SWI/SNF mutations are common across diverse cancer types
To survey the spectrum of SWI/SNF mutations in human cancers, we analyzed data from 24 whole-exome studies 
together spanning 18 diverse cancer types (see Methods
). Selected characteristics of the 24 studies are summarized in . More detailed information, including characteristics of the sequencing platform, fold-sequencing coverage, and genome-wide mutation frequencies (by mutation type and predicted impact) are summarized in Table S1
. The mutational status of the 20 genes encoding canonical subunits of human SWI/SNF is detailed in Table S2.
Given the size of the SWI/SNF genomic “footprint” (spanning 20 genes), it might be argued that SWI/SNF is prone to passenger mutations that could inflate the mutational frequency of the complex. To address this concern, we compared the distribution of mutation types in SWI/SNF subunit genes to that of the whole exome (). Our analysis revealed a notable skew of mutations in SWI/SNF genes, with a significantly increased fraction of predicted deleterious mutations (frameshift, nonsense, rearrangement, splice-site, and missense-damaging) compared to predicted missense-benign and synonymous mutations (P
, chi-square test). This pattern suggests that most observed SWI/SNF mutations are likely driver alterations.
Indeed, rather than overestimating the frequency of SWI/SNF inactivation (due to some level of passenger mutations), the sequencing analysis likely underestimates the true frequency of SWI/SNF inactivation. Evidence exists that genomic DNA deletions, rearrangements, and epigenetic silencing provide alternative mechanisms to inactivate SWI/SNF subunits 
. Moreover, the impact of mutations on protein expression and function has not been adequately explored. Only one of the exome studies analyzed here also evaluated protein levels, finding ARID1A
mutations associated with reduced or lost ARID1A expression (by immunohistochemistry) in gastric cancer 
(and the same had been separately shown for ovarian clear cell carcinoma 
). Systematic efforts to survey all genetic, epigenetic and protein changes would be needed to arrive at the true frequency of SWI/SNF alterations.
SWI/SNF mutations in specific cancer types
Due to its high mutation frequency (36–75%; ), a likely tumor suppressive role of the SWI/SNF complex had been recognized by the respective study authors in ovarian clear cell carcinoma, clear-cell renal cell carcinoma, hepatocellular carcinoma, gastric cancer, and pancreatic cancer 
. Nonetheless, nearly all of those studies highlighted only a single highly-mutated subunit (e.g. ARID1A
mutation in ovarian clear cell carcinoma), whilst our analysis also uncovered less frequent mutations of other SWI/SNF subunits in those same tumor types (Table S2).
Notably, the mutational data implicating a tumor suppressive role of SWI/SNF (mutation frequencies 11–34%; ) are also compelling for melanoma, diffuse large B-cell lymphoma (DLBCL), multiple myeloma, glioblastoma, and head and neck cancers, but had not been appreciated. In these cancers, SWI/SNF mutations likely went unnoticed because they were spread across multiple SWI/SNF subunits, none by itself reaching a critical threshold. Among this group of cancers, melanoma exhibited the highest SWI/SNF mutation rate.
While melanomas have an inherently high mutation rate from UV exposure, the mutations noted here display characteristics of tumor suppressor driver mutations. Among 29 cases sequenced 
, 17 nonsynonymous mutations struck ARID1A
1), and SMARCC1
1) (Table S2). These include a homozygous mutation in ARID2
and three mutations targeting SMARCB1
in the same patient sample, thus likely affecting both alleles. Furthermore, the mutation types included 5 nonsense mutations, 9 probably-damaging missense mutations (as called by polyphen-2 
), 1 possibly-damaging missense mutation, and 2 benign missense mutations. Only one of the three melanoma studies 
reported on synonymous mutations, where there were 2 synonymous mutations (one each for SMARCA4
) compared to 7 nonsynonymous mutations (the nonsynonymous: synonymous mutation ratio in the melanoma exome was 1.9:1). Given the synonymous mutations, and the relatively high mutation rate in melanoma, there is likely some background passenger mutation rate of SWI/SNF in melanoma. Nonetheless, the loss of heterozygosity (LOH; implied by the homozygous mutation and multiple mutations in the same gene and sample), the damaging mutation skew, and the recurrence of mutations for several subunits altogether suggest that many or most of these mutations are drivers.
