Accelerating discovery of functional mutant alleles in cancer

Abstract

Most mutations in cancer are rare, which complicates the identification of therapeutically significant mutations and thus limits the clinical impact of genomic profiling in cancer patients. Here, we analyzed 24,592 cancers including 10,336 prospectively sequenced patients with advanced disease to identify mutant residues arising more frequently than expected in the absence of selection. We identified 1,165 statistically significant hotspot mutations of which 80% arose in 1 in 1000 or fewer patients. Of 55 recurrent in-frame indels, we validated that novel AKT1 duplications induced pathway hyperactivation and conferred AKT inhibitor sensitivity. Cancer genes exhibit different rates of hotspot discovery with increasing sample size, with few approaching saturation. Consequently, 26% of all hotspots in therapeutically actionable oncogenes were novel. Upon matching a subset of affected patients directly to molecularly targeted therapy, we observed radiographic and clinical responses. Population-scale mutant allele discovery illustrates how the identification of driver mutations in cancer is far from complete.

The rapid adoption of prospective clinical tumor sequencing(1–4) has led to the identification of an increasing number of somatic mutations of unknown significance. While a small number of mutations are used to guide treatment selection, the vast majority lack biological or clinical validation, limiting the ability of clinicians to use tumor genomic data to guide therapy selection(5, 6). Indeed, such mutations are often presumed to be passenger mutations with no evidence to support such a classification. As a consequence, many patients whose tumor harbor mutations of unrecognized clinical significance are not being offered potentially beneficial therapies and this knowledge gap represents one of the fundamental hurdles to the broader adoption of precision oncology today.

Given the sheer number of mutations of uncertain significance, there is an urgent need to identify and prioritize for biological and clinical study potentially druggable driver mutations identified within the context of prospective tumor profiling. Unraveling the relationships among different mutant alleles, their co-mutational patterns and the cell types in which they selectively arise, will be critical to defining their function and clinical actionability(7), essential steps to expanding the treatment options for cancer patients. Historically, however, the the incremental laboratory and then clinical validation of novel mutations as sensitizing biomarkers of response or resistance to standard or investigational therapies can take years, preventing current patients from potentially benefiting from such therapies.

Here, we define novel driver mutations in the long right tail of somatic mutations in cancer and developed an exploratory framework by which computational weight-of-evidence alone was utilized in real-time to prioritize treatment-refractory patients harboring novel hotspot mutations of uncertain clinical significance for studies of molecularly targeted therapies. In a subset of patients, clinical response rather than laboratory interrogation was employed as the most expedient approach for clinical validation of mutant alleles of unknown function. No example exists to our knowledge where the identification of a novel mutation of likely significance has taken place in the same population used to validate the mutant allele as sensitizing to therapy, a potential acceleration of the clinical validation of variants of unknown significance as putative sensitizing biomarkers.

Results

To identify novel hotspot mutations of biological and potentially therapeutic importance, we analyzed somatic mutational data from 24,592 patients. This cohort consisted of 14,256 retrospectively sequenced predominantly primary untreated human cancers and 10,336 prospectively sequenced active, advanced cancer patients with recurrent and metastatic disease (43% of specimens were obtained from metastatic sites, see Methods) (8). These samples represent 322 cancer types spanning 32 organ sites, the annotation of which was standardized to conform to an open-source structured disease classification (Supplementary Table 1) (http://oncotree.mskcc.org/). We analyzed each of the 32 organ sites independently as well as the full cohort (pan-cancer) to enhance the probability of identifying hotspots that occur rarely in multiple organ types of different mutational burdens and processes(9). To do this, organ type-specific, gene-specific, and context-specific background mutation rates were computed (Supplementary Table 2, see Methods). We also developed a first-of-its-kind computational approach to identify hotspots of candidate oncogenic small in-frame insertions or deletions, which are more challenging to identify than substitutions due to the variability of mutant allele length and position from tumor to tumor.

