presented above could be incorporated into predictors for neoantigen immunogenicity. Here, we focus on Hints 1 and 3.

We propose that large-scale and in-depth peptidomics MS/MS data of naturally presented ligands can be used to complement existing predictors and to improve HLA-I neoantigen prioritization. We test this concept using the Stronen et al. data (see Materials and Methods for more details). To develop this approach, we first defined three immunopeptidomics MS-based features. The first feature, at the protein level, is the number of HLA-I peptides per protein in our database (nrMatchingPeptides_I). We anticipate that proteins that are highly presented in their wild-type form will have a better chance to present neoantigens once they are mutated. On the peptide level, a match between a predicted peptide and our database can be exact, included or partial (Figure 5A). We expect that if the mutation is in a position in the source protein that is naturally presented, then we have evidence that the neoantigens could also be presented. Obviously, this will not hold if the mutation falls on a HLA-binding site or otherwise strongly weakens HLA binding, but for this proof of concept we will not consider these events. We added two additional features that quantify the overlap of the predicted wild-type (wt) -peptide with other HLA-I peptides in our database, exactMatchScore_I and matchScore_I (Table S3 in Supplementary Material). Please refer to Section Materials and Methods for a more detailed description of these scores. These scores reflect the propensity of the wt-peptide to be presented as a HLA-I peptide. As a simplification, we assume that a mutation does not completely obfuscate this propensity, just as the wt-affinity is still a good predictor for the mutated peptide affinity and immunogenicity.

Applying all predictors to the data published by Stronen et al. (see Materials and Methods), we show that the MS-based features provide valuable information for the prioritization of the neoantigens (Figure 5B). The performance of a single MS-based feature was slightly better than the RNA abundance feature, but lower than the affinity and stability features. Overall, 16 out of 1,034 peptides were immunogenic (1.5%), whereas 3 out of 66 peptides with a positive matchScore I (4.5%) and 1 out of 7 (14.2%) with a positive exactMatchScore I were immunogenic, i.e., the MS-based scores are able to enrich immunogenic peptides. One striking observation was that all 16 immunogenic peptides belonged to proteins that were present in ipMSDB, which is unlikely to happen by chance (872 out of all 1,034 peptides did match ipMSDB, which corresponds to a probability of 0.065 = (872/1,034)¹⁶ that the 16 immunogenic peptides are present in ipMSDB by chance). Therefore, we improve prediction based on binding affinity of mutated peptides if we remove all predicted sequences which do not belong to any protein in our database. Figure 5C shows how many immunogenic peptides are correctly predicted as a function of the rank of the mutAffinity score before (gray line) and after such removal (red line).

Furthermore, we used the Stronen et al. data and the cross-validation scheme outlined in Section Materials and Methods to evaluate how much predictive power the ipMSDB features add in combination with non-MS features. We trained a SVM regression on the set of training peptides in order to rank the test peptides according to their SVM-predicted immunogenicity. We then calculated the number of correctly predicted peptides within the 20 top-ranked peptides (see Materials and Methods for more details). We examined how the incorporation of the MS-based features into the SVM prediction improves the number of correctly predicted peptides. When the non-MS scores were mutAffinity, mutPeptideStability, and rnaExpr, we improved the prediction by 15.9% (group 1). 38.3% prediction improvement was obtained when the non-MS scores were the scores from group 1 plus wtAffinity and wtPeptideStability (group 2). The best improvement of 50.9% was obtained when the non-MS scores were the scores from group 1 plus diffAffinity, diffPeptideStability, which incorporate the differences in the binding affinity and in the binding stability, respectively, of the wt and the mutant peptides (Group 3; where diffAffinity is mutAffinity-wtAffinity and diffPeptideStability is mutPeptideStability-wtPeptideStability) in Figure 5D. These results indicate that our ipMSDB based features are complementary to the affinity based features.

Interesting interactions between the MS-based features and affinity-based features emerged, but given the small size of the dataset the following interpretations are speculative. Figure S9 in Supplementary Material shows that almost all immunogenic peptides (blue dots) had a low mutAffinity score (low values are good), and that MS-based features were able to rescue the two peptides with the higher mutAffinity value (larger orange diamond and red square). Peptides with an exact match to the immunopeptidomics MS database (red points) lied close to the diagonal, i.e., had similar mutAffinity and wtAffinity scores in which the mutation does not seem to change HLA binding. On the other hand, included or partial matches can also identify peptides which had different affinities in their mutational and wild type state. In order to confirm these interactions, refine them or find new ones, many more datasets will be required. Also, more work is needed to extend the prediction to peptides, which are often missed by MS/MS analysis such as very hydrophobic peptides or peptides that fragment very poorly.

One of the advantages of our approach is that it may mitigate the limitation of predicting binding affinity for very rare HLA-I alleles and for HLA-II molecules. We deduced information from hundreds of thousands of ligands detected in tens of different individual donors, representing the overall distribution of HLA allotypes in the human population. As the binding specificities of different HLA allotypes may be redundant, in the future, an exhaustive database will provide a good approximation to the definite immunopeptidome. As we know from many studies, immunogenicity of peptides is not highly correlated with binding affinity (26, 58, 59), and peptides with (measured) low affinity may still induce an immune response, especially upon vaccination. However, the peptides should still be naturally presented on the target cells to induce an effective T-cell response. We envision that the pan-HLA-peptide interaction predictors will provide estimation of the binding affinity, and together with an exhaustive immunopeptidomics database, the prioritization of neoantigens will include hotspot features, which capture other aspects of the natural in vivo presentation.

In this preliminary, proof of concept study we showed how immunopeptidomics database comprising many melanoma tissue samples contained information that enabled prioritization of neoantigens predicted in similar melanoma samples. We anticipate that the same tumor type will have to be adequately represented in the database to overcome and reflect tissue specific expression signatures. The scoring scheme we introduced here may already be implemented providing large in-depth immunopeptidomics data matching the investigated tumor is present. Yet, much more data from T cell based assays of both immunogenic and confirmed non immunogenic neoantigens from multiple patients across different tumor types will be critical to sorely benchmark and optimize this algorithm.