Genomic basis for RNA alterations in cancer

Abstract

Transcript alterations often result from somatic changes in cancer genomes¹. Various forms of RNA alterations have been described in cancer, including overexpression², altered splicing³ and gene fusions⁴; however, it is difficult to attribute these to underlying genomic changes owing to heterogeneity among patients and tumour types, and the relatively small cohorts of patients for whom samples have been analysed by both transcriptome and whole-genome sequencing. Here we present, to our knowledge, the most comprehensive catalogue of cancer-associated gene alterations to date, obtained by characterizing tumour transcriptomes from 1,188 donors of the Pan-Cancer Analysis of Whole Genomes (PCAWG) Consortium of the International Cancer Genome Consortium (ICGC) and The Cancer Genome Atlas (TCGA)⁵. Using matched whole-genome sequencing data, we associated several categories of RNA alterations with germline and somatic DNA alterations, and identified probable genetic mechanisms. Somatic copy-number alterations were the major drivers of variations in total gene and allele-specific expression. We identified 649 associations of somatic single-nucleotide variants with gene expression in cis, of which 68.4% involved associations with flanking non-coding regions of the gene. We found 1,900 splicing alterations associated with somatic mutations, including the formation of exons within introns in proximity to Alu elements. In addition, 82% of gene fusions were associated with structural variants, including 75 of a new class, termed ‘bridged’ fusions, in which a third genomic location bridges two genes. We observed transcriptomic alteration signatures that differ between cancer types and have associations with variations in DNA mutational signatures. This compendium of RNA alterations in the genomic context provides a rich resource for identifying genes and mechanisms that are functionally implicated in cancer.

Main

For a more extensive study of cancer genome alterations, particularly in non-coding regions, the PCAWG project was formed to analyse the large number of whole-genome samples that were contributed to the ICGC and TCGA projects⁵. Individual projects did not use the same methods for key analyses; therefore, a major focus for each of the 16 PCAWG Working Groups was the unified analysis of the PCAWG data. For example, the PCAWG Technical Working Group led raw data collection, realignment of whole-genome sequencing data and implemented core somatic mutation calling pipelines⁵. Other PCAWG working groups focused on unified analyses of copy-number variation⁶, structural variants^7,8, germline variants⁵, mutational signatures⁹ and identification of driver genes⁸, among others⁵. Here, we report the joint analysis of available matched transcriptome and genome profiling for 1,188 samples from 27 tumour types by the PCAWG Transcriptome Working Group⁵, providing the largest, to our knowledge, resource of RNA phenotypes and their underlying genetic changes in cancer so far (Extended Data Fig. 1, Methods, Supplementary Results, Supplementary Table 23). We demonstrate the importance of transcriptomics data in understanding how different dimensions of specific DNA alterations contribute to carcinogenesis and map out the landscape of cancer-related RNA alterations.

Cancer-specific germline cis-eQTLs

To investigate the underlying mechanisms of different types of RNA alteration, we first focused on changes in the gene expression level (Extended Data Fig. 2). We initially considered common germline variants (minor allele frequency ≥ 1%) proximal to individual genes (±100 kb), and mapped expression quantitative trait loci (eQTL) across the cohort (Extended Data Fig. 3, Supplementary Table 1). This pan-cancer analysis identified 3,532 genes with an eQTL (false discovery rate (FDR) ≤ 5%, hereafter denoted eGenes) (Supplementary Table 2), enriched in proximal regions of transcription start sites (TSSs) (Extended Data Fig. 3).

To identify cancer-specific regulatory variants, we compared our eQTLs to eQTLs from the Genotype-Tissue Expression (GTEx) project¹⁰, adopting previous strategies to assess eQTL replication¹¹, and probed lead eQTL variants for marginal significance in GTEx tissues (P ≤ 0.01, Bonferroni-adjusted). Although most lead variants could be detected in GTEx samples (3,110 out of 3,532 eQTL variants), we identified 422 eQTLs that did not correspond to GTEx tissues, which suggests cancer-specific regulation (Extended Data Fig. 4, Supplementary Table 3). The corresponding eQTL lead variants were enriched for heterochromatic regions (Fig. 1a). Overall, this analysis revealed that the germline framework of gene expression regulation is largely conserved in cancer tissues.

Fig. 1: Germline and somatic SNVs associated with expression.

a, Epigenetics Roadmap enrichment analysis, showing the average fold change in Roadmap factors across cell lines in PCAWG-specific eQTLs of the pan-analysis as well as eQTLs that replicate in GTEx tissues. *P < 0.05/25, one-sided Wilcoxon rank-sum test in PCAWG-specific eQTLs corrected for the number of Roadmap factors used (that is, 25). Data are mean and s.d. b, Variance component analysis for gene expression levels, showing the average proportion of variance explained by different germline and somatic factors for different sets of genes including the mean effect across all factors: (1) all genetic factors (germline and somatic); (2) SCNAs; (3) somatic variants in flanking regions; (4) population structure; (5) cis-germline effects; and (6) somatic intron and exon mutation effects. c, Manhattan plot showing nominal P values of association for TEKT5 (highlighted in grey), considering flanking, intronic and exonic intervals. The leading somatic burden is associated with increased TEKT5 expression (P = 1.61 × 10⁻⁶) and overlaps an upstream bivalent promoter (red dots; annotated in 81 Roadmap cell lines, including 8 embryonic stem cells, 9 embryonic-stem-cell-derived and 5 induced pluripotent stem-cell lines). d, Summary of significant associations between mutational signatures (Sig) and gene expression. Top, the total number of associated genes per signature (FDR ≤ 10%). Bottom, enriched GO categories or Reactome pathways for genes associated with each signature (FDR ≤ 10%, significance level encoded in colour, −log₁₀-transformed adjusted P value). e, Standardized effect sizes on the presence of AEI, taking only SCNAs, germline eQTLs, coding and non-coding mutations into account. Data are the estimate and standard error of the estimate of the effect size.

Somatic cis-eQTLs in non-coding regions

Previous studies have described the landscape of non-coding mutations in cancer¹, particularly in promoter regions, and also their regulatory effects on gene expression^12,13. Here, we looked at possible somatic DNA changes, across the whole genome, that underlie alterations in gene expression. We estimated local mutation burdens by aggregating single-nucleotide variants (SNVs) in 2-kb intervals adjacent to genes (flanking), as well as in exons and introns (Extended Data Figs. 2, 5, 6). Next, we decomposed the expression variation of individual genes, considering common mutation burdens in cis, as well as cis germline variants and somatic copy-number alterations (SCNAs). This identified SCNAs as the major driver of expression variation (17%), followed by somatic SNVs in gene flanking regions (1.8%) and germline variants (1.3%) (Fig. 1b).

We also tested for associations between all common mutation burdens and gene expression across the whole genome. We identified 649 genes with a somatic eQTL (FDR ≤ 5%) (Supplementary Table 5). Of these, 11 associations were located in introns or exons of the respective eGene, including genes with known roles in the pathogenesis of specific cancers such as CDK12 in ovarian cancer¹⁴ and IRF4 in chronic lymphocytic leukaemia¹⁵ (Extended Data Figs. 7, 8). Most eQTLs (68.4%) involved associations with flanking non-coding mutation burdens (Extended Data Fig. 6e). Next, we considered eQTLs in flanking regions (n = 556) and tested for enrichment in cell-type-specific annotations from the Epigenetics Roadmap¹⁶. This identified 13 enriched annotations (FDR ≤ 10%) (Extended Data Fig. 9, Supplementary Table 6), including poised promoters, weak and active enhancers, and heterochromatin, but notably no enrichment for transcription-factor-binding sites (Supplementary Table 7). This enrichment in transcriptionally inactive regions may be due to an increased mutation rate in these regions (Extended Data Fig. 9), which has previously been reported in cancer¹⁷.

We also looked at the functional characterization of somatic eGenes and observed an enrichment for somatic eQTLs in bivalent promoters for cancer testis genes (P = 0.04, Fisher’s exact test) such as TEKT5¹⁸ (Fig. 1c, Extended Data Fig. 8h). Furthermore, we found a global enrichment (FDR ≤ 10%) for Gene Ontology (GO) categories related to cell differentiation and developmental processes (Supplementary Table 8). Overall, somatic eQTL analysis identified mostly non-coding regions associated with changes in local gene expression and, similar to cancer-specific germline eQTLs, showed enrichment for transcriptionally inactive regions such as heterochromatin.

Expression and mutational signatures

Global variations in mutational patterns can be quantified using mutational signatures, which tag mutational processes specific to their tissue-of-origin and environmental exposures¹⁹. However, the extraction of mutational signatures is an intrinsically statistical process that requires a posteriori functional annotation. We performed a pan-cancer association analysis between genome-wide mutational signatures and gene expression levels to decipher the molecular processes that accompany the presence of mutational signatures.

We considered 28 mutational signatures derived using non-negative matrix factorization of context-specific mutation frequencies⁹. We tested for association between signature prevalence in donors and total gene expression, accounting for total mutational burden, cancer type, and other technical and biological confounders. This identified 1,176 genes associated with at least one signature (FDR ≤ 10%) (Extended Data Fig. 10, Supplementary Table 19).

We considered 18 signatures with 20 or more associated genes for further annotation (Extended Data Fig. 11) and assessed enrichment using GO categories²⁰ and Reactome pathways²¹. We found that 11 signatures were enriched for at least one category (FDR ≤ 10%) (Supplementary Table 19), revealing associations consistent with known and unknown aetiologies (Fig. 1d). For example, signature 38, which is correlated with the canonical UV signature 7 (r² = 0.375, P = 5 × 10⁻⁴⁰) (Extended Data Fig. 11c), was linked to melanin processes (Fig. 1d). The synthesis of melanin causes oxidative stress to melanocytes²², and we found signature 38 associated with the oxidative-stress-promoting gene TYR²³ (P = 1.0 × 10⁻⁴). A hallmark of signature 38 genes are C>A mutations, a typical product of reactive oxygen species²⁴. This suggests that signature 38 may capture DNA damage that is indirectly caused by UV-induced oxidative damage after direct sun exposure²⁵, with TYR as a possible mediator of the effect.

