Malignant pleural mesothelioma (MPM) is a highly lethal cancer of the lining of the chest cavity. To expand our understanding of MPM, we conducted a comprehensive integrated genomic study, including the most detailed analysis of BAP1 alterations to date. We identified histology-independent molecular prognostic subsets, and defined a novel genomic subtype with TP53 and SETDB1 mutations and extensive loss of heterozygosity. We also report strong expression of the immune-checkpoint gene VISTA in epithelioid MPM, strikingly higher than in other solid cancers, with implications for the immune response to MPM and for its immunotherapy. Our findings highlight new avenues for further investigation of MPM biology and novel therapeutic options.
Through a comprehensive integrated genomic study of 74 MPMs, we provide a deeper understanding of histology-independent determinants of aggressive behavior, define a novel genomic subtype with TP53 and SETDB1 mutations and extensive loss of heterozygosity, and discovered strong expression of the immune-checkpoint gene VISTA in epithelioid MPM.
See related commentary by Aggarwal and Albelda, p. 1508.
This article is highlighted in the In This Issue feature, p. 1494
Malignant pleural mesothelioma (MPM) is a cancer of the mesothelial cells lining the pleural cavity. It was rare until the widespread use of asbestos in the mid-20th century (1). Although reduction and strict regulation of asbestos use may be leading to a leveling off in new cases in Western countries, its long latency, together with continued use of asbestos in non-Western countries, ensures that MPM remains a global problem (2). MPM is almost universally lethal, with only modest survival improvements in the past decade (3), suggesting that standard treatment is reaching a therapeutic plateau. Elucidating oncogenic genomic alterations in MPM is therefore essential for therapeutic progress (4–7). Frequent copy-number loss and recurrent somatic mutations in BAP1, NF2, CDKN2A, and others have been identified (4–6), yet no targeted therapies exploiting these alterations have emerged.
To expand our understanding of the molecular landscape and biological subtypes of MPM, and provide insights that could lead to novel therapies, we have conducted a comprehensive, multiplatform, genomic study of 74 MPM samples, as part of The Cancer Genome Atlas (TCGA). Here, we report how these integrated analyses define prognostic subsets of MPM, characterize a new near-haploid molecular subtype, and identify novel potential therapeutic targets.
We studied 74 samples of primary MPM from patients with no prior systemic therapy. This cohort was predominantly male (82%), with a median age of 64 years, and tumors were of mostly epithelioid histology (70%), a typical profile for MPM. Asbestos exposure history was positive in 62%, negative in 18%, and unavailable or unknown in the remainder. Demographic and clinical details are provided in Supplementary Tables S1A and S1B, as well as Supplementary Fig. S1.
We performed comprehensive molecular profiling, including exome sequencing, copy-number arrays (Supplementary Fig. S2), mRNA sequencing (Supplementary Fig. S3), noncoding RNA profiling, DNA methylation (Supplementary Fig. S4), and reverse-phase protein arrays (RPPA; Supplementary Fig. S5). Methods and detailed results of individual analyses are provided in Supplementary Sections 1–13.
Landscape of Somatic Mutations and Copy-Number Alterations
Whole-exome sequencing (WES) revealed a somatic mutation rate of <2 nonsynonymous mutations per megabase in all samples except for an outlier case with a mutation rate of 8 nonsynonymous mutations per megabase (Fig. 1A). This places MPM at the low end of somatic mutation burden among cancers (8). The outlier tumor with a 10-fold higher mutation rate showed a distinctive pattern of C>T mutations occurring almost exclusively at CpG dinucleotides (Fig. 1B). This relatively hypermutated tumor harbored a homozygous nonsense mutation in MSH2, which would suggest that the tumor lacked mismatch-repair capacity, but the observed mutational spectrum was atypical for mismatch-repair deficiency (9). Otherwise, the observed mutational spectrum was similar across patients and lacked distinctive or novel features (Fig. 1C). Signatures of smoking- or APOBEC-induced mutagenesis were not observed. Asbestos has been proposed to cause genotoxicity via DNA breaks and secondary oxidative damage (10). However, the mutational spectrum and local sequence context was not significantly different between cases with or without known asbestos exposure (χ2, P = 0.3); although this negative finding should be viewed with caution given the limitations of the data set, it is in agreement with prior studies (5).
We sought to identify genes mutated significantly above the background rate using MutSig2CV. Significantly mutated genes (SMG) at a false discovery rate (FDR) of <0.05 included BAP1, NF2, TP53, LATS2, and SETD2, all known cancer genes in MPM (Fig. 1A; Supplementary Table S1C). All five genes showed high rates of nonsense, frameshift, and splice-site mutations, highlighting them as targets of inactivation, consistent with their functions as tumor suppressors. Mutations in these five SMGs did not show associations with histories of asbestos exposure or smoking. For validation in an independent cohort that also used a different algorithm to define SMGs, we compared the results of our SMG analysis with the SMGs previously identified in 99 MPM exomes using the MUSIC algorithm (5) and found a strong overlap for SMGs at an FDR of <0.05 in both studies (Fig. 1D). Notably, among lower confidence SMGs (FDR < 0.15 but > 0.05), only SETDB1, which may define a novel subtype of MPM (discussed below), was identified in both analyses (Fig. 1D).
