Antigen recognition by CD8+ T cells is governed by the pool of peptide antigens presented on the cell surface in the context of HLA class I complexes. Studies have shown not only a high degree of plasticity in the immunopeptidome, but also that a considerable fraction of all presented peptides is generated through proteasome-mediated splicing of noncontiguous regions of proteins to form novel peptide antigens. Here, we used high-resolution mass spectrometry combined with new bioinformatic approaches to characterize the immunopeptidome of melanoma cells in the presence or absence of IFNγ. In total, we identified more than 60,000 peptides from a single patient-derived cell line (LM-MEL-44) and demonstrated that IFNγ induced changes in the peptidome, with an overlap of only approximately 50% between basal and treated cells. Around 6% to 8% of the peptides were identified as cis-spliced peptides, and 2,213 peptides (1,827 linear and 386 cis-spliced peptides) were derived from known melanoma-associated antigens. These peptide antigens were equally distributed between the constitutive- and IFNγ-induced peptidome. We next examined additional HLA-matched patient-derived cell lines to investigate how frequently these peptides were identified and found that a high proportion of both linear and spliced peptides was conserved between individual patient tumors, drawing on data amassing to more than 100,000 peptide sequences. Several of these peptides showed in vitro immunogenicity across multiple patients with melanoma. These observations highlight the breadth and complexity of the repertoire of immunogenic peptides that can be exploited therapeutically and suggest that spliced peptides are a major class of tumor antigens.
Antigen recognition by cytotoxic T cells and subsequent tumor cell destruction is the key component underlying cancer immunotherapy strategies. Its importance has been widely demonstrated, and loss-of-function of elements in the antigen processing and presentation pathways has been shown to confer therapeutic resistance (1). Correlative findings point to neoantigens arising from tumor mutations as an important source of immunogenic antigens in the context of melanoma and other cancers (2). Nonetheless, tumors with low mutational burdens can respond to checkpoint inhibitor therapy, and the presence of a high tumor mutational load does not necessarily correspond to the efficacy of treatment (3). The HLA class I–bound peptides (p-HLA-I) arising from the mutant proteins are mostly heterogeneously expressed and additionally determined by the patient-specific HLA subtypes, making predictions about their presentation and immunogenicity unreliable. Although a number of studies have reported the utility of mass spectrometry combined with exome sequencing in identifying HLA-presented peptides derived from mutated proteins (4–6), the analysis of the contribution of mutational neoantigens to the overall tumor immunogenicity remains complicated and unresolved.
Against this background, the composition of the immunopeptidome, or the repertoire of HLA-bound peptides presented on the surface of the cell and their contribution to tumor immune recognition, becomes significant. The immunopeptidome is largely shaped by antigen processing through the proteasome complex for subsequent presentation of short peptide epitopes on MHC molecules (7). Several forms of the proteasome complex exist, each with differing enzymatic activities (8). In melanoma cells, the constitutive proteasome is expressed under steady state conditions. The expression of an immunoproteasome, the subtype expressed by dendritic cells and other cells of the immune system, may be induced in tumor cells in a cytokine-dependent manner (9), leading ultimately to changes in the peptides presented to the immune system (10, 11). We have previously demonstrated induction of the immunoproteasome in a range of human melanoma cell lines in the presence of the inflammatory cytokine, IFNγ, in vitro, and in melanoma patient inflamed tumors (characterized by presence of tumor-infiltrating lymphocytes) ex vivo (12). Dependent on the proteasome subtype expressed by the cell, we have shown that a single melanoma antigen (NY-ESO-1) can be processed into several different epitopes. These differences in antigen processing lead to concomitant change in the ability of antigen-specific T cells to target the tumor cell. Thus, the potential for a tumor cell to “look” substantially different to CD8+ T cells, depending on inflammation at the tumor site, arises. Studies conducted by our group and others (13, 14) have shown that a significant proportion of p-HLA-I is not genomically templated and results from posttranslational proteasome splicing (ligation of noncontiguous small polypeptide segments from the same or different proteins). To date, these peptides have been missed in most neoantigen discovery studies due to the lack of appropriate bioinformatics tools (15, 16).
In this study, we used high-resolution mass spectrometry approaches combined with a bioinformatics workflow to identify linear and spliced p-HLA-I presented in the melanoma immunopeptidome in the presence or absence of the cytokine, IFNγ. These included a number of linear and cis-spliced peptides derived from melanoma-associated antigens. A series of identified linear and cis-spliced peptides was tested for in vitro immunogenicity across multiple patients with melanoma and healthy donors. Patients with melanoma demonstrated immune responses to peptide pools identified from both treatment conditions (+/– IFNγ), whereas T-lymphocyte responses to pools of IFNγ-upregulated peptides were not observed in healthy donors. We also demonstrated that cis-spliced peptides were widely presented by melanoma cells and were immunogenic in multiple donors. These findings have significant implications for cancer immunotherapy, as well as for fundamental questions such as induction of immune tolerance, T-cell repertoires, and immune recognition.
Materials and Methods
Human ethics approval
Blood samples used in this study were derived from 6 patients with melanoma who provided written informed consent to participate in a clinical research protocol approved by the Austin Health Human Research Ethics Committee (HREC/14/Austin/425 and H2012–04446) and in accordance to the world medical association Declaration of Helsinki and the Australian Code for the Responsible Conduct of Research. Blood samples were collected during routine visits at the Austin Hospital (Heidelberg, Victoria, Australia) for melanoma surgery or systemic treatment, and peripheral blood mononuclear cells (PBMC) were stored in liquid nitrogen with no special requirements for inclusion, despite melanoma diagnosis and informed patient consent. Blood samples from four healthy donors were provided by the Red Cross Blood Service Australia under 16–12VIC-08.
