Landscape of hotspot mutations in primary human cancer

We collected the mutational data from the sequenced exomes and genomes of 11,119 human tumors in 41 tumor types (Supplementary Table 1). These originate from diverse sources including large international consortia and various published studies. This cohort represents a broad range of primary human malignancies with three or more tumor types in each of nine major organ systems (Fig. 1a). The repository consists of 2,007,694 somatic substitutions in protein-coding regions with a median of 57 mutations (25 and 125 mutations; 25th and 75th percentile, respectively) per tumor-normal pair with significant variability in mutation rates among and between tumors and types4,16. In total, 19,223 human genes harbor at least one somatic mutation in this data set.

Figure 1: Mutational data and hotspot detection. (a) The distribution of tumor types included in this analysis. CLL, chronic lymphocytic leukemia; ALL, acute lymphoblastic leukemia; PNET, pancreatic neuroendocrine tumors. (b) Breakdown of known and classified novel hotspots and genes. (c) The number of hotspots in each of 49 genes with two more hotspots detected across the cohort. At right, a summary of hotspots identified. Novel hotspots are in blue boldface. (d) The distribution of mutations and hotspots in six oncogenes refines known patterns and reveals new hotspots. Full size image

Here, we define a mutational hotspot as an amino acid position in a protein-coding gene that is mutated (by substitutions) more frequently than would be expected in the absence of selection. In this analysis, we focus exclusively on individual substitutions rather than other somatic abnormalities such as translocations, amplifications, deletions or epigenetic modifications. To identify mutational hotspots, including low-incidence mutations, we developed a binomial statistical model that incorporates several aspects of underlying mutational processes including nucleotide context mutability, gene-specific mutation rates and major expected patterns of hotspot mutation emergence (Supplementary Figs. 1a and 2, Supplementary Code and Online Methods). As considerable variability exists in the methods and standards for mutation calling used by individual studies and centers, we also developed several evidence-based criteria for eliminating probable false-positive hotspots (Online Methods and Supplementary Fig. 1b). In total, we identified 470 statistically significant hotspots (q < 0.01) affecting 275 protein-coding genes (Supplementary Tables 2 and 3). Overall, more than half of all hotspots were determined to be novel (Fig. 1b, Table 1 and Supplementary Table 2) and 54.8% of all tumors assessed here possessed one or more hotspot mutations.

Table 1: Select new hotspots in cancer genes Full size table

Most affected genes possessed only a single hotspot (Supplementary Fig. 3a). A subset of genes, however, possessed many hotspots of varying frequency. In total, 49 genes possessed two or more hotspots (Fig. 1c), with many of these also arising in the greatest number of tumor types (Supplementary Fig. 3b). TP53 R248 was the most disseminated hotspot, observed in 25 tumor types. Among a subset of even well-characterized oncogenes, a pattern of both known and novel hotspots emerged (Fig. 1d). Moreover, the number of observed mutant amino acids at a given hotspot generally increases with its mutational frequency across tumors types (Supplementary Fig. 3c), though 35% (n = 164) of hotspots mutate to only a single variant amino acid. In most genes, hotspots bear only a fraction of the total mutational burden across the gene, whereas in a subset of cancer genes, the dominant mutational hotspot constitutes the vast majority of mutations independent of total mutational burden (Fig. 1d and Supplementary Fig. 3d). Overall, we identified considerable variability in the patterns of mRNA expression of individual hotspots in even canonical oncogenes (Supplementary Fig. 4), indicating that levels of expression are often not correlated with the biological importance of known activating mutations.

The patterns by which some hotspots emerge support new clinical paradigms for testing targeted agents. Some hotspots that dominate the mutational landscape in one or a few cancer types also arise as uncommon subsets of many others. For instance, IDH1 R132 is most common in low-grade gliomas, glioblastomas, acute myeloid leukemias (AMLs) and cutaneous melanomas; but it is also present in 1 to 6 tumors in each of 11 additional cancer types. AKT1 E17K arises in greatest numbers in breast cancer, but also in 1 to 3 tumors of 10 additional cancer types. The distribution of CREBBP R1446 mutations is qualitatively different. They were originally identified in relapsed acute lymphoblastic leukemias17, but in this cohort of mostly primary disease, we find that they arise in only a small minority (1–3; 0.17–1.7%) of many (11) cancer types. Such patterns reaffirm the value of basket study designs that test mutation-specific inhibitors in early-phase clinical trials, where enrollment is based on specific mutations in patients instead of tissue of origin.