The somatic copy-number alteration (SCNA) landscape was characterized by frequent recurring focal and arm-level deletions, but no recurring amplifications, consistent with the notion that MPM development is driven primarily by loss of tumor suppressor genes, rather than by activation of classic oncogenic driver genes (Supplementary Section 4; Supplementary Fig. S2). Focal deletions affected several tumor suppressor genes known to be altered in MPM (5), most notably CDKN2A with deletions (defined as deep, likely homozygous) in 36 (49%) and losses (defined as shallow, possibly single-copy) in 5 (7%) samples. Likewise, NF2 deletions were confirmed in 25 (34%) samples and losses in 30 (40%) samples; as many of the latter harbored mutations of the remaining allele, evidence of biallelic NF2 inactivation was common (Fig. 1A). CDKN2A deletions often encompass MTAP, the adjacent gene on 9p21 (11), which encodes methylthioadenosine phosphorylase, whose deficiency has recently been reported to lead to reduced PRMT5 enzymatic activity and heightened sensitivity to its pharmacologic inhibition (12, 13). Codeletion of CDKN2A and MTAP, associated with low levels of mRNA expression of both genes, was observed in 20 cases (Supplementary Fig. S2). Although no correlation with overall survival was observed for BAP1 status, loss of CDKN2A was strongly associated with shorter overall survival (Cox proportional hazards, P = 7.3 × 10−6), as previously shown (14, 15).
Finally, an analysis for genomically integrated viral sequences, including SV40, was negative (Supplementary Section 8), as was a screen of exome and RNA-sequencing data for evidence of EWSR1 and ALK fusions, recently reported in rare cases of MPM (16, 17) and peritoneal mesothelioma (18), respectively. As well, no activating mutations in the canonical MAPK or PI3K/AKT signaling pathways were identified in this cohort.
Comprehensive Analysis of BAP1 Status in MPM
BAP1, encoding a nuclear deubiquitinase, is the most frequently mutated cancer gene in MPM, in both our data set and others (5), and is also recurrently inactivated in clear cell renal carcinoma, uveal melanoma, and cholangiocarcinoma (19). Somatic BAP1 mutations have germline counterparts that define the BAP1 hereditary cancer syndrome (20). In contrast to other cancers, BAP1 inactivation in MPM does not correlate with adverse outcomes. Because its role in MPM biology remains unclear, we compared BAP1-inactivated and wild-type cases across multiple platforms. First, to better segregate cases according to BAP1 status, we performed a comprehensive analysis of inactivating alterations through a detailed review of single-nucleotide variants, small and large indels, whole gene deletions, and structural variants; this showed the overall prevalence of BAP1 alterations to be 57% (Fig. 1A; Supplementary Table S2A), in line with recent studies (21). Most MPM with inactivating mutations in BAP1 (25/26, 96%) also had concurrent loss of heterozygosity (LOH) on chromosome 3p21.1, supporting a classic two-hit tumor suppressor mechanism. Overall, BAP1 status was defined as follows: 32 samples with no evidence of BAP1 alteration, 6 with a single (heterozygous) mutation or deletion, and 36 with biallelic inactivation. No germline mutations in BAP1 were identified in this cohort.
BAP1 alterations showed nonrandom patterns of co-occurrence with mutations in other key MPM cancer genes (Fig. 2A). As expected, mRNA expression levels of BAP1 itself were reduced in the presence of genomic BAP1 inactivation or loss (Fig. 2B). We also observed an inverse correlation of BAP1 alterations and SCNAs. BAP1-altered tumors had fewer chromosome arm gains and losses (Fig. 2C, median 9.5 vs. 15.5, Mann–Whitney, P < 0.01), as well as fewer focal SCNAs (Supplementary Fig. S2).
Because BAP1-mediated deubiquitination of histones and transcriptional proteins is thought to regulate gene expression, we compared mRNA expression patterns between tumors that were wild-type for BAP1 and those with inactivation of one or both BAP1 alleles. We identified 1,324 differentially expressed genes with an FDR of <0.01, of which 75% were downregulated in the BAP1-inactivated samples. As previously noted in experimental models (22), BAP1-inactivated tumors in our cohort had lower mRNA expression of several HOXA genes, including HOXA5 and HOXA6, but no difference in EZH2 mRNA expression or its PRC2 partners EED and SUZ12.
As BAP1 is known to regulate the ubiquitination and hence stability of several transcriptional proteins, we examined the association of BAP1 status with inferred activity of transcription factors (TF) using a recently developed computational strategy that integrates phosphoproteomic and transcriptomic data with predicted TF binding sites (23). We used this algorithm to assess significant differences in inferred TF activities between BAP1-inactivated and wild-type MPM (satisfying FDR-corrected P < 0.01, t test), linking BAP1 status to altered activity of TFs. Indeed, many TFs identified in this analysis had highly significant associations with BAP1 status (Fig. 2D and E; Supplementary Table S2B).