In 68 diffuse large B-cell lymphomas 
, nonsynonymous mutations targeted ARID1B
1), and SMARCD3
1). Those mutations can be classified into the following types: nonsense (n
3), frameshift (n
1), splice-site (n
1), probably-damaging missense (n
2), benign-missense (n
1), and missense mutations of undetermined significance (n
4). LOH information was not available from the two DLBCL studies. One study 
reported on synonymous mutations, and only 1 synonymous mutation occurred in ARID2
compared to 9 nonsynonymous mutations across several other SWI/SNF subunit genes. The high frequency of mutations, the recurrence within genes, and the deleterious skew of mutations all suggest a tumor suppressive role of SWI/SNF in DLBCL.
From 38 multiple myelomas that were sequenced 
, six had mutations in six different SWI/SNF subunits, broken down as follows: 2 rearrangements, 2 probably-damaging missense mutations, and 2 benign-missense mutations. The frequency of synonymous mutations and LOH information was not available from this study. Thus, the case for SWI/SNF as a tumor suppressor relies mostly on the frequency of SWI/SNF mutations. Notably, the background mutation rate was not particularly high for multiple myeloma, and consequently the SWI/SNF mutations are unlikely to all represent merely passenger events.
In glioblastoma multiforme (GBM) 
, 4 of 22 samples harbored SWI/SNF mutations. One sample had two different mutations in the same gene (ARID1A
) suggesting mutations striking both alleles and making a strong case that those mutations are driver mutations. SMARCA4
, and SMARCC2
each had a single probably-damaging missense mutation, though SMARCC2
also had a synonymous mutation. Larger validation efforts will be necessary, but the overall mutational pattern is suggestive of driver mutations in GBM.
Head and neck cancers had a total of 12 mutations out of the 106 samples sequenced in two studies 
. The mutational frequency in head and neck cancers is relatively high due to tobacco exposure in a subset of patients. Nonetheless, the 12 mutations hitting ARID1A
, and SMARCC2
can be broken down as follows: 3 nonsense, 1 frameshift, 4 probably-damaging missense, 2 possibly-damaging missense, and 2 benign-missense mutations, representing a skew towards deleterious mutations relative to exome-wide mutational statistics. Some genes were recurrently mutated, including ARID1A
, and PBRM1
. Information regarding LOH and synonymous mutations were not available from these studies.
Medulloblastoma, breast cancer, and chronic lymphocytic leukemia (CLL) all showed lower but likely significant SWI/SNF mutation rates (4–10%; ). In the case of medulloblastoma 
, there was a frameshift mutation in ARID1A
and a possibly-damaging missense mutation in SMARCA4
. In addition to the mutations from the full exome sequencing data, SMARCA4
was found mutated in 2 additional samples from a validation cohort associated with the same study 
. Furthermore, three ARID1A
mutations have been reported in separate validation efforts 
As for breast cancer 
, only a single damaging missense mutation was identified in ARID1B
. However, the sample set was small and not reflective of known breast cancer heterogeneity (all 11 samples from the study were triple negative breast cancer, i.e. negative for estrogen receptor, progesterone receptor and Her2); thus conclusions should be tempered. Nevertheless, other reports have identified ARID1A
mutations in breast cancer 
, suggesting a likely tumor suppressive role of the complex.
