Population-level distribution and putative immunogenicity of cancer neoepitopes

For our analyses, we used somatic mutations identified with MuTect [36] from Mutation Annotation Format (MAF) files for 18 cancer types (see Additional file 1: Table S1) in TCGA, retrieved using gdc-scan (v1.0.0) [37]. The MAF files were then converted to tumor-normal pair variant call format (VCF) files using the maf2vcf tool in the vcf2maf software package [38], with the GRCh38/hg38 genome build available from the Broad Institute resource bundle [39] used as the reference genome. Because these VCFs still contained data for both tumor and normal samples, they were then manipulated to remove data from the paired normal sample, leaving final, tumor-only VCF files for compatibility with pVAC-Seq, which accepts only single-sample VCFs. Each disease type consisted of a variable number of patients (see Additional file 1: Figure S1).

We then annotated these VCF files using Variant Effect Predictor (VEP, v88) [40]. VEP was run according to pVAC-Seq’s recommendations [41] with the Downstream and Wildtype VEP plugins [42] used, gene symbols added to output where available, and mutation consequence terms based on Sequence Ontology annotation guidelines [43]. We also used VEP’s GRCh38 annotation cache (rather than querying remotely) for efficiency.

As the TCGA data we obtained did not allow us to calculate patient-specific HLA types, we assumed each tumor could occur in the setting of any HLA allele type, allowing us to explore neoepitope distributions among a broader theoretical population. To do this, we generated a list of HLA alleles to use for subsequent analysis based on allele frequencies originating from the Allele Frequency Net Database [44] and summarized for use in the software POLYSOLVER (v1.0) [45]. The average frequencies across races (Asian, Black, and Caucasian) of alleles for each HLA gene (HLA-A, HLA-B, and HLA-C) were calculated and normalized to sum to 100%. We then selected the top 145 average-frequency HLA alleles for subsequent analysis (see Additional file 1: Tables S2 – S4), encompassing all HLA alleles among 99% of individuals in the general population.

pVAC-Seq (v4.0.8) was run for each patient and allele combination using 9mer epitopes generated from 17-mer peptides surrounding each missense mutation, and using MHC binding predictions generated by NetMHCpan (v2.8) [46]. For each resulting neoepitope from pVAC-Seq, additional metrics were applied as described below. Note that for the purposes of this study, only epitopes resulting from missense mutations were considered for further analysis, and all peptides were considered to be expressed at equal levels. Note also that neoepitopes from breast cancer (BRCA), cervical cancer (CESC) and melanoma (SKCM) were not assessed for protein sequence similarity against peptides other than their paired normal epitopes.

To assess the degree to which HLA alleles might have overlapping preference for putatively novel binding neoepitopes predicted for mutations across TCGA (see Neoepitope Prioritization Metrics), 1000 random sets of six of the previously described set of HLA alleles (two HLA-A, two HLA-B, and two HLA-C alleles) were chosen using the random.sample function (without replacement) from the random module in Python 2.7.13 [47] (the combinations tested are available in Additional file 2: Table S5). All unique amino acid sequences of neoepitopes that bound to one or more alleles within each random allele set were counted; separate counts were kept for neoepitopes that bound to one, two, three, four, five, or six of the six alleles (i.e. increasing levels of overlap). The script for randomly sampling allele sets and determining overlap is available in our GitHub repository [48].

For comparison, we assessed recurrence rates among 2,813,809 simulated neoepitopes (9mers) mirroring the size of the TCGA data set. These neoepitopes were drawn randomly from the GRCh38 peptidome, with subsequent introduction of a random single amino acid substitution at a random position along each 9mer. These simulated peptides were labeled by patient and disease site to produce a random set of peptides for each patient equivalent in size to that patient’s predicted neoepitope repertoire. We repeated this process again for a smaller set of 1000 simulated neoepitopes to assess trends in peptide similarity scores. The gene of origin of the random peptide and the gene corresponding to its closest peptide match in the human proteome were retained for protein sequence similarity analysis (see “Tumor vs. closest human peptide sequence similarity” below).