We identified 1,165 mutational hotspots in 247 genes (1,110 single-codon and 55 indel; median of 2 hotspots per gene, range 1–120; q-value < 0.1, false discovery rate of 10%) (Supplementary Tables 3–4). This analysis recovered nearly all previously identified hotspots(10) (97%) and identified 840 additional hotspots, reflecting an increased power of detection as well as a more clinically diverse cohort of patients. In total, 5% of these 840 hotspots were identified in cancer types new to this analysis, emphasizing the value of a clinically diverse dataset to power hotspot discovery. The vast majority of hotspots observed here for the first time were due to the large increase in sample size over prior studies, emphasizing how the characterization of the long right tail of the curve of driver mutations in cancer was incomplete. Indeed, the frequency distribution of hotspot-mutant genes had a long right tail(10), the shape of which was independent of the count of unique hotspots in the gene and was different between single-codon and indel hotspots (Fig. 1A and Supplementary Fig. 1). While the majority of hotspots (n=596, 51% of total) were statistically significant both pan-cancer and within individual organ types (Fig. 1B), 20 and 29% of hotspots were significant only within an individual organ type or only in the pan-cancer analysis of the full cohort respectively (Fig. 1C). Many of the mutant alleles identified, both in long-established cancer genes (such as PIK3CA, MTOR, ERBB2, MAP2K1) as well as in genes more recently implicated (such as CYSLTR2),(11) were novel, reflecting the greater sensitivity for rare allele discovery with increasing cohort size in even well-characterized cancer genes (Fig. 1B).

Figure 1. The long tail of mutational hotspots in cancer.

a) The frequency distribution of genes containing one or more single-codon hotspots (top, dark blue) and in-frame indel hotspots (light blue). At bottom, the count of single-codon and in-frame indel hotspots in the same genes. b) Shown is the statistical significance of mutational hotspots inferred from the analysis of the full cohort (pan-cancer, y-axis) and the most significant individual cancer type (x-axis). A subset of hotspots are annotated (circled in black) and include mutations significant in both analyses (upper right), those significant only when combinating all cancer types and data (leftmost) and those significant only within a given cancer type (bottom). c) The proportion of hotspots that were significant only in individual organ types, only in the pan-cancer analysis, or both.

Forty-two percent of the patients in this cohort were prospectively sequenced at our institution as part of their clinical care and had advanced and/or previously treated disease. This clinical profile is distinct from that of patients with untreated primary tumors, from which most of the data in the literature is obtained. The inclusion of such patients facilitated the identification of hotspots present almost exclusively in the metastases of treatment-refractory patients. Eleven hotspots were enriched in metastatic disease compared to the primary tumors of a given cancer type (see Methods), nine of which were therapy-associated arising in specimens after treatment with either anti-androgen, anti-estrogen, or tyrosine kinase inhibitor therapies (Supplementary Fig. 2). Notably, some therapy-resistant hotspots were found to arise in treatment-naïve tumors of other cancer types (such as KIT D820), suggesting that treatment-associated resistance mutations in one cancer type can arise de novo in the absence of therapy as the primary oncogenic driver in another cancer type(12). Other hotspots may reflect new mechanisms of resistance to traditional cytotoxic chemotherapies, such as TP53 N239, which confers paclitaxel resitance in vitro(13) and was the only TP53 hotspot that arose preferentially in metastatic breast cancers (q-value = 0.03), all obtained from tumors that developed resistance to, or rapidly progressed on, taxane-based therapy. Together, these analyses identify a broader range of hotspots than previously recognized and for which biologic and clinical study(14, 15) may accelerate clinical translation.

While substitutions are the most abundant class of mutation in cancer genomes, several recurrent, activating in-frame indels are validated predictive biomarkers of sensitivity to targeted therapies, including indels in exon 19 of EGFR in lung adenocarcinomas and in exon 11 of KIT in gastrointestinal stromal tumors(16, 17). Yet, the unbiased discovery of recurrent oncogenic in-frame indels from population-scale data has not been done. We thus extended our methodology to identify hotspots (clusters) of in-frame indels (see Methods). In total, we identified 55 statistically significant indel hotspots in 36 genes (Supplementary Table 4 and Supplementary Fig. 3). There were 20-fold fewer indel hotspots identified than single-codon hotspots, and while deletions predominated (69%), duplications were enriched in oncogenes (p-value < 0.01, Fig. 2A-B). Multiple indel hotspots in EGFR, ERBB2, KIT, and BRAF were identified(18, 19) as were other novel indel hotspots in AKT, MTOR, PIK3CA, SRSF2, U2AF1, and MYC among others (Supplementary Table 4). Given the recently identified clinical activity of AKT inhibitors in AKT-mutant patients(20), we sought to determine whether the previously uncharacterized AKT indels were activating and potentially drug sensitizing. Specifically, we functionally characterized AKT1 P68_C77dup, one of several paralogous indels found in a hotspot cluster in the pleckstrin homology domain of AKT1 and AKT2 (q-values = 0.09 and 2e⁻⁵, respectively) proximal to known AKT1 hotspots (L52 and Q79; q-values<10^–4) (Fig. 2C and Supplementary Fig. 4). Expression of AKT1 P68_C77dup in MCF10A cells resulted in a higher level of AKT phosphorylation (T308/S473) (Fig. 2D), as well as increased phosphorylation of downstream effectors of AKT such as S6 and PRAS40, as compared to the two most common activating AKT1 missense mutations E17K and Q79K. Isogenic cells expressing AKT1 P68_C77dup also demonstrated greater sensitivity to the ATP-competitive pan-AKT kinase inhibitor AZD5363 than did cells expressing AKT1 E17K or wild-type AKT1 (Fig. 2E).

