Abstract
Discovery of cancer drivers has traditionally focused on the identification of protein-coding genes. Here we present a comprehensive analysis of putative cancer driver mutations in both protein-coding and non-coding genomic regions across >2,500 whole cancer genomes from the Pan-Cancer Analysis of Whole Genomes (PCAWG) Consortium. We developed a statistically rigorous strategy for combining significance levels from multiple driver discovery methods and demonstrate that the integrated results overcome limitations of individual methods. We combined this strategy with careful filtering and applied it to protein-coding genes, promoters, untranslated regions (UTRs), distal enhancers and non-coding RNAs. These analyses redefine the landscape of non-coding driver mutations in cancer genomes, confirming a few previously reported elements and raising doubts about others, while identifying novel candidate elements across 27 cancer types. Novel recurrent events were found in the promoters or 5’UTRs of TP53, RFTN1, RNF34, and MTG2, in the 3’UTRs of NFKBIZ and TOB1, and in the non-coding RNA RMRP. We provide evidence that the previously reported non-coding RNAs NEAT1 and MALAT1 may be subject to a localized mutational process. Perhaps the most striking finding is the relative paucity of point mutations driving cancer in non-coding genes and regulatory elements. Though we have limited power to discover infrequent non-coding drivers in individual cohorts, combined analysis of promoters of known cancer genes show little excess of mutations beyond TERT.
Introduction
Discovery of cancer drivers has traditionally focused on the identification of recurrently mutated protein-coding genes. Large-scale projects such as The Cancer Genome Atlas (TCGA) and the International Cancer Genome Consortium (ICGC) have profiled the genomes of a number of different cancer types to date, leading to the discovery of many putative cancer genes. Whole-genome sequencing (WGS) has made it possible to systematically survey non-coding regions for potential driver events. Recent studies of single-nucleotide variants (SNVs) and insertions/deletions (indels) detected from WGS data from smaller cohorts of patients have revealed putative candidates for regulatory driver events1–8. However, to date, only a few events have been functionally validated as regulatory drivers, affecting expression of one or more target genes by changing their regulation, RNA stability or disruption of normal genome topology6,9–12. Non-coding RNAs play diverse regulatory roles and are enzymatically involved in key steps of transcription and protein synthesis. Though functional evidence for non-coding RNAs in cancer is accumulating13, 14, only few examples have been shown to be the target of recurrent mutation15, 16. In contrast to protein-coding regions, the lower coverage and complexity of DNA sequence in these noncoding regions pose additional challenges for high-quality mutation calling, mutation rate estimation and identification of driver events. Adequate statistical methods that address these issues are needed to identify non-coding drivers from these data.
The ICGC and TCGA Pan-Cancer Analysis of Whole Genomes (PCAWG) effort, which has collected and systematically analyzed >2,700 cancer genome sequences from >2,500 patients representing a variety of cancer types17, offers an unprecedented opportunity to perform a comprehensive analysis of putative coding and non-coding driver events. Here, we describe results from multiple methods for detecting such drivers based on somatic SNVs and indels and a framework for integrating them. Using this approach, we identify significantly mutated genomic elements of various types in individual cancer types and across cancer types. We then use additional sources of information to remove from the initial list of significant elements those likely to result from mapping issues caused by repetitive regions or by inaccurate estimation of the local background mutation rate, and classify recurrently mutated elements as putative cancer-drivers or as likely false positives. Finally, we estimate our power to discover novel drivers in the PCAWG data set and evaluate the overall density of non-coding regulatory driver mutations around known cancer genes.
Overview of recurrent mutations across cancer types
Many protein-coding driver mutations and the two well-studied TERT promoter mutations occur in frequently mutated single base genomic “hotspots”. In order to obtain an initial view of these in the PCAWG data set, we generated a genome-wide list of highly mutated hotspots by ranking all sites in the genome based on the number of cancers that harbor somatic mutations in them. Only 12 genomic sites were recurrently somatically mutated in more than 1% (26) of cancers and 106 in more than 0.5%, while the vast majority (93.19%) of somatic mutations were private to a single patient’s cancer (Methods; Extended Data Fig. 1). Interestingly, although protein-coding regions span only ~1% of the genome, 15 (30%) of the 50 most frequently mutated sites were known amino acid altering hotspots in known cancer genes (in KRAS, BRAF, PIK3CA, TP53, and IDH1) (Fig. 1a). Also among the top hits were the two TERT promoter hotspots. The top 50 list further contained a high proportion of mutations almost exclusively contributed by lymphoid malignancies (Lymph-BNHL and Lymph-CLL) (18 hotspots) and melanomas (6). The melanoma-specific hotspots were located in regions occupied by transcription factors (5/6 in promoters), a phenomenon that has recently been attributed to reduced nucleotide excision repair (NER) of UV-induced DNA damage17–19. Indeed, signature analysis of the melanoma hotspot mutations attributed them to UV radiation (Fig. 1b). The hotspots in lymphoid malignancies reside in the IGH locus on chromosome 14, a known phenomenon in B-cell-derived cancers where IGH has undergone somatic hypermutation by the activation-induced cytosine deaminase (AID), an observation confirmed by mutational signature analysis (Fig. 1b). Hotspots in an intron of GPR126 and in the promoter of PLEKHS13, 20 were located inside palindromic DNA that may fold into hairpin structures and expose single-stranded DNA loops to APOBEC enzymes3; accordingly, these mutations had high APOBEC signature probabilities. The six remaining non-coding hotspots could be attributed to a combination of mutational processes and presumed technical artifacts (Supplementary Note). In contrast to these non-coding and cancer-specific sites, mutations in protein-coding hotspots were attributed to the aging signatures, or to a mixture of signatures. Our observation that, with the exception of the TERT promoter hotspots, all of the top non-coding hotspots are generated by highly localized, cancer-specific mutational processes suggests that these may be passengers.
a, Characteristics of top 50 single-site hotspots in PCAWG SNV data. The stacked bar chart (left) shows the total number of patients mutated across the PCAWG cohorts colored by mutation type. Gene names are given when hotspots overlap functional elements (colour-coded). Known somatic driver sites are marked with a hashtag. Amino acid change is indicated for protein-coding genes. The genomic location (chr:position) is colored by overlap with center of palindromes (orange) or immunoglobulin loci (brown). The table (right) shows the distribution of hotspot SNVs in the PCAWG cohorts sorted by number of hotspots. Tumor-type specific hotspots are indicated with a box. This only includes cohorts with at least 20 patients, and at least 10 patients or 10% of patients with a SNV. The Lymph-BNHL and Lymph-CLL cohorts are shown together as Lymphoid malignancies. See Extended Data Fig. 1 for a complete set of individual and meta cohorts. b, Mutational signature attributions for mutations in each hotspot site. c, Schematic describing definition of functional element types from GENCODE and ENCODE annotation resources. Functional elements (black) are defined based on transcript annotations from various databases. Elements arising from multiple transcripts with the same gene ID are collapsed, as seen here for the protein-coding isoforms. Promoter elements are defined as 200 bases upstream and downstream of the transcription start sites of a gene’s transcripts (marked with red). Splice site elements extend 6 and 20 bases from 3’ and 5’ exonic ends into intronic regions respectively, as indicated. Regions overlapping protein-coding bases and protein-coding splice sites are subtracted from other regions, as indicated for the lncRNA promoter element in gray. d, Organization of meta-cohorts defined by tissue of origin and organ system (Methods). Pan-cancer contains all cancers excluding Skin-Melanoma and lymphoid malignancies.
Next, we sought to systematically identify genomic elements that drive cancer by comprehensively analyzing SNVs and indels in transcribed and regulatory genomic elements, including protein-coding genes, promoters, 5’UTRs, 3’UTRs, splice sites, distal regulatory elements/enhancers, lncRNA genes, short ncRNAs, and miRNAs (in total ~4% of the genome) (Methods; Fig. 1c; Extended Data Fig. 2; Supplementary Table 1). Drivers can be common across many cancer types (such as TP53, PIK3CA, PTEN, and TERT promoter) or highly specific (e.g. CIC and FUBP1 in oligodendroglioma, BCR-ABL in chronic myelogenous leukemia, and FOXA1 in breast and prostate cancers). Analysis of large mixed tumor cohorts increases the power to discover common drivers but dilutes the signal from cancer-specific ones. In order to maximize our ability to detect common and cancer-specific drivers, we performed analyses of individual tumor types, tumors grouped by their tissue of origin or organ site (“meta-cohorts”), as well as a pan-cancer set (Fig. 1d; Methods). This aggregation of tumors by tissue or organ site further allowed us to take advantage of patients from cohorts too small to analyze individually (<20 patients). Overall, we analyzed 2,583 unique patient samples in 27 individual tumor types and 15 meta-cohorts.
Discovery of driver elements
In order to identify bona fide drivers in each cohort, we collected and integrated results from multiple driver discovery methods. These methods identify statistically significant elements based on SNVs and indels and evidence from one or more of three criteria: (i) mutational burden, (ii) functional impact of mutation changes, and (iii) clustering of mutation sites in hotspots (Fig. 2a; Methods; Supplementary Table 2). Generally, mutational burden and clustering of events are independent of the type of genomic element that is tested for significance. The ability to use measures of functional impact, however, may vary greatly depending on the element type. In protein-coding genes, functional impact is assessed through the amino acid change introduced by a given mutation. For some non-coding elements, the impact of mutations on transcription factor binding sites, miRNA binding sites, or functional RNA structure may be predicted. However, in general, the exact consequences of somatic mutations are harder to assess in non-coding regions than in protein-coding regions, and may even vary substantially across cell types.
a, Overview of driver discovery method and their lines of evidence to evaluate candidate gene drivers. Methods employing each feature are marked with blue boxes next to the appropriate track. b, Spearman’s correlation of p-values across the different driver discovery algorithms based on simulated (null model) mutational data. Dendrogram illustrates relatedness of method p-values and algorithm approaches marked by colored boxes on dendrogram leaves. c, P-values are combined with Brown’s method based on the correlation structure calculated in (b). Individual method (left) and integrated (right) log-transformed p-values are shown in a heatmap (grey: missing data). d, Post-filtering used several criteria to identify likely suspicious candidates (shaded grey rectangle). e, Significant driver candidates were identified after controlling for multiple hypothesis testing based on an FDR q-value threshold of 0.1 (blue asterisk). Candidates with q-values below 0.25 (blue dash) were also considered of interest.
Integration of results from multiple driver discovery methods
Most cancer genomic studies to date have employed a single method to identify putative cancer driver elements. This can introduce biases because the number and type of drivers detected by individual methods depends on their underlying assumptions and statistical model. To reach a comprehensive consensus list of candidate drivers, we collected the p-values calculated by up to 16 driver discovery methods for each genomic element (Supplementary Table 2, 3) and derived a general strategy to integrate them (Fig. 2a). Our strategy accounts for the correlation of p-values among driver discovery methods that use similar approaches (burden, clustering, functional impact) when applied to the same data, while accumulating the independent evidence provided by different methods (Fig. 2b). To integrate potentially correlated statistical results, we used Brown’s method for combining p-values21, an extension of the popular Fisher’s method22 (Fig. 2c). We first used simulated data sets that preserve local mutation rates and signatures but do not contain drivers, to assess the specificity of each method and the correlation structure among different methods when no genomic element is expected to have a significant excess or pattern of mutations (Methods; Extended Data Fig. 3). As anticipated, p-values were correlated among methods that share similar approaches. We then used the resulting correlation structure to integrate p-values from the observed (real) mutation data from different methods into a single p-value for each genomic element. Since simulated data sets are based on assumptions, we tested whether we can directly estimate the correlation structure from p-values calculated on the observed data. As this simpler approach yielded analogous results, it was used throughout this study (Methods, Extended Data Fig. 3). After additional filtering steps (see below), we conservatively controlled for multiple hypothesis testing using the Benjamini-Hochberg procedure (BH)23 across the 27 individual and 15 meta-cohorts, for each element type separately (Fig. 2e). Cohort-element combinations (hits) with q<0.1 (10% false discovery rate [FDR]) were considered significant. Overall, the union of all hits from the individual driver discovery methods contained 3,048 cohort-element pairs involving 1,694 unique elements, whereas the integrated results contained 1,406 significant hits with 635 unique elements (Supplementary Table 4, 5).
Flagging and removal of potential false positive hits
Even after careful variant calling, false positive identification of driver loci can arise from residual inaccuracies in background models, residual sequencing and mapping artifacts or from local increases in the burden of mutations generated by as-yet unmodeled mutational processes. To minimize technical issues, we carefully reviewed and filtered all candidate driver elements based on low-confidence mappability regions and site-specific noise in a panel of normal samples from PCAWG (Fig. 2d; Fig. 3a,b; Methods). For example, the lncRNA RN7SK is a small nuclear RNA with many homologous regions in the genome, making it prone to read mapping errors (Fig. 3a), while a fraction of LEPROTL1 mutations fell in positions flagged by the site-specific noise filter (Fig. 3b).
a, Read mappability based on alignability, uniqueness and blacklisted regions. Example shown for RN7SK, which was removed due to low alignability. b, Site-specific noise evaluated in normal samples from PCAWG. c, Increased local mutation density caused by AID and UV mutational processes in lymphoma and melanoma, respectively. d, Palindromic DNA can expose bases to APOBEC enzymes. f, Number of significant unique hits (left) and unique elements (right) removed by each filter.