Melanoma cell line culture
Establishment and characterization of LM-MEL-44, 33, and 53 melanoma cell lines used have been described previously (17, 18). Cell lines were derived in-house from freshly excised melanoma tumors in 2000 (LM-Mel-33), 2004 (LM-MEL-44), and 2006, respectively. Master stocks for all lines were frozen after Mycoplasma testing at passage 10. Cell lines tested negative for Mycoplasma and were reauthenticated by short tandem repeat profiling prior (<10 months) to the experiments. For each replicate, cells were cultured for approximately 28 days in RF10 consisting of RPMI1640, 2 mmol/L GlutaMAX, penicillin (100 IU/mL), streptomycin (100 μg/mL), and 10% heat-inactivated FCS (all Invitrogen). For induction of immunoproteasome catalytic subunits, cells were incubated with IFNγ (100 ng/mL, PeproTech) for 72 hours prior to experiments.
Melanoma cell line sequencing
Whole-exome sequencing of the LM-MEL-44 cell line was performed using the NimbleGen EZ Exome Library v2.0 Kit and run on a Illumina Hiseq2000 Instrument as described previously (19). Sequence reads were aligned to the human genome (hg19 assembly) using the Burrows–Wheeler Aligner program (20). Single-nucleotide variants (SNV) and indels were identified using the GATK Unified Genotyper (21), Somatic Indel Detector (22), and MuTect (Broad Institute, Cambridge, MA; ref. 23).
Isolation of peptides bound to HLA class I molecules
HLA class I peptides were eluted from LM-MEL-44, 33, and 53 cells (prior to or after treatment with IFNγ) as described previously (24–27). In brief, for replicate 1 of LM-MEL-44, 3 × 109 cells were lysed in 0.5% IGEPAL (Sigma), 50 mmol/L Tris-HCl pH 8.0, and 150 mmol/L NaCl supplemented with protease inhibitors diluted to 1 × strength (CompleteProtease Inhibitor Cocktail Tablet; Roche Molecular Biochemicals) for 45 minutes at 4°C. Lysates were cleared by ultracentrifugation at 40,000 x g, and HLA class I complexes were immunoaffinity purified using Poly-Prep Chromatography Columns (Bio-Rad Laboratories) and 10 mg DT9 (anti HLA-C) and W6/32 (pan anti–HLA-I) mAbs (grown and purified in-house in Purcell Laboratory). Wash buffer 1 was 0.005% (v/v) IGEPAL CA-630; 50 mmol/L Tris, pH 8.0; 150 mmol/L NaCl; 5 mmol/L EDTA; 100 μmol/L PMSF; and pepstatin A (1 μg/mL). Wash buffer 2 was 50 mmol/L Tris, pH 8.0 and 150 mmol/L NaCl. Wash buffer 3 was 50 mmol/L Tris, pH 8.0 and 450 mmol/L NaCl. Wash buffer 4 was 50 mmol/L Tris, pH 8.0. Elution buffer was 10% (v/v) acetic acid. For replicates 2 and 3 of LM-MEL-44 and also LM-MEL-33 and LM-MEL-53, 5 × 108 cells (for each sample) were lysed in 0.5% IGEPAL; 50 mmol/L Tris-HCl, pH 8.0; and 150 mmol/L NaCl supplemented with protease inhibitors for 45 minutes at 4°C. Lysates were cleared by ultracentrifugation at 40,000 × g, and HLA class I complexes were immunoaffinity purified using the W6/32 mAb as described.
Fractionation of HLA-bound peptides by reversed-phase high-performance liquid chromatography
The HLA peptide eluates were loaded onto a 4.6 mm internal diameter × 50 mm monolithic C18 Reversed-phase High-performance Liquid Chromatography Column (Chromolith Speed Rod; Merck) at a flow rate of 1 mL/minute using an EttanLC HPLC System (GE Healthcare) with buffer A (0.1% trifluoroacetic acid) and buffer B [80% Acetonitrile (ACN)/0.1% trifluoroacetic acid] as mobile phases. The bound peptides were separated from the class I heavy chains and β2m molecules using increasing concentrations of buffer B, from 0% to 80%. Peptide-containing fractions (500 μL) were collected, vacuum concentrated to approximately 5 μL, combined into nine pools, and reconstituted to 12 μL with 0.1% formic acid. Indexed retention time (iRT) peptides (28) were spiked in for retention time alignment.
Identification of HLA-bound peptides using data-dependent acquisition
For the first replicate of LM-MEL-44, we used a Dionex UltiMate 3000 RSLCnano System equipped with a Dionex UltiMate 3000 RS Autosampler. The samples were loaded via an Acclaim PepMap 100 Trap Column (100 μm × 2 cm, nanoViper, C18, 5 μm, 100 Å; Thermo Fisher Scientific) onto an Acclaim PepMap RSLC Analytical Column (75 μm × 50 cm, nanoViper, C18, 2 μm, 100 Å; Thermo Fisher Scientific). The peptides were separated by increasing concentrations of 80% ACN/0.1% formic acid at a flow of 250 nL/minute for 65 minutes and analyzed with a QExactive Plus Mass Spectrometer (Thermo Fisher Scientific). In each cycle, a full ms1 scan [resolution, 70.000; automatic gain control (AGC) target, 3e6; maximum IT, 120 milliseconds; and scan range, 375–1,800 m/z] preceded up to 12 subsequent ms2 scans (resolution, 17.500; AGC target, 1e5; maximum IT, 120 milliseconds; isolation window, 1.8 m/z; scan range, 200–2,000 m/z; and NCE, 27). To minimize repeated sequencing of the same peptides, dynamic exclusion was set to 15 seconds, and the “exclude isotopes” option was activated.