Figure 2. Oncogenic indel hotspots.

a) The distribution of recurrent indel hotspot types discovered here. b) Duplications were significantly more common than either deletions or insertions in oncogenes (asterisk, p-value < 0.01). c) The paralogous indels are shown defining the AKT1 and AKT2 duplication hotspot. The affected cancer types are similar to those that harbor known activating L52, and Q79 hotspot mutations and include estrogen receptor (ER)-positive HER2-negative breast cancers that lack other PI3K pathway alterations. d) MCF10A cells stably expressing the indicated AKT1 mutations are shown and expression and/or phosphorylation levels were assayed by Western blot indicating the AKT1 P68_C77dup induces elevated levels of phosphorylated Akt and S6 comparable to or exceeding that of known activating E17K or Q79K hotspots. e) Cell survival upon AKT blockade with AZD5363 in AKT1-mutant cells indicated that P68-C77dup-mutant cells were most sensitive to AKT inhibition, more so than the canonical E17K hotspot. f) Schematic of MAP2K1 from amino acids 60 to 140 indicates the position of single-codon hotspots (green arrows) is distal from the position of the indel hotspot (blue lines are individual indels in affected tumors). Arcing red lines reflect the distance in angstroms between the indels and single-codon hotspots in the protein structure. g) The rate of co-mutation with other MAPK effectors varied by MAP2K1 hotspot, with P124 mutations always associated with upstream pathway activation and predominantly in melanomas, while others (E203, G128, F53, C121, and K57) were only partially co-mutated, and the MAP2K1 indel hotspot never arose in tumors with another MAPK driver mutation. h) All but one MAP2K1 P124-mutant tumors possessed another known driver of MAPK signaling, of which most were BRAF V600E (59% of total) and these and others were mostly cutaneous melanomas. Conversely, the MAP2K1 I99_I107 indel hotspot never arose in an otherwise MAPK-altered tumor in a diversity of cancer types.

Other indels identified here may be associated with therapeutic resistance, such as ESR1 V422del (q-value = 0.045), which arose clonally after failure of anti-estrogen therapy in the metastatic site of ER-positive breast cancers that otherwise lacked the most common ligand binding domain hotspots E380, L536, Y537, and D538 (q-values<10^–17) that are known to confer resistance to estrogen deprivation therapies in breast cancer(21) (Supplementary Fig. 3). While most indels, like those in AKT1, spanned or were physically adjacent to single-codon hotspots in the same genes, indel hotspots in three genes were physically distant (greater than 15 angstroms) from subsitution hotspots in their cognate folded protein. These indels included the aforementioned ESR1 V422del, the well-characterized FLT3 internal tandem duplication (ITD), and a cluster of indels spanning I99 to I107 in MAP2K1 (q-value = 3.3×10^–12), which was one of 11 total MAP2K1 hotspots identified here (Supplementary Table 3–4). Structurally distant indels and single-codon hotspots may imply divergent biologic effects, as is the case between ITD and kinase domain mutations in FLT3(22), but the extent of such differences is unknown.