We identified patient-specific neoepitopes in whole exome sequencing data from 12 patients selected from a study exploring genomic features of response to immunotherapy in melanoma patients [49]. Reads were aligned against the GRCh37d5 reference genome using the Sanger cgpmap workflow [50]. This workflow uses bwa-mem (v0.7.15–1140) [51] and biobambam2 (v2.0.69) [52] to generate genome coordinate-sorted alignments with duplicates marked. Realignment around indels and base recalibration were performed using Genome Analysis Toolkit (v3.6) [53]. Variants were called using VarScan (v2.3.9) [54] in accordance with the methods outlined in the workflow [50]. VCF files were annotated using VEP (v88) [40] as described above. For all missense single nucleotide variants identified, the tumor and normal protein epitopes of 8, 9, 10, and 11 amino acids in length were produced by reconstructing the nucleotide sequence surrounding the mutation using its coordinates from the VCF file and the CDS in the hg19 gene transfer format file [55], and translating this sequence into amino acids. Each patient’s HLA type was determined from FASTQ files using Optitype (v1.3.1) [56], and the binding affinity of all predicted tumor and normal epitopes was predicted with NetMHCpan (v2.8) [46] for each epitope and patient-specific HLA allele combination. Additional prioritization metrics were applied as described below.

All neoepitope novelty metrics are summarized in Fig. 1, and scripts for calculating these metrics are available on our GitHub repository [48].

To assess how well our prioritization metrics reflect a neoepitope’s ability to elicit an immune response, we applied our criteria to predicted neoepitopes from six studies in which peptide-specific immune responses were measured [3, 11, 20, 62‐64]. For data from all studies, we used only peptides which had complete information regarding the neoepitope and its paired normal peptide, as well as complete data regarding epitope-level immune response, providing a total cohort of 419 peptides. Because only binary immune response data was available from Ott et al. [11] and Bjerregaard et al. [20], we generated binary response data from the Carreno et al. [62] dataset for compatibility: a neoepitope was considered to have elicited an immune response if it had a percent neoantigen-specific T-cell in lymph+/CD8+ gated cells of greater than 10%. Of the seven peptides from the Le et al. [64] dataset that were tested for clonal T cell expansion, the three peptides that demonstrated clonal T-cell expansion were considered to have elicited an immune response, while those that only demonstrated immune reactivity in an ELISpot assay were considered not to have elicited an immune response. Among peptides evaluated in co-culture experiments from Tran et al. [3] and Gros et al. [63], those that were T-cell reactive were considered to have elicited an immune response. We produced a linear model to determine the relationship between our neoepitope novelty criteria and peptide-specific immune response (see “Statistical analysis” below). Using Scikit-learn for Python [65], SVM and Random Forest models were also trained with 10 fold cross validation for comparison.

Statistical analysis was performed using R (v3.3.2) in RStudio. To test the relationship between per-allele neoepitope burden and neoepitope frequency across the TCGA dataset, we obtained the Pearson’s product-moment correlation and associated p-value using a two-sided test. To determine whether a difference in tumor vs. paired normal peptide binding affinity difference exists between epitopes with and without an amino acid change at an anchor position, we applied a Wilcoxon rank sum test. We also used the Wilcox test to compare tumor vs. paired normal peptide sequence similarity scores between epitopes with novel vs. non-novel binding changes, and to compare the difference in tumor vs. paired normal peptide sequence similarity scores for the neoepitope with its paired normal epitope and its closest matching human peptide from BLAST for matching versus non-matching genes. A Welch’s two sample t-test was used to compare the similarity of neoepitopes to bacterial vs. viral peptides. We used the package pROC [66] to obtain AUROC scores and the lm function to determine the relationship between our continuous predictors and observed peptide-specific immune response data. Our analyses are available as an R script on our GitHub repository [48].