In particular, YY1 had significantly reduced inferred activity in BAP1-inactivated samples (Fig. 2F). BAP1 forms a ternary complex with HCF1 and YY1, which acts as transcriptional repressor of genes involved in cell proliferation (24). YY1 has also been shown to regulate or interact with other TFs whose inferred activity was also altered in our analysis, such as MAX and EGR2. Moreover, there is evidence that YY1 interacts with EZH2, the catalytic subunit of PRC2, and is required for its function in gene silencing (25). Thus, to a degree, the transcriptional consequences of BAP1 inactivation could be attributed to changes in recruitment of specific TFs to their target genes, either directly, as in the loss of ternary YY1–HCF1–BAP1 complexes, or indirectly, as in the case of TFs that are themselves downstream targets of YY1 (Supplementary Table S2C).
Notably, BAP1-inactivated samples demonstrated a significantly increased inferred activity of IRF8 (Fig. 2G; Supplementary Table S2D), a hematopoietic lineage TF involved in interferon signaling and dendritic cell differentiation. Both pathways were upregulated in our BAP1-based supervised gene mRNA expression analysis (see above). Moreover, PARADIGM analysis confirmed that the IRF activation pathways are upregulated in BAP1-inactivated samples. IRF8 has been implicated in the biology of CD103-positive dendritic cells whose antigen-presenting function is highly effective at stimulating cytotoxic T cells in the tumor microenvironment (26). Taken together, these observations suggest an association between BAP1 status and perturbed immune signaling in MPM that warrants further exploration.
MPM with Genomic Near-Haploidization, a Novel Subtype
Allele-specific copy-number analysis of WES data using the FACETS algorithm (27) identified three cases with striking genome-wide LOH that affected more than 80% of the genome, which was confirmed in SNP6.0 copy-number array data (Fig. 3A and B; Supplementary Fig. S2). This phenomenon, well described in acute lymphoblastic leukemia (28), occurs when the nascent cancer clone loses one copy of nearly all chromosomes [which we term “genomic near-haploidization” (GNH)] followed by duplication of the remaining chromosomes. To better assess the prevalence of this unusual genomic phenomenon in MPM, we screened 80 samples from a Japanese International Cancer Genome Consortium (ICGC) MPM cohort, identifying two additional cases (Fig. 3B), representing a combined prevalence of 5 of 154 (3.2%). Neither the three TCGA cases, nor the two ICGC cases, had deletions or point mutations in BAP1, PBRM1, or SETD2. Remarkably, all 5 cases had inactivating point mutations in, or homozygous deletion of, SETDB1, which encodes a histone methyltransferase involved in gene silencing, representing the only SETDB1-mutated MPMs in the combined TCGA–ICGC cohort. Additionally, 4 of 5 cases with evidence of GNH had driver mutations in TP53. Although near-haploid cases have been reported in other cancers (e.g., acute lymphoblastic leukemia, chondrosarcoma, and adrenocortical carcinoma), the strong association of TP53 and SETDB1 comutation with widespread LOH has not been previously reported in other cancers (28, 29).
To further define the clinical features of this novel subset, we obtained data on 16 additional cases with near-haploid karyotypes from a large independent series of MPM studied by cytogenetic analysis of short-term primary cultures (Supplementary Section 2.2; note: cytogenetic analysis can identify only GNH cases that have not undergone genome duplication). Although most of the genome in our 5 cases showed LOH, chromosomes 5 and 7 strikingly retained heterozygosity (Fig. 3C). These two chromosomes also remained disomic in the cases with near-haploid karyotypes in the independent cytogenetic series (Supplementary Table S3). Based on this combined set of 21 near-haploid MPM, a distinctive profile of genomic and clinical features emerged, demarcating a novel molecular subtype of MPM. Strikingly, most patients with MPM showing GNH were female (M:F = 1:4), in sharp contrast to the overall TCGA–ICGC MPM cohort, where males greatly predominated (M:F > 4:1; P < 0.01, Fisher exact test). No difference in distribution of histologic types was observed in the GNH subset.
Because MPM with evidence of GNH have typically undergone a genome-doubling event, we were able to assess the timing of this relative to point mutations (30, 31). Essentially, mutations acquired before genome duplication would be present on both copies of the duplicated chromosome, whereas those occurring after would be heterozygous. In all three TCGA cases with GNH, the majority of the point mutations were present on both copies of duplicated chromosomes, suggesting that the genome duplication occurred relatively late in the evolution of the cancer. In three cases, the TP53 driver mutation was homozygous, suggesting it occurred before genome doubling. In contrast, the SETDB1 mutations were present on only one of the two copies of chromosome 1 in two of the three cases, suggesting they occurred after LOH and genome duplication (Fig. 3D). Overall, these data suggest a model in which TP53 mutations occur early, and presumably permit the steady (or catastrophic) loss of chromosomes. It seems likely that genome reduplication occurs after achieving near-haploidy, with haploinsufficiency of SETDB1 arising later.