Genomic basis of allelic expression

To analyse expression at the level of individual haplotypes, we tested for allelic expression imbalance (AEI) (FDR ≤ 5%, binomial test). We observed substantial differences in the fraction of genes with AEI between different types of cancer (Extended Data Fig. 12), and between cancer and the corresponding healthy tissues, with a high observed concordance between allelic imbalance at the DNA and RNA levels (Extended Data Fig. 13).

We used a logistic regression model to identify the determinants of AEI, accounting for known imprinting status²⁶, the germline eQTL genotype, SCNAs and the weighted mutational burden of proximal somatic SNVs stratified into functional categories (Extended Data Fig. 2). In aggregate, SCNAs accounted for 84.3% of the total explained variation, which confirmed our findings from the somatic eQTL analysis, followed by germline eQTL lead variants (9.1%), somatic SNVs (4.9%) and imprinting status (1.7%) (Extended Data Fig. 14). Although cumulatively, non-coding variants were more relevant than coding variants, somatic protein-truncating variants (‘stop-gained’ variants) that triggered nonsense-mediated decay²⁷ were the most predictive individually. SNVs within splice regions, 5′ untranslated regions (UTRs) and promoters were also strongly associated with the presence of AEI, and we observed a global trend of decreasing relevance of variants with increasing distance from the TSS (Fig. 1e, Extended Data Fig. 14).

Gene-centric attribution of AEI to individual sources of genetic variation (Supplementary Table 9) revealed an enrichment of somatically induced AEI in several known cancer-driver genes, as well as new candidates, such as the mismatch-repair-related gene EXO1 that is associated with survival in colorectal adenocarcinoma (log-rank P = 0.022, hazard ratio = 0.57) (Supplementary Results). We further observed a strong enrichment in the AEI score of cancer testis genes based on somatic SNVs only (χ² test P = 6 × 10⁻³). In summary, we identify somatic and germline genetic variation that is associated with allele-specific dysregulation of genes across cancer types.

Mutations associated with promoter usage

We considered promoter activity^28,29,30 as another molecular phenotype to study the effect of promoter mutations. Although cancer-specific alternative promoter usage has previously been shown²⁸, the association of underlying genomic alterations with promoter activities have not been broadly explored. To estimate the activity of individual gene promoters, we combined the expression of isoforms initiated in TSSs that are identical or nearby, assuming that these are transcribed from the same promoter (Extended Data Fig. 15a–c). We divided promoters into three categories: (1) inactive promoters (activity < 1 fragment per kilobase of transcript per million mapped reads (FPKM)), (2) major promoters (most active per gene) and (3) minor (all remaining) promoters, and examined the rates of mutation across varying activity levels. We observed an increase in the number of mutations near the TSS of major promoters compared with minor or inactive promoters (Extended Data Fig. 15d). This pattern is most prominent in skin melanoma, in which it has been attributed to impaired nucleotide-excision repair (Extended Data Fig. 15e, f, k, l). The cancer type that shows the strongest deviation from this pattern is colorectal adenocarcinoma, which highlights the tissue-specificity of mutational patterns at promoters (Extended Data Fig. 15e, f, m, n). Only 171 promoters show mutations in more than 5 samples per tumour type in a 200-bp window upstream of the promoter (Extended Data Fig. 15g, h). Most mutations occur in skin melanoma and lymphoma, which is expected owing to reduced nucleotide-excision repair and activation-induced cytidine deaminase (Extended Data Fig. 15h). We did not find significant pan-cancer associations between promoter mutational burden and promoter activity (Extended Data Fig. 15i, j). However, TERT has the highest number of promoter mutations^1,5,31 (Extended Data Fig. 16a), and these mutations have previously been reported to be associated with TERT expression¹; therefore, we investigated the TERT locus in more detail (Extended Data Fig. 16b). Although TERT does not show a significant association in the pan-cancer analysis, we found an association with increased promoter activity in individual types of cancer¹ (Extended Data Fig. 16c).

Mutations associated with splicing

Extending the classical hallmarks of cancer, alternative splicing is seen as increasingly relevant to explain cancer heterogeneity³². On the basis of our observations of a globally changing splicing landscape (Extended Data Fig. 17a–c), we sought to specifically understand the relationship between splicing changes and somatic mutations within introns. Focusing on cassette exon events, we integrated the quantification of splice events with somatic variants and identified 5,282 mutations near exon–intron boundaries, 1,800 (34%) of which were associated with a change in splicing (|z-score| ≥ 3) (Supplementary Table 10). Consistent with previous findings using exome sequencing^33,34, most mutations overlapping the essential dinucleotide motifs of the acceptor or donor site are associated with a splicing change—61% or 57%, respectively (Fig. 2a). Nearly one-third of all mutations (226 out of 469) in a 5-nucleotide window downstream of the 5′ site were significantly enriched for splicing changes (Fig. 2a). Almost all changes significantly associated with somatic mutations had a negative effect on splicing (96%) (Extended Data Fig. 17d). For mutations in or near the poly-pyrimidine tract, we found a significant (permutation test, P < 0.05) enrichment for mutations linked to outlier splicing (Fig. 2a). We also found an enrichment (P < 0.05, fold change > 2) of splicing outliers at branch-site adenosines (Fig. 2a middle, Extended Data Fig. 17d, Supplementary Table 11). Together, these results suggest that somatic mutations in the extended splice site region, poly-pyrimidine tract and branch point can affect splicing.

Fig. 2: Position-specific effect of somatic mutations on alternative splicing.

a, Top, proportion of mutations near exon–intron junctions and at branch sites that are associated with exon-skipping events. Mutations with associated splicing changes are those in which the percentage spliced in-derived |z-score| is ≥ 3 (dark blue). Asterisks denote intron positions significantly enriched for splicing changes relative to background based on a permutation test. *P < 0.05, **P < 0.01, ***P < 0.001. Bottom, sequence motifs of regions. b, Example of an exonization event in the tumour-suppressor gene STK11. The RNA-seq read coverage for a part of the gene is shown in red for a donor carrying the alternative allele, and in grey for a random donor with reference allele. The cassette exon event is shown as a schematic below. c, Enrichment of SINE elements in SAVs compared to sequence background (BG). Shown for SINE elements overlapping in sense (middle) and antisense (right) directions.

We also identified 1,900 rare splicing-associated variants (SAVs) that appear in only a small number of samples using the SAVNet approach³⁵ (Extended Data Fig. 17e; see ‘Data availability’ in the Methods). Notably, 862 SAVs affected canonical splice sites, whereas the other 1,038 disrupted non-canonical sites or created new splice sites. Notably, we find a twofold enrichment of cancer genes in SAVs (Extended Data Fig. 17f).

Although we find that those SAVs that create splice sites strongly concentrate near exon–intron boundaries (Extended Data Fig. 17g), 45.9% of SAVs are further than 100 bp away from the nearest annotated exon. Mutations at those sites generally changed the sequences towards the donor or acceptor motif consensus (Extended Data Fig. 17h). Focusing on novel splice sites deep in introns, we analysed the extent of exonizations—that is, the formation of new exons within an intron (Extended Data Fig. 17j, Supplementary Tables 13, 14). More than one-fifth of these new exons (9 out of 43) occur in cancer-related genes, such as the well-known tumour-suppressor gene STK11. As expected, the exonization event would cause a frameshift in STK11 (Fig. 2b, Extended Data Fig. 17k).

Alu elements that are inserted in an antisense direction have sequences that resemble consensus splice sites that, together with activating mutations, can lead to the formation of a new exon³⁶ (Extended Data Fig. 17l). We found a significant enrichment of splice-site-creating SAVs within annotated Alu sequences (P = 2.8 × 10⁻⁹), particularly in the antisense direction (P = 2.6 × 10⁻¹⁵) (Fig. 2c). Our results indicate that the exonization of Alu sequences, which has been extensively studied in the context of primate genome evolution, is also observed in cancer genome evolution.

Patterns of gene fusions across cancer

Gene fusions are an important class of cancer-driving event with therapeutic and diagnostic value³⁷. We identified a total of 925 known and 2,372 new cancer-specific gene fusions by combining the output of two fusion discovery methods as well as genomic rearrangement (structural variants) information and excluding artefacts or fusions in non-cancer samples³⁸ (Fig. 3a). For the 3,540 identified fusion events representing 3,297 unique gene fusions, we categorized them on the basis of novelty, recurrence and known oncogenic gene partners (Fig. 3a).

Fig. 3: Structural rearrangements associated with RNA fusions.

a, The number of all detected and new fusions and their overlap with the cancer census genes. b, Schematic of an example of bridged fusions. Bridged fusions are those composite fusions formed by a third genomic segment that bridges two genes. Only one of the possible orders of genomic arrangement is depicted in each case, with break points highlighted as thunderbolts.

Only 149 (approximately 5%) of the fusions occur in more than one sample, among which 78 are novel. Most of these (46 out of 78) were found across several histotypes. Of the 27 most recurrent gene fusions (Extended Data Fig. 18a), 8 have previously been reported (for example, CCDC6-RET³⁹, FGFR3-TACC3⁴⁰ and PTPRK-RSPO3) or independently detected in the TCGA cohort⁴¹, whereas 6 were new (such as NUMB-HEATR4, ESR1-AKAP12 and TRAF3IP2-FYN). In total, 105 fusion transcripts involved the UTR region of one gene and the complete coding sequences of another gene, possibly resulting from structural variation in promoter regions.

Although most genes involved in fusions engaged with only one fusion partner, 35 genes had more than 5 partners. These ‘promiscuous’ genes tended to be selective in being either a 5′ or a 3′ partner with conserved break points and positions (3′ or 5′), and were overrepresented in cancer census genes and the PCAWG cancer-driver genes (one-tailed Fisher’s exact test, odds ratio = 8.66, P ≤ 1.1 × 10⁻¹⁵, and odds ratio = 12.27, P ≤ 2.2 × 10⁻¹⁶, respectively). Network analysis of promiscuous genes and their partners revealed several large gene clusters containing at least 10 genes (Extended Data Fig. 18b), enriched in cancer-related pathways (Benjamini–Hochberg corrected P ≤ 0.01) and in protein–protein interactions (P ≤ 1.0 × 10⁻⁷), which suggests a possible functional role in cancer.