In BRAF, another effector of aberrant MAPK signaling, V600 mutations induce constitutive kinase activity independent of upstream activation whereas BRAF D594 mutants are kinase dead but cooperatively amplify ERK signaling and tumorigenesis when co-expressed with mutant KRAS(23). Indeed, reflecting this divergent biologic function, D594 mutations typically co-occur with activating RAS mutations in melanoma patients, a pattern missed when all alterations in BRAF and KRAS are considered together. We, therefore, sought to determine whether the structurally distant indel and single-codon hotspots in MAP2K1 may have differing function, as a guide to future functional studies. Using our allele-specific approach, we assessed the statistical significance of co-occurrence among hotspot mutations that arose together in individual tumors more frequently than expected by chance in pairs of genes in MAPK signaling. We identified multiple associations, the most significant of which was between MAP2K1 and BRAF (Supplementary Fig. 5). The pattern and frequency of MAP2K1 co-mutational associations varied in an allele-specific specific manner (Fig. 2G). For example, MAP2K1 P124 was nearly always co-mutated with an upstream activating mutation in the MAPK pathway (95%), most often BRAF V600E (55%) (Fig. 2H). Conversely, the MAP2K1 indel hotspot newly identified here arose in a mutually exclusive pattern with other MAPK lesions in affected tumors independent of cancer type. This pattern was not attributable to acquired resistance to MAPK pathway inhibitors, as only one such tumor was sequenced after RAF or MEK inhibitor failure. Overall, these results illustrate how novel computational methodologies can identify previously occult oncogenic in-frame indels of biological and potential therapeutic importance. Moreover, these data indicate that a broader analysis of co-incident mutational patterns in multiple pathway effectors can uncover potential allele-specific functional differences that may be missed by gene-level analyses and may condition distinctive signaling biology requiring deeper mechanistic investigation.

Reflecting the long right tail of mutation frequencies in human cancer, 82% of all hotspot mutations were identified in 1 in 1,000 or fewer patients. To assess the future impact of larger cohort size on novel hotspot detection, we performed repeated random downsampling of increasing subsets of the cohort to infer the anticipated rate of future hotspot identification. Principal component analysis of gene-specific rates revealed four distinct classes of genes (saturating, fatiguing, linear, accelerating) that accrued their recurrent mutations, independent of their overall mutational burden, in different patterns with considerable variability from gene to gene (Fig. 3A and Supplementary Fig. 6). We also estimated the number of additional tumors that would need to be sequenced from a cohort of similar cancer type composition to identify the next incremental hotspot in each gene and cluster. One cluster was defined by canonical oncogenes (IDH1, K/N/HRAS, GNAQ, MYD88) whose most prevalent hotspots could be identified from the analysis of few samples but, as genes, are approaching saturation and thus additional sequencing is not expected to yield many additional novel hotspots. Indeed, we estimated that an additional ten thousand or more tumors sequenced would be necessary to identify another novel hotspot in many of the genes in this saturating cluster (Fig. 3B). The identification of hotspots in genes in the second cluster initially increased rapidly with increasing cohort size, but their rate is fatiguing yet not saturating, indicating additional rare alleles will continue to be discovered as additional tumor genomic data becomes available. Notably, many of the genes in the fatiguing cluster are therapeutically actionable genes such as BRAF, PIK3CA, ESR1, AKT1, and ERBB2 and thus the identification of novel hotspots in these genes could have immediate clinical implications. The third cluster of genes are still in a linear phase of hotspot identification and additional sequencing should continue to reveal additional new, albeit uncommon hotspots in these genes, many of which are therapeutically targetable oncogenes such as KIT. The fourth cluster is composed of genes (such as MET or MTOR) in which even the enormous quantity of sequencing to date has only begun to reveal rare hotspots of potential clinical significance. In this accelerating cluster, fewer than 1,000 additional specimens would be necessary to identify additional hotspots (Fig. 3B). These patterns have important implications for strategies to prioritize and understand emerging mutations and suggest that many additional hotspots could be identified by pooling existing prospectively sequenced tumors that currently reside in siloed institutional or commercial repositories.