For the second and third replicates of LM-MEL-44, LM-MEL-33, and LM-MEL53, we used a Dionex UltiMate 3000 RSLCnano System equipped with a Dionex UltiMate 3000 RS Autosampler. The samples were loaded via an Acclaim PepMap 100 Trap Column (100 μm × 2 cm, nanoViper, C18, 5 μm, 100 Å; Thermo Fisher Scientific) onto an Acclaim PepMap RSLC Analytical Column (75 μm × 50 cm, nanoViper, C18, 2 μm, 100 Å; Thermo Fisher Scientific). The peptides were separated by increasing concentrations of 80% ACN/0.1% formic acid at a flow of 250 nL/minute for 158 minutes and analyzed with an Orbitrap Fusion Tribrid Mass Spectrometer (Thermo Fisher Scientific). Six microliters of each sample fraction were loaded onto the trap column at a flow rate of 15 μL/minute.
Orbitrap Fusion Tribrid Mass Spectrometer (Thermo Fisher Scientific) was set to data-dependent acquisition (DDA) mode with the following settings: all MS spectra (MS1) profiles were recorded from full ion scan mode 375 to 1,800 m/z in the Orbitrap at 120,000 resolution with AGC target of 400,000 and dynamic exclusion of 15 seconds. The top 12 precursor ions were selected using top speed mode at a cycle time of 2 seconds. For tandem mass spectrometry (MS-MS), a decision tree was made that helped in selecting peptides of charge state 1 and 2 to 6 separately. For single charged analytes, only ions falling within the range of m/z 800 to 1,800 were selected. For +2 to +6 charge states, no such parameter was set. The c-trap was loaded with a target of 200,000 ions with an accumulation time of 120 milliseconds and isolation width of 1.2 amu. Normalized collision energy was set to 32 (high energy collisional dissociation) and fragments were analyzed in the Orbitrap at 30,000 resolution.
DDA data analysis
Linear and cis-spliced peptide sequences were identified as described previously (14). In brief, the acquired .raw files from six LM-MEL-44 lines were searched with PEAKS Studio X (Bioinformatics Solutions Inc.) against the human UniProtKB/SwissProt (v2017_10) database, which was manually corrected for the SNVs characteristic of the LM-MEL-44 cell line as identified by whole-exome sequencing. The parent mass error tolerance was set to 10 ppm for de novo sequencing and database search, and the fragment mass error tolerance to 0.02 Da for both searches. Oxidation of M and deamidation of N and Q were set as variable modifications, and an FDR cutoff of 1% was applied. High confidence de novo peptide sequences without any linear peptide match in the provided database were further interrogated with the “Hybrid finder” algorithm (14), and the identified cis-spliced candidate sequences from all six samples were combined together and added back to the original UniProtKB/SwissProt database and all data were researched using PEAKS DB. Linear and cis-spliced peptides in this search were extracted at 5% FDR to create the final list of identified peptides (Supplementary Table S1). For identification of linear and cis-spliced peptides from LM-MEL-33 and 53, we used this combined database and the same setting on PEAKS studio.
We used the NetMHC4 (29, 30) algorithms for binding predictions for both spliced and linear peptides. A default rank cutoff of 2 was implemented as a binder peptide (Supplementary Table S2).
We used the Immune Epitope Database and Analysis Resource (www.iedb.org), Class I Immunogenicity algorithm for immunogenicity predictions of linear peptides (31). The peptides used in functional assays indicated as “predicted immunogenic” had an immunogenicity score between 0.37 and 0.55 (Supplementary Table S3).
T-cell stimulation assay
To assess T-cell responses, a panel of peptides derived from cancer/melanoma antigens was synthesized (Mimotopes; Supplementary Tables S4 and S5). PBMCs from healthy donors (Australian Red Cross Lifeblood) or patients with melanoma were purified by density centrifugation over Ficoll Hi-Paque (GE Healthcare). Cells were cultured in TCRPMI consisting of RPMI1640, 2 mmol/L GlutaMAX, penicillin (100 IU/mL), streptomycin (100 μg/mL), 20 mmol/L HEPES, 1% nonessential amino acids, 1 mmol/L sodium pyruvate, 55 μmol/L β-mercaptoethanol (all Invitrogen), and 10% Human AB Serum (Australian Red Cross). Peptides were combined into pools of five to nine peptides. PBMCs (106 cells/mL) were incubated with 10 μmol/L of each peptide in pools for 10 to 12 days at 37°C. IL2 (100 IU/mL, PeproTech) was added and replaced every 3 days. DMSO (Sigma) served as negative control and FEC (a mixture of Flu, EBV, and CMV T-cell epitope peptides) served as positive control.
Statistical analyses were performed using GraphPad Prism. For peptide responses, a repeated measures two-way ANOVA was used, and for cis-spliced peptide assays, a matched-mixed effect analysis with Dunnett multiple comparison test was used to compare each column with DMSO.