Notably, a large number of fusions, including known fusions, could not be associated with only a single structural-variation event. For example, the ETV6-NTRK3 gene fusion⁴² was present in a head and neck thyroid carcinoma sample, linking exon 4 of ETV6 to exon 12 of NTRK3. We found three separate structural variants in the same sample: (1) a translocation of ETV6 to chromosome 6; (2) a translocation of NTRK3 also to chromosome 6; and (3) an additional copy-number loss spanning from intron 5 of ETV6 to the exact structural variant break points, jointly bringing ETV6 within 45 kb upstream of NTRK3—a distance that would allow transcriptional read-through⁴³ or splicing⁴⁴ to yield the ETV6-NTRK3 fusion⁴⁵ (Fig. 3b). Thus, the short chromosome-6 segment appeared to function as a bridge, which linked two genomic locations to facilitate a gene fusion. We term such products ‘bridged fusions’. This class of fusion is not uncommon. Out of a total of 436 gene fusions supported by 2 separate structural variants, 75 are bridged fusions (Supplementary Table 15).

On the basis of the nature of the underlying genomic rearrangements, we propose a unified fusion classification system (Extended Data Fig. 19a). Aside from bridged fusions, 344 additional fusions are linked to more than one structural variant in the same sample. These multi- structural variant fusions are collectively termed ‘composite fusions’ (Extended Data Fig. 19a, b). We find 284 intercomposite fusions (interchromosomal translocation) and 124 intracomposite fusions (intrachromosomal rearrangement), exemplified by ERC1-RET1 and NUMB-HEATR4 fusions, respectively (Extended Data Fig. 19b). Composite rearrangements bring the fusion partners significantly closer to each other, from the median natural distance of 6.8 Mb to the median of 7.9 kb (Wilcoxon rank-sum test, P ≤ 2.2 × 10⁻¹⁶; Extended Data Fig. 19c) after translocation. For 18% of fusions, no evidence of structural variation was found. Given that 340 structural-variant-independent, intrachromosomal fusions had significantly closer break points than those with structural variation (Extended Data Fig. 19d), it is possible that they could result from RNA read-through events. The other possibility is that the underlying supporting structural variants escaped detection, as shown by the observation that known gene fusions that are driven by structural variation, such as TMPRSS2-ERG⁴⁶, did not have consistent evidence for structural variation in matching samples.

Landscape of RNA alterations in cancer

Given our comprehensive set of RNA alterations, we sought to characterize the heterogeneous mechanisms of cancer genome and transcriptome alterations. To enable joint analyses of RNA and DNA alterations, we created a gene-level table, which indicates the presence or absence of possible functional changes to RNA or DNA for each gene and donor. After stringent filtering, we identified 1,523,098 alteration events, in which an event is a gene–sample–alteration triplet (Extended Data Table 1, Supplementary Table 14). It should be noted that we chose to include only RNA alterations with potential functional effects or with the strongest quantitative affect, resembling similar strategies for filtering DNA alterations⁴⁷. Recurrence analysis across several alteration types helped us to further enrich for functionally relevant genes. Building on the gene-centric table, we characterized gene alterations at the RNA level and contrasted these with DNA alterations (non-synonymous SNVs or SCNAs)⁵. On the basis of the calculated association between each RNA- and DNA-level alteration across all histotypes, we found that half of the RNA alterations significantly correlated with DNA alterations (likelihood ratio test, FDR < 1 × 10⁻⁴) (Extended Data Fig. 20).

When comparing gene alteration frequencies across all histotypes (Fig. 4a), we note that different types of cancer contain distinct combinations of DNA- and RNA-level alterations (Fig. 4a, Supplementary Table 17). Although, as expected, skin melanoma significantly exceeds other cancers in the number of non-synonymous SNVs⁴⁸ (Wilcoxon rank-sum test, P < 0.012), lymphatic cancers have low numbers of SNVs (Wilcoxon rank-sum test, P = 5.3 × 10⁻¹⁵), but high incidences of alternative splicing outliers (Wilcoxon rank-sum test, P = 4.9 × 10⁻⁴⁷), which suggests that transcriptomic alterations can be relatively more pronounced in certain cancer types.

Fig. 4: Global view of DNA and RNA alterations that affect tumours.

a, The median numbers of different alterations across histotypes. Histotypes are ordered by hierarchical clustering based on the pattern of different types of alteration. Only histotypes with more than 10 donors are shown. Alt., alternative; non-syn, non-synonymous. Cancer-type abbreviations are listed in Supplementary Table 23. b, c, Circular representations of the selected genes significantly co-occurred with B2M (b) and PCBP2 (c). Connecting lines indicate the specific types of co-occurrence of alteration pairs. The inner histograms indicate the frequencies of incidences of different alteration types shown in different colours. d, All 74 Catalogue of Somatic Mutations in Cancer (COSMIC) cancer census genes or PCAWG driver genes that are both frequently and heterogeneously altered across both RNA- and DNA-level alterations. Yellow bars indicate the proportion of samples that had DNA-level alterations, and green bars indicate the proportion of samples with RNA-level alterations. Middle column is the proportion of each alteration type observed for that gene. e, The enrichment of cancer genes within our list of significantly recurrent genes.

To evaluate to which extent RNA changes provide additional mechanisms for cancer gene alterations, we examined DNA- and RNA-level alterations both in sets of genes in pathways (Extended Data Fig. 21) and in individual genes with known roles in cancer (Extended Data Fig. 22). We found that RNA alterations occur at a high proportion in many pathways, including the NOTCH and TGF-β pathways. In addition, KRAS exhibits more RNA alterations than DNA alterations in some types of cancer. Given the recent finding that alternative splicing of KRAS expanded the prognostic affect beyond mutation status in colorectal cancer⁴⁹, our data further support several modes of alteration for KRAS in tumours.

Co-occurrence of RNA and DNA alterations

The diverse types of alteration in this study enabled us to investigate trans-associations between different genetic and expression characteristics involving cancer-related genes (FDR < 5%) (Supplementary Table 18). By investigating whether somatic mutations of known cancer genes are associated with the expression of other genes, we found IDH1 and NFKBIE to be widely linked to the dysregulation of many genes (Extended Data Fig. 23a, b). Notable co-occurrences were present in several types of cancer. For example, B2M and EIF4G2 alterations were simultaneously observed in both B-cell non-Hodgkin lymphoma and lung squamous cell carcinoma. Pathway enrichment analysis of the top 100 genes associated with all B2M alterations indicates that the most affected genes are involved in DNA repair (FDR ≤ 1%), and approximately two-thirds of those associations were significant in more than one cancer type (Fig. 4b, Extended Data Fig. 23c).

We also examined how cancer genes could be affected by other genes by co-occurrence analyses. Expression outliers of PCBP2 co-occurred with aberrant splicing of a large number of cancer-related genes, including CTNNB1 and CDK4 (Fig. 4c). PCBP2 has been reported to enhance the splicing of cassette exons⁵⁰. Our results thus further support the possible role of PCBP2 in regulating the splicing of cancer-related genes.

Recurrent RNA alterations in driver genes

In our analyses of cis-acting mutations that are associated with these individual RNA phenotypes, the vast majority were observed rarely in the PCAWG cohort. Many cancer genes (such as MET^51,52) are known to be somatically altered by heterogeneous mechanisms such as gene fusions, splicing mutations and non-synonymous mutations; therefore, examining genes that are altered by several cis-acting mechanisms may help to identify cancer genes in which an individual alteration type is rare. A total of 5,413 genes were altered by gene expression, allele-specific expression (ASE), splicing and/or gene fusion, and had an associated DNA-level mutation in cis (Supplementary Table 20). PCAWG-defined driver genes⁸ tended to have more diverse mechanisms of RNA-level alterations when compared to genes that have not previously been identified as a cancer gene (P < 0.001) (Extended Data Fig. 24a). We identified, for example, a somatic eQTL, a splicing-associated variant and fusions in the known tumour-suppressor NF1 in the MAPK pathway (Extended Data Fig. 24b).

Owing to the fact that most somatic mutations are rare⁵, it is difficult to statistically distinguish functionally relevant, potential driver alterations from passenger alterations. Therefore, we aimed to identify genes that are both recurrently and heterogeneously altered, under the hypothesis that these genes have increased functional relevance. This analysis identified 731 genes with significant recurrent aberrations (FDR < 5%) (Extended Data Fig. 25a), with the top-ranking genes carrying both RNA and DNA alterations. RNA alterations accounted for 0.05–99.14% (mean: 78.23%) of all identified alterations in each gene (Extended Data Fig. 25a, Supplementary Table 21). This ranking is enriched for the union of cancer census genes⁵³ (60 out of 603) and PCAWG-defined driver genes (33 out of 157, unioned: 74 out of 674 P = 4.6 × 10⁻¹³, enrichment: 2.45) (Fig. 4d, e).

Among the top 10% of our ranked genes is CDK12 (rank 55). We find 91 samples that have an alteration involving its protein kinase domain, which has been implicated in DNA repair dysregulation⁵⁴. Many of these samples have no DNA-level alterations in CDK12 (46%) (Extended Data Fig. 26a). Furthermore, splicing, alternative promoter, SNV, RNA-editing and fusion alterations in this gene are mutually exclusive (adjusted P = 4.8 × 10⁻³) (Extended Data Fig. 26b, c). Upon further investigation, we find that somatic eQTL mutations in CDK12 are associated with a tandem duplicator phenotype⁵⁵. Although this association was not replicated with other RNA alterations, it provides evidence that somatic CDK12 mutations may alter its function through gene expression changes. This example illustrates that performing a recurrence analysis over diverse RNA and DNA alterations can help to identify genes known to be important in tumorigenesis.

Discussion

Here we present a comprehensive catalogue of RNA-level alterations in cancer, spanning 27 different tumour types, and provide a harmonized resource of matched transcriptome and whole-genome sequences. We identified 731 genes that were recurrently altered by several mechanisms, jointly enriched for known cancer census and PCAWG driver genes⁸. The list includes genes that are primarily altered at the DNA level (such as TP53), but also genes for which the alteration most frequently manifests in RNA (such as GAS7). Out of 87 samples from the PCAWG study that did not have a driver alteration at the DNA level⁵, and had RNA-sequencing (RNA-seq) data, every sample had an RNA-level alteration identified. Although cancer is thought to be driven by changes in DNA primarily, some driver alterations may manifest themselves via changes in RNA rather than DNA sequence mutations.