Consistent with prior analyses of TCGA data [4, 67], we demonstrate a varied spectrum of neoepitope burden across diseases, with skin cutaneous melanoma and pheochromocytoma/paraganglioma having the highest and lowest median neoepitope burdens, respectively (see Fig. 2a). These differences are likely due to known differences in somatic mutation burden as a function of disease type, as the ratio of neoepitopes to somatic missense mutations per patient was relatively constant across diseases with strong correlation between both metrics (Pearson’s product-moment correlation of 0.99; p < 2.2 × 10^− 16; see Additional file 1: Figure S2). Across disease types, an average of 36.8% (sd = 0.7%) of all predicted neoepitopes were from HLA-A alleles, 42.4% (sd = 14.8%) from HLA-B alleles, and 21.0% (sd = 0.7%) from HLA-C alleles.

We next sought to explore the repertoire of shared neoepitopes across TCGA as a function of HLA subtypes. Among the ten HLA alleles with the greatest number of high-affinity epitopes, there was surprisingly little repetition of epitopes binding to that allele across TCGA (mean 1.06 epitopes repeated); however, each allele had at least one epitope encountered multiple times across patients and diseases (31–168 occurrences). The most frequently repeated neoepitope for HLA-B*15:03, KQMNDARHG, was found most often in breast carcinoma patients. In all cases, this neoepitope originated from a H1047R substitution caused by a single nucleotide variant in the gene PIK3CA, a known driver mutation [68]. Two other recurrent epitopes, LSKITEQEK and STRDPLSKI, were identified 168 times for HLA-A*30:01 and HLA-B*15:17, respectively, with both originating from the same oncogenic mutation in PIK3CA (E542K substitutions) [68], and were found exclusively in breast carcinoma, cervical squamous cell carcinoma, and prostate adenocarcinoma patients. There were 5175 other occurrences of PIK3CA neoepitopes, as well as 7139 TP53 (a known cancer driver gene [69]) neoepitopes, and 35,872 occurrences of neoepitopes originating from MUC16, the gene encoding the CA-125 cancer biomarker [70]. On average, 4.7% of patient epitopes were repeated across patients within their own disease site and 7.9% of patient epitopes were repeated across all of TCGA, a rate significantly higher than that anticipated by random chance alone (see Additional file 1: Figure S3 and Additional file 1: Table S6), and likely attributable to common cancer mutations. It is, finally, important to note that the true number of shared neoepitopes among cancers is likely to be smaller due to the random assortment of actual HLA alleles across the population.

We were also interested in the degree to which an overlap of neoepitope preferences existed between the HLA alleles studied, as an epitope that binds strongly to more than one of a patient’s suite of HLA alleles would likely be a better candidate for applications such as peptide vaccines. For 1000 randomly sampled sets of six HLA alleles (see Methods), on average, 3.1% of neoepitopes across TCGA had affinity for at least one of the six alleles in each set, emphasizing the importance of a patient’s unique HLA repertoire in neoepitope presentation. Of these epitopes, the majority (87.5% on average) only had affinity for one allele, and 11.0% on average had affinity for two alleles, but there were some cases where epitopes had affinity for all six alleles (0.0007% of epitopes on average, and up to 0.2% of epitopes for one allele set; see Additional file 1: Figure S4).

We then examined the distribution of paired tumor and normal epitope binding affinities across a cohort of melanoma patients to understand the landscape of differential HLA-specific binding affinities [49] (see Fig. 3a). While most mutations did not have a large effect on epitope binding affinity (median 71.1 nM binding affinity difference), we were particularly interested in those mutations that changed the binding affinity significantly in favor of the neoepitope (> 5-fold increased affinity, see Fig. 3a). We applied these criteria to identify neoepitopes with putatively novel binding changes (see Methods), and noted only a small fraction of qualifying neoepitopes from each patient (0.09% on average, see Fig. 3b). This dramatic refinement in neoepitopes led us to consider how these criteria might affect neoepitope distribution across a larger cancer cohort from TCGA (see Additional file 1: Table S1) and among a broad population of HLA types.