Figure 3. Saturation analysis and the discovery of actionability of mutational hotspots.

a) Downsampling and clustering analysis revealed four distinct classes of genes with different rates of hotspot acquisition (light and dark grey and light and dark blue) from the number of sequenced samples necessary to identify a given fraction of all hotspots in affected genes. Shown in gray are all genes. In red and purple are genes that are either saturating in their hotspot discovery (green) or were rapidly increasing and now fatiguing (purple). In red and blue are those genes in either their still linear and accelerating phases of hotspot discovery. b) An estimate of the number of additional specimens to be sequenced to identify an additional hotspot in each gene in each of the four aforementioned classes (clinically actionable genes are identified). c) Of hotspot mutations identified in one of 18 clinically actionable cancer genes (see panel b for genes), the fraction of hotspots used to guide the use of standard-of-care or investigational therapies at present (see Methods) versus those that were identified here but are clinically uncharacterized. d) Initial response of a triple-negative breast cancer patient to neratinib treatment whose tumor harbored a novel ERBB2 V697 hotspot mutation. e) A complete response observed in a patient with gallbladder cancer harboring a novel BRAF L485W hotspot mutation to the ERK inhibitor BVD-523. f) A model by which advanced treatment-refractory patients can be directed to molecularly driven therapies based on computational weight-of-evidence alone as an efficient means for determining mutant allele function and expand biomarkers of drug response.

The above analysis indicates that we are far from completing the identification of potentially actionable hotspots. As such, patients with functional but rare mutant alleles in targetable cancer genes are not being offered potentially beneficial matched therapies as a result of the unrecognized clinical significance of the mutations identified in their tumor. To determine the scope of such occult actionability, we utilized a curated knowledgebase of the oncogenic effects and treatment implications of mutations (http://oncokb.org/)(24) in 18 genes in which one or more mutations are used in current clinical practice to guide routine (FDA-approved or part of established practice guidelines) prescribing of targeted therapy or are being evaluated as investigational biomarkers (see Methods). Of the 196 hotspot mutations identified in these genes, only a minority have been investigated clinically (Fig. 3C), though patterns vary by gene (Supplementary Fig. 7). Fifty hotspots (26%) were newly discovered here, being neither annotated in OncoKB nor identified in further detailed literature review. Because these novel hotspots arise in genes for which targeted therapies are already available, we sought to test the therapeutic hypothesis that these mutations may be similarly sensitizing biomarkers by matching a subset of the affected, prospectively sequenced patients to molecularly targeted therapies. This patient-to-drug matching was performed in the absence of laboratory data confirming that such mutaitons were activating or sensitizing alleles.

As a proof of this concept, we identified seven active patients at the time of this study analysis that harbored one of the novel rare hotspots identified here and enrolled them on existing clinical trials where the therapy was targeting the affected oncogene. All seven patients derived clinical benefit from therapy including two patients with a novel ERBB2 V697 hotspot that were treated with the pan-HER tyrosine kinase inhibitor neratinib(25). One ERBB2 V697 patient was a heavily pre-treated triple negative breast cancer patient (Fig. 3D), while the other had a cancer of unknown primary involving the head and neck that responded to neratinib monotherapy for 13 months. At the time of progression on therapy, we biopsied and sequenced a cutaneous metastasis, which revealed a clonally related post-progression tumor that lacked any evidence of the ERBB2 V697 mutation, indicating loss of the sensitizing mutation was sufficient to confer drug resistance. Four other patients had tumors harboring different novel or previously uncharacterized PIK3CA hotspot mutations (P104, T1025, V344, R38), each of which had durable responses to either an isoform-selective PIK3CA inhibitor or an mTORC1/2 catalytic inhibitor. The final patient had a gallbladder cancer with a BRAF L485W missense hotspot who achieved a durable complete response to the ERK inhibitor BVD-523 lasting nearly a year (Fig. 3E)(26). While these seven patients harboring one of six novel hotspot mutations represent only an exploratory proof-of-principle, further studies of other hotspots are needed. However, these results indicate that, in some genes, mutation recurrence alone could be used as the initial screen to select otherwise treatment-refractory patients for targeted therapy when biological data does not exist. When affected patients are identified prospectively, such clinical reponses to molecularly targeted therapy may be the most efficient way to determine functionality and expand the pool of mutant alleles within a targetable gene that are considered sensitizing biomarkers (Fig. 3F).

Discussion

In sum, we identified 1,165 hotspot mutations across a spectrum of primary and advanced cancers. The rate at which hotspots were identified with increasing cohort size varied widely among genes. In some genes, potentially actionable mutational hotspots were still being identified at a rapid rate with increasing cohort size. As many of the novel hotspots identied here were not previously recognized as functional variants, patients whose tumor harbored such mutations were unlikely to have been offered matched molecularly targeted therapies to which patients with other previously characterized alleles in the same genes have had a profound clinical benefit. The implications of this are especially urgent for those patients with metastatic disease most in need of novel therapeutic approaches.