Intracellular cytokine staining of antigen-activated T lymphocytes
To assess antigen responses, T lymphocytes were restimulated (following 10–12 days incubation as outlined above) with peptide pools (10 μmol/L of each peptide) for 4 to 8 hours in TCRPMI in the presence of brefeldin A (10 μg/mL; BFA, Golgi plug, BD Biosciences). Cells were washed with PBS (Invitrogen), labeled with live/dead fixable Violet Stain (Invitrogen), and then incubated with antibodies against CD3 (SK7) and CD8 (SK1; all BD biosciences) for 15 minutes at 4°C. Samples were washed, fixed for 20 minutes at 4°C, and permeabilized (Cytofix/Cytoperm Kit, BD Biosciences), followed by staining with anti-TNFα (Mab11; Thermo Fisher Scientific) for 25 minutes at 4°C. Intracellular cytokine staining (ICS) for spliced peptides was performed accordingly, but the Foxp3 Transcription Factor Staining Kit (Thermo Fisher Scientific) was used according to the manufacturer's protocol for fixation/permeabilization of T cells prior to cytokine staining. The gating strategy was: SSC/FSC; Singlets; SSC/LD−; CD3+CD8+; CD8+TNFα+. Data were acquired on a FACSCanto (BD Biosciences) and analyzed with FlowJo Software (version 10, FlowJo). To account for the variation in DMSO background CD8+ T-cell activity across multiple donors, signals were normalized by subtracting the background from the DMSO control–treated samples in each case.
HLA-A2 stabilization assays
The binding activity of the peptides was assayed by measuring peptide-induced stabilization of HLA-A2 on TAP-deficient T2 cells by flow cytometry. T2 cells were purchased from the ATCC (ATCC CRL-1992) and cultured in RF-10 [RPMI with 10% FBS, 5% Glutamine, and 5% Penicillin–Streptomycin (all Invitrogen)] in T25 flasks for 2 to 3 days before the assay. Cells were Mycoplasma tested prior to use and expanded from frozen stocks stored within 6 months of purchase. T2 cells (2 × 105 cells/well) were cultured for 16 hours at 37°C in 5% CO2 in 200 μL RF10 in 96-well U-bottomed plates in presence or absence of synthetic peptides (10 μg/mL). Peptides from Melan-A (modified aa 26–35 ELAGIGILTV and aa 60–72) served as positive or negative controls, respectively. All peptides were tested in triplicate. After 16 hours of stimulation, the cells were washed and stained with HLA-A2 mAb, BB7.2 (BioLegend), for 30 minutes at 4°C. Cells were subsequently stained with Fixable Viability Kit (Zombie NIR, BioLegend) for 15 minutes at 4°C before flow cytometry on a FACSCanto (BD Biosciences). Data were analyzed using FlowJo Software (version 10, FlowJo).
Validation of spliced peptides using retrospectively synthesized peptides and retention time prediction
We validated the identity of a panel of spliced peptides (including peptides that were tested for immunogenicity) using 28 synthetic peptides (Mimotopes) by comparing chromatographic retention and MS-MS spectra with the original p-HLA-I. The PKL files of the synthetic peptide and corresponding eluted peptides were exported from PEAKS X studio software. For evaluating the similarity between two spectra, we predicted all b- and y-ions for each sequence by using PEAKS studio, and then extracted the corresponding intensity for each ion (with a fragment mass error tolerance of 0.1 Da). The Pearson correlation coefficient and the corresponding P value (Prism version 8.01, GraphPad) between the log10 intensities of identified b- and y-ions in the synthetic and sample-derived spectra were calculated (32). The closer a correlation coefficient to 1, the more identical the spectra. All tested peptides were found to have a P value of less than 0.05. We spiked a standard set of reference peptides (iRT Kit; Biognosys) into all samples. A linear regression equation was calculated on the basis of the retention time and hydrophobicity of each reference peptide. By using this equation, the iRT indices of each synthetic peptide and corresponding eluted peptide were calculated (28). We also used the GPTime tool to compare the predicted versus actual chromatographic retention time of both the identified linear and spliced peptides. For each dataset generated from LM-MEL-44 cell lines (six replicates in total), we sorted peptides (8–12mer peptides without modification) on the basis of the –logP score (high to low) from the PEAKS Studio Software. We used the first 1,000 peptides for training the algorithm, and then used the trained algorithm to predict the retention times of all linear and spliced peptides in the corresponding dataset (32). We also predicted the hydrophobicity of all identified 8–12mer peptides by using Grand Average of Hydropathy Score equation (33).
Quality control of biological replicates
The LC/MS-MS instrument and antibody used for immunoprecipitation of p-HLA-I for replicate one of LM-Mel-44 cell were not identical with replicates (rep2) 2 and (rep3) 3. To assess the effect of this difference on our conclusions, we compared the overlap in peptides identified in replicate 2 and 3 (for which we used the same methodology). First, we compared biological replicates (rep2 vs. rep3 untreated and rep2 vs. rep3 IFNγ treated). In both the comparisons, >70% of peptides were the same. We then compared identified peptides in the presence and absence of IFNγ in each replicate and found in both replicates, >60% of peptides present after IFNγ were not detected in the untreated sample. This aligned with our conclusion (observation when we compared all three replicates) that IFNγ treatment significantly diversified and expanded the HLA-I immunopeptidome.
The Cancer Genome Atlas analysis
Investigation of an association between proteasome subunit gene expression and overall survival was made using patient data from The Cancer Genome Atlas (TCGA) program, hosted by the NCI (https://portal.gdc.cancer.gov/). Within this program, the TCGA-SKCM Project (dbGaP study accession no. phs000178), incorporating 470 cases of disease in skin and nevi, was used for the analysis. Survival time and RSEM RNASeqV2 read counts were downloaded for each of the proteasome subunits PSMB5, 6, 7, 8, 9, and 10, and for CD3g. The top quartile of patients expressing CD3g (read count cutoff at 69.21) was removed from each of the proteasome datasets. The remainder of samples (n = 342) in each group were plotted on a Kaplan–Meier curve comparing survival time between the top quartile (n = 85) and bottom quartile (n = 85) of patients expressing each gene, read count cutoffs were as follows: PSMB5 top quartile >3,235.6, bottom quartile <1,899.8; PSMB6 top quartile >2,271.86, bottom quartile <1,389.66; PSMB7 top quartile >3,856.04, bottom quartile <2,001.48; PSMB8 top quartile >3,775.6; bottom quartile <1,370.3; PSMB9 top quartile >1,572.34, bottom quartile <325.36; and PSMB10 top quartile >2,111.3, bottom quartile <701.38. Kaplan–Meir curves were generated using online tools at https://astatsa.com/LogRankTest/, which also performed a log-rank test for significance in survival differences between the groups. A P < 0.05 in this test was considered significant.