In addition, mutational processes not properly captured by current background models can cause local accumulation of mutations. As shown in the hotspot analysis above, events in the promoters of PIM1 and RPL13A are strongly associated with the AID and UV mutational processes, respectively, as previously noted18, 19, 24 (Fig. 3c). Moreover, DNA palindromes may form hairpin structures that expose loop bases to APOBEC enzymes, as has been proposed for PLEKHS1, GPR126, TBC1D12 and LEPROTL12, 3 (Fig. 3d). Non-coding RNAs that form secondary structures often contain palindromic sequences, which could be preferential targets for APOBEC enzymes. However, only a small fraction of the mutations in structural RNAs appear to be APOBEC derived overall (4.8%) (Supplementary Note). In total, 569 out of 1,406 hits (element-cohort combinations with q<0.1) and 383 out of 635 unique elements that were initially significant, were filtered out, resulting in 837 hits (Fig. 3e; Supplementary Tables 4, 5). Finally, multiple hypothesis correction was repeated after assigning a non-significant p-value (P = 1) to the filtered out hits, resulting in further removal of 91 candidate element-cohort combinations.
Performance of the integration method
To evaluate the sensitivity and validity of the integration approach, we compared the performance of individual methods and of the integration on protein-coding genes using, as a gold-standard, a list of 603 known recurrently-mutated cancer genes in the Cancer Gene Census25 (CGC v80). These analyses revealed that, typically, the integration of different methods outperformed individual methods, both in terms of sensitivity and specificity, yielding lists of significant genes that were longer and more enriched in high-confidence cancer genes (Extended Data Fig. 4).
Discovery of recurrently mutated elements
Our conservative integration and filtering strategy yielded 746 total hits, which include 646 hits in 157 protein-coding genes, 30 hits in seven protein-coding promoters, 26 hits in eight long non-coding RNAs, 18 hits in six non-coding RNA promoters, ten hits in six 3’UTRs, six hits in four 5’UTRs, six hits in one enhancer, three hits in one microRNA and one hit in a small RNA candidate (Fig 4a; Supplementary Table 5). The number of significant elements varied from just one in clear-cell renal cancer to 78 in the Carcinoma meta-cohort (Fig. 4b) and depended strongly on cohort size (r = 0.84; P = 1.9x10−8; also see power analysis below), reflecting that the landscape of driver candidates, particularly in small tumor cohorts, is still incomplete.
a, Number of total hits (elements significant in a certain patient cohort; left) and number of unique significant elements (right). b, Number of significant elements by type and by cohort. c, Significant non-coding elements identified in this study. Element types are indicated with colors as in Fig. 1c.
Comparison of the p-values obtained from individual tumor types and meta-cohorts revealed that although most candidate drivers gained significance in larger meta-cohorts, the tumor-type specific genes DAXX (Panc-Endocrine), BRAF (Skin-Melanoma), NRAS (Skin-Melanoma), SPOP (Prost-AdenoCa) and hsa-mir-142 (Lymph-BNHL) scored higher in their respective tumor types (Extended Data Fig. 5). These results emphasize the need for careful tradeoff between tumor-type specificity and cohort size when searching for drivers.
Protein-coding mutations
The ability to discover known protein-coding cancer genes from WGS data provided the opportunity to evaluate our strategy and to put non-coding driver hits in context (Extended Data Fig. 6). Overall, candidate coding drivers were concordant with previous results: of the 157 genes significant in at least one patient cohort, 64% are listed in the CGC and 87% are in a more comprehensive list of cancer genes (https://doi.org/10.1101/190330). In contrast to prior studies based on large exome sequencing data sets, the moderate number of patients per cancer type in this data set provided sufficient power to detect only the genes with the strongest signal. Indeed, when we relaxed the significance threshold (q<0.25) to detect hits “near significance”, we found 93 additional hits in 62 unique genes. Half (31) of these genes were already discovered as significant (q<0.1) in at least one cohort in this study, and now became significant in additional cohorts. Of the 31 genes not previously identified by our analyses, 19% were in the CGC and 32% in the more comprehensive list of drivers. These results confirm that many cancer genes were just beyond our significance threshold and would likely have been discovered with larger cohorts (Supplementary Table 4).
Non-coding driver candidates
In contrast to protein-coding elements, there was far less agreement among significance methods when analyzing non-coding elements, and most hits were strongly supported by only a few methods (Extended Data Fig. 7). The limited agreement between driver discovery methods in non-coding elements is likely due to a weaker signal of selection and different underlying assumptions that imperfectly model background mutation frequencies in non-coding regions. Thus, in order to nominate a significant element as a candidate driver, we carefully reviewed the supporting evidence from the genomic data and sought additional evidence from PCAWG (chromosomal breakpoints, copy-number, loss-of-heterozygosity and expression data), cancer gene databases and the literature (Supplementary Table 6).
Promoters and 5’UTRs
Due to the overlap between the downstream regions of promoters and 5’UTRs, we reviewed the 11 significant elements from our promoter and 5’UTR analyses together, yielding several interesting candidates. Our integrated results confirmed that the TERT promoter is the most significantly mutated promoter in cancer, being significant in 8 individual tumor types and 11 meta-cohorts (Fig. 4c; Supplementary Table 5). As previously reported, we observed significantly higher TERT transcript expression levels in mutated compared to non-mutated cases (Extended Data Fig. 8).
Among the remaining candidates, we found recurrent mutations in the promoters of PAX5 and RFTN1 in lymphoma. PAX5 is a known off-target of AID hypermutation24 and, indeed, the percentage of its mutations attributable to AID activity (46%) was just below our filtering threshold. Consistently, mutations in its promoter were not associated with gene expression changes, suggesting that these mutations may be passengers (Extended Data Fig. 8). In contrast, mutations in the promoter of RFTN1 were weakly associated with an increase in RFTN1 expression levels (P = 0.03; mutant/wild type fold difference (FD) = 1.2; P = 0.02, after excluding 8/21 mutations attributed to the AID signature) (Fig. 5a). The protein encoded by RFTN1, Raftlin, associates with B-cell receptor complexes26 but its function in lymphoma is unclear.
a, RFTN1 promoter locus (left) and gene expression for mutant (mut.) and wild-type (wt) (right) in Lymph-BNHL tumors. Expression values are colored by copy number status; AID mutations are highlighted in yellow. b, A mutational hotspot inside RNF34 overlapping a DNAse I site (left) correlates with gene expression changes of KDM2B in Liver-HCC (right). Coloring of mutations as in a. c, MTG2 promoter locus (left) and associated gene expression changes in Carcinoma tumors (right). Coloring of mutations as in a. d, An enhancer associated with TP53TG1 (Methods) (left) contains mutations mostly attributed to an esophageal cancer-associated signature.
Mutations called in the 5’UTR of RNF34 in Liver-HCC overlap an intronic DNAse I hypersensitive region in multiple cell types27, suggesting that events may affect expression of RNF34 or a neighboring gene through an intragenic enhancer. Indeed, expression of the histone demethylase KDM2B located downstream of RNF34 shows weak correlation with these mutations (P = 0.03; FD = 1.3; Fig 5b) with no effect on RNF34 mRNA (P = 0.41). Mutations in the 5’ region of MTG2 were concentrated in a hotspot and showed marginally significant decreased expression in the Pan-cancer (P = 0.036; FD = 0.8) and Carcinoma (P = 0.029; FD = 0.8) meta-cohorts (Fig. 5c; Extended Data Fig. 8). MTG2 encodes a little-studied GTPase that associates with the mitochondrial ribosome28. HES1 promoter mutations were significant in Carcinoma and Pan-cancer, but showed no association with gene expression (Extended Data Fig. 8). HES1 is a NOTCH signalling target29, and is focally amplified in gastric cancers (Extended Data Fig. 9).
PTDSS1, DTL (and INTS7, which shares a bidirectional promoter), IFI44L and POLR3E showed trends towards increased or decreased expression in mutated tumors, although the small number of samples prevented us from drawing definitive conclusions (Extended Data Fig. 8). None of these genes were present in significantly amplified or deleted focal peaks (Methods). Validation of these hits with additional data in further studies will be needed to evaluate whether they are genuine drivers.
As previously reported in a number of studies1, 3, 7, we also found recurrent mutations in the promoter of WDR74, which was significant in multiple cohorts. The mutations concentrate in a small region overlapping an evolutionarily conserved spliceosomal U2 snRNA and therefore potentially affect splice patterns. However, no significant associations with either WDR74 or transcriptome-wide splicing were seen (Supplementary Note; Methods). We did, however, find that the repetitive U2 sequence was frequently affected by recurrent artifacts in SNP databases, raising concerns about potential mapping errors for this repetitive RNA type (Extended Data Fig. 10).
Finally, restricted hypothesis testing of the promoters of CGC genes (n = 603) revealed significant recurrence of mutations in the promoter region of TP53 (11 patients in the pancancer cohort; q = 0.044, NBR method). Most of these mutations were substitutions and deletions affecting either the TSS or the donor splice site of the first non-coding exon of TP53. For the 6 samples with expression data, these mutations were associated with a dramatic reduction of expression (Fig. 5d). In 8 of the 11 mutant cases, the mutation occurred in combination with copy-neutral loss of heterozygosity. This is the first report of a relatively infrequent, but impactful, form of TP53 inactivation by non-coding mutations.
Enhancers
Two enhancer regions were found significant in our integration analysis; one at the IGH locus in Lymph-CLL and the other near TP53TG1 in multiple cohorts. However, many of the recurrent mutations in the enhancer overlapping the IGH locus are due to characteristic somatic hypermutation; 48% of mutations in this enhancer element matched the AID signature, which is just below our filtering threshold. Nine of 18 mutations in the enhancer near TP53TG1 were contributed by esophageal cancers. However, 89% of mutations in these esophageal tumors were attributed to a mutational signature common in this cancer type (mostly T>G substitutions in NTT context), raising the concern that these reflect a localized mutational process rather than a driver event30 (PCAWG7 publication). These mutations were concentrated around a conserved region (Fig. 5e) and overlapped sites bound by NFIC and ZBTB7 transcription factors in HepG2 cells27. TP53TG1 is a noncoding RNA suggested as a tumor suppressor involved in p53 response to DNA damage and has been reported to be epigenetically silenced in cancer31.
3’UTRs
Recurrent somatic events were identified in the 3’UTRs of six genes: FOXA1 in prostate cancer, ALB in liver cancer, SFTPB in lung adenocarcinoma, NFKBIZ and TMEM107 in lymphomas, and TOB1 in Carcinoma, most of which contained a large number of indels (Fig. 4c). High rates of indels in the 3’UTR of ALB in liver cancer and SFTPB in lung cancer have been recently reported8 but it is unclear whether they are cancer drivers or the result of a poorly-understood mechanism of localized indel hypermutation. To determine whether indels affect function and accumulate specifically in the 3’UTRs, we compared indel rates in 3’UTRs with other regions around these genes. Consistent with local hypermutation, rather than selection, we observed similar indel rates in 3’UTR and 1 kb downstream of the polyadenylation site in ALB, SFTPB and FOXA1 (Fig. 6a). This, together with additional observations presented below, strongly suggest that indel recurrence at these loci is caused by a novel localized hypermutation process.
a, Quantification of indel rates for functional elements associated with significant 3’UTRs. b, Overview of ALB genomic locus and PCAWG variants illustrates local density of indels. Coloring of enlarged ALB gene elements corresponds to functional elements in barplots in a. Copy number changes from 314 Liver-HCC cases are compared to other samples from PCAWG. c, Genomic locus of TOB1 3’UTR. Note extreme conservation (PhyloP) for the 3’UTR. d, Genomic locus of NFKBIZ 3’UTR (left) and associated gene expression changes in Lymph-BNHL (right). e, Genomic locus of the RMRP transcript and promoter region (left), and its RNA secondary structure, tertiary structure interactions, protein and substrate interactions, and mutations with their predicted structural impact (right; Extended Data Figure 9d); lymphoma and melanoma are excluded.
On the other hand, although ALB is subject to localized hypermutation, protein-coding SNVs in ALB are enriched in missense and truncating events relative to synonymous mutations in liver cancer (P = 1.5x10−7; Fig. 6b). Furthermore, the ALB locus is lost in a fraction of liver cases (Fig. 6b), with copy losses having a tendency to occur in samples without somatic mutations (Fisher’s exact test, P = 0.0099). These findings, suggest that loss of ALB may be a genuine driver event in liver cancer. Similarly, the well-known driver role of FOXA1 in prostate tumors is supported by the functional impact of coding mutations (P = 4.4x10−7) and focal amplifications, raising the possibility that the mutation enrichment observed in the 3’UTR might also be functional.