We identified germline eQTLs for around 20% of expressed genes. The number of eGenes found is generally low compared with some other studies, reflecting the heterogeneity of our samples. Only 422 genes appeared to be specific to cancer; this is likely to be an underestimate owing to the heterogeneity, small sample numbers and the rather conservative strategy chosen. We have also mapped linkages between genes and somatic aberrations in cis, in which 68.4% of associations were between non-coding somatic variants and gene expression. Allelic copy-number imbalance is a major determinant of ASE dysregulation in cancer. We found mutations associated with splicing changes including novel cancer-specific exons that can be partially explained by mutation-driven exonization. We systematically compared gene fusions with whole-genome rearrangements across many tumour types and found 82% of detected fusions were associated with specific genomic rearrangements. For the remaining fusions, it is possible that the relevant genomic rearrangements have not been detected, or that some fusions happen directly at the RNA level, as trans-splicing or read-through events. The availability of whole-genome sequences allowed us to develop a systematic classification of fusion events and to propose a new bridged fusion mechanism.

Because global differences in RNA expression phenotypes are largely tissue-specific, our ability to associate mutations in cis or trans are limited by the small and variable sample sizes within each histotype. Further work is needed to investigate other mechanisms of genome alteration that can lead to changes in RNA such as epigenetic changes⁵⁶ or enhancer hijacking⁵⁷. Our work will help to prioritize further investigations.

Overall, our analyses show diverse modes of alteration of cancer genes and pathways at the DNA and RNA levels, and demonstrate that RNA analyses reveal cancer-associated pathway alterations that have not yet been detected via DNA-only approaches. These insights illustrate the power of integrated transcriptome and whole-genome sequencing analysis for cancer studies.

Methods

RNA-seq alignment and quality-control analysis

Tumour and healthy ICGC RNA-seq data, included in the PCAWG cohort⁵, was aligned to the human reference genome (GRCh37.p13) using two read aligners: STAR⁵⁸ (v.2.4.0i, two-pass), performed at MSKCC and ETH Zürich, and TopHat2⁵⁹ (v.2.0.12), performed at the European Bioinformatics Institute. Both tools used Gencode (release 19)⁶⁰ as the reference gene annotation. For the STAR two-pass alignment, an initial alignment run was performed on each sample to generate a list of splice junctions derived from the RNA-seq data. These junctions were then used to build an augmented index of the reference genome per sample. In a second pass, the augmented index was used for a more sensitive alignment. Alignment parameters have been fixed to the values reported in https://github.com/ICGC-TCGA-PanCancer/pcawg3-rnaseq-align-star. The TopHat2 alignment strategies also followed the two-pass alignment principle, but was performed in a single alignment step with the respective parameter set. For the TopHat2 alignments, the irap analysis suite⁶¹ was used. The full set of parameters is available along with the alignment code in https://hub.docker.com/r/nunofonseca/irap_pcawg/. For both aligners, the resulting files in BAM format were sorted by alignment position, indexed and are available for download in the GDC portal (https://portal.gdc.cancer.gov/) and the ICGC Data Portal (https://dcc.icgc.org/). The individual accession numbers and download links can be found in the PCAWG data release table: http://pancancer.info/data_releases/may2016/release_may2016.v1.4.tsv. Cancer-type abbreviations are listed in Supplementary Table 23. Histology was derived from an older version released by the PCAWG Pathology and Clinical Correlates Working Group. Assignments of donor to histology used in this study can be found in the file rnaseq.extended.metadata.aliquot_id.V4.tsv.gz at https://dcc.icgc.org/releases/PCAWG/transcriptome/metadata/.

Quality control of all datasets was performed at three main levels: (1) assessment of initial raw data using FastQC⁶² (v.0.11.3) (Supplementary Fig. 4); (2) assessment of aligned data (percentage of mapped and unmapped reads for both alignment approaches); and (3) quantification (by correlating the expression values produced by the STAR and TopHat2 based expression pipelines) (Supplementary Fig. 2). In total, we defined six quality-control criteria to assess the quality of the samples. We marked a sample as a candidate for exclusion if: (1) 3 out of 5 main FastQC measures (base-wise quality, k-mer overrepresentation, guanine-cytosine content, content of N bases and sequence quality) did not pass; (2) more than 50% of reads were unmapped or fewer than 1 million reads could be mapped in total using the STAR pipeline; (3) more than 50% of reads were unmapped or fewer than 1 million reads could be mapped in total using the TopHat2 pipeline; (4) we measured a degradation score⁶³ greater than 10; (5) the fragment count in the aligned sample (averaged over STAR and TopHat2) was <5 million; and (6) the correlation between the expression counts of both pipelines was <0.95. If a sample did not pass one of these six criteria it was marked as problematic and placed on a greylist. If more than two criteria were not passed, we excluded the sample.

A subset of 722 libraries from the projects ESAD-UK, OV-AU, PACA-AU and STAD-US were identified as technical replicates generated from the same sample aliquot. These libraries were integrated post-alignment for both the STAR and the TopHat2 pipelines using samtools⁶⁴ into combined alignment files. Further analysis was based on these files. Read counts of the individual libraries were integrated to a sample-level count by adding the read counts of the technical replicates.

Initially, a total of 2,217 RNA-seq libraries were fully processed by the pipeline. Quality-control filtering and integration of technical replicates (722 libraries) gave a final number of 1,359 fully processed RNA-seq sample aliquots from 1,188 donors.

GTEx data analysis

For a panel of RNA-seq data from a variety of healthy tissues, data from 3,274 samples from GTEx (phs000424.v4.p1) were used and analysed with the same pipeline as PCAWG data for quantifying gene expression. A list of GTEx identifiers are provided at https://dcc.icgc.org/releases/PCAWG/transcriptome/metadata.

Quantification and normalization of transcript and gene expression

STAR and TopHat2 alignments were used as input for HTSeq⁶⁵ (v.0.6.1p1) to produce gene expression counts. Gencode v.19⁶⁰ was used as the gene annotation reference. Quantification on a per-transcript level was performed with Kallisto⁶⁶ (v.0.42.1). This implementation is available as a Docker container at https://hub.docker.com/r/nunofonseca/irap_pcawg. The implementation of the STAR and TopHat2 quantification is available as docker containers in: https://github.com/ICGC-TCGA-PanCancer/pcawg3-rnaseq-align-star and https://hub.docker.com/r/nunofonseca/irap_pcawg/, respectively. Quantification of consensus expression was performed by taking the average expression based on STAR and TopHat2 alignments. Gene counts were normalized by adjusting the counts to FPKM⁶⁷ as well as FPKM with upper quartile normalization (FPKM-UQ) in which the total read counts in the FPKM definition has been replaced by the upper quartile of the read count distribution multiplied by the total number of protein-coding genes.

The FPKM and FPKM-UQ calculations were as follows. FPKM = (C × 10⁹)/(NL), in which N denotes the total fragment count to protein-coding genes, L denotes the length of the gene and C denotes the fragment count. FPKM-UQ = (C × 10⁹)/(ULG), in which U denotes the upper quartile of fragment counts to protein-coding genes on autosomes unequal to zero, and G denotes the number of protein-coding genes on autosomes.

t-Distributed stochastic neighbour embedding analysis

The t-distributed stochastic neighbour embedding (t-SNE) plots in Supplementary Figs. 5 and 6 were produced using the RTsne package⁶⁸ (with a perplexity value of 3) based on the Pearson correlation of the aggregated expression (log + 1) of the 1,500 most variable genes. FPKM expression values per gene were aggregated (median) by tissue (GTEx) and study (PCAWG). Coefficient of variation for each gene was also computed per tissue (GTEx) and study (PCAWG) to determine the 1,500 most variable genes. Purity values were previously described⁶⁹.

The t-SNE plot in Extended Data Fig. 17c is based on all exon-skipping events in protein-coding genes confirmed by SplAdder⁷⁰. Each event was quantified in both the PCAWG and GTEx cohort. All events with more than 1% of missing percentage spliced in (PSI) values across the concatenated PCAWG and GTEx samples were removed. The remaining missing values were imputed as the mean over the non-missing samples. The centred data were then visualized using the TSNE package from the Scikit Learn toolkit⁷¹ with a perplexity value of 100, random state 0 and an initialization with PCA.

Associations between genetic variation and gene expression: patient cohort

To associate genetic variation with gene expression, we analysed whole-genome sequencing (WGS) of the 1,188 donors with matched whitelisted RNA-seq data from the PCAWG cohort. Germline genotypes, SNV calls and segmented allele-specific SCNA calls were previously reported⁵. We matched 1,188 tumour RNA-seq IDs⁵ to WGS whitelist tumour IDs (synapse entry syn10389164). For patients with multiple WGS IDs (2 out of 1,188) or RNA-seq aliquot IDs (17 out of 1,188), we resolved the matching by pairing samples with the same ‘tumor_wgs_submitter_specimen_id’ (Supplementary Table 1). The 1,188 patients are spread across 27 types of cancer and 29 project codes and include 899 carcinomas; 34 patients are metastatic and 13 recurrent with the remaining patients being primary tumours (Supplementary Table 1).

We used the data of these 1,188 patients for performing somatic and germline eQTL mapping, ASE analysis and association studies between gene expression and mutational signatures.

Gene expression filtering

Gene expression values (measured in FPKM; https://dcc.icgc.org/releases/PCAWG/transcriptome/gene_expression) from consensus expression quantification as described above were used for this analysis.

Genes with FPKM ≥ 0.1 in at least 1% of the patients (12 patients) were retained, resulting in 47,730 genes. Only 18,898 protein-coding genes (according to the ‘gene_type’ biotype reported in Gencode v.19⁶⁰) were used for the subsequent QTL analyses. The log₂-transformed expression values (FPKM + 1) were subjected to peer analysis⁷² to account for hidden covariates (syn7850427; https://dcc.icgc.org/releases/PCAWG/transcriptome/eQTL/phenotype). To balance the number of covariates, statistical power and available sample sizes per cancer type, we followed the GTEx protocol and estimated 15, 30 and 35 hidden covariates to be used depending on sample size⁷³ (n < 150, 150 ≤ n < 250, n ≥ 250). Peer residuals were then rank-standardized across patients. The FPKM cut-off values and peer correction were also applied to the subset of 899 patients with carcinoma, yielding 18,837 protein-coding genes after filtering. Furthermore, we used ordinary least-squares regression to correlate each of the 35 peer factors with per-sample covariates, including cancer project codes, gender, tumour purity, somatic burden and several sequence metrics (Supplementary Notes), to understand the proportion of variance explained by known biological and technical covariates.