Integrative Multiplatform Analysis Defines Novel Prognostic Subsets
Although the current classification of MPM into epithelioid, sarcomatoid, and biphasic histologies is prognostically useful, there remains variability in clinical features and patient outcomes within histologic subtypes. Previous analyses (5, 7) based on mRNA expression alone have defined unsupervised clusters largely recapitulating these histologic distinctions. To find out whether multiplatform molecular profiling might provide additional resolution to define prognostic subsets of MPM, we performed integrative clustering across multiple assay platforms using two algorithms: iCluster (32) and PARADIGM (33). Both identified four distinct integrated subtypes of MPM. There was a strong concordance in subtype assignments between the two algorithms (Fig. 4A; Supplementary Figs. S6 and S7), especially for the best (cluster 1) and worst (cluster 4) prognosis clusters, indicating that integration of molecular data can identify distinct subgroups of MPM, independent of the specific statistical methodology. Survival was significantly different across the 4 clusters (P < 0.0001 for either algorithm; Fig. 4B; Supplementary Fig. S7). This survival difference remained significant (P = 0.008) after adjusting for histology (epithelioid vs. nonepithelioid) and CDKN2A homozygous deletion (Fig. 4C; Supplementary Fig. S6), a known molecular prognostic factor in MPM (14, 15). Although they did provide additional independent prognostic information, the iClusters nonetheless did show correlation with consensus histology (P = 0.002), with iCluster cluster 1 being enriched for epithelioid histology (similar finding for PARADIGM cluster 1). They also exhibited differences in immune cell infiltrates (Supplementary Table S4A), as discussed below, but showed no significant correlation with clinical variables such as T stage, N stage, asbestos exposure history, or smoking (Supplementary Table S4B). Molecularly, these tumors had low SCNA, relatively few CDKN2A homozygous deletions (11%), and a high level of methylation (Supplementary Fig. S4). All but one (95%) had BAP1 alterations: 26% had homozygous deletions and 53% had heterozygous loss with mutations.
The poor prognosis cluster (cluster 4; red) had a high score for epithelial–mesenchymal transition (EMT) based on gene expression (P < 0.001; Fig. 4D), which was distinguished by high mRNA expression of VIM, PECAM1, and TGFB1, and low miR-200 family expression. These tumors also displayed MSLN promoter methylation and consequent low mRNA expression of mesothelin, a marker of differentiated mesothelial cells, as noted previously in sarcomatoid MPM and the sarcomatoid components of biphasic MPM (15, 34). Overall, this poor prognosis cluster also showed enrichment of LATS2 mutations (30% compared with 4% in the rest of the cohort) and CDKN2A homozygous deletions (66%). Moreover, this cluster showed higher AURKA mRNA expression, higher leukocyte fraction (based on DNA methylation), and elevated mRNA expression of E2F targets, G2–M checkpoints, and DNA damage response genes. PI3K–mTOR and RAS–MAPK signaling were upregulated, based on both mRNA and protein expression (Supplementary Fig. S7). Additionally, several miRNAs were differentially expressed between the good and poor prognostic clusters, including miR-193a-3p, which has been proposed as a potential tumor suppressor (35). Finally, a comparison of immune gene mRNA expression signatures (36) across the four clusters revealed a significantly higher score for the Th2 cell signature in the poor prognosis cluster 4 compared with the other clusters (Fig. 4E; Supplementary Table S4A). Coincidentally, it has been reported that Th2 cytokines secreted by immune cells upon exposure to asbestos may promote MPM (37). The analyses of other immune signatures are shown in Supplementary Fig. S5.
Although biphasic and sarcomatoid MPM are more aggressive, there remains a need for improved risk stratification of epithelioid MPM, for which clinical outcomes are more heterogeneous. Therefore, we conducted an integrative clustering analysis restricted to epithelioid MPM. The results for the 4-cluster epithelioid-only solution were highly similar to the 4-cluster all-MPM solution (Fig. 5A), with only 7 of the 52 epithelioid samples reassigned to other clusters. This stability indicates that the features driving the all-MPM clustering are largely independent of histology. The epithelioid-only clusters share many of the features defining the corresponding clusters in the all-MPM solution (Fig. 5B). The survival analysis also paralleled the all-MPM solution, with cluster 1 having the best outcomes and cluster 4 having the worst (Fig. 5C). PARADIGM analysis of the epithelioid-only subset confirmed upregulation of AURKA mRNA expression in the poor-prognosis epithelioid-only cluster 4 (Fig. 5D), corroborating the results from the all-MPM analysis.
Finally, we sought to independently validate the clinical correlations of clusters identified in the TCGA epithelioid cases using mRNA expression profiles from two published studies: 211 MPM analyzed by RNA sequencing (5) and 52 MPM samples analyzed by mRNA expression microarrays (14). Specifically, we assigned each mRNA expression profile to one of the integrative clusters based on the rules derived from the TCGA mRNA data set. For the larger validation cohort (henceforth referred to as Bueno), we restricted our analysis to epithelioid samples and used the epithelioid-only gene signature to cluster samples. We found that the epithelioid-only samples assigned to iCluster 1 (good prognosis) had significantly better survival, even after adjusting for age (P = 3.9 × 10−4; Fig. 5E; Supplementary Fig. S6). In the smaller cohort (referred to as Lopez-Rios), patient numbers were too small to split by histology. However, this analysis provided independent validation of the survival differences for the four all-MPM clusters (P = 0.01; Supplementary Fig. S6). Taken together, these results suggest that the prognostically relevant molecular profiles defined by our analysis are robust and reproducible, and could potentially be used to improve risk stratification of patients with epithelioid MPM. The core mRNA gene lists are provided in Supplementary Table S4C–S4G, which also include reduced classifiers based on methylation, lncRNA, miRNA, and a reduced classifier combining mRNA and methylation data to facilitate practical application of these data and independent validation of these clusters.