A lineage map of all hotspots in genes with at least one common hotspot (Fig. 2a and Supplementary Fig. 5) indicates most hotspots are defined more by the tissue types rather than the organ systems in which they arise. Of all hotspots, 81% arise in two or more tumor types, suggesting that many hotspot mutations may confer a growth advantage across diverse lineages. Indeed, of hotspots present in multiple tumor types, only 7.6% (n = 36) are confined to a single organ system (Table 2). Thus, hotspot mutations that arise in a single tumor type may reflect organ-specific growth advantages, but they represent only a small minority of all hotspot mutations in cancer. Likewise, a subset of hotspots arises in a cell type–specific manner. Twenty-seven hotspots (5.7%) were more frequently mutated in tumors of a squamous cell lineage (Supplementary Fig. 6), the most significant of which were MAPK1 E322 and EP300 D1399 (q = 6 × 10−13 and 1 × 10−11, respectively, χ2) and may potentially confer a squamous cell type–specific growth advantage.

Figure 2: Lineage landscape of hotspot mutations. (a) Both common and rare hotspots are largely disseminated across a broad range of malignancies. All hotspots detected in genes with at least one hotspot affecting >5% of tumors of one or more tumor types are shown. Novel hotspots are in blue boldface. Genes are grouped broadly by functional similarity, hotspots are ordered by amino acid position, and tumor types (columns, labeled at bottom) are sorted according to the fraction of tumors affected by one or more hotspots overall (b). The percent of samples altered is represented by colored squares and indicated text. Hotspots in tumor suppressors TP53, PTEN, APC and FBXW7 were excluded here (Supplementary Fig. 5). (b) The fraction of tumors of a given type (as indicated) affected by one or more hotspots. Black circles represent the median mutation rate (right axis) in the indicated tumor type (bar is the median absolute deviation). Shown at top is the number of tumors of each type with a hotspot mutation affecting a known or candidate oncogene1. Full size image

Table 2: Organ system-specific hotspots Full size table

Overall, the presence, type and frequency of hotspots by tumor type vary widely (Fig. 2b). In some tumor types, a large proportion of tumors possess one or more hotspot mutations including a substantial fraction of tumors with a hotspot in a candidate oncogene (Fig. 2b, top). Conversely, other tumor types never or rarely possess a tumor defined by a hotspot identified here. Some of these differences are certainly attributable to the fact that hotspots are only one of many possible driver genomic aberrations, including specific gene fusions or focal amplifications and deletions. These other aberrations may define tumors of a given type, but they are not mutually exclusive with hotspots in many cancers. Other differences could not, alone, be explained by the overall mutational burden in these tumor types. For instance, uterine carcinosarcomas and prostate cancers have a similar mutation rate whereas there is a threefold greater frequency of hotspot-bearing tumors among the former. Likewise, whereas papillary thyroid and high-grade pontine gliomas have mutations rates similar to nasopharyngeal tumors and neuroblastomas, the former far more commonly bear hotspot mutations (Fig. 2b).

Unconventional hotspots

In addition to missense mutations, we identified a variety of unconventional hotspot mutations with varied impact. Among these were 13 splice-site hotspots. For each of these hotspots, an associated transcript abnormality was identified from RNA sequencing of affected tumors (exon skipping, intron retention, in-frame deletions; Supplementary Fig. 7a), including two previously characterized in-frame activating mutations (MET D1010_splice and PIK3R1 M582_splice, both exon 14 skipping events). We also identified 70 hotspots in 34 genes for which a nonsense mutation was among a diversity of changes at the affected residue, including 28 hotspots in which only a nonsense mutation was present (Supplementary Fig. 7b). Whereas nonsense mutations scattered throughout a gene may reflect a pattern of loss-of-function consistent with tumor-suppressor activity, a nonsense hotspot would appear to indicate the selection for the selective truncation of specific functional domains. Such events are consistent with the loss of some functions and the retention of others, as has been observed previously in genes such as PIK3R1, NOTCH1 and MET18,19. These hotspots aside, there was a depletion of nonsense mutations in hotspots in constitutively essential genes (P < 10−16, those genes predicted or experimentally verified to be essential across all cell and tissue types and developmental states20). Otherwise, the specific impact of nonsense hotspots is generally unknown and belies the disseminated pattern of truncating mutations in likely or proven tumor suppressors (Supplementary Fig. 7c).