Covariates

In all linear models, we accounted for known confounding factors by modelling them as fixed effects. In all association studies, we accounted for sex, project code (describing cancer type and country of origin) and per-gene copy-number status (Supplementary Table 1 for the list of per patient covariates; syn7253568 and syn7253569 for sex and project codes; syn9661460 for per gene copy number). Per-gene copy-number alterations were derived as the average copy number across all copy-number aberrations called within the annotated gene boundaries based on syn8042988.

The somatic eQTL, ASE and mutational signature analyses also accounted for total somatic mutation burden (number of SNVs and short insertions and deletions (indels)) and sample purity (Supplementary Table 1). Purity was estimated based on copy-number segmentation. In addition, the somatic eQTL and ASE analyses accounted for local SNV burden calculated in a 1-Mb window from the gene coordinates (https://dcc.icgc.org/api/v1/download?fn=/PCAWG/transcriptome/eQTL/covariates/pergene.somatic.snv.cis.burden.1188.wl.donors.tsv.gz).

The germline eQTL analysis also modelled the population structure as random effect. The population structure was assessed by a kinship matrix that was calculated based on every twentieth germline variant, processed as described below (see ‘Germline eQTL variants’). The kinship matrix was then calculated as an empirical patient-by-patient covariance matrix.

Different covariates were accounted for per-analysis method (Supplementary Table 1). The project code describes cancer type and country-of-origin. Somatic burden is the total number of SNVs and indels. Purity was estimated based on copy-number segmentation. Local somatic burden is the number of SNVs in a 1-Mb window around the gene coordinates. Local copy number was defined as the average copy-number state across all SCNAs called within the annotated gene boundaries.

GO and Reactome pathway enrichment

We performed GO^74,75 and Reactome pathway^20,21 enrichment with the Bioconductor packages biomaRt^76,77, clusterProfiler⁷⁸ and ReactomePA⁷⁹ (FDR ≤ 10%). The number of genes used as background set is described per analysis method.

Germline eQTL variants

PCAWG variant calls v.0.1⁵ were downloaded from GNOS and processed following the PCAWG-8 protocol: (1) VCF files were indexed and merged using bcftools⁸⁰. (2) All variants were filtered for ‘PASS’ flag. (3) All variants were filtered for quality larger than 20. (4) Only bi-allelic sites were considered.

HDF5 files for each 100-kb chunk of the VCF files were generated, assuming additivity that was numerically encoded as 0, 1 or 2 for homozygous reference, heterozygous or homozygous alternative state, respectively. For indels, we encoded the presence or absence of the variant as 0 or 1, respectively. Each variant was normalized to mean 0 and standard deviation 1. Missing variants were mean-imputed. To create our eQTL release set v.1.0, the resulting HDF5 files were subsequently merged into a global HDF5 file and all variants which follow any of the following conditions were removed: (1) minor allele frequency ≤ 1%; and (2) missing values ≥ 5%

Germline eQTL analysis

In the germline eQTL analyses, we used the processed gene expression dataset from 1,178 patients for which germline variant calls (eQTL release set v.1.0, see ‘Germline eQTL variants’) were available. Linear mixed models were used to model the correlation between germline variants (within 100 kb of gene boundaries) and gene expression values (see ‘Gene expression filtering’) using the limix package⁸¹. Known covariates were modelled as fixed effects and population structure as random effect (see ‘Covariates’).

A two-step approach was used to adjust for multiple testing. First, for each gene, we adjusted for the number of independent tests estimated based on local linkage disequilibrium⁸². Second, we performed a global correction across the lead variants, that is, the most significant SNPs, per eQTL. Germline eGenes were defined as genes with an eQTL with global FDR ≤ 5%.

GTEx comparative analysis

The GTEx comparative eQTL analysis was based on the eQTL maps v.6p¹⁰. We mapped the positions and alleles of our PCAWG-specific eQTL to the eQTL in all GTEx tissues. To determine whether a lead eQTL variant is replicated in a given GTEx tissue, we followed the previously described strategy¹⁰. For each eGene, we considered the eQTL lead variant and assessed the replicability of the signal in the GTEx cohort based on marginal association statistics using 42 GTEx tissues without cell lines (P < 0.00024 = 0.01/42, corrected for the number of GTEx tissues—that is, 42)). If the lead variant did not replicate or was not tested, we determined replication based on the variant with the smallest P value within the linkage disequilibrium block (r² ≥ 0.8 estimated based on UK10K project) of the lead variant across 25 (or 42) tissue-matched GTEx analyses. If neither lead nor any variant within the linkage disequilibrium block was tested, we determined replication based on the smallest P value of any variant within the 100-kb window tested within the GTEx cohort. We also derived less stringent sets of PCAWG-specific eGenes by allowing replication in up to 1, 5 or 10 GTEx tissues.

Tissue sharing of germline eGenes between histotypes

Using the R package qvalue (https://github.com/StoreyLab/qvalue, v.2.14.0), we generated π₁ statistics comparing the lead variants of one histotype against their P value distribution in the other histotypes. Because π₁ statistics are known to be confounded by sample size and number of eQTL found, we subsampled the eQTL lead variants to a randomly selected set of 100 variants. After 20 rounds of subsampling, we derived the same π₁ statistics as mentioned earlier and reported the average.

Roadmap enrichment of germline eGenes

For each lead variant, we generated a matching background set of 1,000 variants using SNPsnap⁸³. Each variant (background and foreground) was intersected with the location of 25 Roadmap factors¹⁶ in 127 cell types. From this we derived fold change and P values. Significant changes of fold change between PCAWG-specific and unspecific eQTLs is based on a one-sided Wilcoxon rank-sum test.

Enrichment analysis

Enrichment of Reactome pathways of PCAWG-specific eGenes was performed using the Bioconductor package ReactomePA⁷⁹.

Somatic calls and mutational burden

We used the set of consensus SNVs somatic calls provided by PCAWG (syn7357330) based on three core caller pipelines and MuSE⁸⁴. On average, we counted 22,144 somatic SNVs per patient, with different median numbers of SNVs per cancer type, ranging from 1,139 in thyroid adenocarcinoma to 72,804 SNVs in skin melanoma (Extended Data Fig. 5a). Owing to the low frequency of somatic SNVs across the cohort (Extended Data Fig. 5b), we collapsed the variants by genomic regions defined by gene annotations (Gencode v.19⁶⁰). Specifically, we generated a set of disjoint gene exons by collapsing overlapping exon annotations into single features using bedtools⁸⁵. The set of disjoint introns was generated using bedtools by subtracting the collapsed exonic regions from the gene regions. To map local effects of somatic mutations in flanking features outside the gene body, we binned the surrounding regions (plus and minus 1 Mb from the gene boundaries) into 2-kb windows (flanking) overlapping by 1 kb.

We defined three different types of aggregated somatic burden to assess differences in power in detecting somatic eGenes and P value calibration. The burden in a genomic region was defined as (1) a binary value that indicates presence or absence of SNVs; (2) the aggregated burden as sum of SNVs; or as (3) weighted burden, that is, sum of variant allele frequencies of the SNVs (Supplementary Fig. 10a) to take into account their clonality (https://dcc.icgc.org/releases/PCAWG/transcriptome/eQTL/genotypes). We assessed calibration of all three analyses with Q–Q plots of nominal and permuted P values (permutation of the patients in the gene expression matrix) (Supplementary Fig. 10b–d). Moreover, for the linear regression analysis, genotypes were standardized across patients (to mean zero and standard deviation one) and standardized effect sizes are provided in Supplementary Table 5.

Overall, somatic burden within flanking regions was the most prevalent type of burden tested per gene (Extended Data Fig. 6a). We found similar average relative mutation density per type of genomic region (flanking = 0.008 mutations per kb; introns = 0.007 mutations per kb; exons = 0.006 mutations per kb) (Extended Data Fig. 6b) and average recurrence of the same mutated region across the cohort was rather low (flanking = 1.4%; exons = 1.7%; introns = 4%) (Extended Data Fig. 6c).

Somatic eQTL analysis

Linear models were used to model the correlation between recurrent somatic burden and gene expression of up to 18,898 protein-coding genes, using the limix package⁸¹ (see ‘Gene expression filtering’). Gene expression was corrected for 35 hidden Peer factors. Known covariates were modelled as fixed effects (see ‘Covariates’). We considered only somatic burdens with frequency greater than 1%, including exonic and intronic burdens, as well as flanking burdens, within 1 Mb from gene boundaries.

The somatic eQTL analysis was performed on all 1,188 patients and on the subset of 899 patients with carcinoma (representing 20 of the 27 types of cancer) to replicate the analysis on a more homogeneous set of tumours. A cis window of 1 Mb from the gene boundaries was used to find mutated genomic intervals with a burden frequency ≥ 1% in the cohort (at least 12 patients in the full cohort and 9 patients in the carcinoma cohort). Together, 18,708 of the genes had at least one mutated interval at that frequency and were included in the analysis and 1,049,102 regions showed a burden frequency ≥ 1%

Bonferroni correction was applied to correct for multiple cis windows tested within the same gene. Then, Benjamini–Hochberg correction was applied to adjust the P values of the lead genomic regions across genes. Somatic eGenes were defined as genes with an eQTL at a FDR ≤ 5%.

Somatic cis-eQTL comparative analysis

We compared our 649 somatic eQTL set with three previous cancer studies^86,87,88 to identify independent evidence of interaction between our eGenes and the associated cis-genomic regions with somatic burden. Studies were chosen if they provided lists of cancer regulatory elements linked to genes or regulatory elements with somatic mutations linked to gene expression deregulation in cancer. All the three studies examined were based on TCGA cancers. For this, we checked perfect overlaps with both the somatic burden location and the eGene. Moreover, we looked at the overlap between somatic eQTL and 72,987 GeneHancer⁸⁹ enhancers-to-genes interactions, with at least two independent supporting methods (called ‘double-elite’), downloaded from the UCSC hg19 GeneHancer track⁹⁰. We then compared this overlap with a set of nulls generated by 1,000 random permutations of the GeneHancer regulatory elements with nearby genes located within 1 Mb. We then retrieved an empirical P value of enrichment by counting the number of random nulls (N) showing greater number of overlaps than those found between the somatic eQTL set and the GeneHancer set (P = (N + 1)/(1,000 + 1)).