The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE (34) partner repository with the dataset identifier PXD014397 and 10.6019/PXD014397. The exome sequencing data have been deposited to the NCBI Sequence Read Archive database with the accession code PRJNA655753.
The melanoma immunopeptidome is composed of linear and spliced peptides
We established a comprehensive repository of HLA class I peptide ligands presented by a patient-derived melanoma cell line (LM-MEL-44) utilizing either the constitutive proteasome (–IFNγ) or the immunoproteasome (+IFNγ). Using three biological replicates for each condition, we identified around 60,000 peptides presented across all HLA class I allotypes expressed by LM-MEL-44 cells (Supplementary Table S1). After several steps of validation and quality control (Supplementary Figs. S1–S4), approximately 6% to 8% of the peptides in each sample were conservatively assigned as cis-spliced in origin (Fig. 1A; Supplementary Table S1), being derived from noncontiguous sequences of the same protein. This proportion of peptides of cis-spliced origin is in agreement with previous studies (14, 35, 36). As is expected for HLA class I epitopes, the majority of peptides were nine amino acids in length with no apparent difference between linear and cis-spliced sequences (Fig. 1B).
Using the NetMHC 4.0 binding prediction algorithm, more than 80% of the linear peptides were assigned to at least one of the HLA class I alleles expressed on the surface of the LM-MEL-44 cells (HLA-A*02:01, B*40:01/*44:02, and C*03:04/*05:01; ref. 37), suggesting that the majority of the identified peptide sequences could be considered genuine HLA class I ligands (Fig. 1C; Supplementary Table S2). The percentage of cis-spliced peptides predicted to bind to HLA-A or HLA-C molecules was found to be comparable with linear epitopes, but substantially fewer cis-spliced peptides were predicted to bind to HLA-B*40:01 and HLA-B*44:02, suggesting that LM-MEL-44 cells generated a lower number of cis-spliced peptides that conformed to the consensus-binding motif of these HLA-B allotypes (Fig. 1C). A significantly higher percentage of unassigned sequences was observed among the cis-spliced epitopes, which is in agreement with previous reports (13, 14) and can be attributed to the fact that binding algorithms, such as NetMHC 4.0, are exclusively trained on linear peptide sequences. We did not identify in our entire dataset any of the mutational neoantigens that have been described for the LM-MEL-44 cell line based on exome sequencing data (38). However, this was not surprising because this cell line had a relatively low mutational load (Supplementary Table S3).
The generation of spliced peptides is not a random process
To address whether spliced peptides were randomly generated, we comparatively analyzed the overlap of both linear and spliced peptides across our three biological replicates. A total of 1,399 and 1,795 cis-spliced peptides were identified in at least two of the three biological replicates for the untreated and IFNγ-treated samples (Fig. 2A), which suggested that the generation of cis-spliced peptides was not a random process. A similar overlap was observed for the linear epitopes. The comparatively low overlap between the replicates could be attributed to the stochastic nature of DDA mass spectrometry (DDA-MS), which is particularly pronounced when acquiring complex samples that contain individual analytes of low abundance (such as HLA peptide samples).
To investigate whether the identified cis-spliced peptides were also expressed on other cell lines with a similar HLA signature, we analyzed the cell lines LM-MEL-53 and LM-MEL-33 by DDA-MS. LM-MEL-53 cells were derived from the same patient as LM-MEL-44 cells, but from another metastasis at a different point in time (18). In contrast, LM-MEL-33 (HLA-A*02:01/A*03:01, B*40:02/*47, and C*03:04/*06:02) cells were isolated from a different patient that shares three HLA alleles with LM-MEL-44 (18). A total of 47% and 28% of the identified cis-spliced peptides from LM-MEL-44 were also identified on LM-MEL-53 and LM-MEL-33 cells, respectively (Fig. 2B; Supplementary Table S1), which further confirmed that the generation of cis-spliced peptides was not a random process and indicated that cis-spliced peptides are potential targets for cancer immunotherapy.
IFNγ treatment alters the melanoma HLA class I immunopeptidome
Considering the well-described clinical relevance of so called “hot” versus “cold” tumor microenvironments and previous work demonstrating the influence of cytokine exposure on antigen presentation pathways (10), we wanted to examine the impact of IFNγ exposure on the immunopeptidome. IFNγ treatment led to the identification of a significantly higher number of HLA epitopes than that from untreated cells, consistent with the upregulation of HLA molecules at the cell surface (Fig. 3A; Supplementary Fig. S5; ref. 10). For peptides that were identified in all six samples of LM-MEL-44 cell lines, we observed, on average, an increase in peptide abundance in the cytokine-treated samples. However, only 44.7% of the linear and 52.5% of the spliced peptides were identified under both conditions, suggesting that the addition of IFNγ significantly impacted the composition of the immunopeptidome.