Pooling prospective genomic data from many sources may quickly achieve the scale needed to saturate the discovery of hotspots in most of the genes targetable with current drugs(27), expanding the reach of precision oncology. To accelerate the identification of novel clinically actionable hotspots, we have deposited all of the data and results at http://cancerhotspots.org for query, visualization, and download to facilitate their dissemination and use by the wider biomedical community. Despite our functional and clinical validation of select novel hotspots, we recognize that hotspot mutations in individual genes may have varying drug sensitivities and potentially allele-specific neomorphic functions. By making all hotspots discovered here available in an easily searchable portal, we aim to catalyze broader functional and clinical validation of individual mutant allels, results from which we already curate in a knowledgebase of oncogenic effects and treatement implications(24). Together, our findings provide a means to prioritize the experimental validation and clinical cross-validation of long-tail driver mutations which will expand the treatment options for molecularly defined populations of cancer patients.

Methods

Mutational Data

Retrospective mutational data were obtained from three publicly available sources: 1) The Cancer Genome Atlas (TCGA), 2) International Cancer Genome Consortium (ICGC), and 3) independent published sequencing projects(10). The subset of this cohort that was prospectively sequenced consists of 10,945 samples from 10,336 unique advanced cancer patients and whose tumors were profiled as part of their active care between January 2014 and July 2016 at Memorial Sloan Kettering Cancer Center (MSKCC). The consent of these patients, acquisition of specimens, sequencing, analysis, and reporting are described in an accompanying manuscript(8). All such patients provided written and informed consent for sequencing and review of patient medical records for detailed demographic, pathologic, and treatment information (NCT01775072). This study was approved by the Memorial Sloan Kettering Cancer Center Institutional review board and the studies were conducted in accordance with the Declaration of Helsinki, International Ethical Guidelines for Biomedical Research Involving Human Subjects (CIOMS), Belmont Report, or U.S. Common Rule. Briefly, matched tumor and normal specimens were sequenced (to 500–1000-fold sequence coverage) with a validated capture-based next-generation sequencing assay called MSK-IMPACT that is New York state-approved for clinical use. This assay captures the coding exons and select introns of oncogenes, tumor suppressor genes, all genes targeted by either approved therapies or those investigational drugs being studied in clinical trials at our Center, and significantly mutated genes reported by large-scale cancer sequencing efforts (Supplementary Table 2). These sequencing data are analyzed as previously described(1) to detect somatic mutations, small insertions and deletions (indels), DNA copy number alterations (CNAs) and select translocations using DNA from both frozen and formalin fixed-paraffin embedded tissue. An IRB protocol facilitates this prospective genomic characterization (IRB #12–245, ClinicalTrials.gov NCT01775072) and enables the return of results to patients. All genomic data generated as part of routine standard-of-care therapy is deposited, along with relevant clinical data, in a HIPAA-compliant manner, in the cBioPortal for Cancer Genomics(28, 29). All somatic nonsynonymous mutations reported were manually reviewed in primary sequencing data as described in ref. (8) and combined with synonymous mutations in the same samples and utilized in this analysis. All mutations in any one of 469 genes that overlap among the retrospective and prospective subsets of the final cohort were uniformly re-annotated using vcf2maf ver. 1.6.10 (https://github.com/mskcc/vcf2maf). Variants identified by the Exome Aggregation Consortium (ExAC)(30) as having a minor allele frequency greater than 0.0004 in any subpopulation were excluded as presumed germline unless they were annotated by ClinVar(31) as either pathogenic, a risk factor, or protective.