In contrast to the genes discussed above, the number of indels in the 3’UTRs of NFKBIZ and TOB1 was significantly higher than in other parts of these genes, suggesting that indels in the 3’UTRs could be under functional selection (Fig. 6a). TOB1 encodes an anti-proliferation regulator that associates with ERBB2, and also affects migration and invasion in gastric cancer32. As part of the CCR4-NOT complex, TOB1 regulates other mRNAs through binding to their 3’UTR and promoting deadenylation33. Tumors with 3’UTR mutations in TOB1 showed a trend towards decreased expression (P = 0.053; FD = 0.7). Mutations did not concentrate in known miRNA binding sites, possibly due to incomplete annotation (Fig. 6c). However, the extreme conservation of this region (5th most conserved among all tested 3’UTRs with an average PhyloP score of 5.6) indicates that these events likely disrupt functionally important sites (Fig. 6c). Interestingly, TOB1 and its neighboring gene WFIKKN2 are focally amplified in breast cancer and pan-cancer, suggesting a complex role in cancer (Extended Data Fig. 9). NFKBIZ is a transcription factor that is mutated in recurrent/relapsed DLBCL34 and amplified in primary lymphomas34, 35. 3’UTR mutations accumulated in a hotspot proximal to the stop codon and upstream of conserved miRNA binding sites which might be affected by these mutations (Fig. 6d). Only two events in these regions are SNVs, suggesting that mutations in the NFKBIZ 3’UTR are not the consequence of AID off-target activity. Lymphomas with 3’UTR mutations showed a trend towards increased mRNA levels of NFKBIZ (P = 0.035; FD = 3.2; the trend remains after correction for copy number, P = 0.03; Fig. 6d). Functional studies will be required to understand the exact function of 3’UTR mutations on transcript regulation of TOB1 and NFKBIZ.
Mutations in the significant 3’UTR of TMEM107 in lymphoma overlap the U8 small RNA SNORD118 (Extended Data Fig. 10). Similar to WDR74, overlap with a repetitive RNA element suggests that these mutations might be caused by mapping artifacts.
Non-coding RNAs
ncRNAs are often part of large gene families with multiple copies in the genome. Sequence similarity with other genomic regions may therefore complicate their analysis. Similarly to the U2 RNA upstream of WDR74, we observed high levels of germline polymorphisms in normal samples in some of the significant ncRNAs and their promoters (Extended Data Fig. 10). Though interesting candidates, caution should thus be exercised in interpreting these hits.
The non-coding RNA RMRP is significantly mutated in multiple cancer types, in both its gene body and promoter (Fig 4c; Fig 6e; Supplementary Table 5). RMRP is the RNA component of the endoribonuclease RNase MRP, an enzymatically active ribonucleoprotein36, 37, 38. Its catalytic function depends on the RNA secondary and tertiary structure and its interactions with proteins39. Germline mutations in RMRP cause cartilage-hair hypoplasia, and somatic promoter mutations have been reported to be functional9. In addition, the RMRP locus is focally amplified in several tumor types, including epithelial cancer (Extended Data Fig. 9c). SNVs in the gene body (7 in pan-cancer) are significantly biased towards secondary structure impact (P = 0.011, permutation test)40, 41. Three of these are individually significant (each with P < 0.1, sample level permutation tests), with two affecting the same position and one located in a tertiary interaction site (Fig. 6e). Of the four gene-body indels, three are located in or near protein-binding sites (P = 0.08), including a deletion that is predicted to affect the secondary structure (Extended Data Fig. 9d). Given RMRPs role in replication of the mitochodrial genome36, we tested whether mutations in this locus were associated with altered mitochondrial genome copy number. Indeed, mutated samples showed a trend towards higher mitochondrial copy number (two-sided rank-sum test, P = 0.1). However, RMRP also appears to be a target of potential artifacts (Extended Data Fig. 10) and so its relevance to cancer requires further scrutiny.
The microRNA precursor (pre-miRNA) for miR-142 was found significant in Lymph-BNHL, Lymphatic system and Hematopoietic system (Fig. 4c; Supplementary Table 5). The locus is a known AID off-target in lymphoma5, 42(Extended Data Fig. 10). However, five of seven mutations (71%) in the dominantly expressed mature miRNA mir-142-p3, where the largest functional impact is expected, were not assigned to AID, raising the possibility that they may be under selection5.
Eleven additional ncRNA-related elements (RNU6-573P, RPPH1, RNU12, TRAM2-AS1, G025135, G029190,RP11-92C4.6 and promoters of RNU12, MIR663A, RP11-440L14.1, and LINC00963) passed our stringent post-filtering, but had limited supporting evidence after manual inspection. These are discussed in a Supplementary Note. We further checked whether any of our non-coding candidates were associated with known germline cancer risk variants, and found no significant associations, with the exception of TERT (Methods).
NEAT1 and MALAT1, two neighboring lncRNAs previously reported as recurrently mutated in liver43 and breast cancer3, were found significant in multiple cohorts (Fig. 4c), mutated in a large number of patients (369 for NEAT1 and 158 for MALAT1; Fig 7a). In addition, these genes are significantly associated with nearby chromosomal breakpoints (NEAT1: P = 5.3x10−9; MALAT1: P = 0.0012; Methods; Extended Data Fig. 11). The pattern of breakpoints, indels and SNVs scattered throughout their sequence would suggest a possible tumor suppressor role. To evaluate this hypothesis, we tested whether MALAT1 and NEAT1 mutations are associated to loss of heterozygosity (LOH). Neither gene showed biallelic loss, in contrast to many known canonical tumor suppressors (Fig. 7b; Extended Data Fig. 11). Mutations in MALAT1 and NEAT1 also do not exhibit higher cancer allele fractions compared to mutations in flanking regions, a feature of early driver mutations (Fig 7c; Extended Data Fig. 11). Furthermore, NEAT1 and MALAT1 mutations are not associated with altered expression levels, suggesting that they do not affect post-transcriptional stability nor show increased expression like many mutated oncogenes (Fig 7d; Extended Data Fig. 11).
a, Genomic locus overview showing the distribution of indels and SNVs in NEAT1 and MALAT1. b, Analysis of enrichment in LOH associated with mutation (“double-hit”) for canonical tumor suppressors (TSG), oncogenes (OG), and NEAT1/MALAT1. c, Comparison of cancer allelic fractions within those same genes and within flanking regions (including 2 kbp upstream, 2 kbp downstream and introns). d, Difference in expression between mutated and wild-type alleles of these genes. e, Percentages of different groups of indel sizes for all protein-coding and lncRNA genes, ALB, NEAT1, MALAT1 and the set of genes enriched in 2-5 bp indels. f, SNV and indel rates (total events/bp) in different functional regions of 18 protein-coding genes enriched in 2-5 bps indels (without ALB, which contributed 47% of indels). Red lines indicate background indel and SNV rates estimated from all protein-coding genes.
The high enrichment of indels throughout the gene body of NEAT1 and MALAT1, which have very high expression levels across many cancer types, resembles the phenomenon described above for ALB and SFTPB. If these indels are generated by a specific mutational process, we might expect distinct features from indels found elsewhere in the genome. Indeed, we find that indels in NEAT1, MALAT1, ALB and SFTPB are strongly enriched in events longer than 1 bp and particularly in indels of length 2-5 bp, compared to the genomic background (least significant Fisher’s P = 6.8x10−5, for MALAT1; Fig. 7e). The association of these indels with highly expressed genes suggests a transcription-coupled mutagenic mechanism, possibly transcription-replication collisions44–46. A systematic search of genes with increased rates of 2-5 bp indels reveals that this yet-unknown mutational process affects other highly expressed genes (Extended Data Fig. 11), some of which were reported in a recent study8, indicating that these indel sizes are a feature of the reported mutational process. Interestingly, SNVs also occur at higher frequencies in these genes, suggesting that they may be generated by the same mechanism, although less frequently (Fig. 7f).
Overall, the discovery of a localized mutational signature and the lack of association of the mutations in MALAT1 and NEAT1 with loss of heterozygosity or higher allele frequencies suggests that indels in these genes are most likely passenger events. The previously reported oncogenic phenotypes associated with both ncRNAs43, 47, 48 may thus be related to other types of alterations.
Localized lack of power to call mutations
Certain genomic regions, especially those with high GC content, are subject to systematic low coverage in next-generation sequencing49. We used power analysis to quantify the number of possibly missed mutations, especially in non-coding regions. These calculations attempt to account for variations in local sequencing depth, purity and overall ploidy of individual tumor samples, background mutation rates across cohorts and elements, and the size of patient cohorts9,50–52. We found that mutation detection sensitivity was high across the element types, (median d.s. = 0.98) but slightly lower for the typically GC-rich promoters (median d.s. = 0.96) and 5’UTR elements (0.96; Fig. 8a; Supplementary Table 4,5). However, for some individual genomic regions the detection sensitivity is dramatically reduced and it is therefore important to evaluate it for elements of interest. For example, only 43% of tested cases have at least 50% average detection sensitivity across the third exon of TCF7L2 and 20% in the 5’UTR of AKT1 (Extended Data Fig. 12a). Positional detection sensitivities for the two canonical TERT promoter hotspot sites were highly variable among patients and cohorts, ranging from 4% of sufficiently powered (≥90%) patients in CNS-PiloAstro to 100% of patients in Thy-AdenoCa (Fig. 8b; Extended Data Fig. 12b). This means that we have nearly no information about possible TERT promoter mutations in CNS-PiloAstro and incomplete information in other cohorts. Using the detection sensitivity and observed mutation count at the TERT hotspot sites, we inferred the expected total number of events in each tumor type (Fig 8c). This analysis revealed that ~216 (CI95 = [188,245]) TERT hotspot mutations were likely missed in this study due to lack of power. Likewise, about four additional mutations would be expected at the recurrent FOXA1 promoter hotspot shown to be mutated in hormone-receptor positive breast cancers (Extended Data Fig. 12)9.
a, Cumulative distribution of average detection sensitivity (d.s.) from 60 PCAWG pilot samples in different functional element types. b, Distribution of TERT promoter hotspot (chr5:1295228; hg19) detection sensitivity for each each patient, by cohort. Grey dots indicate values for individual patients inside estimated distribution (areas colored by cohort). Horizontal black bars mark the medians. Numbers above distributions indicate the percentage of patients powered (d.s. >90%) in each cohort. See Extended Data Fig. 12b for the second TERT hotspot. c, Percentage of patients with observed (blue) and inferred missed (red) mutations at the chr5:1,295,228 and chr5:1,295,250 TERT promoter hotspot sites. Error bars indicate 95% confidence interval. Numbers above bars show the total inferred number of TERT promoter mutations for each site in this cohort. Red numbers indicate the absolute number of inferred missed mutations (due to lack of read coverage). d, Heatmap shows minimal frequency of a driver element in a cohort that is powered (≥90%) to be discovered. Cell numbers indicate the number of patients in a given cohort that would need to harbor a mutation in a given element. For example, the pan-cancer cohort is powered to discover a driver gene (CDS) present in <1% or 18 patients, while the Bladder-TCC cohort is only powered to discover drivers present in at least 27% or 6 patients. Power to discover driver elements is dependent on the background mutation frequency (shown above the heatmap) and element length (shown to the right). e, The ratio between observed numbers of mutations (SNVs and indels) in regulatory and coding regions of 142 protein-coding cancer genes. The absolute number of driver mutations predicted, with CI95% in brackets, is shown above each bar.
These calculations suggest that a considerable number of potentially impactful somatic mutations were not detected due to systematic low sequencing coverage, and that specific positions can be unpowered in elements with overall sufficient coverage.
We further evaluated the discovery power for recurrent events identified in mutational burden tests in the tumor cohorts analyzed in this study9, 50. As expected, discovery power was highest in cohorts with many patients and low background mutation densities allowing discovery of typical-sized driver elements mutated in fewer than 1% of patients in the Carcinoma, Adenocarcinoma and pan-cancer meta-cohorts with 90% power (Fig. 8d). In contrast, the small Bladder-TCC cohort (n=23) with relatively high background mutation density (~2.7 mutations/Mb) is only powered to discover drivers that occur in at least 25% of patients (Fig. 8d). Overall, power differences between element types for individual tumor cohorts were small, suggesting that lack of power alone cannot explain the observed paucity of regulatory driver elements.
Relative paucity of non-coding driver point mutations
To further analyze the relative paucity of driver mutations in non-coding elements, we sought to estimate the overall number of driver mutations, in coding and non-coding regions of known cancer genes. Given the limited statistical power to detect individually significant elements, we combined the signal from multiple elements, as recently described53. The difference between the total number of mutations observed across a combined set of elements and the number of passenger mutations that is expected by chance, i.e. the excess of mutations, approximates the total number of driver mutations in the set of elements. To estimate the expected number of background mutations, we fit the NBR model to presumed passenger genes, controlling for sequence composition, element size and regional mutation densities (Methods; Supplementary Table 7; Supplementary Note)53. We focused on 142 known cancer genes that include the significantly mutated cancer genes in this study and frequently amplified or deleted cancer genes (Methods). We specifically excluded TERT from this analysis because of its high frequency of promoter driver mutations and the incomplete detection sensitivity reported above.
