Pan-cancer analysis of whole genomes

1Cancer is the second most-frequent cause of death worldwide, killing more than 8 million people every year; the incidence of cancer is expected to increase by more than 50% over the coming decades,. ‘Cancer’ is a catch-all term used to denote a set of diseases characterized by autonomous expansion and spread of a somatic clone. To achieve this behaviour, the cancer clone must co-opt multiple cellular pathways that enable it to disregard the normal constraints on cell growth, modify the local microenvironment to favour its own proliferation, invade through tissue barriers, spread to other organs and evade immune surveillance. No single cellular program directs these behaviours. Rather, there is a large pool of potential pathogenic abnormalities from which individual cancers draw their own combinations: the commonalities of macroscopic features across tumours belie a vastly heterogeneous landscape of cellular abnormalities.

2This heterogeneity arises from the stochastic nature of Darwinian evolution. There are three preconditions for Darwinian evolution: characteristics must vary within a population; this variation must be heritable from parent to offspring; and there must be competition for survival within the population. In the context of somatic cells, heritable variation arises from mutations acquired stochastically throughout life, notwithstanding additional contributions from germline and epigenetic variation. A subset of these mutations alter the cellular phenotype, and a small subset of those variants confer an advantage on clones during the competition to escape the tight physiological controls wired into somatic cells. Mutations that provide a selective advantage to the clone are termed driver mutations, as opposed to selectively neutral passenger mutations.

3Initial studies using massively parallel sequencing demonstrated the feasibility of identifying every somatic point mutation, copy-number change and structural variant (SV) in a given cancer–. In 2008, recognizing the opportunity that this advance in technology provided, the global cancer genomics community established the ICGC with the goal of systematically documenting the somatic mutations that drive common tumour types.

4The expansion of whole-genome sequencing studies from individual ICGC and TCGA working groups presented the opportunity to undertake a meta-analysis of genomic features across tumour types. To achieve this, the PCAWG Consortium was established. A Technical Working Group implemented the informatics analyses by aggregating the raw sequencing data from different working groups that studied individual tumour types, aligning the sequences to the human genome and delivering a set of high-quality somatic mutation calls for downstream analysis (Extended Data Fig. 1). Given the recent meta-analysis of exome data from the TCGA Pan-Cancer Atlas–, scientific working groups concentrated their efforts on analyses best-informed by whole-genome sequencing data.

5We collected genome data from 2,834 donors (Extended Data Table 1), of which 176 were excluded after quality assurance. A further 75 had minor issues that could affect some of the analyses (grey-listed donors) and 2,583 had data of optimal quality (white-listed donors) (Supplementary Table 1). Across the 2,658 white- and grey-listed donors, whole-genome sequencing data were available from 2,605 primary tumours and 173 metastases or local recurrences. Mean read coverage was 39× for normal samples, whereas tumours had a bimodal coverage distribution with modes at 38× and 60× (Supplementary Fig. 1). RNA-sequencing data were available for 1,222 donors. The final cohort comprised 1,469 men (55%) and 1,189 women (45%), with a mean age of 56 years (range, 1–90 years) across 38 tumour types (Extended Data Table 1 and Supplementary Table 1).

6To identify somatic mutations, we analysed all 6,835 samples using a uniform set of algorithms for alignment, variant calling and quality control (Extended Data Fig. 1, Supplementary Fig. 2 and Supplementary Methods 2). We used three established pipelines to call somatic single-nucleotide variations (SNVs), small insertions and deletions (indels), copy-number alterations (CNAs) and SVs. Somatic retrotransposition events, mitochondrial DNA mutations and telomere lengths were also called by bespoke algorithms. RNA-sequencing data were uniformly processed to call transcriptomic alterations. Germline variants identified by the three separate pipelines included single-nucleotide polymorphisms, indels, SVs and mobile-element insertions (Supplementary Table 2).

7The requirement to uniformly realign and call variants on approximately 5,800 whole genomes presented considerable computational challenges, and raised ethical issues owing to the use of data from different jurisdictions (Extended Data Table 2). We used cloud computing, to distribute alignment and variant calling across 13 data centres on 3 continents (Supplementary Table 3). Core pipelines were packaged into Docker containers as reproducible, stand-alone packages, which we have made available for download. Data repositories for raw and derived datasets, together with portals for data visualization and exploration, have also been created (Box 1 and Supplementary Table 4).