We next assessed novelty of MHC tumor vs. paired normal binding affinity change across TCGA, noting that a minority of neoepitopes (20.3% on average) met this criterion, with a similar distribution across all TCGA cohorts (average proportion per patient of 18.0–24.7%, see Fig. 2a). This finding was dependent upon HLA type, with a median 5.6-fold difference in the number of neoepitopes with novel binding changes between the 25th and 75th percentile HLA alleles among diseases (see Fig. 2b). All diseases had the greatest number of neoepitopes associated with the allele HLA-B*15:03; however, there was no statistical association between HLA allele frequency in the general population and that allele’s corresponding number of novel binding change neoepitopes (Pearson’s product-moment correlation of 0.1; p = 0.1; see Fig. 2b).

Importantly, we note that the mutation of amino acid residues at MHC anchor positions (i.e. the second residue and last residue of each 9mer epitope) tends to result in more dramatic predicted binding affinity differences between the tumor and paired normal peptides compared to mutation of non-anchor residues (median of 2935.1 nM vs. 28.1 nM, respectively; p < 2.2 × 10^− 16; see Additional file 1: Figure S5).

While anchor residue mutations may influence differential peptide binding, anchor residues are anticipated to be more directly engaged with the MHC complex and thus less accessible for T-cell recognition [57]. We therefore sought to investigate the differences in peptide sequence between tumor neoepitopes with mutations in T-cell exposed (i.e. non-anchor) residues and paired normal epitopes. The average protein sequence similarity score (see Methods) between paired epitopes with single nucleotide variants at non-anchor positions across all diseases was 83.5% (ranging from 60.0% to 98.1%, sd = 6.3%). When we assessed these similarity scores in conjunction with our novel binding criteria (see Methods), we observed that neoepitopes which displayed a putatively novel binding change tended to have lower similarity to their paired normal counterpart than those without such a binding affinity change (mean 81.5% vs. 83.7%, respectively; p < 2.2 × 10^− 16). This level of significance held even when controlling for tumor neoepitope binding affinity (see Additional file 1: Table S7).

We reasoned that regardless of how similar or dissimilar a neoepitope may be to its paired normal epitope, it may closely mimic a different normal epitope present within the human proteome (see Fig. 1c). Comprehensive blastp analysis of all neoepitopes from all disease types generated human proteome matches for more than 99.9% of peptide queries, with an average protein sequence similarity score of 84.3% (sd = 10.7%). The majority of neoepitopes (77.3% on average) mapped most closely to one or more normal peptides from the same gene (see Figs. 4a and b). However, 22.7% of neoepitopes matched more closely to one or more unrelated human peptides; in 3.5% of these cases, the unrelated human peptide was an exact match to the tumor neoepitope across all 9 amino acid positions. This phenomenon is likely stochastic in nature, as 24.9% of simulated neoepitopes (see Methods) matched most closely to unrelated human peptides, with an average protein sequence similarity score of 81.9% (sd = 10.6%), significantly different from that of the TCGA neoepitopes (p = 4.927 × 10^− 13, Welch Two Sample t-test).