Overall, this approach predicted an excess of 3,133 driver mutations (CI95% [2,987-3,273]; 2,258 SNVs and 875 indels) in the protein-coding sequences of these genes across the pancancer meta-cohort (Fig. 8e). In contrast to coding regions, the observed number of mutations across the combined set of non-coding elements associated with these genes is very close to the expected number of passenger mutations,, with an excess of 71 (CI95%:22-133) mutations in promoters, 32 (0-79) in 5’UTRs, and 103 (25-184) in 3’UTRs. These results indicate that coding genes contribute the vast majority (>90%) of driver point mutations in these 142 cancer genes. Importantly, these estimates are conservative, since the estimation of the background number of mutations from putative passenger genes may include yet undetected driver mutations. In addition, non-coding mutations in promoters of cancer genes were also not generally associated with LOH nor with altered expression, as one would expect if they were enriched with driver mutations (Supplementary Note). Altogether, our results suggest that mutations in protein-coding regions dominate the landscape of driver point mutations in known cancer genes.
Discussion
If we are to fulfill the ambitions of precision medicine, we need a detailed understanding of the genetic changes that drive each person’s cancer, including those in non-coding regions (https://doi.org/10.1101/190330). Most cancer genomic studies have focused on protein-coding genes, leaving non-coding regulatory regions and non-coding RNA genes largely unexplored. The unprecedented availability of high-quality mutation calls from whole-genomes of >2,500 patients across 27 cancer types has enabled us to comprehensively search for functional elements with cancer-driver mutations across the genome. To obtain reliable results and avoid common pitfalls in detecting drivers54, we have benchmarked a large number of methods and developed a novel and rigorous statistical strategy for integrating their results.
Among the most interesting candidate non-coding driver elements identified in this analysis, we have uncovered promoter or 5’UTR mutations in TP53, RFTN1 and MTG2, a putative intragenic regulatory element in RNF34 that possibly regulates KDM2B; 3’UTR mutations in NFKBIZ and TOB1; and recurrent mutations in the non-coding RNA RMRP. We have also found evidence suggesting that a number of previously reported and frequently mutated non-coding elements may not be genuine cancer drivers. Particularly, the non-coding RNAs NEAT1 and MALAT1 are subject to a high density of passenger indels, seemingly due to a transcription-associated mutational process that targets some of the most highly expressed genes.
Our study has yielded an unexpectedly low number of non-coding driver candidates. The results from four analyses - genomic hotspot recurrence, driver element discovery, discovery power and mutational excess - suggest that the regulatory elements studied here contribute a much smaller number of recurrent cancer-driving mutations than protein-coding genes. This contrasts the distribution of germline polymorphisms associated with heritability of complex traits, which are most frequently located outside of protein-coding genes 55.
Technical shortcomings, such as severe localized lack of sequence coverage in GC-rich promoters (as in the case of the TERT promoter), may lead to considerable underestimation of true drivers in certain regions. The yield of driver events will thus benefit from technical improvements, including less variable sequence coverage (e.g. PCR-free library preparation), longer sequencing reads and advanced methods for read mapping and variant calling that can overcome common artifacts. Moreover, studying larger tumor cohorts will provide increased power to discover infrequently mutated driver elements.
Our analyses have focused on currently-annotated non-coding regulatory elements and noncoding genes, comprising 4% of the genome, and exclusively on SNVs and indels. Although our genome-wide hotspot analysis did not detect novel highly recurrent non-coding mutation sites, it is possible that additional non-coding drivers reside outside of the regions tested here. Improvements in the functional annotation of non-coding elements, our understanding of their tissue-specificity, and the impact of mutations in them, will refine current driver discovery models and enable more comprehensive screens.
A challenge in the identification of non-coding drivers is distinguishing real novel driver events from yet-unidentified mutational mechanisms targeting certain genomic regions, such as the recently described hypermutation of transcription factor binding sites in melanoma56, 57. Better understanding of mutational processes and their activity along the genome will be critical to improve the sensitivity and specificity of statistical methods for driver discovery.
One potential explanation for the relative paucity of non-coding drivers is the smaller functional territory size of many regulatory elements, and hence smaller chance of being mutated, compared to protein-coding genes (Extended Data Fig.1; Fig. 8d). The presence of TERT promoter mutations at just two sites suggests that non-coding driver mutations may be confined to specific positions, similar to the small number of impactful sites in the oncogenes BRAF, KRAS and IDH1.
SNVs and small indels may not easily alter the function of non-coding regulatory elements. Directly mutating protein-coding sequences or altering expression levels by copy number changes, epigenetic changes or repurposing of distal enhancers through genomic rearrangements58, 59 may be more likely to provide large phenotypic effects. While protein-truncating or frameshift mutations can be created by a single nucleotide change, silencing regulatory regions of tumor suppressors may depend on large deletions to achieve a similareffect. Although the very high frequency of TERT promoter mutations in cancer demonstrates that some promoter point mutations can be potent driver events, our analyses suggest that the number of regulatory sites where a point mutation can lead to major phenotypic effects may be smaller than anticipated across the genome. This highlights TERT as an unusual example, perhaps because even a modest increase in the expression level of TERT may be sufficient for circumventing normal telomere shortening in cancer cells.
Comprehensive and reliable discovery of non-coding driver mutations in cancer genomes will be an integral part of cancer research and precision medicine in the coming years. We anticipate that the approaches developed here will provide a solid foundation for the incipient era of driver discovery from ever-larger numbers of cancer whole-genome sequences.
Author contributions
This work was carried out by the Discovery of Driver Elements Subgroup of the PCAWG Drivers and Functional Interpretation Group based on data from the ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Network. Contributions spanned a range of categories, as defined in Supplementary Note. The Driver discovery category includes method development, driver significance evaluation and results integration. The Candidate vetting and filtering category covers removal of potential false positive hits. The Case-based analysis category includes driver potential and performed follow-up studies using additional evidence. The Power analysis and driver mutations at known cancer genes category includes power evaluations and estimation of the number of true drivers. The Leadership and organisational work category includes organization and workgroup leadership. E.R., M.M.N., F.A., G.T., H.H., J.M.H., R.I.P., I.M., J.S.P. and G.G. wrote the manuscript. All authors have had access to read and comment on the manuscript.
Methods
Table of contents:
Patient cohorts
Mutational hotspot analysis
Mutational signatures
Genomic element definition
Driver discovery methods
Simulated data sets
Statistical framework for the integration of results from multiple driver discovery methods
Post-filtering of candidates
Sensitivity and specificity analysis
Gene expression analyses
Normalization for copy number variation
Mutation to expression association
Copy number analyses
Power analysis
Associations between mutation and signatures of selection: loss of heterozygosity and cancer allelic fractions
Signals of selection in aggregates of non-coding regions of known cancer genes
Mutational process and indel enrichment
Mutation association with splicing
Structural variation analysis
RNA structural analysis
Cancer associated germline variant distance to non-coding driver candidates
1. Patient cohorts
Generation of high-quality tumor set (Loris Mularoni)
We selected a total of 2583 samples to be included in the driver detection analyses. This list contains all the samples that were not flagged as problematic by the PCAWG-TECH group.
A single aliquot was assigned to each sample; in cases where multiple aliquots were present, we selected a single aliquot based on the following criteria, in order of importance:
- we prioritized primary tumors over metastatic or recurrent tumors
- we selected aliquots with an OxoG score higher than 40
- we prioritized aliquots with the highest quality (as indicated by the Stars values)
- we prioritized aliquots with RNA-seq data availability
- we prioritized aliquots with the lowest contamination (as indicated by the ContEst values)
- if a selection could not be made after applying the above filters we selected an aliquot randomly
Selection of tumor cohorts for analysis (Esther Rheinbay)
Individual tumor type cohorts from the high-quality tumor set were selected for analysis if they met a minimum size. This size was determined based on the cumulative number of patients, such that no more than 2.5% of total patients were excluded. This led to a minimum cohort size criterion of 20 patients, and removed the Bone-Cart (9 donors), Bone-Epith (11), Bone-Osteoblast (5), Breast-DCIS (3), Breast-LobularCa (13), Cervix-AdenoCa (2), Cervix-SCC (18), CNS-Oligo (18), Lymph-NOS (2), Myeloid-AML (13) and Myeloid-MDS (2) individual cohorts. Samples from these cohorts were still included in meta-cohort analysis (see below).
Tumor meta-cohorts (Esther Rheinbay)
Tumor meta-cohorts were assembled for identification of drivers and increase of discovery power across cell lineages and organ systems. The following meta cohorts were used in driver analyses:
By cell type of origin:
Epithelial: Carcinoma (comprised of tumor cohorts Bladder-TCC, Biliary-AdenoCa, Breast-AdenoCa, Breast-LobularCa, Cervix-AdenoCa, ColoRect-AdenoCa, Eso-AdenoCa, Kidney-ChRCC, Kidney-RCC, Liver-HCC, Lung-AdenoCa, Ovary-AdenoCa, Panc-AdenoCa, Panc-Endocrine, Prost-AdenoCa, Stomach-AdenoCa, Thy-AdenoCa, Uterus-AdenoCa,Head-SCC, Cervix-SCC, Lung-SCC), Adenocarcinoma (Biliary-AdenoCa, Breast-AdenoCa, Breast-LobularCa, Cervix-AdenoCa, ColoRect-AdenoCa, Eso-AdenoCa, Kidney-ChRCC, Kidney-RCC, Liver-HCC, Lung-AdenoCa, Ovary-AdenoCa, Panc-AdenoCa, Prost-AdenoCa, Stomach-AdenoCa, Thy-AdenoCa, Uterus-AdenoCa), squamous epithelium (Head-SCC, Cervix-SCC, Lung-SCC)
Mesenchymal cells/sarcoma (Bone-Cart, Bone-Epith, Bone-Leiomyo, Bone-Osteosarc)
Glioma (CNS-PiloAstro, CNS-Oligo, CNS-GBM)
Hematopoietic system (Lymph-BNHL, Lymph-CLL, Lymph-NOS, Myeloid-AML, Myeloid-MDS, Myeloid-MPN)
By organ system:
digestive tract (Liver-HCC, ColoRect-AdenoCa, Panc-AdenoCa, Eso-AdenoCa, Stomach-AdenoCa, Biliary-AdenoCa), kidney (Kidney-RCC, Kidney-ChRCC), lung (Lung-AdenoCa, Lung-SCC), lymphatic system (Lymph-BNHL, Lymph-CLL, Lymph-NOS), myeloid (Myeloid-AML, Myeloid-MDS, Myeloid-MPN), breast (Breast-AdenoCa, Breast-LobularCa), female_reproductive_system (Breast-AdenoCa, Breast-LobularCa, Cervix-AdenoCa, Cervix-SCC, Ovary-AdenoCa, Uterus-AdenoCa), central nervous system (CNS-PiloAstro, CNS-Oligo, CNS-Medullo, CNS-GBM)
Pan-cancer:
Two “Pan-cancer” cohorts were created: “Pancan-no-skin-melanoma” containing all tumor types with the exception of Skin-Melanoma to remove issues caused by very high mutation rate tumors; and “Pancan-no-skin-melanoma-lymph” with the additional removal of lymphoid tumors (Lymph-BNHL, Lymph-CLL, Lymph-NOS) that have local somatic hypermutation caused by AID.
2. Mutational hotspot analysis (Randi Istrup Pedersen)
We selected the top 50 single position hotspots based on the number of patients with an SNV mutation. The individual positions marked as problematic by the site-specific noise filter (see below) analysis were excluded.
Each hotspot was defined by its genomic position and annotated by the number of patients with an SNV mutation in the given hotspot. We also annotated each hotspot with whether it falled in one of the genomic element types analyzed in the driver discovery. We further overlapped with loop-regions of palindromes, which are hypothesized to fold into DNA-level hairpins, and with location in immunoglobulin loci. When a hotspot overlapped a protein-coding gene we extracted the corresponding amino acids changes from Oncotator1 (http://portals.broadinstitute.org/oncotator/).
We identified known driver hotspots, by overlap with the somatic driver positions compiled in the Cancer Genome Interpreter repository (https://www.cancergenomeinterpreter.org/mutations), which among others include mutations from ClinVar, DoCM and the literature (ref: https://doi.org/10.1101/140475).
For each hotspot we calculated the proportion of mutations in the defined cohorts and metacohorts. Only cohorts with at least 20 patients, and at least 10 patients or 10% of patients with an SNV were included in Fig. 1a (for the distribution in all cohorts and meta-cohorts see Extended Data Fig. 1). Lymph-BNHL and Lymph-CLL were shown together as Lymphoid malignancies.