To understand whether IFNγ exposure changed the abundance of individual peptides, independent of HLA expression, we identified species that were present across all replicates (a total of 4,942 peptides) and calculated their log2 fold change between IFNγ-treated and untreated samples after median normalization of their ms1 intensities to remove any bias introduced through varying IFNγ-induced HLA expression (Fig. 3B; Supplementary Table S1). A considerable number of epitopes changed in abundance by a factor of at least 2 (both up and down), confirming that the addition of IFNγ altered HLA class I presentation, while not affecting the proportion of presented cis-spliced epitopes on LM-MEL-44 melanoma cells. Despite median normalized ms1 intensities, most of the peptides predicted to bind to HLA-B*40:01 and HLA-B*44:02 were still upregulated upon IFNγ exposure, which correlated to the enhanced upregulation of HLA-B molecules in response to IFNγ compared with HLA-A and HLA-C molecules (10).
Identification of cancer-specific peptides in the melanoma immunopeptidome
We identified a total of 2,213 peptides in our dataset (1,827 linear and 386 spliced peptides) derived from 142 different melanoma-associated antigens (MAA; refs. 39–43; Fig. 4A; Supplementary Table S4). A large proportion (∼45%) of linear peptides has not been reported previously (Fig. 4B; ref. 41). Almost all of the peptides generated by splicing constitute potential epitopes. Of the previously reported epitopes, the majority were detected in both the presence and absence of IFNγ, whereas >32% of spliced peptides were exclusive to IFNγ-treated samples, demonstrating the importance of experimental conditions for epitope discovery (Fig. 4C).
Patients expressing immunoproteasome genes have a survival advantage
Tumor recognition in vivo relies on the processing and generation of cognate peptides within the tumor cells. We mined gene expression data generated by TCGA research network (http://cancergenome.nih.gov/; ref. 44) for correlation of both immuno- and constitutive proteasome–specific genes with survival in patients with melanoma. We found that expression of all three immunoproteasome-specific subunits was significantly associated with increased survival in patients with melanoma. Conversely, constitutive proteasome–specific subunits were associated with decreased melanoma patient survival (Supplementary Fig. S6).
Because immunoproteasome subunits are also expressed by immune cells, including intratumoral T cells that associated with better prognosis, we removed the top quartile of samples with the highest CD3 expression. Following removal of these samples, we found that a significant survival benefit, associated with two of three immunoproteasome subunits, was maintained (Fig. 5). Presence of the three constitutive proteasome subunits was associated with a trend toward decreased survival, indicating that patients whose tumors expressed an immunoproteasome had a survival benefit, which specifically associated with this proteasome type. This effect persisted, albeit to a lesser extent, when we removed two tumors that showed the highest CD45+ infiltration, thus including non–T-cell lineage immune cells and antigen-presenting cells (Fig. 5). Because immunoproteasome expression in tumors is largely driven by cytokine exposure, it remains unclear whether this was merely a footprint of a (previous) successful immune recognition or whether it was part of the preconditioning to allow for such an immune response.
CD8+ T cells frequently recognize linear melanoma-specific epitopes
In this study, we identified several linear melanoma-specific peptides predicted to bind HLA allotypes presented by a tumor-derived melanoma cell line (HLA-A*02:01, B*40:01/*44:02, and C*03:04/*05:01). This included HLA-A*02:01, one of the most prevalent HLA types, and, therefore, a common target for peptide identification and therapeutic focus. We addressed functional immunogenicity of selected peptides by using them to stimulate CD8+ T lymphocytes in PBMCs derived from healthy donors or patients with melanoma (Fig. 6; Supplementary Fig. S7; Supplementary Table S5). In doing these studies, selected donors were matched for at least two HLA allotypes (across HLA-A/B/C), and 3 patients with melanoma were matched across all three. Of note, the LM-MEL-44 cell line was derived from melanoma patient 2, and melanoma patient 6 shared none of the HLA alleles from this cell line, serving as a negative control. We also assessed differences in functional immunogenicity of immune or constitutive proteasome–processed epitopes. This was done by pooling a selection of the most up- or downregulated peptides following IFNγ treatment (Fig. 6A and B; Supplementary Table S5). Although, also observed in healthy donors, we found that CD8+ T-lymphocyte responses to the peptides were more frequently seen in patients with melanoma (Fig. 6A and B, individual and combined donors, respectively). Peptides derived from 15 of the melanoma antigens identified in our screen stimulated specific CD8+ T-lymphocyte responses (more than 2% TNFα+ cells) in three or more donors, demonstrating functional melanoma T-cell epitopes (Fig. 6A; Supplementary Tables S5 and S6; representative examples; Supplementary Fig. S8A). Of those where a clear HLA-binding prediction could be determined, 54.5% (n = 6) were predicted to bind to HLA-A*02:01, 36.4% (n = 4) to HLA-B*44:02, and 9.1% (n = 1) to HLA-C*05:01. The strongest responses were induced by the peptides derived from SART1 (U4/U6.U5 tri-snRNP–associated protein 1) and PGK1 (phosphoglycerate kinase 1), both of which stimulated responses in 2% to 7% of T lymphocytes from 4 patients with melanoma. Both peptides were predicted to bind to HLA-A*02:01. When the selection of peptides was pooled in groups of those up- or downregulated or unchanged following IFNγ treatment, no appreciable difference in functional immunogenicity in patients with melanoma was observed between groups. One pool in each group was made on the basis of higher in silico–predicted immunogenicity (www.iedb.org; ref. 45; Fig. 6A and B, asterisks). However, these groups did not display enhanced ability to activate CD8+ T lymphocytes in either patients with melanoma or healthy donors.