Analysis of Noncoding RNAs in MPM
We next assessed two types of noncoding RNAs not extensively studied in MPM that may also represent a source of robust biomarkers: lncRNAs, which often show higher expression specificity for cell type than coding genes, and miRNAs, which are relatively stable in biological fluids, so potentially suitable as noninvasive biomarkers.
In both the TCGA and Bueno (5) cohorts, lncRNAs returned four unsupervised consensus subtypes that were associated with 5-year survival (P = 1.4 × 10−10 and P = 1.1 × 10−3, respectively). The TCGA lncRNA subtypes were highly concordant with iCluster (P = 8 × 10−14) and PARADIGM (P = 2 × 10−21) subtypes and were associated with EMT scores (P = 2.5 × 10−6; Figs. 4A, 6A–E; Supplementary Table S5A.1). For both cohorts, lncRNAs that were differentially abundant between the good-prognosis subtype and other samples included those associated with cancer (e.g., H19, LINC00152, and MEG3), or with MPM in particular: NEAT1 and SNHG8 (38) and GAS5 (ref. 39; Fig. 6F and G; Supplementary Table S5B.3-6). We noted that a number of lncRNAs distinguished the good-prognosis cluster in both TCGA and Bueno cohorts (Supplementary Table S5B.1-2,5-6).
Unsupervised clustering based on miRNA mature strands resolved five consensus subtypes in the TCGA cohort (Fig. 6H); these were associated with 5-year survival (P = 7 × 10−4), iCluster (P = 1 × 10−12) and PARADIGM (P = 3 × 10−13) clusters, and EMT scores (P = 2 × 10−7; Fig. 6H–K; Supplementary Table S5A.2). For the good-survival cases, the miRNA subtypes were strikingly concordant across multiple analysis platforms, and many cancer-associated miRNAs were differentially abundant in the good-survival cluster (Fig. 6L; Supplementary Table S5B.3-4). Taken together, these results suggest that lncRNAs and miRNAs may be important predictors of survival in MPM.
EMT and VISTA Expression
Because mRNA expression of EMT-associated genes was a key differentiating feature between prognostically distinct integrative clusters, we performed a detailed analysis of EMT in MPM using a previously published EMT-related mRNA signature (40). Increasing EMT scores significantly correlated with clusters defined by integrative algorithms (iCluster and PARADIGM), as well as with individual genomic features including miRNA, lncRNA, methylation, RPPA, and overall gene expression (Fig. 7A). Of all tumor types included in the TCGA pan-cancer analysis, MPM had the second-highest average EMT score (Fig. 7B; after soft-tissue sarcomas), consistent with previous reports that EMT is a frequent phenomenon in MPM (7, 41).
Although EMT score positively correlated with the mRNA expression of many immune-regulatory genes, such as OX40L, TGFB1, CD276, OX40, and PD-L2 (P < 0.001; Fig. 7A and C), mRNA expression of VISTA (42), a negative immune-checkpoint regulator primarily expressed on hematopoietic cells (43), was strongly inversely correlated with EMT score (Fig. 7C) and was expressed at levels higher than in any other TCGA tumor type analyzed (Fig. 7D). In the MPM cohort, VISTA mRNA levels were highest in the epithelioid subtype (Fig. 7C and E). Using Regulome Explorer (Supplementary Section 13), we found that VISTA mRNA expression was highly correlated with MSLN mRNA expression (Spearman correlation = 0.81; P = 6.3 × 10−19), but not with mRNA expression of PD-1 or PD-L1. Moreover, there was no significant correlation between overall mutation burden and VISTA expression levels (P = 0.64).
VISTA (V-domain Ig suppressor of T-cell activation; also known as c10orf54, PD-1H, and B7-H5) is a member of the B7 family of negative checkpoint regulators, expressed on the surface of several immune cell types. It can function as both receptor and ligand (44). As a ligand, VISTA is present on the surface of antigen-presenting cells (APC) and inhibits early-stage T-cell activation (45). The normal mesothelium has APC properties (46), which are retained upon malignant transformation (47, 48). We thus performed IHC staining of two epithelioid TCGA MPM cases, as well as normal and reactive mesothelium (Supplementary Section 14), to define the cellular compartment expressing VISTA in MPM tumor samples. Remarkably, VISTA protein expression was not restricted to infiltrating immune cells, but was present in tumor cells in MPM, as well as in normal and reactive mesothelium (Fig. 7F and G), suggesting that its expression in epithelioid MPM may reflect retention of APC properties in this more differentiated subset of MPM.