Based on mutational signature analysis of all the cancer samples, we extracted the posterior probability that each hotspot mutation from a given patient was generated by one of 37 identified mutational signatures. In lymphoid malignancies somatic hyper-mutations generated by AID come in clusters along the genome. Posterior probabilities for the ten signatures relevant for the lymph cohorts were therefore derived from models that consider the correlation of AID mutations along the genome. For each hotspot the collected posterior probabilities were averaged.
3. Mutational Signatures (Jaegil Kim)
We performed a de novo global signature discovery to identify mutational signatures operating in PCAWG WGS cohort (N = 2,583). All single nucleotide variants (SNVs) were stratified into M = 4608 mutation channels according to the combination of six base substitutions at pyrimidine bases (C>A/G/T and T>A/C/G) in penta-nucleotide sequence contexts and a transcription strand direction, transcribed (-), non-transcribed (+), non-coding regions. All insertions and deletions were classified into additional 20 channels depending on the number of inserted or deleted bases and any indels beyond 9 bases were grouped into a single channel. The resulting mutation frequency matrix X (4,628 channels across 2,583 samples) were ingested to the Broad Institute’s signature analysis pipeline, SignatureAnalyzer, to determine the optimal number of signatures (K) and the signature profiles by de-convoluting X into a product of two non-negative matrices as X ~ WH, W (M × K) and H (K × N) being a signature-loading and an activity-loading matrix, respectively. The most distinguished feature of SignatureAnalyzer is that it suggests an optimal number of signatures best explaining X at the balance between the error measure and the model complexity exploiting Bayesian non-negative matrix algorithm (NMF)2,3,4. Our de-novo signature discovery identified 37 signatures including a split of 4 UV-related signatures, 3 APOBEC-related signatures, 5 POLE-related signatures, 3 MSI-related signatures, 2 COSMIC17 signatures, and other 21 singleton signatures (see PCAWG7 paper for details).
Although the matrix H contains the most representative activity across samples it also has spurious activity assignments intrinsic to the global signature discovery (i.e., applying SignatureAnalyzer on samples with different histological subtypes and mutational processes). To minimize this contamination and interference we re-assigned activity separately in each histological subtype using a more refined projection approach. We first identified a subset of ultra-mutant samples with a dominant activity of POLE (n = 8), MSI (n = 18), and alkylating signatures (n = 1) from the original H, which are usually very exclusive to samples with a specific DNA repair or replication defect, or a treatment. To prevent a false-positive assignment of these signatures in remaining samples we introduced a binary matrix Z (37 × N), where the row elements corresponding to POLE, MSI, and alkylating signatures, are set to zero except for samples with a dominant activity, while all other elements are set to one. More specifically, the projection was done by keeping the signature-loading matrix W’ (M × 37) frozen (W’ represents the normalized signature profiles of 37 global signatures), while allowing SignatureAnalyzer to determine a subset of signatures and infer the activity-loading matrix H’ (37 × Nr) that best approximates the mutation frequency matrix X’ (M × N), where N’ represents the number of samples in each histological subtype, as X’ ~ W’ (Z¤H), Here “¤” denotes element-wise multiplications.
4. Definition of genomic elements
(Morten Muhlig Nielsen; Nicholas Sinnott-Armstrong) For coding elements and elements pertaining to protein coding genes (3’UTR, 5’UTR and protein coding promoters) regions were defined based on GENCODE annotations (v.19)5:
Coding elements (CDS)
The set of coding bases collapsed across all coding transcripts with a given GENCODE gene ID.
Protein coding splice site elements (pcSS)
Intronic regions extending 6 bases from donor splice sites and 20 bases from acceptor splice sites were collected for all coding transcripts. Bases were collapsed across all coding transcripts with a given GENCODE gene ID. The global set of CDS bases were subtracted.
5’UTR elements (5UTR)
The set of 5’UTR bases collapsed across all coding transcripts with a given GENCODE gene ID. The global set of CDS and pcSS bases were subtracted.
3’UTR elements (3UTR)
The set of 3’UTR bases collapsed across all coding transcripts with a given GENCODE gene ID. The global set of CDS, pcSS and 5UTR bases were subtracted.
Protein coding promoter elements (pcPROM)
Regions extending 200 bases in both directions from all protein coding transcripts’ transcription start sites (5’ ends). Bases were collapsed across all coding transcripts with a given GENCODE gene ID. The global set of CDS and pcSS bases were subtracted.
IncRNA elements
IncRNA transcripts were defined based on annotations from GENCODE (v.19) and MiTranscriptome (v.2)6 not overlapping GENCODE. Transcripts were included if fulfilling criteria 1-5 and 6 or 7 below:
No sense overlap to protein coding gene regions
More than 5kb away from protein coding genes on sense strand.
Longer than 200 bases.
Not annotated as the following biotypes: immunoglobulin, T-cell receptor, Mt_rRNA, Mt_tRNA, miRNA, misc_RNA, rRNA, scRNA, snRNA, snoRNA, ribozyme, sRNA or scaRNA.
Not overlapping genomic regions aligning back to the human genome (self chained regions).
More than 20% of bases overlap conserved elements (except if annotated as pseudogene)
Expressed in more than 10% of PCAWG samples with RNAseq data.
Genes corresponding to the selected transcripts were supplemented with a set of known functional lncRNA genes from the literature in addition to GENCODE annotated non-coding snoRNA and miRNA host genes. The elements were made by collapsing bases across transcripts with given gene ID. The global set of CDS, pc_SS, 5UTR, 3UTR, pcPROM and lncRNA_SS bases were subtracted.
lncRNA splice site elements (lncRNASS)
Intronic regions extending 6 bases from donor splice sites and 20 bases from acceptor splice sites were collected for all lncRNA transcripts. Bases were collapsed across all lncRNA transcripts with a given gene ID. The global set of CDS, pc_SS, 5UTR, 3UTR and pcPROM bases were subtracted.
lncRNA promoter elements
Regions extending 200 bases in both directions from all lncRNA transcripts’ transcription start sites (5’ends). Bases were collapsed across all lncRNA transcripts with a given gene ID. The global set of CDS, pc_SS, 5UTR, 3UTR, pcPROM and lncRNA_SS bases were subtracted.
Short RNA elements
Short RNA transcripts were defined based on annotations from databases Rfam (v.11)7, tRNAscanSE (v.2.0)8 and snoRNAdb (v.3)9 in addition to GENCODE transcripts with biotype annotations mt_rRNA, mt_tRNA, misc_RNA, rRNA and snoRNA.
Bases were collapsed across all smallRNA transcripts with a given gene ID. The global set of CDS, pc_SS, 5UTR, 3UTR and pcPROM bases were subtracted.
microRNA elements
Mature miRNAs were defined based on mirBase (v.20)10 and a set of potential novel miRNAs11
Enhancers
Contiguous 15-state ChromHMM called enhancers correlated between H3K4me1 and RNA-seq across 57 human tissues were downloaded from Roadmap Epigenomics Consortium extended data12. Associated links, defined by co-occurring activity in a given cell type, were merged across cell types at FDR = 0.1. HoneyBadger213 p10 calls for all DNase I sites were filtered to peaks with signal strength 0.8 or greater and intersected with enhancer elements. The union of all DNase I peaks which overlapped with a given element, with all CDS regions filtered out, were used as the input to driver detection.
5. Candidate driver identification methods
A summary of approaches used by each method is listed in Supplementary Table 2.
ActiveDriverWGS (Juri Reimand)
Driver analysis with ActiveDriverWGS (ref) was performed after discarding hypermutated samples (>90,000 mutations) from the PCAWG cancer cohort. To avoid leakage of signals from known cancer drivers, we removed missense mutations in analyses of non-coding regions. ActiveDriverWGS is a local mutation enrichment method for genome-wide discovery of cancer driver mutations with increased mutation burden of single nucleotide variants (SNVs) and indels. ActiveDriverWGS performs a model-based test whether a given genomic element is significantly more mutated than adjacent background genomic sequence (+/-10kb and introns). Statistical significance of mutations is computed with a Poisson-linked generalised linear regression model. The null model treats all SNVs with trinucleotide context as cofactor, while indels are modelled with a separate cofactor for all nucleotides. Mutation counts per nucleotide are presented as the response variable The alternative model tests whether the element has different mutation burden than the background sequence. The null and alternative models are compared with chi-square tests and confidence intervals of expected mutations were derived from the null model using resampling. If the confidence intervals indicated significant excess of mutation in the background and depletion in the element of interest, we inverted corresponding small p-values (p= 1-p if p<0.5). Elements with no mutations were automatically assigned p=1.
CompositeDriver0.2 (Eric Minwei Liu, Ekta Khurana)
We have developed CompositeDriver (ref) – a computational method that combines signals of mutation recurrence and the functional impact score derived from FunSeq2 scheme14 to identify coding and non-coding elements under positive selection. CompositeDriver assigns a score to each region of interest (i.e., CDS, promoter, UTR, enhancer or ncRNA) through summation of positional mutation recurrence multiplied by the functional impact score for all mutations within the region. A null CompositeDriver score distribution is built to calculate the p-value for a region of interest. Mutations in the same element type but outside the region of interest are defined as background mutations. To build the null distribution, the same numbers of mutated positions are repeatedly drawn (default is 105 times) from background mutations with similar replication timing and similar mutation context15. By drawing random mutations from the same element type, CompositeDriver incorporates DNase I hypersensitive sites and histone modification marks as covariates into the null model16. Finally, the Benjamini-Hochberg method is used for multiple hypothesis correction17.
dNdScv (Inigo Martincorena)
dNdScv is a maximum-likelihood algorithm designed to test for positive or negative selection in cancer genomes or other sparse resequencing studies. dNdScv models somatic mutations in a given gene as a Poisson process, accounting for sequence composition and mutational signatures using 192 trinucleotide substitution rates. Mutation rates are also known to vary across genes, often co-varying with functional features of the human genome, such as replication time and chromatin state. This information is exploited by dNdScv to refine the estimates of the background mutation rate of each gene, using a negative binomial regression. This regression removes known sources of variation of the mutation rates and models the remaining unexplained variation of the mutation rate across genes as being Gamma distributed, which protects the method against overconfidence in the estimated background mutation rate for a gene. Overall, the local mutation rate for a gene is estimated accounting for mutational signatures in the samples analysed, the sequence composition of a gene in a trinucleotide context, 20 epigenomic covariates and the local number of synonymous mutations in the gene. Inferences on selection are carried out separately for missense substitutions, truncating substitutions (nonsense and essential splice site mutations) and indels, and then combined into a global P-value per gene. dNdScv has been described in much greater detail elsewhere18.
DriverPower (Shimin Shuai)
DriverPower is a combined burden and functional impact test for coding and non-coding cancer driver elements. In the DriverPower framework, randomized non-coding genome elements are used as training set. In total 1373 reference features covering nucleotide compositions, conservation, replication timing, expression levels, epigenomic marks and compartments are collected for downstream modelling. For the modelling, a feature selection step by randomized Lasso is performed at first. Then the expected background mutation rate is estimated with selected highly important features by binomial generalized linear model. The predicted mutation rate is further calibrated with functional impact scores measured by CADD and Eigen scores. Finally, a p-value is generated for each test element by binomial test with the alternative hypothesis that the observed mutation rate is higher than the adjusted mutation rate.
ExInAtor (Andres Lanzos; Rory Johnson)
ExInAtor was specifically created for prediction of cancer driver lncRNAs, but is agnostic to gene type and can also be used for protein-coding genes. The exons of each gene are identified and collapsed across transcript isoforms. For each gene, the trinucleotide content of the exonic region is calculated. The remaining intronic regions, along with 10 kb of sequence upstream and downstream, are defined as the background region. From this background, a new background region is created by randomly sampling the maximum number of nucleotides, such that the trinucleotide content exactly matches that of the exonic region. Next, the number of mutations in the exonic and sampled background regions are compared by hypergeometric test. Genes with elevated exonic mutational density are considered candidate driver genes. ExInAtor was used with randomisation seed of 256. Otherwise ExInAtor was run exactly as described in Lanzós et al, 201719.
LARVA (Jing Zhang; Lucas Lochovsky)
LARVA20, or Large-scale Analysis of Recurrent Variants in noncoding Annotations, is a computational method that detects significantly elevated somatic mutation burdens in genomic elements—both coding and non-coding—to identify putative cancer-driving elements. Given a cancer cohort variant call set, and a list of genomic elements, LARVA models the expected background somatic mutation rate by fitting a beta-binomial distribution to the elements’ variant counts. This model properly accounts for the high mutation rate variability seen throughout the genome, which improves over some previous models’ assumption of a constant mutation rate. LARVA’s model also incorporates the influence of mutation rate covariates, such as DNA replication timing. LARVA’s output lists each genomic element from the input, along with a p-value based on the deviation of the element’s observed variant count from the expected variant count under LARVA’s model.