Lineage diversity and mutant allele-specificity

The majority of hotspot mutations arose in diverse tumor types and organ systems, yet widespread differences exist among individual residues and mutant amino acids in hotspots, genes and tumor types (Fig. 3a). Examining the spectrum of KRAS mutations, which includes the most frequently mutated hotspot overall in our study (KRAS G12; n = 736 mutant tumors; Figs. 1d and 2a), clarified patterns only incidentally observed in the past. We found that gastric cancers were more similar to multiple myeloma in the preponderance of non-G12 mutations compared to endometrial, lung, colorectal and pancreatic tumors (P = 5.3 × 10−18; Supplementary Table 4). Only colorectal tumors had KRAS A146 mutations, whereas pancreatic tumors lacked G13 mutations (P s = 4 × 10−7 and 2.8 × 10−15, respectively). Many of these lineage-specific patterns were present at finer resolution as well. Among KRAS G12 mutations, the abundance of G12C mutations are highest in lung adenocarcinomas (P = 4 × 10−42), an event that may be associated with prognostic differences compared with non-G12C KRAS mutations21,22,23. Such mutant amino acid specificity was also apparent in pancreatic tumors, where KRAS G12R was more common than in any other tumor type (21% versus between 0 and 2.6%; χ2 P = 4.8 × 10−19). Gastric cancers, on the other hand, had the fewest G12V mutations among all KRAS G12-mutant tumor types, but the highest proportion of G12S (P = 0.007, Fig. 3b). There is a different balance among hotspots in the other Ras genes. Whereas papillary thyroid cancers nearly exclusively possessed codon Q61 mutations in HRAS and NRAS (P = 4 × 10−7), there was a higher prevalence of G12 and G13 codon mutations in these genes in AMLs, colorectal, bladder, and head and neck cancers, which together share few mutational processes in common (P = 4 × 10−10, Fig. 3a).

Figure 3: Lineage diversity and mutant allele specificity. (a) The fraction of cases mutated for each of the most common hotspots in eight frequently mutated genes in the most commonly mutated lineages indicate substantial lineage diversity and hotspot specificity. (b) Same as in a, but for KRAS G12 and IDH1 R132 mutations, showing that mutant amino acid specificity exists within individual hotspots across affected tumor types. (c) The fraction of clonal mutations, those present in 80% or more of the tumor cells of affected samples, was higher among mutations in hotspots versus all other nonrecurrent mutations in the same genes (χ2 P = 1 × 10−14). (d) The fraction of tumor cells mutated for PIK3CA E545 and PIK3CA H1047 hotspots in affected colorectal and uterine endometrial cancers indicates a pattern of allele-specific subclonality for E545 mutations in colorectal cancer. Full size image