Our comprehensive integrative analysis of 74 cases of MPM further defines a cancer driven primarily by inactivation of tumor suppressor genes. Indeed, we confirm the high frequency of BAP1 inactivation by mutation and copy-number loss, as well as recurrent inactivating alterations in CDKN2A, NF2, TP53, LATS2, and SETD2. In addition to this landscape of known loss-of-function events, we have genomically characterized a novel molecular subtype of MPM accounting for approximately 3% of MPM in our data sets, defined by evidence of genomic near-haploidization and recurrent TP53 and SETDB1 mutations, with a distinctive clinical phenotype showing female predominance and younger age at diagnosis. Our findings should facilitate systematic clinical studies of this subset to better define its survival and its association with asbestos exposure, which so far appears weak or unclear. Isolated cases with similar molecular profiles (GNH and SETDB1 mutation) have been anecdotally reported (5, 49), but had not, until now, been recognized as a distinct molecular subtype of MPM. The genomic data in our cases support a model in which early TP53 mutations are permissive for a loss of chromosomes to a near-haploid state, followed by genome reduplication and SETDB1 inactivation. As H3K9 methylation by SETDB1 is a repressive chromatin mark (50), it is tempting to speculate that inactivation of SETDB1 allows a general increase in transcription activity in a cancer cell that has sacrificed nearly half of its genome, presumably including several genes that are imprinted or otherwise monoallelically expressed. Why chromosomes 5 and 7 are spared from the LOH in these MPM remains mysterious—interestingly, the same two chromosomes have recently been shown to also retain heterozygosity in thyroid Hürthle cell carcinomas with evidence of genomic near-haploidization (51). Overall, the identification of this novel subset of MPM highlights how, as the proportion of asbestos-related MPM plateaus and hopefully begins declining in Western countries, cases of possibly different etiology may become more apparent.
A better understanding of the determinants of aggressive behavior and predictors of poor clinical outcomes in MPM remains an unmet clinical need. To this end, integrative analyses with iCluster and PARADIGM revealed a set of molecular features that define MPM subsets with better and worse prognosis, and might point to candidate therapeutic targets. For instance, cases in the poor-prognosis subset showed higher AURKA mRNA expression, consistent with prior studies (14, 52). Treatment of MPM cell lines with an Aurora kinase inhibitor leads to growth arrest (53), and several Aurora kinase inhibitors are under investigation in patients with MPM. AURKA mRNA expression could be used to help identify patients with poor-prognosis epithelioid MPM so that they could be directed toward experimental therapies early in their treatment course. It is presently unclear whether Aurora kinase inhibitors will be active in MPM, and the ongoing phase II study of alisertib (NCT02293005) includes all patients irrespective of molecular signature. Although the value of targeting the AURKA pathway with currently available clinical compounds in MPM remains unknown, mRNA expression of AURKA is a prognostic marker that could be used in clinical practice to help stratify patients with epithelioid MPM.
Additionally, the poor-prognosis group also exhibited upregulation of the PI3K and mTOR signaling pathways. Preclinical studies have reported this finding, leading to clinical trials with a low response rate (2%) and no significant survival benefit in the salvage setting (54). These data suggest that combination therapies with mTOR inhibition may be necessary. Our analysis of epigenetic alterations revealed some associations between BAP1 status and DNA methylation (Supplementary Table S6). A recently established functional link between BAP1 and EZH2 (22) provided the rationale for the recently completed clinical trial of the EZH2 inhibitor tazemetostat in BAP1-null MPM (NCT028602). By contrast, no relevant findings resulted from our analysis of viral and microbial sequences (Supplementary Table S7). Finally, another genomically driven potential vulnerability in MPM is very frequent CDKN2A deletion associated with codeletion of MTAP, the latter recently shown to metabolically lead to impaired activity, and therefore sensitization to further inhibition, of the arginine methyltransferase PRMT5 (12, 13), an inhibitor of which is currently being evaluated in a phase I trial (NCT02783300).
Harnessing current immunotherapy approaches to improve outcomes of patients with MPM is an area of intense clinical interest. Although our study confirms that MPMs show a low tumor mutation burden and therefore may present a more challenging setting for immunotherapy, a remarkable and novel finding of the present study is that of strong expression of the immune-checkpoint gene VISTA in epithelioid MPM, on the tumor cells themselves, unlike other cancer types where it is more often expressed on infiltrating reactive cells (42). VISTA is a member of the B7 family of negative checkpoint regulators that is expressed primarily on infiltrating tumor macrophages, and whose immune restraining effects on T cells may be similar to those of PD-1 (43, 55). However, unique structural features mean that VISTA can repress activation of T cells both as a ligand present on the surface of APC cells and as a receptor on the surface of T cells (56). Because VISTA is expressed on MPM cells, and its mRNA expression levels do not correlate with overall mutation load, our results raise the possibility that VISTA expression may be restraining antitumor immune responses in a subset of MPM cases. As we find that nonneoplastic mesothelium also expresses VISTA protein, we speculate that VISTA expression in MPM is retained or possibly selected for by immune pressure. This is consistent with previous publications demonstrating that MPM is an “immunogenic” tumor (57), including recent trials showing some responses to immune-checkpoint blockade therapy (58, 59). Indeed, VISTA has recently been reported as a possible compensatory immune-inhibitory pathway in prostate cancers that fail to respond to ipilimumab (60). Our findings thus provide both a rationale and a candidate biomarker for clinical trials of emerging anti-VISTA therapy (refs. 42, 44; NCT02812875) in epithelioid MPM. Taken together, our findings point to new lines of investigation into the biology of MPM with the potential to lead to new therapeutic strategies.