Functional enrichment in somatic cis-eQTL

To identify putative regulatory sites enriched for somatic eQTL, we retrieved functional annotations of the lead genomic flanking intervals of the somatic eQTL (556 intervals linked to 638 somatic eQTL). Therefore, we mapped somatic eQTL to 25 Roadmap Epigenomics chromatin marks of 127 different cell types¹⁶ and ENCODE transcription-factor binding site annotations in 9 cell types (including 8 cancer and one embryonic stem-cell lines⁹¹) (Supplementary Tables 6 and 7). We compared annotations in the significant set of eQTLs with a null distribution based on 1,000 random samplings of a matched set of genomic intervals. To define the matched sets of genomic intervals, we selected flanking genomic intervals from the whole set of tested genes that showed a similar distance from the gene start (exact distance ± 2 kb) and that matched the exact burden frequency of the corresponding interval in the significant associations. We then overlapped the 1,000 matched sets with Roadmap Epigenomics and ENCODE annotations. To avoid ambiguous overlaps (with multiple annotations), we retained only genomic intervals showing a minimum overlap of 10% of their length.

We retrieved an empirical P value of enrichment for each annotation by counting the number of randomly sampled flanking intervals (N) showing greater number of overlaps compared to the eQTL set (P = (N + 1)/(1,000 + 1)). Benjamini–Hochberg correction was applied to the empirical P values (over 25 marks in 127 cell lines for Roadmap Epigenomics annotations and over 149 transcription-factor-binding sites for 9 ENCODE cell lines). We then computed the fold change per annotation and cell line as a ratio of annotated lead flanking intervals and mean number of annotated matched random flanking intervals over the 1,000 samplings.

Furthermore, we performed GO^74,75 and Reactome pathway^20,21 enrichment with the Bioconductor packages biomaRt^76,77, clusterProfiler⁷⁸ and ReactomePA⁷⁹ (FDR ≤ 10%) and also looked at enrichment within high-confidence cancer testis genes previously described⁹², using 18,708 genes with at least one mutated interval as background.

Variance component analysis

Limix was used to perform variance decomposition using the same covariates as in the somatic variant analyses except for local copy-number state (see ‘Covariates’). The random effects were based on the following common germline variants and somatic burden (frequency > 1%) (see ‘Somatic calls and mutational burden’ for detailed description of burden): (1) cis-somatic intronic: weighted burden in introns; (2) cis-somatic exonic: weighted burden in exons; (3) cis-somatic flanking: weighted burden in 1-kb-overlapping regions of 2 kb within 1 Mb from gene boundaries; (4) somatic intergenic: weighted burden in 1-kb-overlapping regions of 2 kb outside the 1 Mb window; (5) cis-germline: germline variants within 100 kb from gene boundaries; (6) trans-germline: genome-wide population structure (see ‘Covariates’); and (7) local copy-number variation (see ‘Covariates’).

All the data was mean-centred and standardized. For each of the random effects, a linear kernel was computed and used as covariance matrix. The resulting variance components were normalized to add up to one.

Mutational signature associations

We obtained 39 mutational signatures from PCAWG-7 beta 2 release⁹ and used linear models to associate the mutational signatures with gene expression of up to 18,898 protein-coding genes across 1,159 patients while accounting for known covariates (see ‘Covariates’) (quality control) (Extended Data Fig. 10a–e). The 1,159 patients were a subset of the total 1,188 patients, for whom mutational signature profiles were available. Gene expression was corrected for 35 hidden peer factors (see ‘Gene expression filtering’).

We retained 18,888 genes that showed a minimum FPKM of 0.1 in at least 1% of 1,159 the patients (see ‘Gene expression filtering’). Signatures with zero variance and a prevalence below 1% were filtered, and we obtained 28 signatures. We applied linear models to associate expression of these genes with the signatures across all 1,159 patients, a subset of 877 patients with carcinoma or a subset of 891 European patients to assess consistency of the associations (Extended Data Fig. 10f, g).

Across all patients, we found 1,176 significantly associated genes after Benjamini–Hochberg correction (we used an FDR ≤ 10% for enrichment analyses, multiple testing was applied across all signature–gene pairs) (Supplementary Tables 19a–c). We performed gene enrichment analyses of the significant genes per signature (see ‘GO and Reactome pathway enrichment’) (here 18,831 background genes, multiple testing correction across all ontologies per signature FDR ≤ 10%) (Supplementary Table 19d). Whereas most signatures were associated with only few genes, 18 showed recurrent trans effects and affected expression of over 20 genes (Extended Data Fig. 11d, Supplementary Table 19e). We further found that the vast majority of genes (85.8%) were associated with only one signature (1,009 genes); 129 genes were associated with two, 32 with three, 5 with four and 1 with five signatures.

To assess how tissue-specific both mutational signatures and their associations with gene expression are, we analysed the occurrence of each signature in each of the types of cancer. We assessed the presence (at least one SNV of a signature in at least one patient with a specific cancer type) and mean prevalence (mean number of SNVs of a certain signature across all patients of a specific cancer type) of the signatures in the types of cancer (Extended Data Fig. 13c, d). We defined cancer-type-specific signatures to occur in up to four types of cancer (signatures 4, 7, 9, 12, 16, 38 and 39) and common signatures to be missing in up to five types of cancer (signatures 2, 13 and 18). For each of these signatures, we performed cancer-type-specific analyses, that is, we assessed the association between the respective signature and gene expression in just the patients who are of a cancer type that shows mutations of the respective signature (Extended Data Fig. 13c, left heat map). We then correlated the P values of these cancer-type-specific analyses with the P values of the analysis across all patients and calculated the Pearson correlation coefficients (Supplementary Fig. 24a–e). We show that the correlation between cancer-type-specific and whole-cohort P values is dependent on the sample size of the respective analysis (r² = 0.671) (Supplementary Fig. 1f).

We further performed PCA on the signatures across both, patients (PCA on signature-specific SNVs per patient) and genes (PCA on adjusted P values of signature-gene expression associations) (Extended Data Fig. 11a, b).

To assess significance of the functional annotation of SNVs by mutational signatures, we also associated gene expression with the total number of SNVs and correlated the P values (−log₁₀(P)) of the associations with the respective signature-specific P values. The absolute Pearson correlation coefficients remain below 0.1 (Supplementary Table 19f).

To establish causality of signature–gene expression associations, we included the germline eQTL into the analysis using linear mixed models; 197 of our 1,176 signature-associated genes were also germline eGenes. These 197 associations involved 26 of the 28 mutational signatures. We associated the lead variants of these eGenes with the rank-standardized signature SNVs across 2,507 patients. We used the subset of the 2,818 WGS patients for which mutational signature profiles and all known covariates were available. We accounted for the same fixed covariates as in the mutational signature–gene expression association studies and, in addition, for kinship as a random effect (see ‘Covariates’).

We then performed proportional colocalization analysis with Bayesian model averaging using the R package coloc⁹³ to test whether gene expression and mutational signatures share common causal genetic variants in a given gene region. A proportional colocalization analysis tests the null hypothesis of colocalization by assuming that two phenotypes that share causal variants will have proportional regression coefficients for either phenotype with any variant selection in the vicinity of the causal variant. We applied the Bayesian model averaging approach, with each tested model consisting of a selection of two variants. The P values are then averaged over all models to generate posterior predictive P values⁹³. We filtered variants so that no pair of variants showed r² > 0.95 and each variant’s marginal posterior probability of inclusion with one of the phenotypes was greater than 0.01. The nominal P values of rejecting the null hypothesis of colocalization are listed in Supplementary Table 19e.

We then performed mediation analysis^94,95 to assess directionality of the effect between germline eQTL, gene expression and mutational signature. First, causal mediation analysis was applied to each of the triples of eQTL lead variant, gene and mutational signature using a structural equation model from the R package lavaan⁹⁶. Then, we used the R package mediation⁹⁷ to assess significance of mediation and estimate the proportion of mediated effect by non-parametric bootstrap confidence intervals (1,000 simulations).

ASE analysis: assembling phased germline and somatic variants

To understand the precise effect of somatic variations in their genomic context and for subsequent allele-specific analyses, both germline and somatic variants were phased. For assembling phased germline genotypes, we used the Sanger 1000G callset⁶, and applied IMPUTE2⁹⁸ for phasing of heterozygous germline variants. The IMPUTE2 output was corrected using results from the Battenberg CN calling algorithm⁹⁹ to ascertain that no haplotype switches occur within regions of consecutive copy-number gain. The resulting phased germline genotypes were arranged such that haplotype 1 always corresponded to the amplified alleles in regions with SCNAs (major allele). In cases in which both co-occur on the same NGS read (approximately 10 million variants, 20% of all SNVs), we phased individual somatic variants to the nearest germline heterozygous site. For downstream analyses, we considered only SNVs that were phased by at least three reads to the respective germline variant (approximately 6 million out of 10 million SNVs).

All phased SNVs were aggregated into functional categories based on their genomic regions defined by gene annotations (upstream, downstream, promoter, 5′ UTR, intron, synonymous, missense, stop gain and 3′ UTR) and mapped to the nearest gene within a cis window of 100 kb using the Variant Effect Predictor (VEP) tool¹⁰⁰. Promoter variants were defined as 1-kb upstream of the TSS. We included flanking regions by using the VEP ‘UpDownDistance’ plugin with a maximum range parameter of 100 kb. We divided the upstream and downstream variant categories into disjoint categories using 10-kb windows from 10 to 100 kb. We integrated ‘splice donor’ and ‘splice acceptor’ variants into the general ‘splice region’ variant category and mapped ‘stop retained’ variants to the ‘synonymous’ variant category. We averaged transcript-level annotations to gene-level annotations to retrieve the expected functional effect of a variant for a given gene. We analysed the relationship between SNV variant allele frequency and SCNAs at the same locus to determine whether variants occurred before (‘early’) or after (‘late’) the corresponding SCNA (PCAWG-11). We computed a weighted cis-mutational burden per category by estimating the cancer cell fraction of each SNV and aggregating SNVs to a total localized burden weighted by their respective cancer cell fraction.

ASE read counts

The positional information of the heterozygous germline variants was used together with the RNA-seq BAM files as input to the GATK ASEReadCounter¹⁰¹ algorithms for counting ASE reads. We considered reads with a minimum mapping quality of 20 and a minimum base quality of 10. Only heterozygous variants with a minimum coverage of eight RNA-seq reads were considered for all further analyses.