MutSig (Julian Hess, Esther Rheinbay)
The MutSig suite21 classifies whether genomics features, both coding and non-coding, are highly mutated relative to a predicted background mutation rate (BMR), which varies on a macroscopic-level across patients (patient-specific mutation rates can span orders of magnitude across pan-cancer cohorts) and genes (known covariates such as replication timing are strongly correlated with mutation rate) and on a microscopic level across sequence contexts (since mutational signatures are heterogeneous across a cohort and highly context-dependent). MutSig accounts for all three of these to compute the joint BMR distribution across genes/patients/contexts, and then convolves across the latter two dimensions to estimate the expected distribution of total background burden for a given gene across a whole cohort. Genes are then scored by how their total non-background burden exceeds this null distribution.
MutSig estimates a gene’s BMR by its synonymous mutation rate for coding genes, and by its mutation rate at nonconserved positions for non-coding genes. If the number of background mutations in a given gene is insufficient to provide a confident estimate of its BMR, MutSig will incorporate the background counts from other genes with similar covariate profiles into its estimator.
MutSig (MutSig2CV) was originally designed for coding regions only21. Modifications to this version of the algorithm to run on non-coding regions are novel to PCAWG. Coding MutSig also incorporates per-gene functional impact and clustering tests, which were not run on non-coding regions.
NBR (Inigo Martincorena)
NBR is a method that test for evidence of higher mutation density than expected by chance in a given region of the genome, while accounting for trinucleotide mutational signatures, sequence composition and the local density of mutations around each element. This method has been described in detail in a previous publication22, where it was used to identify candidate driver noncoding elements across 560 breast cancer whole-genomes.
Based on some of the features of dNdScv, NBR involves two main steps. First, all mutations across all elements tested are used to obtain maximum-likelihood estimates for the 192 rate parameters (rj) describing each of the possible trinucleotide substitutions in a strand-specific manner. rj = nj/Lj, where nj is the total number of mutations observed across samples of a given trinucleotide class (j) and Lj is the number of available sites for each trinucleotide. These rates are used to estimate the total number of mutations across samples expected under neutrality in each element considering the mutational signatures active in the cohort and the sequence of the elements (Eh = Sj rLj,h). This estimate assumes no variation of the mutation rate across elements in the genome. Second, a negative binomial regression is used to refine this estimate of the background mutation rate of an element, using covariates and Eh as an offset. In this study, the local density of somatic mutations (normalized by sequence composition) was used as a covariate, using a window around the element of a variable size across cohorts to ensure sufficient numbers of mutations in each window around each element and excluding coding sequences and previously identified candidate noncoding driver regions. Replication time and average gene expression level for 100 kb genomic bins were also used as covariates. The negative binomial regression models mutation counts as Poisson-distributed within an element with mutation rates varying across elements according to a Gamma distribution. As in dNdScv, this provides a refined estimate of the background mutation rate for each element (Eh*) as well as a data-driven measure of uncertainty around this estimate (-q-the overdispersion parameter of the negative binomial regression). P-values for each element are calculated using a cumulative negative binomial distribution with the mean (Eh*) and dispersion (q) parameters estimated by the negative binomial regression.
To protect against neutral indel hotspots or indel artefacts, unique indel sites rather than total indels per element were used. To protect against misannotation of a mutation clusters as sets of independent events, a maximum of two mutations per region and per sample were considered in the analysis.
ncdDetect (Malene Juul)
ncdDetect23 is a driver detection method tailored for the non-coding part of the genome. It uses a burden based approach, in which the frequency of mutations is considered, to reveal signs of recurrent positive selection across cancer genomes. For each candidate region, the observed mutation frequency is compared to a sample- and position-specific background mutation rate. A scoring scheme is applied to further account for functional impact in the significance evaluation of a candidate cancer driver element. In the present application, the scoring scheme is defined as log-likelihoods, i.e. minus the natural logarithm of the sample- and position-specific probabilities of mutation.
The position- and sample-specific probabilities of mutation used by ncdDetect are obtained by a statistical null model, inferred from somatic mutation calls of a collection of cancer samples (https://doi.org/10.1101/122879). The model includes a set of genomic annotations, known to correlate with the mutation rate in cancer. These are replication timing, trinucleotides (the nucleotide under consideration and its left and right flanking bases), genomic segment (a variable segmenting the genome into regulatory element types) and a position-specific measure of the local mutation rate (a weighted average of the mutation rate, calculated across samples in a 40 kb window flanking each specific position plus/minus 10 kb).
ncDriver (Henrik Hornshøj)
The ncDriver method (doi: https://doi.org/10.1101/182642) provides separate evaluations of the significance for two functional mutation properties, the level of conservation and the level of cancer type specificity. In the ncDriverConservation test, the conservation level of mutated positions were evaluated locally for being surprisingly high, given the distribution of conservation within the element. The p-value of the mean mutation phyloP conservation score for an element was obtained by Monte Carlo simulation of 100,000 mean phyloP scores based on the same number of mutations. Each mutated element was also evaluated globally by looking up the rank of the element mean phyloP conservation score among all elements annotated as the same type. This provided p-values for both local and global mutation conservation level, which were combined into a single conservation p-value using Fisher’s method. In the ncDriverCancerType test, the distribution of observed mutation counts of an element across the cancer types were evaluated for being surprising compared to expected counts estimated from a background null model (as described for the ncdDetect method) that accounts for cancer type specific mutation signatures and other co-variates. A Goodness-of-fit test with Monte Carlo simulation was used to determine whether the distribution of observed mutation counts across cancer types within the element is surprising given the expected mutation counts based on cancer types, mutation contexts and element type. For indels, the expected mutation counts were estimated solely from the mutation rates calculated from the mutation context, cancer type and element type.
OncodriveFML (Loris Mularoni)
OncodriveFML24 is a method designed to estimate the accumulated functional impact bias of tumor somatic mutations in genomic regions of interest, both coding and non-coding, based on a local simulation of the mutational process affecting it. The rationale behind OncodriveFML is that the observation of somatic mutations on a genomic element across tumors, whose average impact score is significantly greater than expected for said element constitutes a signal that these mutations have undergone positive selection during tumorigenesis. This, in turn is considered as a direct indication that this element drives tumorigenesis.
OncodriveFML first computes the average functional impact score of the observed mutations in the element of interest. The functional impact scores of mutations have been calculated using both CADD25 (coding and non-coding regions) and VEST326 (only coding regions). Then the method randomly samples the same number of observed mutations following the probability of mutation of different tri-nucleotides, computed from the mutations observed in each cohort. The randomization step is repeated many times (1,000,000 in these analyses) and each time an average functional impact score is calculated. Finally, OncodriveFML derives an empirical p-value for each element by comparing the average functional impact score observed in the element to its local expected average functional impact score resulting from the random sampling. The empirical p-values are then corrected for false discovery rate and genomic elements that after the correction are still significant are considered candidate drivers.
regDriver (Husen M. Umer)
regDriver assesses the significance of mutations affecting transcription factor motifs using tissue-specific functional annotations (https://doi.org/10.1002/humu.23014). For each tumor cohort, functional annotations from the cell lines most similar to the respective tumor type are gathered. A functionality score is computed for each mutation based on its overlapping functional annotations. regDriver, collects highly-scored mutations in each of the defined elements and assesses the elements’ significance by comparing its accumulative score to a background score distribution obtained from the simulated sets. Therefore, only candidate regulatory mutations are considered in evaluating mutation enrichment per element.
6. Simulated datasets
Broad simulations
(Yosef Maruvka, Gad Getz)
Due to their differing context characteristics, we simulated SNVs and indels with different approaches. For SNVs, we divided the genome into 50kb regions. For each region, we counted the number of mutations in it across all the PCAWG patients and divided it by the total number of mutations. Every mutation was randomly assigned into a new region based on the region’s rate. The position inside the region was chosen to maintain the trinucleotide context of each mutation (the 5’ and 3’ nearest neighbors and the mutated position itself) and the alternate allele. In addition, for every base we counted how many times it was covered sufficiently, in 401 tumor and normal WGS pairs, in order to enable calling of a mutation27. The fraction of patients with enough coverage at a given site was used as the position’s probability for being mutated inside the new current region.
For indels, a new, randomized position was chosen in a region of 50,000 bases around the indel. The position of the new indel was chosen to match the indel 5’ and 3’ neighboring reference bases. For insertions, the inserted motif was the same as the original insertion, but for deletions only the length of the indels was kept but not the exact sequence.
DKFZ simulations (Carl Hermann, Calvin Chan)
This simulation utilizes the SNV calls to perform a localised randomisation. The original SNV entries which do not map to chromosome 1-22, X, Y are first filtered and excluded from randomization. All SNVs located in the protein coding regions (CDS) corresponding to GENCODE19 definition are erased before performing randomisation. The trinucleotide centered at each SNV position is determined and an identical trinucleotide is randomly sampled within the 50kb window. In case of insertion, instead of the mutated trinucleotide, the neighboring nucleotide of the insertion site is scanned within the randomisation window. For deletion and multi-nucleotides variants, the altered sequence is scanned within the randomization window with a ranked probability assigned for each position. The randomised sample is then selected from the top 100 matched positions with scaled probability.
Sanger simulations (Inigo Martincorena)
This simulation aimed to generate datasets of neutral somatic mutations that retain key sources of variation in mutation rates known to exist in cancer genomes, including mutational signatures, and variable mutation rates across the genome and among individuals and cancer types. To do so while minimizing the number of assumptions in the simulation, we used a simple local randomization approach. First, all coding mutations as well as mutations in the TERT promoter, MALAT1 or NEAT1 were excluded. Second, each mutation in each patient was randomly moved to an identical trinucleotide within a 50 kb window, while retaining the patient ID. Third, mutations falling within 50 bp of their original position were filtered out. This simple randomization retains the variation of the mutation rate and mutational signatures across large regions of the genome, across individuals and across cancer types.
7. Statistical framework for the integration of results from multiple driver discovery methods
(Grace Tiao, Gad Getz)
The classical approach for combining p-values obtained from independent tests of a given null hypothesis was described by R. A. Fisher in 1948. He noted that for a set of k p-values, the sum X of the transformed p-values, where
and pi is the p-value for the ith test, follows a chi-square distribution with 2k degrees of freedom28. Thus, to obtain a single combined p-value for a set of independent tests, the new test statistic X is computed from the p-values obtained from the tests and scored against a chi-square distribution with 2k degrees of freedom. Fisher’s test is asymptotically optimal among all methods of combining independent tests29; however, in cases where tests exhibit dependence, the Fisher combined p-value is generally too small (anti-conservative).
In this study, we combine p-values from several driver detection methods, many of which share similar approaches and whose results are therefore not independent. To address this issue, we used an extension of the Fisher method developed by Morten Brown for cases in which there is dependence among a set of tests29. Using the same test statistic, renamed Ψ to indicate the difference in the independence assumption, Brown observed that if Ψ were assumed to have a scaled chi-square distribution - i.e.,
then
Note that E[Ψ] = 2k irrespective of the independence requirement, and that
Thus when the pi are independent, var(Ψ) = 4k, which gives f = k and c ≠ 1, and the test statistic follows the chi-square distribution with 2k degrees of freedom described by Fisher. However, when the independence condition is relaxed, var(Ψ) Φ 4k, and the test statistic generally follows a different, scaled, chi-square distribution whose scaling parameter c and degrees of freedom 2f are determined by the covariances of the pi’s. The covariances can be computed via numerical integration over the joint distributions of all pi and pj pairs, but this requires knowledge of the joint distribution; and even in cases where the joint distribution is known, the integration may not be computationally feasible for large and complex datasets30.
In this study, following the example of Poole et. al31, we computed the empirical covariance of pi and pi, using the samples wi and wj, where wi is the set of all reported p-values for method i, and used the empirical covariance to approximate the Brown scaled chi-square distribution. The advantage to this approach is that the empirical covariance estimation is non-parametric – it does not assume an underlying joint distribution of pi and pi – and is thus applicable to complex and interrelated biological datasets where data is noisy and not regularly Gaussian. Poole et. al showed that the empirical covariance estimation approach is accurate, robust, and efficient for such datasets.
Implementing and evaluating the integration method on simulated and observed data
To evaluate the efficacy of the empirical Brown’s method of dependent p-value integration, we generated three sets of simulated mutation data (see above) and ran the driver detection algorithms on each of the simulated datasets. We checked that the p-value results from the various driver detection algorithms followed the expected null (uniform) distribution (Extended Data Fig. 3a). Then, for each simulated data set, we calculated the empirical covariance for each pair of driver algorithm results. We then used these covariance values over simulated datasets to compute the combined Brown p-values on observed data: for each gene in the observed PCAWG somatic mutation dataset, we computed the Brown test statistic from the set of p-values reported by the various driver detection algorithms. The Brown test statistic was then evaluated against the appropriate chi-square distribution, whose scale and degree parameters were approximated by the covariance values calculated on the simulated data (see above).