In the case of CLL 
, 4.5% of cases harbored SWI/SNF mutations. Although relatively low, this frequency is likely meaningful for several reasons. First, there were 196 CLL cases sequenced between the two studies, making this one of the higher powered cancer types included in this analysis, and 8 of these cases had a mutation in a SWI/SNF subunit. Two mutations each hit ARID1A
, suggesting some level of recurrence within subunits. Of the eight total mutations, 1 was a nonsense mutation, 5 were probably-damaging missense mutations, and 2 were predicted benign-missense mutations, again suggesting a skew towards damaging mutations (synonymous mutations were not reported in these two studies). Importantly, the overall mutation rate for CLL is relatively low; the average CLL case had only 15 mutations, corresponding to a mutation rate of less than 1 mutation/Mb of exome sequenced. Furthermore, the single-most mutated gene in CLL, SF3B1
, was itself mutated in only 15% of cases. Thus, the observed SWI/SNF mutations, albeit uncommon, are likely meaningful.
No SWI/SNF mutations were identified in colon cancer, myelodysplasia, oligodendroglioma, pancreatic neuroendocrine tumors, and pancreatic cysts 
. It is worth noting that these neoplasms, with the exception of colon cancer, tend to be less aggressive or even benign. However, it is possible that SWI/SNF mutations do occur but were not evident because the studies were under-powered, or because the sample sets were biased. In that regard, the colon cancers sequenced appear to all be microsatellite stable (based on the low overall mutation frequencies), and thus not representative of all colon cancer subtypes. Intriguingly, SWI/SNF mutations in gastric cancers tended to occur in microsatellite instable (MSI) tumors 
. Indeed, targeted resequencing of ARID1A
across several cancer types suggests that it is inactivated in a high fraction of MSI colon cancer cases 
. Furthermore, SMARCC1
has been reported to be deficient in a single colon cancer cell line 
. Thus the evidence overall suggests that SWI/SNF may play a role in colon tumorigenesis.
The pancreatic cyst study 
sequenced eight samples of serous cystadenomas (SCAs), intraductal papillary mucinous neoplasms (IPMNs), mucinous cystic neoplasms (MCNs), and solid pseudopapillary neoplasms (SPNs). IPMNs and MCNs are the only cysts with the capacity to evolve to frank adenocarcinoma; yet they are not the canonical pancreatic intraepithelial neoplasia (PanIN) lesions that typically precede pancreatic ductal adenocarcinoma (PDAC) 
. In contrast to the exome data, loss of SMARCA4 expression at the protein level has been reported in a subset of IPMNs 
. More work is needed to characterize the status of SWI/SNF in pancreatic cancer precursor lesions, and for that matter the timing of SWI/SNF mutations during the development and progression of other human cancers.
One remarkable finding of our analysis is the breadth of tumor types in which SWI/SNF is mutated – from brain, to hematopoietic, to various solid epithelial cancers. Albeit different tumor types exhibit different SWI/SNF mutation frequencies and subunit preferences (discussed more below), SWI/SNF mutations on the whole are not confined to any tissue origin, histologic, or molecular subtype of cancer. Mechanistically, this raises the possibility that SWI/SNF could work through a general tumor suppressor pathway(s) rather than any lineage-specific pathway.
SWI/SNF mutations preferentially target certain subunits
The above data indicate that SWI/SNF subunits are frequently mutated in human cancers. We next asked whether certain subunits in the complex were preferentially targeted. The SWI/SNF complex subunits can be assigned to roughly three functionally distinct groups – an enzymatic subunit (SMARCA2 or SMARCA4), a subunit thought to confer functional specificity to the complex (hereafter referred to as targeting subunits; ARID1A, ARID1B or PBRM1), and the remaining core and variant subunits (hereafter referred to as scaffolding subunits) 
. Across the 13 cancer types with SWI/SNF mutations, the preponderance of mutations occurred in the SMARCA4 enzymatic subunit and in the three targeting subunits (). Mutations did occur but less commonly in scaffolding subunits.
Some SWI/SNF subunits are preferentially mutated.