Next, we assessed neoepitope sequence homology with peptides from pathogenic and commensal microorganisms. Almost all neoepitopes were found to have at least one matching bacterial or viral peptide by blastp (87.6% and > 99.9%, respectively). Overall, tumor neoepitopes were more similar to bacterial peptides compared to viral peptides (mean percent peptide sequence similarity score of 91.4% (sd = 6.6%) and 76.7% (sd = 9.1%), respectively, (p < 2.2 × 10^− 16)). Interestingly, in 56.9% of cases where a neoepitope had a bacterial blastp hit, the bacterial peptide was more similar to the neoepitope than either its normal counterpart or its most similar normal protein as determined by blastp; this was only true for 15.8% of the viral peptide matches for neoepitopes (see Fig. 1d for example). More strikingly, when considering protein sequence similarity scores across all residues, 59.6% of neoepitopes with bacterial blastp hits had higher similarity to these peptides than to either of their normal peptide matches; only 5.8% of viral epitopes showed this phenomenon. This was true despite the fact that neoepitopes had significantly more mismatches in sequence with bacterial peptides than they did with either their paired normal epitopes (mean 1.5 vs 1; p < 2.2 × 10^− 16) or their closest matching peptides from blastp (mean 1.5 vs 1.4; p < 2.2 × 10^− 16). However, in terms of total amino acid length, the bacterial peptide data base was 577.5 times larger than the human peptide database, and the viral peptide data base was only 2.0 times larger, so these phenomena may be in part reflective of these differences.

Additional file 1: Figure S6 shows the distribution of protein percent similarity scores for bacterial and viral hits for each TCGA disease site and for 1000 randomly simulated neoepitopes (see Methods) across non-anchor residues. Predicted TCGA neoepitopes were significantly more similar than random peptides to their closest matching bacterial peptide (mean 91.4% vs. 82.3%, p < 2.2 × 10^− 16, Welch Two Sample t-test) and their closest matching viral peptide (mean 76.7% vs. 58.7%, p < 2.2 × 10^− 16, Welch Two Sample t-test), indicating that this phenomenon of sequence similarity to microorganism peptides may be specific to cancer neoepitopes. We determined the top 10 most frequently occurring bacterial genera in cases where a bacterial peptide was a closer match to a neoepitope than either of its human peptide counterparts (see Fig. 5), which includes, of particular interest, frequently pathogenic genera such as Clostridium [71], Mycobacterium [72, 73], and Vibrio [74], and the frequently commensal genus Lactobacillus [75].

Finally, we applied our criteria to a cohort of neoepitopes with paired immune response data. Applying any single criterion to predict immune response to a neoepitope in a linear model did not lead to significant prediction in any case, except for the percent protein sequence similarity between a neoepitope and its closest viral peptide (p = 0.046; see Table 1, Figs. 6a-f). We also observed how well immune response was predicted by neoepitope binding affinity, paired normal epitope binding affinity, and the number of mismatches in amino acid sequence between the neoepitope and its paired normal epitope. Only the number of mismatches was alone able to predict neoepitope immunogenicity, favoring those neoepitopes with multiple amino acid changes (p = 0.03; see Table 1). Our putatively novel binding change criteria alone was able to predict true immunogenicity with an AUROC of 0.53 (see Additional file 1: Figure S7). A linear model incorporating 1) neoepitope binding affinity, 2) putatively novel binding change status of the neoepitope, 3) binding affinity difference between the neoepitope and both its paired normal epitope and 4) its closest BLAST peptide match, 5) number of amino acid sequence mismatches between the neoepitope and its paired normal epitope, and 6) percent protein sequence similarity between the neoepitope and its paired normal epitope, 7) its closest human peptide match, 8) its closest bacterial peptide match, and 9) its closest viral peptide match was able to significantly predict immune response to neoepitopes (p = 0.02). However, only three individual predictors contributed significantly to the model: neoepitope binding affinity (p = 0.003), percent protein sequence similarity of the neoepitope to its closest viral peptide match (p = 0.048), and binding affinity difference between the neoepitope and its closest human peptide match (p = 0.002). The contribution of the number of amino acid sequence mismatches between the neoepitope and its paired normal epitope approached significance (p = 0.075). A reduced, multiplicative version of our linear model incorporating only these four predictors was able to predict immune response to neoepitopes with greater significance (p = 3.3 × 10^− 6; AUROC = 0.66; see Fig. 6g; mathematical representation available in supplementary materials). For comparison, we also trained SVM and Random Forest models to predict peptide-specific immune response; however, the simpler linear model remained the best predictor (see Additional file 1: Figure S8).