The raw ASE read counts were post-processed as follows: (1) ASE sites were converted to BED files and aligned against the ENCODE 50-mer mappability track (wgEncodeCrgMapabilityAlign50mer.bigWig) to extract mappability scores for all sites. All sites with mappability scores unequal to 1 were removed. (2) All sites with allelic read counts less or equal to 1 were removed to prevent genotyping error to influence ASE quantification. (3) All sex chromosomes were dropped for further analysis. (4) We estimated sequencing error per patient as the sum of non-reference and non-alternative bases over the total number of bases. We assessed statistical mono-allelicity through a binomial test using the estimated sequencing error probabilities, corrected using the Benjamini–Hochberg step down procedure. All sites that appeared to be statistically mono-allelic were removed. (5) For each ASE site, copy-number states were retrieved from the Sanger copy-number consensus callset (PCAWG-11). Purity estimates for each patients were retrieved from the accompanying purity tables.

To aggregate site-level ASE to a gene-level readout and to allow for estimation of effect directionality, we used the phased germline genotypes. Gene mapping was performed against ENSEMBL release 75 using the pyEnsembl Python library. We retrieved all genes at each ASE site and summed up the read counts on the respective haplotypes to gene-level haplotype-specific read counts. We further averaged haplotype-specific copy-number states to a mean haplotype-specific copy-number state per gene and computed the gene-level copy-number ratio as the major over total ratio of those averages. To allow for a robust assessment of gene-level ASE, we considered only genes with at least 15 reads total, yielding 4,379,378 gene–patient pairs of 1,120 patients and 17,009 unique genes across 12,441,502 accessible sites in total. Every remaining gene was tested for AEI using a binomial test against an expected read ratio of 0.5 to derive nominal P values, and a binomial test against the expected copy-number ratio modified by tumour purity to derive copy-number-corrected P values. Nominal and copy-number-corrected P values were adjusted separately for multiple testing using the Benjamini–Hochberg procedure. Significant AEI was called at FDR ≤ 5%. We further annotated each gene with the number of ASE sites used for aggregation. For all downstream analyses, we considered only genes annotated as protein coding (ENSEMBL biotype = ‘protein_coding’).

Generalized linear models

Across all 4,379,378 gene–patient pairs, we trained multivariate linear models using (i) logistic regression against a binary indicator of AEI absence or presence in a gene, or (ii) standard linear regression against the phased ASE ratio of a gene to assess the directionality of the regulatory change. For (i), haplotype-specific mutations were summed up to a total burden per category, whereas for (ii) we used the difference in burden between the haplotypes 1 and 2. The consistency of the phasing map between somatic variants and ASE sites ensured that model coefficients kept their directionality independent of the arbitrary labelling of haplotypes as 1 or 2. The full set of considered factors is as follows: (1) copy-number ratio at the gene locus (0.5 ≤ x ≤ 1); (2) sample purity (0 < x < 1); (3) natural logarithm of total gene length (x > 0); (4) natural logarithm of the length of the canonical transcript (x > 0); (5) heterozygosity of the lead eQTL variant (x = 0 if homozygous, x = 1 if not homozygous); (6) all mutational burden categories as determined by VEP annotations (upstream in 10-kb windows, downstream in 10-kb windows, promoter, 5′ UTR, intron, synonymous, missense, stop gain and 3′ UTR; x ≥ 0 for logistic model, x ∈ ℝ for directed model).

To compare global effects and different contributions of SCNA, germline eQTL, coding and non-coding SNVs, a simplified logistic model was trained after accumulating all coding and non-coding variants to separate categories and reporting standardised effect sizes (Fig. 1e).

Cancer gene enrichment

Cancer gene enrichment was conducted on the COSMIC census⁵³ using Fisher’s exact test and gene set enrichment analysis as previously described¹⁰². For enrichment, the average score of a gene was computed across the cohort and only genes with at least five replicates in the cohort were kept, yielding a total of 16,078 genes.

Chromosomal distribution of ASE

We calculated the recurrence of ASE genes in each tumour type. To examine the chromosomal distribution of ASE genes, we calculated the average recurrence of all genes for every 200-gene window with a 10-gene step, and then subtracted the average ASE occurrence in each tumour type to obtain the peaks of ASE surplus across all chromosomes. The recurrence of copy-number genes was calculated in an analogous manner.

Estimation of alternative promoter activity

We estimated promoter activities using RNA-seq data and Gencode (release 19) annotations for 70,937 promoters in 20,738 genes. We grouped transcripts with overlapping first exons under the assumption that they are regulated by the same promoter¹⁰³. TSSs that are located within internal exons, or which overlap with splice acceptor sites, were removed from this analysis as these promoters are difficult to estimate from RNA-seq data²⁸. Promoter activity can be estimated using exon usage²⁹, spliced reads²⁸ or isoform-based estimates³⁰. Here we used an isoform-based approach to quantify promoter activity. We quantified the expression of each transcript from the RNA-seq data using Kallisto⁶⁶ and calculated the sum of expression of the transcripts initiated at each promoter to obtain an estimate of promoter activity. To obtain the relative activity for each promoter, we normalized each promoter’s activity by the overall gene’s expression. We divided the promoters of each gene into three categories based on their average pan-cancer promoter activity. The promoters with <1 FPKM average activity are called inactive promoters, and the most active promoter of each gene is called the major promoter. The remaining active promoters of the gene are called minor promoters.

The association between promoter activities and promoter mutation burden was estimated using the same framework as the somatic eQTL analysis. We examined associations for the promoters of expressed multi-promoter genes with a burden frequency ≥ 1% in the cohort (at least 12 patients in the full cohort). The weighted burden of the region 1-kb upstream of the TSS—that is, the sum of variant allele frequencies of the SNVs for each gene—was used as the genotype for the promoters of the respective genes. We used linear models to study the associations between the recurrent somatic burden and the promoter activity (both for the relative activity and the log₂-transformed absolute activity). Similar to the somatic eQTL analysis, the known covariates and the 35 hidden peer factors were provided as cofactors to the linear models. We adjusted the P values using Benjamini–Hochberg correction method and looked for associations with FDR ≤ 5%.

Identification of alternative splicing

We used the alignments based on the STAR pipeline to collect and quantify alternative splicing events with SplAdder⁷⁰. The software has been run with its default parameters with confidence level 3. We generated individual splicing graphs for each RNA-seq sample for both tumour samples as well as matched healthy samples (when available). All graphs were then integrated into a merged graph to comprehensively reflect all splice junctions observed in all samples together. On the basis of this combined graph, SplAdder was used to extract alternative splicing events of the following types: alternative 3′ splice site, alternative 5′ splice site, cassette exon, intron retention, mutually exclusive exons, coordinated exon skip (see supplementary figure 3 in ref. ⁷⁰). Each identified event was then quantified in all samples by counting split alignments for each splice junction in any previously identified event and the average read coverage of each exonic segment involved in the event was determined. We then computed a PSI value for each event that was then used for further analysis. We further generated different subsets of events, filtered at different levels of confidence, in which confidence is defined by the SplAdder confidence level (generally 2), the number of aligned reads supporting each event, the number of samples that were found to support the event by SplAdder, and the number of samples that passed the minimum aligned read threshold.

Enrichment of outlier splicing associated with splice sites and branchpoint motifs

We assessed the significance of mutational enrichment for 5′ and 3′ splice sites, and branch-point^104,105 intronic regions using a permutation-based approach. Impactful mutations were defined as mutations overlapping exons and introns involved in cassette exon events, in which the PSI-derived z-score was ≥ 3 or ≤ −3. For each intronic site, we compared the frequency of observed impactful mutations against frequencies of randomly sampled intronic regions (number of iterations = 1,000). For exonic sites, the null distribution was established from randomly sampled exonic sites. Randomly sampled sites were within a 100-bp window around the 5′ and 3′ splice site. For branch-point regions, sampled sites were within a 50-bp window around the branch-point sequence. The P value was computed as the number of randomly sampled frequencies greater or equal to the observed frequency.

SAVNet analysis for identifying rare SAVs

The SAVNet approach³⁵ was designed for identifying somatic variants associated with local aberrant splicing alterations from matched genome and transcriptome sequencing data. It uses permutations to calculate an FDR and by restricting to two classes of relationships between somatic mutations and splicing alterations to focus: (1) splice site disruption, in which exon skipping, alternative 5′ or 3′ splice site, or intron retention is associated with a mutation in a splice site motif; and (2) splice site creation, in which alternative 5′ or 3′ splice sites are associated with mutations that create a novel splice motif (FDR ≤ 10%) (Extended Data Fig. 17e).

Identification of RNA fusions

Gene fusions between any two genes were identified based on two gene fusions detection pipelines: FusionMap (v.2015-03-31) pipeline¹⁰⁶ and FusionCatcher (v.0.99.6a)/STAR-Fusion (v.0.8.0) pipeline¹⁰⁷. ChimerDB 3.0 was used as a reference of previously reported gene fusions. The database contains 32,949 fusion genes split into three groups: (1) KB: 1,067 fusion genes manually curated based on public resources of fusion genes with experimental evidences; (2) Pub: 2,770 fusion genes obtained from text mining of PubMed abstracts; and (3) Seq: archive with 30,001 fusion gene candidates from deep-sequencing data. This set includes fusions found by re-analysing the RNA-seq data of the TCGA project encompassing 4,569 patients from 23 types of cancer.

In brief, FusionMap was applied to all unaligned reads from the PCAWG aligned TopHat2 RNA-seq BAM files for each aliquot to detect gene fusions. In the FusionCatcher/STAR-Fusion pipeline, for each aliquot with paired-end RNA-seq reads FusionCatcher was applied to the raw reads, with the genome reference. Specifically, for each aliquot with paired-end RNA-seq reads FusionCatcher was applied to the raw reads. The ‘-U True; -V True’ runtime options were used. For each aliquot with single-end RNA-seq reads, STAR-Fusion was applied to the raw reads, with the same reference genome and gene models as FusionCatcher and with default settings. In parallel, FusionMap was applied to all unaligned reads from the PCAWG aligned TopHat2 RNA-seq BAM files for each aliquot to detect gene fusions with the following non-default options values: MinimalHit = 4; OutputFusionReads = True; RnaMode = True; FileFormat = BAM.