The finding that mutations occur across several different SWI/SNF subunits suggests that the main impact of mutations may be to compromise in part or whole the functional activity of the complex. The preponderance of mutations in the enzymatic and targeting subunits suggests that those subunits may be most critical to SWI/SNF function. Consistent with this interpretation, germline mutations of multiple SWI/SNF subunits were recently found to underlie Coffin-Siris syndrome (CSS; a rare developmental disorder) 
, implying a genetic equivalence of different subunits. Of 16 SWI/SNF subunits sequenced across 23 individuals with CSS, mutations were found in SMARCA4
(4%), and SMARCE1
. Notably, the set of affected SWI/SNF subunits to a large extent mirrors that of human cancers, supporting that certain enzymatic and targeting subunits are likely most critical to the function of the complex. Nonetheless, there is likely to be additional subtlety with regard to possible distinct functions of SWI/SNF complexes with different subunit compositions, cell and tissue-type specificity of those complexes, and in the case of mutations, possible compensatory activity of residual SWI/SNF complexes (containing non-mutated alternative subunits).
Indeed, viewing each of the 13 cancer types separately, interesting mutational patterns emerge (). Some cancer types exhibit mutations predominantly in a single SWI/SNF subunit, including (and as also noted by the authors in those studies) ARID1A
in ovarian clear cell carcinoma and gastric cancer, and PBRM1
in renal cell carcinoma. Most others, including melanoma, pancreatic cancer, and DLBCL, exhibit a more balanced spectrum of mutations among the commonly mutated subunits. For those tumor types where a single subunit is predominantly affected, it is possible that the subunit (and the complexes containing it) has cell or tissue-type specific functions that account for its selective inactivation. Such is almost certainly the case for the finding of SMARCB1
(SNF5) mutations in all rhabdoid tumors 
. Alternatively, cell or tissue-type specific mutational processes (e.g., relating to access of genomic loci) might be operating.
An interesting question is whether within any particular SWI/SNF subunit gene, mutations affect specific residues or structural/functional domains. Data from the exome studies analyzed here did not reveal obvious mutation “hotspots” (Figure S1
). However, the data are too sparse to draw firm conclusions. We note that a few validation studies, evaluating single SWI/SNF subunits (e.g. ARID1A
) in much larger cohorts, have also not observed mutation hotspots 
. In this respect, SWI/SNF appears to differ from TP53
, where mutations disproportionally target a small number of codons, and mostly occur within a single (in this case, DNA-binding) domain 
We also investigated, across the cancer types, whether mutations of different SWI/SNF subunits were mutually-exclusive of one another. Somewhat surprisingly, mutations in two different SWI/SNF subunits occurred within the same patient tumor about as often as would be expected by chance (i.e. as estimated by the square of SWI/SNF mutation frequency in a given tumor type) (). This finding suggests that mutational hits in two different SWI/SNF subunits are not functionally redundant, but rather that each might provide incremental perturbation or disruption of the complex.
Co-occurrence of mutated SWI/SNF subunits.
SWI/SNF mutations are not mutually-exclusive of other cancer gene mutations
The tumor suppressive function of SWI/SNF has been proposed to operate by controlling the expression or activity of specific genes and pathways, including Rb, TP53, Polycomb, sonic hedgehog, Myc, stem cell programs, and nuclear hormone receptor signaling 
. Our exome analysis provided a unique opportunity to try to systematically identify the key pathways mediating SWI/SNF tumor suppression, by mutual exclusivity analysis. Specifically, two different genes operating along in the same linear pathway, e.g. KRAS
, are thought less likely to be mutated in the same tumor sample because the mutations would be functionally redundant. Thus, identifying cancer genes that are mutated only in tumors without SWI/SNF mutation would imply a shared pathway. Likewise, identifying cancer genes that are always mutated in tumors with SWI/SNF mutation (mutually inclusivity) might suggest necessary cooperating pathways.