8The PCAWG landing page (http://docs.icgc.org/pcawg) provides links to several data resources for interactive online browsing, analysis and download of PCAWG data and results (Supplementary Table 4).

9Direct download of PCAWG data

10Aligned PCAWG read data in BAM format are also available at the European Genome Phenome Archive (EGA; https://www.ebi.ac.uk/ega/search/site/pcawg under accession number EGAS00001001692). In addition, all open-tier PCAWG genomics data, as well as reference datasets used for analysis, can be downloaded from the ICGC Data Portal at http://docs.icgc.org/pcawg/data/. Controlled-tier genomic data, including SNVs and indels that originated from TCGA projects (in VCF format) and aligned reads (in BAM format) can be downloaded using the Score (https://www.overture.bio/) software package, which has accelerated and secure file transfer, as well as BAM slicing facilities to selectively download defined regions of genomic alignments.

11PCAWG computational pipelines

12The core alignment, somatic variant-calling, quality-control and variant consensus-generation pipelines used by PCAWG have each been packaged into portable cross-platform images using the Dockstore system and released under an Open Source licence that enables unrestricted use and redistribution. All PCAWG Dockstore images are available to the public at https://dockstore.org/organizations/PCAWG/collections/PCAWG.

13ICGC Data Portal

14The ICGC Data Portal (https://dcc.icgc.org) serves as the main entry point for accessing PCAWG datasets with a single uniform web interface and a high-performance data-download client. This uniform interface provides users with easy access to the myriad of PCAWG sequencing data and variant calls that reside in many repositories and compute clouds worldwide. Streaming technology provides users with high-level visualizations in real time of BAM and VCF files stored remotely on the Cancer Genome Collaboratory.

15UCSC Xena

16UCSC Xena (https://pcawg.xenahubs.net) visualizes all PCAWG primary results, including copy-number, gene-expression, gene-fusion and promoter-usage alterations, simple somatic mutations, large somatic structural variations, mutational signatures and phenotypic data. These open-access data are available through a public Xena hub, and consensus simple somatic mutations can be loaded to the local computer of a user via a private Xena hub. Kaplan–Meier plots, histograms, box plots, scatter plots and transcript-specific views offer additional visualization options and statistical analyses.

17The Expression Atlas

18The Expression Atlas (https://www.ebi.ac.uk/gxa/home) contains RNA-sequencing and expression microarray data for querying gene expression across tissues, cell types, developmental stages and/or experimental conditions. Two different views of the data are provided: summarized expression levels for each tumour type and gene expression at the level of individual samples, including reference-gene expression datasets for matching normal tissues.

19PCAWG Scout

20PCAWG Scout (http://pcawgscout.bsc.es/) provides a framework for -omics workflow and website templating to generate on-demand, in-depth analyses of the PCAWG data that are openly available to the whole research community. Views of protected data are available that still safeguard sensitive data. Through the PCAWG Scout web interface, users can access an array of reports and visualizations that leverage on-demand bioinformatic computing infrastructure to produce results in real time, allowing users to discover trends as well as form and test hypotheses.

21Chromothripsis Explorer

22Chromothripsis Explorer (http://compbio.med.harvard.edu/chromothripsis/) is a portal that allows structural variation in the PCAWG dataset to be explored on an individual patient basis through the use of circos plots. Patterns of chromothripsis can also be explored in aggregated formats.

23To benchmark mutation calling, we ran the 3 core pipelines, together with 10 additional pipelines, on 63 representative tumour–normal genome pairs (Supplementary Note 1). For 50 of these cases, we performed validation by hybridization of tumour and matched normal DNA to a custom bait set with deep sequencing. The 3 core somatic variant-calling pipelines had individual estimates of sensitivity of 80–90% to detect a true somatic SNV called by any of the 13 pipelines; more than 95% of SNV calls made by each of the core pipelines were genuine somatic variants (Fig. 1a). For indels—a more-challenging class of variants to identify with short-read sequencing—the 3 core algorithms had individual sensitivity estimates in the range of 40–50%, with precision of 70–95% (Fig. 1b). For individual SV algorithms, we estimated precision to be in the range 80–95% for samples in the 63-sample pilot dataset.