We ran this procedure, as well as the Fisher method, for six representative tumortype cohorts (Breast-AdenoCa, CNS-GBM, ColoRect-AdenoCa, Lung-AdenoCa, Uterus-AdenoCa, and meta-Carcinoma) and found that the Brown combined p-values generally followed the null distribution as expected (Extended Data Fig. 3b). The Fisher combined p-values were significantly inflated (Extended Data Fig. 3b), confirming that dependencies existed between the results reported by the various driver detection algorithms.
We next explored whether it was possible to reduce the number of algorithm runs required to complete these calculations for all tumor-type cohorts by computing the covariance values on observed data instead of simulated data. In each of the six representative tumor-type cohorts, we calculated the empirical covariances on the observed data only and then computed the integrated Brown p-values on the observed data using the observed covariances. Significant genes identified using only observed covariances remained mostly unchanged from the significant genes identified using the simulated covariances (Extended Data Fig. 3d), and examination of the differences in the covariance values between the simulated estimations and the observed estimations revealed only minor differences in values (Extended Data Fig. 3c). The significant drivers presented in this study were identified using this final approach – e.g., by computing integrated Brown p-values using estimations of covariance on observed data only.
Integration of p-values from observed data was performed for 42 tumor-type cohorts and 13 target element types. Methods were selected for each given data set (see “Selecting methods to include in the integration of observed p-values”, below) and raw p-values smaller than 10−16 were trimmed to that value before proceeding with the integration. Methods with missing data for a given element (i.e., ones that failed to report a p-value for a given element) were excluded from the calculation for that element, and therefore in some cases the integrated Brown p-value was computed from p-values reported by only a subset of all the driver detection algorithms contributing results for that data set.
Selecting methods to include in the integration of observed p-values
In some cases, individual driver detection algorithms reported p-values for a given data set that deviated strongly from the expected uniform null distribution. These were methods for which the quantile-quantile (QQ) plots demonstrated considerable inflation. We removed results that reported an unusual number of significant hits by calculating, for each set of results, the number of significant elements found by each individual method using the Benjamini-Hochberg FDR with q<0.1 as the significance threshold. Any single method that reported four times the median number of significant elements identified by individual methods was discarded from the integration. In a separate analysis, we found that removing methods that yielded fewer hits than the median (i.e., methods with deflated QQ-plots) did not affect the number of significant genes identified through the integration of the reported p-values (Extended Data Fig. 3d); hence we did not remove such methods.
8. Post-filtering of candidates
(Esther Rheinbay, Morten Muhlig Nielsen, Lars Feuerbach, Henrik Tobias Madsen)
Post-filtering of significant hits was performed to remove those with accumulation of mutations caused by sequencing problems or mutational processes. In particular, we applied the following: (i) at least three mutations are present in the element, (ii) mutations are present in at least three patients of the tested cohort, (iii) less than 50% of mutations are located in palindromic DNA sequence22, (iv) more than 50% of mutations are located in mappable genomic regions (CRG alignability, DAC blacklisted regions and DUKE uniqueness32; (v), less than 50% of mutations occur near indels, (vi) a site-specific noise filter (see below); and (vii) manual review of sequence evidence for novel drivers. For lymphomas, which contain regions of somatic hypermutation caused by AID enzyme activity, we (viii) further required less than 50% of mutations contributed by this process; and for Skin-melanoma, we (ix) excluded mutations occurring in the extended (CTTCCG) context that contributes to promoter hotspot mutations in this tumor type33,34,35. Even after this motif-based filter, a large number of promoter and 5’UTR regions remained in the list of candidates. We, therefore, marked these as likely due to failed repair in TF occupied sites since it is unclear how many of them are true or false drivers. Additionally, elements that failed manual mutation review were filtered out at this stage.
Site-specific noise filter
Genomic positions of mutations in each significant hit were analyzed in all normal control samples to assess the position-specific noise-level. Therefore, for each of the three nonreference nucleotides a and cohort c ϵ C the relative frequency of normal samples which had at least two reads supporting the alternative allele a were calculated as p(a,c) ϵ [0;100]. The noise score was then computed as ΣcϵC log10(p(a,c) + 1). Mutations at positions with a score > 20 for at least one of the non-reference nucleotides were flagged. Elements for which the number of mutations at flagged positions exceeded 20% were removed.
DNA palindromes
We define a palindrome as a sequence of DNA followed by its complementary reverse with a sequence of variable length in between (Fig. 3d). It is hypothesized that these palindromes can temporarily form DNA hairpins36. While in the hairpin state the loop region is singlestranded and open to attack by APOBEC enzymes. Based on observations in breast cancer whole genome sequences37, we decided to consider palindromes with a minimal repeat length of 6 bp and an intervening sequence (loop) length of 4-8 bp. We call these regions genome-wide using the algorithm described in Ye et al.38 however using our own implementation (https://github.com/TobiasMadsen/detectIR). In total, we find 7.3 M palindrome regions covering a total of 135.2 MB of which 33.6 MB are loop sequence.
Computing the false discovery rate
We controlled the false discovery rate (FDR) within each of the sets of tested genomic elements by concatenating all integrated Brown p-values from across all tumor-type cohorts and applying the Benjamini-Hochberg procedure17 to the integrated Brown p-values. A q-value threshold of 0.1 was chosen to designate cohort-element combinations as significant hits. In addition, we defined cohort-element combinations in the range 0.1 ≤ q<0.25 as “near signif cance”. We next applied several additional, mutation-based filtering criteria to each significant or near-significant candidate and assigned p-values of 1 to candidates that failed these filtering criteria. Final Benjamini-Hochberg FDR values were then re-calculated on the adjusted sets of integrated Brown p-values to arrive at a list of candidate driver cohort-element combinations.
9. Sensitivity and precision analysis of driver predictions (Iñigo Martincorena)
All methods employed in this study were shown to have a low rate of false positives when run on a series of neutral simulated datasets without driver mutations. To evaluate the sensitivity and precision of different methods and particularly of our approach for p-value integration, we compared their relative performance in detecting known cancer genes when applied to protein-coding genes. As a reference gold-standard set of known cancer genes we used a list of 603 genes from the manually-curated Cancer Gene Census v80 database. Results are shown in Extended Data Fig. 4.
10. Gene expression analyses
(Samir B. Amin, Morten M. Nielsen, Andre Kahles, Nuno Fonseca, Lehmann Kjong, members of the PCAWG Transcriptome Working Group and Jakob Skou Pedersen)
To extend the RNAseq-based expression profiling of GENCODE annotations provided by the PCAWG Transcriptome Working Group39, we profiled an extended set of gene annotations, including a comprehensive set of non-coding RNAs (described above and at Synapse:syn5325435).
The profiling used the docker-based workflow described in 39 for 1,180 RNA-seq donor libraries, matched to WGS data across 27 different cancer types39. In brief, raw sequence reads from donor libraries were uniformly evaluated for QC using FastQC tool, and subsequent alignment was performed on QC-passed libraries using two methods: STAR (v2.4.0i)40 and TopHat2 (v2.0.12)41. Resulting QC-passed bam files were independently used to quantify extended RNA-seq annotations at the gene-level counts using htseq-count method with following parameters: -m intersection-nonempty --stranded=no —idattr gene_id. This step resulted in two sets of gene-level counts files per donor library which were independently normalized using FPKM normalization and upper quartile normalization (FPKM-UQ). The final expression values were provided as a gene-centric table (rows as genes, columns as samples) with each value representing an average of the TopHat2 and STAR-based alignments FPKM values. Gene-centric tables based on both, GENCODE and extended RNA-seq annotations are available at Synapse data portal: syn2364711. Docker-based workflow for quantifying extended RNA-seq annotations is at https://github.com/dyndna/pcawg14_htseq
11. Normalization for copy number variation
(Henrik Tobias Madsen, Morten Muhlig Nielsen, and Jakob Skou Pedersen)
To account for the effects of somatic copy number alterations (SCNAs) on expression, we used two different approaches to create two additional versions of the expression profiles: First a conservative approach where we remove all samples not having the regular bi-allelic copy number for the gene in question. Second a less conservative approach where we build a regression model of expression data based oncopy number (CN) data, and then tested for an effect of somatic mutations on the residual (i.e. the expression that is not explained by copy number).
Generally the higher the copy number of a particular gene the higher expression. The relationship between copy number and gene expression is not strictly linear, as various feedback mechanisms in the cell try to compensate for the mostly deleterious effects of SCNAs. This is known as dosage compensation and has been studied extensively in the context of mammalian sex chromosomes, but also in evolution of yeast and in diseases caused by aneuploidy42–45. We therefore fit a linear regression model between the logarithm of expression and the logarithm of CN. This effectively amounts to a power-regression model.
A number of factors makes it hard to learn the regression parameters for each gene and cancer type in isolation. (i) for some cancer types we have only a limited number of samples; (ii) for some genes there is not much variation in CN; and (iii) the variation in expression between samples is generally high. We overcome these problems by employing a mixed model strategy, that allows sharing of information between genes, effectively regularizing the parameter estimates for gene cancer-type combinations that carry little information on their own.
LetFPKMg,c,i andSCNAg,c,i denote the expression and SCNA measurement respectively for gene, g, in cancer-type, c, and sample i respectively. We then define:
where αg,c and β are fixed effects, whereas γg and λg,c are random effects, with 
and 
Finally the residual is 
.
Using this model we infer a global SCNA to expression regression, β, but allow some regularized gene-specific and gene/cancer type-specific variation: γg and λg,c
Thus we exploit the similarity across genes and similarity within genes across cancer types.
Since the variance increases with the absolute value of the explanatory variable associated with a random slope, this kind of mixed model display heteroskedasticity. Furthermore the model is not invariant under scaling of the explanatory variable, in this case SCNA. We centralise l°g SCNA, such that normal diploid regions have the least variance.
12. Mutation to expression association
(Morten Muhlig Nielsen, Henrik Tobias Madsen, and Jakob Skou Pedersen)
Mutation to expression association was calculated using non-parametric rank sum based statistics on z-score normalized expression values. This equalizes expression mean and variance for each cancer type and is a way to allow for comparison across cancer types. For comparisons of expression in mutated vs non-mutated (wild-type) cases within the same cancer type, the test reduces to the Wilcoxon rank sum test. For comparisons involving samples in multiple cancer types, such as meta-cohorts and pan-cancer cohorts, the statistic and associated p-value is still the non-parametric Wilcoxon rank sum test, and thus no assumptions regarding distribution of z-score expressions are made. Tied expression values were broken by adding a small random rank robust value. Association estimates were performed based on original expression values as well as the two copy-number normalized expression sets mentioned above. Fold difference values were calculated per mutation as the log2 ratio of the expression of the mutated tumor to the median of all wild-type tumors of the same cancer type. Reported Fold Difference values for an element with multiple mutations represent the median fold difference of all mutations in that element.
13. Copy number analyses
(Esther Rheinbay)
We surveyed significant focal copy number alterations for candidate driver genes as orthogonal evidence for their “driverness”. Significant copy number alterations were obtained from the TCGA Copy Number Portal (http://portals.broadinstitute.org/tcga/home), analysis 2015-06-01-stddata-2015_04_02 regular peel-off, a database of recurrent copy number alterations calculated by the GISTIC2 algorithm46 across >10,000 samples and 33 tumor types from TCGA. GISTIC2 results were included for candidate drivers if a gene was significant (residual q<0.1) and was located within a peak with ≤ 10 genes. Visualization was performed with the Integrative Genomics Viewer (IGV)47.
14. Power Calculations
(Esther Rheinbay)
Calculation of mutation detection sensitivity
Average detection sensitivity, the power to detect a true somatic variant, was calculated using a binomial model across all exon-like regions for different genomic element types as previously published27, 48. Detection sensitivity was based on sequencing coverage and estimated clonal variant allele fraction from 60 representative PCAWG alignments from the pilot cases (https://doi.org/10.1101/161562; Supplementary Table 7). Clonal allele fraction was estimated based on PCAWG consensus purity and ploidy estimates (doi: https://doi.org/10.1101/161562) as purity/(average ploidy)48. Element-wise averages were calculated as average across all exons for a given element.
Estimation of total number of promoter hotspot mutations
Detection sensitivity for all patients was calculated for the two most recurrent TERT promoter hotspot sites (chr5:1295228, chr5:1295250; hg19) using total read depth at these positions, sample purity and average ploidy. For each cohort, the number and percentage of powered (>90%) patients was obtained. The number of total expected mutations was then inferred as number of observed (called) mutations divided by the fraction of patients powered. The number of “missed” mutations is the difference between the total expected and observed mutations. Percentages of these numbers were calculated relative to the size of individual patient cohorts. Confidence intervals (95%) on the total percentage of patients with a TERT hotspot mutation were calculated using the beta distribution. Poisson confidence intervals were calculated for the number of missed mutations in the PCAWG cohort. Note that the inference of TERT mutations assumes exactly one mutation per patient. Estimates for the FOXA1 promoter hotspot mutation (chr14:38064406; hg19) were conducted using the same procedure.