Single-codon hotspot significance calculations

The statistical significance of single-codon hotspots was determined in each of 32 separate organ types as well as pan-cancer (full cohort) using an extended version of our previously described method(10). Briefly, statistical significance of every codon was assessed with a truncated binomial probability model in which the expected probability incorporates underlying features of gene-specific rather than genome-wide mutation rates including gene length, gene- and position-specific mutability, and overall mutational burden of the gene(10). This background model is gene-specific and assesses the significance of individual mutant alleles relative to the background of all mutations in the gene in which it emerges rather than across genes. Unlike in our prior study, here we calculated gene- and position-specific mutability on a per-organ type basis to reflect their differences in background mutability and mutational processes. The mutability of each of 32 possible trinucleotides was calculated independently for each organ type as the fraction of mutations affecting the central position of the given trinucleotide t across all samples from cancer types belonging to the given organ type (Supplementary Table 1). The mutability of each codon, expected mutability of each gene, and the final binomial probability was calculated as before(10). For 7 of 32 organ types, insufficient whole exome sequencing data existed to robustly estimate trinucleotide mutability (<50 samples per organ type), so a pan-cancer mutability was calculated as above and utilized. Multiple hypothesis correction for both pan-cancer and organ-specific analyses were performed using the method Benjamini and Yekutieli method. Mutational hotspots corresponding to a q-value < 0.1 were considered statistically significant (False Discovery Rate < 10%).

Small in-frame insertion/deletion significance

We assessed the statistical significance of in-frame small insertions and/or deletions (indels) in a manner similar to single-codon hotspots using the truncated binomial probability model. For this analysis we excluded frameshift mutations as presumed truncating loss-of-function mutations. As a background model of indel mutability in both normal and disease human genomes is poorly understood, none was utilized here (neither gene nor position-specific mutability). Also when calculating the expected probability at each site, we allowed the minimum probability to decrease beyond the 20th percentile of all probabilities dataset-wide used for single-codon hotspot detection(10). Due to the allelic variability of indels, in-frame indels were grouped using a maximal common region defined as the contiguous genomic region spanned by overlapping indels. The mutation count for each such region is the sum of all spanning (single bp or more) in-frame indels. Significance was assessed, as with single-codon hotspots, with the binomial model described above. Statistically significant indels that exclusively arose in samples from retrospective data (published or consortial studies) were manually reviewed in aligned sequencing data of representative cases to identify and exclude potential false positives.

Simulating mutational acquisition rates

To assess hotspot acquisition rates within genes, we performed the hotspot analysis on repeated random downsampling of samples in the dataset starting from 100 patients to the final total number of patients in the dataset in 100-sample increments. Only statistically significant hotspots in each downsample were considered if significant in the final analysis. For each gene, we then fit a locally weighted polynomial regression to the distribution of downsamples to estimate the rate of hotspot acquisition for each gene. To infer broader patterns of hotspot acquisition, these fits were then clustered using fuzzy c-means clustering (R package e1071 v1.6–7) and the optimal number of clusters (four) was determined based on reduction of sum of squared error for between 1 to 15 clusters.

Mutational annotation

Hotspots identified here were considered novel if they were absent from the results of prior hotspot studies or upon detailed literature review, no prior publication describe the mutation or its biochemical or biological validation. All mutations were further annotated for their potential prognostic and therapeutic significance utilizing OncoKB, a curated knowledgebase of the oncogenic effects and treatment implications of mutations at the individual allele resolution (http://www.oncokb.org/)(24). The potential therapeutic actionability of each mutation (sensitizing to either standard-of-care or investigational therapies) was defined as having one of four levels of evidence based on published clinical or laboratory evidence. Levels are: 1) genomic alterations that are FDA-approved biomarkers in patients of the indicated cancer type; 2A) mutations that were deemed to be standard-of-care biomarkers for FDA-approved drugs in the indicated cancer type based on currently accepted practice guidelines such as those issued by the National Comprehensive Cancer Network (NCCN); 2B) alterations that are FDA-approved biomarkers in another cancer indication, but not in patients with the affected cancer type; 3) alterations for which clinical evidence links the biomarker to drug response in patients, but use of the biomarker is not currently a standard-of-care in any cancer type; and finally 4) alterations for which compelling preclinical data associates the biomarker with drug response. Only levels 1, 2A, and 3A were utilized for the analyses and results described here.

Enrichment and clinical analyses

To test the enrichment of hotspots in either primary or metastatic disease within cancer types, we required that a given hotspot be present in at least 15 samples or 5 metastatic samples in each cancer type. Only samples and cancer types for which we could confirm their primary or metastatic disease status were included in the analysis (TCGA, SU2C prostate(32)), and the prospective MSK-IMPACT series). The significance of enrichment for individual hotspots was assessed on a per-cancer type basis and determined by two-sided Fisher exact test comparing the number of primary samples of a given cancer type that possess the hotspot to metastatic samples of that same type. Both cutaneous melanoma and gliomas were excluded from this analysis due to the high rate of presentation with metastatic disease in the former, and the absence of distant metastasis (local recurrence only) of the latter. Resulting p-values were corrected for multiple hypothesis testing with Benjamini and Hochberg method on a per-cancer type basis.