Similar differences emerged in other driver cancer genes with multiple hotspots. V600E mutations describe nearly all BRAF hotspot mutations in melanoma, papillary thyroid and colorectal carcinomas, whereas multiple myelomas are similar to lung adenocarcinoma in which non-V600E hotspots predominate (P = 1.9 × 10−32). The balance between extracellular and kinase domain mutations in EGFR between brain tumors and lung adenocarcinoma (P = 3.3 × 10−12), respectively, have been documented previously and affect their biological impact and the efficacy of genotype-directed therapy10. ERBB2 followed a similar pattern, where extracellular domain mutations typified by S310F are far more common than are kinase domain mutations in bladder cancers compared to breast cancers (P = 0.006, Fig. 3a). Another notable gene was PIK3CA. Whereas bladder and cervical cancers are similar in their distribution of PIK3CA hotspot mutations, they vary significantly from breast cancers in the overall balance of helical to kinase domain mutations, possessing far fewer H1047R mutations among PIK3CA-mutated cases (P = 4.8 × 10−19). Endometrial and colorectal cancers also have a similar pattern of PIK3CA hotspots, but both have a higher prevalence of R88Q mutations than any other tumor type (P = 1.3 × 10−11; Fig. 3a). Such patterns extend beyond essential MAPK or PI3K signaling components, such as with SF3B1 K700 mutations that predominate in breast cancers and chronic lymphocytic leukemias whereas melanomas more frequently possess SF3B1 R625 mutations (P = 0.0001). Finally, mutant amino acid specificity was not limited to hotspots in Ras genes. The IDH1 R132H hotspot mutation predominated in multiple brain tumor types, but cysteine was the most common IDH1 R132 mutant amino acid in melanoma, which is unlikely to be exclusively related to UV light exposure, as this is also true in AMLs that lack a UV-driven etiology (P = 3.9 × 10−21). Together, these results indicate that substantial mutant amino acid specificity exists among hotspot mutations across highly diverse tumor lineages. Two related conclusions may be drawn from these data. First, different hotspots in the same gene may possess in many cases different functions, much of which may be lineage-dependent, while not excluding the possibility that some may still arise as a function of differing underlying mutational mechanisms. Second, that perhaps different mutant amino acids within the same hotspot can be functionally different, support for which idea is growing8,11.

Timing of individual hotspots

We next sought to determine if hotspot mutations, many of which are likely driver mutations and in some cases may serve as the initiating lesion, typically arise earlier than do nonrecurrent mutations in the same genes and are therefore more often clonal. Overall, mutations at hotspot residues more often resided in a greater fraction of tumor cells and therefore arose earlier (presumptive clonal), than non-hotspot mutations in the same genes (Fig. 3c). So, whereas prior work has shown that driver genes in lung adenocarcinomas were enriched for clonal mutations24, we found that this was true of hotspot mutations across a broad class of cancer genes and tumor types. However, there was considerable variability among hotspots. Whereas colorectal and endometrial cancers have a similar pattern of PIK3CA hotspot mutations (Fig. 3a) and share hypermutated subtypes of tumors driven by MSI and POLE exonuclease domain mutations25,26, colorectal tumors were unique in the clonality of the E545 and H1047 mutations. The majority of PIK3CA E545 helical domain mutations in colorectal cancers were subclonal, whereas H1047 kinase domain mutations were clonal, a difference that was not apparent in endometrial tumors, in which both are early clonal mutations (Fig. 3d). This may be a function of the pattern of oncogenic co-mutation in these tumors as PIK3CA E545, but not H1047, mutations were significantly associated with KRAS mutations in these colorectal cancers (χ2 P = 0.0004) and in previous cohorts27. Overall, these differences in the molecular timing of specific hotspots augurs potentially important differences in their function in tumor initiation versus progression that requires further study.

Population-level hotspots in the long tail

Consistent with the so-called long tail of the frequency distribution of somatically mutated genes across cancer2, we found that 85% of all hotspots identified here were mutated in less than 5% of tumors of all cancer types in which they were found (Fig. 4a). Such findings have led to calls for sequencing up to many thousands of additional specimens from every tumor type28. However, many hotspots present at low frequency across cancers are not mutated commonly or significantly in even a single cancer type. Indeed, 23% of all hotspots identified here were present in only one or two samples in the tumor types in which they were observed. This included 19 hotspots arising in only one sample of each affected cancer type such as U2AF1 I24, MYC T58, the hyperactivating MTOR I2500 (ref. 29), PIK3CB D1067, EP300 H1451 and ERBB3 M60. Conversely, population-level analysis, rather than by individual cancer type or organ system, allows identification of hotspots that arise as even private mutations in rare malignancies, for which additional broad-scale sequencing is most challenging. Although rare, such recurrent alleles are evidence of selection and may be associated with specific phenotypes, such as exceptional responses30,31 or de novo resistance to cancer therapy, or may reveal specific facets of pathway biology. Consequently, we found that notable long-tail hotspots affect a broad spectrum of abnormal molecular function including macromolecular transport and transcriptional regulation (Table 1, Supplementary Note and Supplementary Fig. 8), as well as essential components of key signaling pathways.