To reduce the number of false-positive fusions, the two sets of fusions were filtered to exclude fusions based on the number of supporting junction reads, sequence homology, and occurrence in normal samples (from the GTEx and the PCAWG cohort). To get a high-confident consensus fusion call set from these two pipelines, a fusion to be included in the final set of fusions had to: (i) be detected by both fusion detection tools in at least one sample; and/or (ii) be detected by one of the methods and have a matched structural variant in at least one sample. The consensus WGS-based somatic structural variants (v.1.6) were obtained from the PCAWG repository in https://dcc.icgc.org/releases/PCAWG.

For integration with matched structural variant evidence, a fusion was considered to match a structural variant if the absolute distance between the fusion break points and structural variant break points did not exceed 500 kb (the distance was considered infinite when the chromosomes of the fusion and structural variant break point differ). When there was no evidence for a direct structural variant fusion, the search was expanded to look for composite fusions. In this case, an exhaustive search was performed to look for two structural variants with break points close to the fusion break points and with an effective distance smaller than 250 kb.

Finally, 3,540 fusion events were included as the consensus fusion call set, from these 2,268 were detected by both FusionCatcher/STAR-Fusion and FusionMap (from these, 1,821 had matched structural variant evidence) and 1,112 were detected by only one method and had matched structural variance evidence.

In total, approximately 36% of all detected fusion transcripts were predicted to be in-frame, several UTR-mediated fusion transcripts preserve complete coding sequences of one fusion partner. These include a known fusion TBL1XR1-PIK3CA in a breast tumour and a notable new example CTBP2-CTNNB1 in a gastric tumour.

All fusions are available in Synapse: https://dcc.icgc.org/releases/PCAWG/transcriptome/fusion.

Identification of RNA-editing events

We used an RNA-editing events calling pipeline, which is an improved version of that previously published¹⁰⁸. First, we summarized the base calls of pre-processed aligned RNA reads to the human reference in pileup format. Second, the initially identified editing sites were then filtered by the following quality-aware steps: (1) the depth of candidate editing site, base quality, mapping quality and the frequency of variation were taken into account to do a basic filter: the candidate variant sites should be with base-quality ≥ 20, mapping quality ≥ 50, mapped reads ≥ 4, variant-supporting reads ≥ 3, and mismatch frequencies (variant-supporting-reads/mapped-reads) ≥ 0.1. (2) Statistical tests based on the binomial distribution B(n, p) were used to distinguish true variants from sequencing errors on every mismatch site¹⁰⁹, in which p denotes the background mismatch rate of each transcriptome sequencing, and n denotes sequencing depth on this site. (3) Discard the sites present in combined DNA SNP datasets (dbSNP v.138, 1000 Genome SNP phase 3, human Dutch populations¹¹⁰, and BGI in-house data; combined datasets deposited at: ftp://ftp.genomics.org.cn/pub/icgc-pcawg3). (4) Estimate strand bias and filter out variants with strand bias based on two-tailed Fisher’s exact test. (5) Estimate and filter out variants with position bias, such as sites only found at the 3′ end or at 5′ end of a read. (6) Discard the variation site in simple repeat region or homopolymer region or <5 bp from splicing site. (7) To reduce false positives introduced by misalignment of reads to highly similar regions of the reference genome, we performed a realignment filtering. Specifically, we extracted variant-supporting reads on candidate variant sites and realign them against a combination reference (hg19 genome plus Ensembl transcript reference v.75) by bwa0.5.9-r16. We retain a candidate variant site if at least 90% of its variant-supporting reads are realigned to this site. Finally, all high confident RNA-editing sites were annotated by ANNOVAR¹¹¹. (8) To remove the possibility of an RNA-editing variant being a somatic variant, the variant sites are positionally filtered against PCAWG WGS somatic variant calls (9). The final two steps of filtering are designed to enrich the number of functional RNA editing sites. First, we keep only events that occur more than two times in at least one cancer type. Second, we keep only events that occur in exonic regions with a predicted function of missense, nonsense or stop-loss. The final step of filtering within exonic regions with a specific predicted function induces the largest difference in observed frequencies of RNA-editing events between our analysis and the published one¹⁰⁸. A comparative depiction of the frequencies of RNA-editing events identified in our analysis (Supplementary Table 24) and the previously published analysis¹⁰⁸ is seen in Supplementary Fig. 23.

Gene-centric table creation

To perform joint analysis across RNA and DNA alterations, each alteration type was condensed into a binary gene-centric format. Because alterations occur at many different scales (nucleotide, exonic, gene or transcript), to make them comparable we projected each alteration type onto the gene body. We summarized each alteration type by its presence or absence within a single gene, yielding a binary value per type for each gene-sample pair.

The events we included in this analysis were: RNA editing, non-synonymous variants, expression, splicing alterations, copy-number alterations, fusions and alternative promoters. Each alteration type was summarized differently owing to their inherent differences.

RNA-editing events and non-synonymous variants can occur several times within a single gene body, so these events were denoted as 1 if they occurred at least once within a gene–sample pair.

For copy number, to obtain a single numerical value per gene-sample pair, the copy-number alteration was averaged over the gene body. Because we do not have matched normal samples against which to compare, we instead consider outlying events within each histotype as significant. Thus, a value of 1 was given to average copy-number alterations larger than 6 or smaller than 1.

Similar to non-synonymous variants, multiple splice events can occur within a gene body. The event with the most extreme PSI value within the gene body is selected as the candidate event for the gene. The candidate’s PSI value for a gene is compared over all samples within a histotype and it is set to 1 (that is, significant) only if it the absolute value of its z-score is larger than 6 and the standard deviation is larger than 0.01 within that histotype.

Similar to expression outliers, we calculate a z-score using the log-transformed upper-quartile normalized FPKM values with a pseudo-count of 1. All genes within a histotype with a standard deviation larger than zero and an absolute value larger than three were identified as an outlier. Alternative promoter outliers were calculated based on relative promoter activity within each cancer type. To binarize the promoter activity, a z-score cut-off of two over the relative expression distribution within each cancer type was used.

For ASE outliers, only genes with significant allelic imbalance (FDR ≤ 5% and allelic imbalance > 0.2, binomial test) were denoted as 1. All ASE events that were identified were further filtered to keep only genes that have not been identified as imprinted²⁶.

In addition to the z-score-filtering mentioned above, we further filtered non-synonymous SNVs, RNA-editing events and splicing events such that they either induce a frameshift or the alternative region contains an HGMD variant¹¹² of the category ‘damaging’.

It must be noted that in many cases, the z-score calculated is not from a Gaussian distribution, so some events may be missed or falsely included. Through our choice of very stringent z-score thresholds and functional filters, we hope that spurious outlier events are minimized.

Pathway analysis

For our pathway analysis, we used the TCGA pathway definitions to examine genes and pathways that have several alterations at both the DNA and RNA level¹¹³.

Co-occurrence analysis

The co-occurrence analysis was also performed on the aforementioned binarized gene-centric table, but only including variants, expression outliers, alternative promoters, alternative splicing and fusions. SCNA and ASE are excluded owing to a large number of anticipated co-occurrence. In this analysis, we required at least one gene of a given alteration pair to be a COSMIC gene. For each alteration pair, based on the number of donors with both alterations, one alteration only and neither alterations in a set of cancer samples, we performed Fisher’s exact test to determine whether the alteration pair was independent of each other. Such tests were followed by Benjamini–Hochberg multiple testing correction to obtain the FDR (or q values). To rule out the potential false-positive association caused by tissue-specific alterations, we performed the same analysis for each of the tumour types with at least 50 patients, and retained only those alteration pairs that were significantly associated in both the pan-cancer analysis and in at least one specific cancer indication. Among the significantly associated alteration pairs, the co-occurred pairs were those with odds ratio greater than 1. Pathway enrichment and visualization^21,114 were conducted using the R package ReactomePA⁷⁹. The circos plots were generated using the R package circlize¹¹⁵. The splicing related genes were derived from the genes annotated as ‘REACTOME_MRNA_SPLICING’ or ‘REACTOME_MRNA_SPLICING_MINOR_PATHWAY’ in the Molecular Signatures Database (MSigDB)¹¹⁶.

Identifying genes with heterogeneous mechanisms of alterations in cis

Genes with multiple heterogeneous mechanisms of RNA alteration were identified from associations of cis variants with gene expression, ASE, fusions and splicing. For gene expression, genes associated with somatic eQTL with FDR < 5% were selected. For ASE, the top 5% of genes ranked by the predicted contribution of somatic variants on ASE. For fusions, all RNA fusions with structural variant support were selected. For splicing, genes having somatic mutations within 10 bp of an annotated splice site or 3 bp of a branch point and associated splicing were selected. These associated splicing events also had to have a |z-score| greater than or equal to 3 and the difference of percent spliced in the outlier event was greater than or equal to 10%.

Recurrence analysis

The recurrence analysis was performed on the binarized gene-centric table for all nine alteration types. The recurrence analysis was performed in three main steps: (1) Aggregate within each alteration type across all samples. This results in a sum for each gene-alteration pair. (2) Convert the counts to ranks within each alteration. The smallest rank goes to the most frequently altered genes. Ranks are split evenly across ties. (3) To generate a single score for each gene, the second smallest rank across alterations is used as the score. To identify a score cut-off value for significantly altered genes, a null distribution was generated through permutation. The permutations were performed over the samples within each gene-alteration pair, this was done over all genes and samples 1,000 times, concatenating together all observations, results in 16.8 million permuted scores. P < 0.05 as derived from the null distribution was defined as significant, resulting in a score greater than or equal to 774 considered as significant.

WExT¹¹⁷ was used to test the significance of mutually exclusivity of RNA and DNA alterations. As further evidence that CDK12 alterations may have a functional affect, we find evidence of the previously detected link⁵⁵ between a large tandem duplicator phenotype (here defined as more than 10 tandem duplications of size greater than 100 kb) and CDK12 somatic eQTL mutation (7 out of 18 somatic eQTL carriers are also among the 215 large tandem duplicator cases, P = 0.032, hypergeometric test).

Statistical tests

All common statistical tests are two-sided unless otherwise specified. No statistical methods were used to predetermine sample size. The experiments were not randomized and investigators were not blinded to allocation during experiments and outcome assessment.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this paper.