Calculation of the minimum powered mutation frequency in a population
Power to discover driver elements mutated at a certain frequency in the population were conducted as described before21, 49, but solving for the lowest frequency for a driver element in the patient population that is powered (>90%) for discovery. The calculation of this lowest frequency takes into account the average background mutation frequencies for each cohort/element combination, the median length and average detection sensitivity for each element type, patient cohort size, and a global desired false positive rate of 10%.
15. Associations between mutation and signatures of selection: loss of heterozygosity and cancer allelic fractions
(Federico Abascal, Iñigo Martincorena)
For protein-coding sequences mutation recurrence can be analysed in the context of the functional impact of mutations (e.g. missense, truncating) to better distinguish the signal of selection. In contrast, estimating the functional impact of mutations in non-coding elements of the genome is a difficult, yet unsolved problem. To overcome this limitation and be able to compare selection signatures for both coding and non-coding elements under a similar framework, we developed two measures of selection which are agnostic to the functional impact of mutations.
Association between mutation and loss of heterozygosity
When a tumour carries a driver mutation in one allele of a given gene, it may be the case that a second hit on the other allele confers a growth advantage and is positively selected. When one of the events involves the loss of one of the alleles the process is referred to as loss of heterozygosity (LOH). This kind of biallelic losses are typical of, but not exclusive to, tumour suppressor genes (TSGs).
For each gene, we build a 2x2 contingency table indicating the number of cases in which the gene was mutated or not and the number of cases in which the gene was subject to LOH or not. We applied a Fisher’s exact test of proportions to identify which genes showed an excess of LOH associated to mutation. P-values were corrected with the FDR method to account for multiple hypotheses testing. This analysis was applied to each tumour type and cohort separately and proved very successful in identifying TSG as well as some oncogenes (OGs) as well.
Association between mutation and cancer allelic fractions
Driver mutations that provide an advantage for tumour cells are expected to show higher allelic fractions based on different interacting processes, including: early selection; amplification of the locus carrying the driver mutation; loss of the non-mutated locus (LOH). Comparing cancer allelic fractions (CAF) can be informative to detect signatures of selection, both for TSGs and OGs.
CAFs are defined here as the proportion of reads coming from the tumour and carrying the mutation. To transform observed fractions (VAFs) into CAFs, tumour purity and local ploidy needs to be taken into account according to the following formula:
Where Lp corresponds to the local ploidy for the mutated locus, and Pt denotes the tumour purity. Ploidy and tumour purity predictions were obtained from Dentro et al (in preparation).
To determine whether CAFs for a given gene or element were higher than expected we compared them to the CAFs observed in flanking regions. To define flanking regions we took 2 kb at each side of the gene/element, excluding any eventually overlapping coding exon, and also included introns (if present). The two sets of CAFs associated to each gene/element, i.e. those CAFs lying within the gene and those flanking it, were compared with a t-test to detect significant deviations. P-values were corrected with the FDR method. This approach was able to identify most known TSGs and OGs.
16. Signals of selection in aggregates of non-coding regions of known cancer genes
(Federico Abascal, Iñigo Martincorena)
We conducted a series of analyses on regions combined across genes to determine whether the paucity of driver mutations found in non-coding regions was related to lack of statistical power in single-gene analyses. This analysis also aimed to estimate how many driver mutations present in cancer genes were missed in this study. For protein-coding sequences, the number of driver mutations was estimated using dN/dS ratios as described in (50). For non-coding, regulatory regions of protein-coding genes (Promoter and UTRs), we relied on a modified version of the NBR negative binomial regression model described above (Section 5) to quantify the overall excess of driver mutations. We applied a second approach to determine whether there was an enrichment of LOH associated to mutations in the different types of non-coding regions associated to protein-coding genes.
Observed vs. expected numbers of mutations based on the NBR mutation model
NBR was used to estimate the background mutation rate expected across cancer genes, using a conservative list of 19,082 putative passenger genes as background. The resulting model is used to predict the numbers of expected SNVs and indels per element type per gene, and aggregate sums across genes. For this analysis we selected a reduced but diverse set of 142 cancer genes, encompassing 27 focally amplified and 22 deleted genes found by TCGA and present in the Cancer Gene Census51, and 112 genes with a significant signal of selection in this study (q<0.1) and present in the Cancer Gene Census. The list of 142 genes can be found in Supplementary. Table 7. To be as accurate as possible, we used a diverse set of covariates in the NBR model, including: local mutation rate (estimated on neutral regions +/-100 kb around each gene), detection sensitivity (defined as the element-average proportion of callable samples per site according to MuTect), gene expression covariates (first 8 principal components of the matrix of average gene expression values in each tumour type, as well as two binary variables marking the 500 genes with highest expression values in any tumour and 1,229 genes with a maximum FPKM lower than 0.1 across tumour types), and averaged copy-number calls for each gene across all samples (see Supplementary Note for more details). For each element type, the sum of observed mutations across the 142 cancer genes were compared to the sum of the expected rates to estimate the excess of mutations in regulatory and coding regions of cancer genes. An excess of observed mutations provides an estimate of the number of driver events18. Confidence intervals were calculated using the equation for the ratio of two Poisson observations, which are the number of mutations in the list of known cancer genes and in the list of passenger genes. It is important to know that these confidence intervals do not capture uncertainty in the background model and should be interpreted with caution. For this reason, we systematically evaluated the impact of a diverse array of covariates on our estimates (see Supplementary Note). We also note that this test can underestimate the number of non-coding drivers since some driver mutations can be present in the list of putative passenger genes, although this effect is expected to be quantitatively small if the density of driver mutations in regulatory regions of known cancer genes is higher than in those of putative passenger genes.
Mutation-LOH association for aggregates of genes
For this analysis we combined data across known cancer genes, including 603 genes in the Cancer Gene Census v80 and 154 additional significantly mutated genes found by exome studies18, 21. To estimate whether there was an excess of LOH associated to mutation in regulatory and coding regions of cancer genes, we calculated the fold change in LOH for the aggregate of cancer genes and normalized it dividing by the fold change observed in passenger genes. Confidence intervals were estimated using parametric bootstrapping (100,000 pseudoreplicates) for both cancer and passenger genes.
17. Mutational process and indel enrichment
(Federico Abascal, Iñigo Martincorena)
After noticing the skewed distribution of indel lengths in genes like ALB, SFTBP, NEAT1 and MALAT1, we carried a search for other genes showing the same pattern, which may be subject to the same mutational process. For every gene we record the proportion of indels of length 2-5 bp out of the total number of indels and compared this proportion with the background proportion using a binomial test. The background proportion was calculated using all protein-coding and lncRNAs genes. For every gene we also calculated the indel rate and compared it to the background indel rate using a binomial test. Both sets of p-values were independently corrected with the FDR method. The analysis was done for each tumour type separately. Genes with a q-value < 0.1 both for enrichment in 2-5 bp indels and for higher indel rates were further analyzed as candidates to be under the process of localized indel hypermutation described in this study. The levels of expression of these genes were analysed across all tumour types.
18. Mutations association to splicing (Andre Kahles)
For the assessment of the relationship between mutations in the U2 locus upstream of WDR74 and local changes in alternative splicing, we analysed changes in the percent-spliced-in (PSI) value52 of alternative splicing events located in WDR74 relative to the given genotype. The analysis was based on four different alternative splicing event types extracted and quantified by the PCAWG Transcriptome Working Group (exon skip, intron retention, alternative 3’ splice site, alternative 5’ splice site)39. In total we analysed 116, 103, 27 and 98 events, respectively, for the above event types. We then filtered the events for a minimum of expression evidence (non-NaN PSI) in at least 10 samples, presence of the alternate allele together with expression evidence in at least 5 samples, and a minimum absolute distance of mean PSI values of alternate and reference group of 0.05. Based on the selected events, we retained 3, 3, 1 and 0 events for analysis. Using an ordinary least squares model with the genotypes as factors and the the PSI values as responses, we computed p-values and the variance explained by presence of mutations. Multiple testing correction followed the Bonferroni method.
To assess the relationship between mutations in the U2 locus upstream of WDR74 and global changes in splicing, we used two global splicing statistics. We measured the amount of splicing as the number of edges in the splicing graph of a gene in a given samples. All splicing graphs were taken from the analyses of the PCAWG Transcriptome Working Group39. For each sample, we computed the mean number of splicing edges over all genes, resulting in the extent of splicing per sample. We measured the extent of splicing as how far each event is outlying from the mean over all event in the same cancer type (encoded by project code x histotype). The splicing outlier values per sample and gene were taken from the PCAWG Transcriptome Working Group analyses39. The mean over all genes of a sample was taken as a statistic for the extent of splicing.
For both statistics, we again used an ordinary least squares model as described above to model the relationship between genotype and splicing. As splicing is highly tissue dependent, we included the project codes as additional factors, accounting for both tissue specificity as well as possible underlying batch effects. To extract the amount of additional variance explained through the presence of mutations, we computed one model on project codes and presence of mutations and one model on project codes alone. For the model, we only included samples of project codes, where we had at least one sample with a mutation present, resulting in a total of 618 samples over 14 project codes.
19. Structural variation analysis (Morten Muhlig Nielsen, Lars Feuerbach)
Structural variant data was provided by the PCAWG Structural Variation Working Group. The data provide p-values for the observed breakpoint counts in 50kb bins along the genome. Candidate elements were overlapped with the bins, and fisher’s method was used to calculate a single p-value for each element. The set of element p-values were corrected with the FDR method.
20. RNA structural analysis
(Radhakrishnan Sabarinathan, Ciyue Shen, Chris Sander, Jakob Skou Pedersen)
In order to test if the observed mutations (SNVs) in the RMRP gene are biased towards high RNA secondary structure impact, we performed a permutation test by following the steps used in oncodriveFML24 together with the predicted structural impact scores from RNAsnp53. At first, the RNAsnp was run with the options -m 1 -w 300 and other default parameters to obtain the minimum correlation coefficient (r_min) score for each possible mutations in the RMRP gene. The r_min scores were then transformed, 1-((r_min+1)/2), to range between 0 and 1, where 1 indicates high structural impact score. Further, we followed the steps of oncodriveFML (see section ‘oncodriveFML’ for more details) with 1,000,000 randomizations and by using per sample mutational signatures (i.e., the probability of observing a mutation in a particular tri-nucleotide context in a given sample) to compute the p-value at the cohort and sample level.
Furthermore, the RNA secondary structure impact scores (r_min) of indels (insertions/deletion) were computed by using a modified version of RNAsnp (since the current version of RNAsnp is limited to substitutions only). Briefly, we first computed the base pair probability matrices of wild-type and mutant sequences (by taking into account the insertion or deletion) and then adjusted the size of matrices to be equal (by introducing additional rows and columns with zeros in one of the matrices with respect to insertion or deletion). Further, by following the steps of RNAsnp, we computed the r_min score. The structure shown in Fig. 6d is based on the conserved secondary structure annotation obtained from Rfam (RF00030)54.
Tertiary structure contacts in RMRP were predicted using evolutionary couplings co-variation analysis (EC analysis 55) of the multiple sequence alignment of 933 eukaryotic RMRP sequences from Rfam (RF00030). The EC analysis (software available at https://github.com/debbiemarkslab/plmc) was run with the options -le 20.0, -lh 0.01, -t 0.2, -m 100 and the top 100 interactions were chosen as predicted contacts, either in secondary or tertiary structure, depending on local context. As no experimental 3D structure or crosslinking experiments of the mammalian RMRP are available, interaction sites were inferred by homology to the partially known yeast RMRP crystal structure. We (1) aligned the human RMRP sequence with the Saccharomyces cerevisiae RMRP sequence using the sequence family covariance model from Rfam and (2) mapped the locations of RNA-protein interactions within 4Å56 from the crystal structure and the experimentally determined RNA-protein crosslinking sites57, and RNA substrate crosslinking sites58 from the yeast sequence to the human RMRP sequence. For the crosslinking sites, a ±3 nucleotide window is reported as the interaction site. In order to test if the locations of the observed indels are biased towards tertiary structure, protein- or substrate-interaction sites, 1,000,000 randomizations of five indels were performed assuming uniform distribution of indels across the RMRP gene body, and an empirical p-value was calculated (P = .08), showing a potential for functional impact.
Two different overlapping deletion calls in the RMRP gene body were observed in the same thyroid cancer patient. After manual inspection of the tumor and normal bam files, it was found that these calls were based on the same mutational event.
21. Cancer associated germline variant distance to non-coding driver candidates
(Morten Muhlig Nielsen)
We used a set of genome-wide significant cancer associated germline SNPs (n=650) from the NHGRI-EBI GWAS catalog59 as collected by Sud et al60. We evaluated the genomic distance from candidate non-coding drivers to the closest germline variant. All distances were above 50 kb with the exception of the TERT promoter which was 1 kb away from a coding variant (rs2736098) in the TERT gene.
Acknowledgements
We thank Kirsten Kübler for assistance with meta cohort generation. We are grateful to Matthew Meyerson, Eric S. Lander and the PCAWG steering committee for helpful feedback on the manuscript.
References
References
7. References