Figure 4: Candidate Ras-related small GTPase driver mutations in the long tail. (a) The frequency distribution of hotspot mutations in cancer has a long right tail of mutated residues that, although recurrent, are not common in any cancer type. (b) There is a statistically significant difference in the pattern of Q61 codon mutations in KRAS, HRAS and NRAS (χ2 P-value = 0.016). (c) The sequence of Gly60-Glu62 of KRAS, HRAS and NRAS are shown along with mutant alleles from affected cases indicating the GQ60GK dinucleotide mutation was the only source of KRAS Q61K mutation, whereas the far more common HRAS and NRAS Q61K mutations arose almost exclusively from single nucleotide events. The KRAS G60G synonymous mutation also creates a G60 codon in sequence (ACC > TCC) identical to wild-type sequence of NRAS G60, where Q61 mutations are the most common. (d) RAC1, RRAS2 and KRAS are shown in schematic form indicating the position of novel hotspots RAC1 A159V and RRAS2 Q72L/H at paralogous residues in the Ras domain to known activating mutations in KRAS (A146 and Q61, respectively). (e) The pattern of RAC1 (left) and RRAS2 (right) mutations along with those in BRAF and Ras genes in affected tumor types. (f) Western blot analysis of RAC1 activation (GTP-bound RAC1) by PAK1 pull down (right). RAC1 A159V was associated with significant RAC1 activation at levels equal to or exceeding the positive control GTPγS and greater than those of the known oncogenic RAC1 P29S. Full size image

Long-tail hotspots in Ras superfamily members

Mutations in the Ras family of small GTPases occur widely in human cancers. As expected, these were among the most significant hotspots detected here (Supplementary Table 2), affecting 1,335 tumors (12% of all cases). Whereas G12, G13 and Q61 codon hotspots predominate in KRAS, NRAS and HRAS, albeit at varying frequencies in different tumor types (Figs. 2a and 3a), we also identified GQ60GK, K117 and A146 hotspots in KRAS. Both K117 and A146 are known activating hotspots in the long tail, but we also identified a previously occult GQ60GK dinucleotide substitution (q = 2.3 × 10−6) in 11 tumors. This dinucleotide substitution results in a Q61K mutation accompanied by a G60 synonymous mutation that are present in cis (in concomitant RNA sequencing; Supplementary Fig. 9). Although Q > K mutations at codon 61 can result from 3′ G > T single-nucleotide mutations in KRAS, 100% of these tumors harbored the dinucleotide substitution, a rare spontaneous event in human genomes. Overall, the distribution of codon 61 mutations in KRAS, NRAS and HRAS are very different, with Q > K mutations occurring significantly less frequently in KRAS (P = 0.016; Fig. 4b). GA > TT mutations were the most common dinucleotide substitution producing GQ60GK (Fig. 4c) and converts the ACC codon at KRAS G60 to TCC, which is the sequence of the G60 codon in NRAS, in which Q61K mutations are far more common and arise nearly exclusively from single-nucleotide mutations. It remains to be determined whether KRAS GQ60GK is therefore driven by a pattern of codon usage at the −1 position. Notably, only one tumor had evidence of a non-KRAS GQ60GK mutation, an NRAS-mutant cutaneous melanoma (Fig. 4c and Supplementary Table 5).