Co-mutational analysis

To assess the statistical significance of observed co-mutational frequency, we first construct a 2-by-j binary matrix M where each entry m_ij refers to the status of the gene i in the sample j and whose value is 1 if sample j has a hotspot alteration in gene i. Co-occurrence analysis was performed for all unique pairwise combinations of genes within a given pathway (whose members were curated from OncoKB, see above). Other than hotspots identified here, for the purposes of this analysis presumed loss-of-function mutations in tumor suppressor genes in these pathways (NF1) were considered altered (nonsense, frameshift insertions or deletions, splice site, nonstop, or translation start site). We generated a null model of random co-occurrence by permuting the observed alterations (10⁶ permutations) while preserving the overall frequency of the alterations observed in our cohort. Empirically derived p-values were generated as the number of times co-occurrence was observed equal to or more often in this null distribution compared to that of the observed data. Multiple hypothesis correction was performed using Benjamini and Hochberg approach and significant co-occurrence were those pairwise combination of genes within pathway of q-value < 0.01.

AKT1 duplication indel validation

293-FT cells were obtained from ATCC and maintained on DMEM supplemented with 10% FBS and 2mM glutamine. MCF10a cells were similarly acquired from ATCC (and generously provided by the Solit laboratory), and maintained in DMEM/F-12 base medium containing 5% horse serum and other supplements (20ng/ml EGF, 0.5mg/ml hydrocortisone, 100ng/ml cholera toxin and 10mg/ml insulin) (complete growth medium). Both cell lines obtained from ATCC were mycoplasma-free and authenticated by ATCC using karyotyping, morphology and PCR-based methods. For experiments, growth factors were withdrawn from the media, and an “assay medium” was used (DMEM/F-12 base medium containing 2% horse serum, hydrocortisone, and cholera toxin). Plasmids, cloning, and stable line generation was performed as follows. AKT1-wildtype (WT) and AKT1-E17K in pDONR223 vector were provided by the Baselga laboratory. AKT1 point and indel mutants were generated by site-directed mutagenesis using KAPA HiFi polymerase (KAPA Biosystems) or Q5 mutagenesis kit (New England Biolabs) and verified by Sanger sequencing. AKT1-WT and all the other mutants were subsequently sub-cloned into gateway lentiviral vector pLX302 using LR Clonase II enzyme mix (Invitrogen). Lentiviruses encoding WT or mutant AKT1 were packaged in 293FT cells and the supernatant media containing viral particles was filtered through 0.45μm filters and used to infect MCF10a cells. Cells stably expressing the lentiviral constructs were selected with Puromycin (2.5μg/ml).

For western blot assays, MCF10a cells stably expressing WT and mutant AKT1 were seeded on 6-well plates. After overnight exposure to the assay medium the cells were lysed, sonicated, and 30μg protein was loaded onto SDS-PAGE gels, transferred to nitrocellulose membranes, and immunoblotted for p-Akt and other downstream molecular targets of Akt pathway activation. Antibodies for p-Akt (T308) (D25E6), p-Akt (S473), p-S6RP (S240/244), p-PRAS40 (T246), total PRAS40, and total S6 were obtained from Cell Signaling Technology. V5 probe (E10) and actin antibodies were purchased from Santa Cruz Biotechnology. Drug treatment and cell viability assays were performed as follows. AZD5363 was generously provided by AstraZeneca, dissolved in DMSO to yield a 10mM stock, and diluted in assay medium to achieve the desired concentrations. MCF10A stable lines expressing WT or mutant AKT1 were seeded in 96-well plates, treated with a range of drug concentrations, and cell viability was assessed 72 hours post treatment using the CellTiter-Glo luminescent cell viability assay (Promega).

Significance.

Our systematic computational, experimental, and clinical analysis of hotspot mutations in ~25,000 human cancers demonstrates that the long right tail of biologically and therapeutically significant mutant alleles is still incompletely characterized. Sharing prospective genomic data will accelerate hotspot identification, thereby expanding the reach of precision oncology in cancer patients.