Spliced peptides are immunogenic and represent targets for immunotherapy
The potential implications of the presence of spliced peptides for all facets of immunity have sparked intense discussions in the past 4 years (14, 46–48). In cancer, their presence widens the repertoire of potentially targetable epitopes and may allow for many more tumor-specific antigens (including mutational-derived neoantigens) being presented in various HLA contexts (16). So far, only six immunogenic cis-spliced HLA-I–bound peptides derived from four different proteins (46) have been described, and most have been discovered by T-cell assays rather than by mass spectrometry (49–54). To test some of the identified spliced peptides for their ability to activate CD8+ T cells in vitro, we synthesized 26 cis-spliced peptides on the basis of (i) their de novo sequencing confidence score, (ii) their binding prediction score for HLA alleles expressed on LM-MEL-44 cells (HLA-A*02:01, HLA-B*40:01, HLA-B*44:02, or HLA-C*05:01), and (iii) the quality of their peptide spectrum matches. When employed as pools of eight to nine individual peptides, all three pools evoked immune responses, as measured by intracellular TNFα production in CD8+ T cells (Fig. 7A) in multiple melanoma patient– and healthy donor–derived PBMCs (example shown in Supplementary Fig. S8B).
Given the differences in the potential to stimulate HLA-A2–positive versus –negative patient and healthy donor samples, most of the immunogenic peptides derived from pools I and III in our assays were HLA-A2 associated. To identify specific immunogenic peptides, PBMCs were stimulated with the listed peptide pools (Fig. 7A) for 10 to 12 days followed by single peptide restimulation. Six of 26 peptides induced a TNFα response above DMSO background (Fig. 7B) in more than one patient sample. The peptide demonstrating the highest immunogenicity (1832, shown as an example in Supplementary Fig. S8C) was a spliced peptide derived from the cancer testis antigen, MAGE-C2 (LILGLLTKV), and showed CD8+ T-cell activation across all 4 patients. Matched-mixed effect analysis showed significant differences across peptides, and peptide 1832 and FEC represented the treatments with significant differences to DMSO. However, the other shown peptides displayed higher immunogenic potential compared with their respective DMSO controls, but with high patient-to-patient variability. Peptide 1832 was identified across all replicates of LM-MEL-44, -33, and -53 (Supplementary Table S6). All spliced peptides that tested positive for immunogenicity were subjected to T2 peptide-binding assays to examine HLA-A2 binding. These peptides all stabilized HLA-A2, albeit to a lesser extent than the well-described modified ELAGIGILTV HLA-A2 peptide (aa 26–35) from the melanoma antigen, Melan-A (55), with some just showing minor stabilization (Supplementary Fig. S9). We did not find any particular pattern in the length of the N- and C-terminal segments of these spliced peptides nor in the distance between these segments on the protein level (Supplementary Fig. S10). Taken together, these data show that these spliced peptides could serve as bona fide anticancer targets and provide a large number of additional targets that would have not been considered using previous mass spectrometry–based epitope discovery strategies.
In this study, we described a detailed and in-depth immunopeptidome analysis on a patient-derived melanoma cell line (LM-MEL-44) generated from a lymph node metastasis. Our qualitative assessment of the immunopeptidome yielded around 60,000 high-confidence peptide identifications that encompassed two culture conditions (+/– IFNγ) to gain insights into the influence of differences in the microenvironment of the cells on the global immunopeptidome. We demonstrated consistence of more than 50% of these peptides with a temporally distinct autologous tumor sample and 37% with a tumor sample from a different donor. The well-described effect of IFNγ in mediating changes to the composition of the antigen-processing machinery, coupled with reports of differences in antigen processing between the constitutive and the immunoproteasome (12), led us to expect a degree of difference between the two immunopeptidomes. Nevertheless, we found that approximately 55% of linear and 47% of spliced HLA class I epitopes were exclusive to either IFNγ-treated or untreated conditions. Our observations are also consistent with studies in ovarian and lung cancer (10, 11) that demonstrate profound changes between cytokine treatment conditions.
To have a closer look at the tumor-specific immunopeptidome landscape, we focused on MAA-derived peptides. More than 50% of the previously unreported peptides that we identified were exclusively presented in the presence of IFNγ. Of the MAA epitopes that have been described previously in other studies (45, 56, 57), only 27.7% were present uniquely in IFNγ-treated conditions. This observation suggests that many immunoproteasome-processed epitopes have not yet been described because traditional approaches to identify tumor-associated antigens have largely been undertaken using cells lines under steady state conditions (i.e., express only the constitutive form of the proteasome). It is evident from our study that the steady state immunopeptidome may vary from the in vivo tumor scenario depending on the tumor microenvironment at any given time. Although our functional studies did not reveal a difference in the immunogenicity of peptides derived from either IFNγ-treated or untreated conditions, in the in vivo setting, a T-cell response to IFNγ-related epitopes is likely to be aided by correlative IFNγ influences, such as upregulation of surface HLA (58). The potential for tumor escape from CD8+ T-lymphocyte killing due to whole-scale changes to the immunopeptidome upon initiation of an antitumor response, and corresponding induction of IFNγ, is clear from our studies. These data become particularly significant in the context of studies demonstrating that tumors with an IFNγ-inflamed, or “hot” microenvironment, are associated with better prognosis and are more likely to be amenable to treatment with immune checkpoint inhibitors (59). It seems conceivable that in vivo, the difference between immunopeptidomes is indeed of immunologic relevance to disease progression and overall patient prognosis. Taken together, it is tempting to speculate that antigens processed via the immunoproteasome may represent an untapped resource of IFNγ-associated neoepitopes.
This plasticity in the peptide landscape of melanoma is further increased by the presence of spliced peptides. The identification of spliced peptides as tumor antigens in cancer was first described in 2004 in both the FGF5 protein in renal cancer (49) and the gp100 protein in melanoma (50), and since then, only four more cis-spliced peptides have been described in cancer (46). Of these six spliced peptides, three have been shown to be processed exclusively by the constitutive proteasome, and two by both the constitutive and immunoproteasome (one remains undetermined; ref. 46).