TCGA Project Management has collected necessary human subjects documentation to ensure the project complies with 45-CFR-46 (the “Common Rule”). The program has obtained documentation from every contributing clinical site to verify that Institutional Review Board (IRB) approval has been obtained to participate in TCGA. Such documented approval may include one or more of the following:
An IRB-approved protocol with informed consent specific to TCGA or a substantially similar program. In the latter case, if the protocol was not TCGA-specific, the clinical site principal investigator provided a further finding from the IRB that the already-approved protocol is sufficient to participate in TCGA.
A TCGA-specific IRB waiver has been granted.
A TCGA-specific letter that the IRB considers one of the exemptions in 45-CFR-46 applicable. The two most common exemptions cited were that the research falls under 46.102(f)(2) or 46.101(b)(4). Both exempt requirements for informed consent, because the received data and material do not contain directly identifiable private information.
A TCGA-specific letter that the IRB does not consider the use of these data and materials to be human subjects research. This was most common for collections in which the donors were deceased.
A detailed description of the sample acquisition and pathology review process, as well as the experimental and computational methods used for the different analyses presented in our study, is provided as Supplementary Sections 1–13.
Disclosure of Potential Conflicts of Interest
L.R. Chirieac has received remuneration for medicolegal work related to mesothelioma. V.W. Rusch reports receiving a commercial research grant from Genentech and is a consultant/advisory board member for BMS. H. Pass reports receiving a commercial research grant from Genentech. M.G. Zauderer is a member of the Board of Directors of the Mesothelioma Applied Research Foundation, reports receiving commercial research grants from Bristol-Meyers Squibb and Millennium, and is a consultant/advisory board member for AstraZeneca, Sellas Life Science, and Epizyme. D.J. Kwiatkowski is a consultant/advisory board member for Novartis. R. Bueno has ownership interest (including stock, patents, etc.) in Navigation Sciences. A.D. Cherniack reports receiving commercial research support from Bayer. No potential conflicts of interest were disclosed by the other authors.
Co-chairs: Marc Ladanyi, Peter J. Campbell, Bruce W. S. Robinson
Manuscript Coordinator: Julija Hmeljak
Analysis Coordinator: Ina Felau
Data Coordinator: Francisco Sanchez-Vega
Project Manager: Ina Felau
Copy-number: Andrew D. Cherniack, Juliann Shih
mRNA: Katherine A. Hoadley,
Methylation: Luda Danilova
miRNA: Reanne Bowlby, Gordon Robertson, Ewan Gibb
Exomes: Manaswi Gupta, Chip Stewart, Patrick Tarpey
RPPA: Rupa Kanchi, Rehan Akbani
Transcription factor analysis: Hatice U. Osmanbeyoglu
PARADIGM: Yulia Newton, Kiley Graim,
Firehose: David Heiman, Michael Noble
iCluster: Esther Drill, Ronglai Shen; David T. Severson, Raphael Bueno
Regulome explorer: Lisa Iype, Vésteinn Thorsson, David L. Gibbs,
EMT analysis: Lauren A. Byers, Carl Gay, Lixia Dao
Tissue Source Sites: Michael Feldman, Harvey Pass, Nico van Zandwijk, Raphael Bueno, Anne Tsao, Marc Ladanyi, Robert Rintoul, Hedy Kindler, Rajiv Dhir, Bruce Robinson, Andre Carvalho, Troy Shelton
Pathology Subgroup: William D. Travis, Sanja Dacic, Lucian Chirieac, Françoise Galateau-Sallé, Junya Fujimoto, Aliya Husain, Marc Ladanyi
VISTA IHC: Patrice Desmeules, Achim Jungbluth
Clinical Subgroup: Marjorie Zauderer, Anne Tsao, David J. Kwiatkowski, Hedy Kindler, Harvey Pass, Jenette Creaney, Valerie Rusch, Robert C. Rintoul, Tara Lichtenberg, Bruce W. S. Robinson, Nathan Pennell, Henrique Silveira,
Genomic near-haploidization validation datasets: Gloria Ravegnini, Jonathan Fletcher, Hiroyuki Aburatani
The TCGA research network contributed collectively to this study. M.L., B.W.R.S. and P.J.C. oversaw data analyses and interpretation. J.H., F.S.-V., M.L., P.J.C., M.Z., R.S., B.W.S.R. and A.T. wrote the manuscript.
We thank Drs. Elaine Mardis and Matthew Meyerson for invaluable feedback and Dr. Lee Spraggon for helpful comments and assistance with figures. This work was supported by the following grants from the USA National Institutes of Health: U54 HG003273, U54 HG003067, U54 HG003079, U24 CA143799, U24 CA143835, U24 CA143840, U24 CA143843, U24 CA143845, U24 CA143848, U24 CA143858, U24 CA143866, U24 CA143867, U24 CA143882, U24 CA143883, U24 CA144025, P30 CA016672, and P30 CA008748. Whole-exome sequencing and analysis was funded by the British Lung Foundation grant APG11-7 (“The genomic architecture of mesothelioma”). Publication charges were funded by the Memorial Sloan Kettering Mesothelioma Program Fund.