To address mutual exclusivity, we first focused on the relation between SWI/SNF and mutations of TP53
. Recent studies reported mutual exclusivity of ARID1A
mutations in both ovarian clear cell carcinoma and gastric cancer 
. Our analysis of exome datasets affirmed a mutually-exclusive relationship between SWI/SNF and TP53
mutations in ovarian clear cell carcinoma and gastric cancer (); though statistical significance was only reached for gastric cancer (P
0.018; Fisher's exact test). Notably, however, no such mutually exclusive relationship was apparent for other tumor types, including pancreatic cancer, melanoma, hepatocellular carcinoma, and DLBCL (). Indeed, in pancreatic cancer, all cases with SWI/SNF mutations actually carried TP53
mutations, suggesting a trend towards mutual inclusivity (P
0.085; Fisher's exact test).
Furthermore, the data suggest a need for caution in interpreting the mutual-exclusivity of SWI/SNF and TP53
mutations. Ovarian and gastric cancers are both histological and genetically diverse diseases, and mutual exclusivity may rather correlate with tumor subtypes rather than reflect a mechanistic relation. Indeed, in gastric cancer SWI/SNF mutations tend to occur in MSI tumors, whereas TP53
mutations generally occur in microsatellite stable tumors 
. Thus, mutual exclusivity here may relate more to distinct mutagenic processes.
SWI/SNF has also been proposed to suppress tumor growth by antagonizing the oncogenic effects of polycomb repressive complex 2 (PRC2), mirroring its role in development 
. Approximately 15% of DLBCL cases harbor activating mutations of EZH2
, the enzymatic component of PRC2. Thus, DLBCL provides an opportunity to assess the mutual exclusivity of SWI/SNF and PRC2 alterations. Notably, our analysis revealed several patient samples having both SWI/SNF and EZH2
mutations (), not supportive of mutual exclusivity.
We next sought to take a more systematic approach to identify cancer gene mutations exhibiting mutual exclusivity with SWI/SNF mutation. To this effect, we analyzed the top 189 mutated genes (all genes with ≥13 mutations) across the 24 exome studies. The 189 genes included other well-known cancer genes (e.g. KRAS
, etc) and represented many of the canonical signaling pathways (e.g. Ras, PI3K, Wnt, Notch, etc) in cancer. In spite of this, no significant mutually-exclusive (or mutually inclusive) relationships with SWI/SNF mutations were identified ( and Text S1
SWI/SNF mutations are not mutually exclusive of mutations in other commonly mutated genes.
It is possible that our analysis was under-powered to identify true mutually-exclusive relationships. Alternatively, it is possible (and we favor the explanation that) SWI/SNF rather effectuates tumor suppression by impacting multiple pathways, including Rb, TP53, Polycomb, sonic hedgehog, Myc, stem cell programs, nuclear hormone receptor signaling, and likely others that remain to be discovered. This “one-to-many” relationship would obscure mutual-exclusivity analysis. For example, mutations might occur in both PRC2 (EZH2) and SWI/SNF because they occur first in PRC2 and the later-occurring SWI/SNF mutation also impacts other pathways, or because the SWI/SNF mutation only partially activates PRC2 (which is further activated by EZH2 mutation).
Several lines of evidence support the multi-functioning of SWI/SNF in tumor suppression. First, SWI/SNF mutations occur across tumors from diverse tissue types, suggesting they impact one or more broadly conserved processes. Second, recent studies seeking to map the genome binding sites of SWI/SNF (for example by chromatin-immunoprecipitation and deep-sequencing) have revealed on the order of 5,000–10,000 binding sites 
, and thus SWI/SNF likely impacts many genes and pathways. Finally, it has long been known that SWI/SNF controls diverse biological processes (e.g. mating type switching and sugar fermentation in yeast), and it is thus reasonable to think that the same might be true of relevant processes (e.g. growth, survival, metabolism) in cancer cells.