24Next, we defined a strategy to merge results from the three pipelines into one final call-set to be used for downstream scientific analyses (Methods and Supplementary Note 2). Sensitivity and precision of consensus somatic variant calls were 95% (90% confidence interval, 88–98%) and 95% (90% confidence interval, 71–99%), respectively, for SNVs (Extended Data Fig. 2). For somatic indels, sensitivity and precision were 60% (34–72%) and 91% (73–96%), respectively (Extended Data Fig. 2). Regarding somatic SVs, we estimate the sensitivity of merged calls to be 90% for true calls generated by any one pipeline; precision was estimated as 97.5%. The improvement in calling accuracy from combining different pipelines was most noticeable in variants with low variant allele fractions, which probably originate from tumour subclones (Fig. 1c, d). Germline variant calls, phased using a haplotype-reference panel, displayed a precision of more than 99% and a sensitivity of 92–98% (Supplementary Note 2).

25The uniformly generated, high-quality set of variant calls across more than 2,500 donors provided the springboard for a series of scientific working groups to explore the biology of cancer. A comprehensive suite of companion papers that describe the analyses and discoveries across these thematic areas is copublished with this paper– (Extended Data Table 3).

26Across the 2,583 white-listed PCAWG donors, we called 43,778,859 somatic SNVs, 410,123 somatic multinucleotide variants, 2,418,247 somatic indels, 288,416 somatic SVs, 19,166 somatic retrotransposition events and 8,185 de novo mitochondrial DNA mutations (Supplementary Table 1). There was considerable heterogeneity in the burden of somatic mutations across patients and tumour types, with a broad correlation in mutation burden among different classes of somatic variation (Extended Data Fig. 3). Analysed at a per-patient level, this correlation held, even when considering tumours with similar purity and ploidy (Supplementary Fig. 3). Why such correlation should apply on a pan-cancer basis is unclear. It is likely that age has some role, as we observe a correlation between most classes of somatic mutation and age at diagnosis (around 190 SNVs per year, P = 0.02; about 22 indels per year, P = 5 × 10−5; 1.5 SVs per year, P < 2 × 10−16; linear regression with likelihood ratio tests; Supplementary Fig. 4). Other factors are also likely to contribute to the correlations among classes of somatic mutation, as there is evidence that some DNA-repair defects can cause multiple types of somatic mutation, and a single carcinogen can cause a range of DNA lesions.

27We extracted the subset of somatic mutations in PCAWG tumours that have high confidence to be driver events on the basis of current knowledge. One challenge to pinpointing the specific driver mutations in an individual tumour is that not all point mutations in recurrently mutated cancer-associated genes are drivers. For genomic elements significantly mutated in PCAWG data, we developed a ‘rank-and-cut’ approach to identify the probable drivers (Supplementary Methods 8.1). This approach works by ranking the observed mutations in a given genomic element based on recurrence, estimated functional consequence and expected pattern of drivers in that element. We then estimate the excess burden of somatic mutations in that genomic element above that expected for the background mutation rate, and cut the ranked mutations at this level. Mutations in each element with the highest driver ranking were then assigned as probable drivers; those below the threshold will probably have arisen through chance and were assigned as probable passengers. Improvements to features that are used to rank the mutations and the methods used to measure them will contribute to further development of the rank-and-cut approach.

28We also needed to account for the fact that some bona fide cancer genomic elements were not rediscovered in PCAWG data because of low statistical power. We therefore added previously known cancer-associated genes to the discovery set, creating a ‘compendium of mutational driver elements’ (Supplementary Methods 8.2). Then, using stringent rules to nominate driver point mutations that affect these genomic elements on the basis of prior knowledge, we separated probable driver from passenger point mutations. To cover all classes of variant, we also created a compendium of known driver SVs, using analogous rules to identify which somatic CNAs and SVs are most likely to act as drivers in each tumour. For probable pathogenic germline variants, we identified all truncating germline point mutations and SVs that affect high-penetrance germline cancer-associated genes.

29This analysis defined a set of mutations that we could confidently assert, based on current knowledge, drove tumorigenesis in the more than 2,500 tumours of PCAWG. We found that 91% of tumours had at least one identified driver mutation, with an average of 4.6 drivers per tumour identified, showing extensive variation across cancer types (Fig. 2a). For coding point mutations, the average was 2.6 drivers per tumour, similar to numbers estimated in known cancer-associated genes in tumours in the TCGA using analogous approaches.