The members of The Cancer Genome Atlas Research Network are Hiroyuki Aburatani, Rehan Akbani, Adrian Ally, Pavana Anur, Joshua Armenia, J. Todd Auman, Miruna Balasundaram, Saianand Balu, Stephen B. Baylin, Michael Becich, Carmen Behrens, Rameen Beroukhim, Craig Bielski, Tom Bodenheimer, Moiz S. Bootwalla, Jay Bowen, Reanne Bowlby, Denise Brooks, Raphael Bueno, Lauren Averett Byers, Flávio M. Cárcano, Rebecca Carlsen, Andre L. Carvalho, Andrew D. Cherniack, Dorothy Cheung, Lucian Chirieac, Juok Cho, Eric Chuah, Sudha Chudamani, Carrie Cibulskis, Leslie Cope, Daniel Crain, Jenette Creaney, Erin Curley, Sanja Dacic, Ludmila Danilova, Assunta De Rienzo, Timothy DeFreitas, John A. Demchok, Noreen Dhalla, Rajiv Dhir, Lixia Diao, Esther Drill, Ina Felau, Michael Feldman, Martin L. Ferguson, Jonathan A. Fletcher, Junya Fujimoto, Junya Fujimoto, Shiro Fukuda, Stacey B. Gabriel, Françoise Galateau Sallé, Jianjiong Gao, Johanna Gardner, Julie M. Gastier-Foster, Carl M. Gay, Nils Gehlenborg, Mark Gerken, Gad Getz, Ewan A. Gibb, David L. Gibbs, Chandra Goparaju, Kiley Graim, Benjamin Gross, Guangwu Guo, Manaswi Gupta, Seiki Hasegawa, David Haussler, D. Neil Hayes, David I. Heiman, Zachary Heins, Julija Hmeljak, Katherine A. Hoadley, Robert A. Holt, Alan P. Hoyle, Aliya Husain, Carolyn M. Hutter, Lisa Iype, Stuart R. Jefferys, Steven J.M. Jones, Corbin D. Jones, Rupa S. Kanchi, Katayoon Kasaian, Jaegil Kim, Hedy Kindler, Nobuyuki Kondo, Thomas Krausz, Ritika Kundra, Kozo Kuribayashi, David J. Kwiatkowski, Marc Ladanyi, Phillip H. Lai, Peter W. Laird, Michael S. Lawrence, Darlene Lee, Kristen M. Leraas, Tara M. Lichtenberg, Pei Lin, Jia Liu, Wenbin Liu, Eric Minwei Liu, Laxmi Lolla, Adhemar Longatto-Filho, Yiling Lu, James Luketich, Yussanne Ma, Dennis T. Maglinte, David Mallory, Marco A. Marra, Michael Mayo, Sam Meier, Jonathan Melamed, Shaowu Meng, Matthew Meyerson, Piotr A. Mieczkowski, Gordon B. Mills, Richard A. Moore, Cesar Moran, Scott Morris, Lisle E. Mose, Andrew J. Mungall, Karen Mungall, Takashi Nakano, Rashi Naresh, Yulia Newton, Michael S. Noble, Angelica Ochoa, Hatice Osmanbeyoglu, Joel S. Parker, Harvey I. Pass, Joseph Paulauskis, Arjun Pennathur, Nathan A. Pennell, Robert Penny, Charles M. Perou, Todd Pihl, Nilsa C. Ramirez, Doris M. Rassl, Gloria Ravegnini, Glen Reid, Rui M. Reis, Sheila M. Reynolds, David Rice, William G Richards, Robert C. Rintoul, Jeffrey Roach, A. Gordon Robertson, Valerie Rusch, Sara Sadeghi, Gordon Saksena, Francisco Sanchez-Vega, Chris Sander, Ayuko Sato, Cristovam Scapulatempo-Neto, Jacqueline E. Schein, Nikolaus Schultz, Steven E. Schumacher, Tanguy Seiwert, Yoshitaka Sekido, David T Severson, Candace Shelton, Troy Shelton, Ronglai Shen, Robert Sheridan, Yan Shi, Juliann Shih, Yuichi Shiraishi, Ilya Shmulevich, Henrique C. S. Silveira, Janae V. Simons, Payal Sipahimalani, Tara Skelly, Heidi J. Sofia, Matthew G. Soloway, Paul Spellman, Chip Stewart, Josh Stuart, Qiang Sun, Jumpei Takeshita, Angela Tam, Donghui Tan, Roy Tarnuzzer, Kenji Tatsuno, Barry S Taylor, Nina Thiessen, Eric Thompson, Vesteinn Thorsson, William D. Travis, Anne Tsao, Kane Tse, Tohru Tsujimura, Federico Valdivieso, David J. Van Den Berg, Nico van Zandwijk, Umadevi Veluvolu, Luciano S. Viana, Douglas Voet, Yunhu Wan, Zhining Wang, Jing Wang, Joellen Weaver, John N. Weinstein, Daniel J. Weisenberger, Matthew D. Wilkerson, Lisa Wise, Ignacio Wistuba, Tina Wong, Ye Wu, Shogo Yamamoto, Liming Yang, Marjorie G. Zauderer, Jean C. Zenklusen, Jiashan Zhang, Hailei Zhang, Hongxin Zhang, and Erik Zmuda.