We next explored whether KRAS GQ60GK may serve as a driver of Ras pathway activity as do conventional KRAS hotspots. GQ60GK is indeed present in diverse tumor types that all have well-established Ras-driven subsets (Supplementary Table 5). Reasoning that if GQ60GK were a passenger mutation in Ras-driven tumors, alternative MAPK-activating mutations may be present in these tumors. Instead, we found that in every GQ60GK-mutant sample where another putative driver of MAPK signaling was present, that lesion was either (i) subclonal, defining a different clone than did GQ60GK; (ii) low activity; or (iii) a passenger mutation (Supplementary Table 5). Also, despite the frequency of GA > TT, there was no evidence that a common underlying mutational process or exogenous mutagen was the source of GQ60GK. There was no evidence of UV light exposure in the clinical histories or nucleotide contexts of most affected cases, only one of which was a cutaneous melanoma. Moreover, GQ60GK arose in both hypermutated (MSI-H colon lacking BRAF V600E) and nonhypermutated tumors. Finally, rare G60 missense mutations were evident in KRAS and HRAS in this data set and in the literature (Supplementary Table 5) (ref. 32). So, although we cannot exclude the possibility that the GQ60GK dinucleotide substitution is simply an alternative mechanism to achieve Q61K, the accompanying KRAS-specific G60 synonymous mutation may potentiate a different class of Q61-mutant tumors or cause signaling differences among Q61K-mutant tumors between KRAS, NRAS or HRAS. Although further studies will need to explore the molecular properties of KRAS GQ60GK, this allele represents the most common dinucleotide substitution spanning two codons in human cancer and a mutation more common than other known hotspots in KRAS.

Novel long-tail hotspots were also identified in two other genes that encode members of the Ras superfamily of small GTPases. RAC1, in which we identified two hotspots, is a Rho subfamily member that plays a vital role in various cellular functions. RAC1 P29S is an oncogenic hotspot in melanomas12,33, that we also identified in head and neck, and endometrial cancers (Fig. 4d). This mutation can confer resistance to RAF inhibitor treatment in vitro34, and may underlie early resistance in patients35. We also identified a novel RAC1 A159V hotspot present in 10 tumors (q = 2.27 × 10−6; Fig. 4d). Notably, RAC1 A159V is paralogous to KRAS A146, a known activating mutation36. Whereas activating KRAS A146T mutations arise predominantly in colorectal carcinomas (Supplementary Table 2), RAC1 A159V mutations are most common in head and neck cancers and were not present in any melanomas, despite the frequency of RAC1 P29S in this cancer type. Moreover, similar to P29S mutations, we observed RAC1 A159V mutations in tumors that are both Ras/Raf wild-type and mutant (Fig. 4e). To determine whether RAC1 A159V is an activating mutation, we assessed its effect in vitro. Active RAC1 is GTP-bound, interacting with PAK1 to activate downstream effectors. Therefore, to quantify RAC1 activation in vitro, we used a PAK1 pull-down assay. In HEK293T cells expressing RAC1 A159V, there was substantial RAC1 activation to levels equal to or exceeding positive-control RAC1 GTPγS cells and greater than even those levels induced by the known RAC1 P29S oncogenic mutation (Fig. 4f). Moreover, cells expressing RAC1 Q61R, a mutation we identified in a primary prostate cancer that is paralogous to KRAS Q61, also potently induced RAC1 activation (Fig. 4d,f).

RRAS2 is a Ras-related small GTPase37. RRAS2 is overexpressed or mutated in a small number of cancer cell lines of various origins38,39,40, and is oncogenic in vitro with transforming ability similar to that of established Ras oncoproteins41. However, it has not been documented as somatically mutated in human tumor specimens. Here, we identified a RRAS2 Q72 hotspot present in nine tumors (q = 8 × 10−15). Similar to RAC1 A159V, the RRAS2 Q72 hotspot is paralogous to KRAS Q61 (Fig. 4d). However, unlike RAC1, RRAS2 Q72 does not predominate in any individual tumor type. Also unlike RAC1, the RRAS2 Q72 mutation was present in Ras/Raf wild-type tumors among the affected types (Fig. 4e). This result suggests that RRAS2 activation may be an alternative avenue for tumors to acquire Ras-like activation as previous studies have shown that RRAS2 shares many Ras downstream signaling elements including phosphatidylinositol-3 kinase (PI3K)42,43, the Ral GDP dissociation pathway42, and Raf kinases44. Beyond these hotspots, several less common RAC1 and RRAS2 mutations affect paralogous residues of highly recurrent alleles in KRAS (Fig. 4d); some which we validated were also activating in vitro (Fig. 4f), indicating that the landscape of potentially functional mutations in these genes extends beyond even these less common long-tail hotspots to private mutations as well.