Several bioinformatics tools are now available to reliably identify spliced peptides (13, 14, 35, 36, 60). Nevertheless, the contribution of spliced peptides to the overall immunopeptidome has been reported in a range from 2% to 40% and is still heavily debated. Studies have identified cis-spliced peptides in the cancer and viral infection context (16, 35, 60), but few have provided experimental evidence of their immunogenicity (41). In this study, we identified 386 cis-spliced peptides that were potentially derived from MAAs and, therefore, considered as potential candidates to induce therapeutic immune responses. We demonstrated that generation and presentation of spliced peptides are not random processes because within three biological replicates, we found comparable reproducibility of both linear and spliced peptides. More than 50% of spliced peptides identified in the LM-MEL-44 cell line were present on at least one of the other two distinct cell lines (LM-MEL-53 and 33). We also demonstrated the immunogenicity of six cis-spliced epitopes tested across multiple patients, strengthening the argument that cis-splicing is a random, functional process leading to diversification of the antigenic pool of peptides. How far increased potential to evoke CD8+ T-cell activation reflects meaningful and translatable antitumor effects remains to be tested in much larger patient cohorts with additional clinical data and in a more formal setting.
This study and other publications, which focus on the identification of spliced peptides (16, 61, 62) and their impact on the plasticity of the immunopeptidome, will beyond doubt open up new questions and opportunities in the field. These will range from the basic understanding of immune tolerance, autoimmunity, and thymic selection to opportunities for development of novel peptide-based therapeutics. This includes vaccines in an infectious and cancer setting where predictability and HLA-binding characteristics of linear and constitutive proteasome–derived peptides are potentially limiting factors.
Disclosure of Potential Conflicts of Interest
N.P. Croft is a specialist scientific advisory board member for Evaxion Biotech outside the submitted work and has a patent (22153EP00, “Method for Identifying MHC Binding Peptides Using Mass Spectrometry”) pending for a new method for predicting immunogenicity using peptidomics datasets. C. Li reports grants and personal fees from National Health and Medical Research Council of Australia (NHMRC early career fellowship) during the conduct of the study. J.S. Cebon reports an advisory board role with Merck and an advisory board role with and clinical trial support from Bristol-Myers Squibb outside the submitted work. A.W. Purcell reports grants from National Health and Medical Research Council of Australia (project grants 1007381 and 1165490 as well as a personal fellowship) that were active during the conduct of the study; grant funds for unrelated projects from Janssen Pharmaceuticals and Cue Biopharmaceuticals; an appointment as a member of the scientific advisory board for Evaxion Biotech outside the submitted work; and provisional patents 2019903421 (“Methods of Identifying MHC-Bound Peptides” for methods associated with ultrasensitive detection of HLA-bound peptides) and 22153EP00 (“Method for Identifying MHC Binding Peptides Using Mass Spectrometry” for a new method for predicting immunogenicity using peptidomics datasets) pending. A. Behren reports grants from La Trobe University (RFA UD grant GR 2000002970) and Department of Health and Human Services acting through the Victorian Cancer Agency (Mid-Career Fellowship) during the conduct of the study. No potential conflicts of interest were disclosed by the other authors.
P. Faridi: Conceptualization, data curation, formal analysis, supervision, validation, investigation, methodology, writing–original draft, writing–review and editing. K. Woods: Conceptualization, investigation, methodology, writing–original draft, writing–review and editing. S. Ostrouska: Investigation. C. Deceneux: Investigation. R. Aranha: Investigation. D. Duscharla: Supervision, investigation. S.Q. Wong: Investigation. W. Chen: Investigation. S.H. Ramarathinam: Supervision, investigation. T.C.C. Lim Kam Sian: Investigation. N.P. Croft: Supervision, investigation. C. Li: Data curation, software, investigation. R. Ayala: Investigation. J.S. Cebon: Resources. A.W. Purcell: Conceptualization, resources, formal analysis, supervision, funding acquisition, investigation, writing–original draft, project administration, writing–review and editing. R.B. Schittenhelm: Conceptualization, resources, formal analysis, supervision, investigation, writing–original draft, project administration, writing–review and editing. A. Behren: Conceptualization, resources, data curation, formal analysis, supervision, investigation, writing–original draft, project administration, writing–review and editing.
The authors acknowledge the Monash Proteomics & Metabolomics Facility for the provision of mass spectrometry instrumentation, training, and technical support, as well as the Monash University Flowcore for flow cytometry instrumentation and assistance. This project was funded, in part, by Ludwig Cancer Research, Melanoma Research Alliance, the Victorian Cancer Agency–supported Melbourne Melanoma Project, Australian National Health and Medical Research Council (NHMRC) project grants 1007381 and 1165490 (to A.W. Purcell), and the Victorian State Government Operational Infrastructure Support Program. A. Behren, P. Faridi, and A.W. Purcell were supported by a grant from La Trobe University (RFA Understanding Disease). C. Li was supported by an Australian NHMRC CJ Martin Early Career Fellowship 1143366. J.S. Cebon was supported by an Australian NHMRC Practitioner Fellowship 487905. A.W. Purcell was supported by an NHMRC Principal Research Fellowship 1137739. A. Behren is the recipient of a Fellowship from the Victorian Government Department of Health and Human Services acting through the Victorian Cancer Agency. Computational resources were supported by the R@CMon/Monash Node of the NeCTAR Research Cloud, an initiative of the Australian Government's Super Science Scheme and the Education Investment Fund.
The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.