30To address the frequency of non-coding driver point mutations, we combined promoters and enhancers that are known targets of non-coding drivers– with those newly discovered in PCAWG data; this is reported in a companion paper. Using this approach, only 13% (785 out of 5,913) of driver point mutations were non-coding in PCAWG. Nonetheless, 25% of PCAWG tumours bear at least one putative non-coding driver point mutation, and one third (237 out of 785) affected the TERT promoter (9% of PCAWG tumours). Overall, non-coding driver point mutations are less frequent than coding driver mutations. With the exception of the TERT promoter, individual enhancers and promoters are only infrequent targets of driver mutations.

31Across tumour types, SVs and point mutations have different relative contributions to tumorigenesis. Driver SVs are more prevalent in breast adenocarcinomas (6.4 ± 3.7 SVs (mean ± s.d.) compared with 2.2 ± 1.3 point mutations; P < 1 × 10−16, Mann–Whitney U-test) and ovary adenocarcinomas (5.8 ± 2.6 SVs compared with 1.9 ± 1.0 point mutations; P < 1 × 10−16), whereas driver point mutations have a larger contribution in colorectal adenocarcinomas (2.4 ± 1.4 SVs compared with 7.4 ± 7.0 point mutations; P = 4 × 10−10) and mature B cell lymphomas (2.2 ± 1.3 SVs compared with 6 ± 3.8 point mutations; P < 1 × 10−16), as previously shown. Across tumour types, there are differences in which classes of mutation affect a given genomic element (Fig. 2b).

32We confirmed that many driver mutations that affect tumour-suppressor genes are two-hit inactivation events (Fig. 2c). For example, of the 954 tumours in the cohort with driver mutations in TP53, 736 (77%) had both alleles mutated, 96% of which (707 out of 736) combined a somatic point mutation that affected one allele with somatic deletion of the other allele. Overall, 17% of patients had rare germline protein-truncating variants (PTVs) in cancer-predisposition genes, DNA-damage response genes and somatic driver genes. Biallelic inactivation due to somatic alteration on top of a germline PTV was observed in 4.5% of patients overall, with 81% of these affecting known cancer-predisposition genes (such as BRCA1, BRCA2 and ATM).

33Although more than 90% of PCAWG cases had identified drivers, we found none in 181 tumours (Extended Data Fig. 4a). Reasons for missing drivers have not yet been systematically evaluated in a pan-cancer cohort, and could arise from either technical or biological causes.

34Technical explanations could include poor-quality samples, inadequate sequencing or failures in the bioinformatic algorithms used. We assessed the quality of the samples and found that 4 of the 181 cases with no known drivers had more than 5% tumour DNA contamination in their matched normal sample (Fig. 3a). Using an algorithm designed to correct for this contamination, we identified previously missed mutations in genes relevant to the respective cancer types. Similarly, if the fraction of tumour cells in the cancer sample is low through stromal contamination, the detection of driver mutations can be impaired. Most tumours with no known drivers had an average power to detect mutations close to 100%; however, a few had power in the 70–90% range (Fig. 3b and Extended Data Fig. 4b). Even in adequately sequenced genomes, lack of read depth at specific driver loci can impair mutation detection. For example, only around 50% of PCAWG tumours had sufficient coverage to call a mutation (≥90% power) at the two TERT promoter hotspots, probably because the high GC content of this region causes biased coverage (Fig. 3c). In fact, 6 hepatocellular carcinomas and 2 biliary cholangiocarcinomas among the 181 cases with no known drivers actually did contain TERT mutations, which were discovered after deep targeted sequencing.

35Finally, technical reasons for missing driver mutations include failures in the bioinformatic algorithms. This affected 35 myeloproliferative neoplasms in PCAWG, in which the JAK2V617F driver mutation should have been called. Our somatic variant-calling algorithms rely on ‘panels of normals’, typically from blood samples, to remove recurrent sequencing artefacts. As 2–5% of healthy individuals carry occult haematopoietic clones, recurrent driver mutations in these clones can enter panels of normals.

36With regard to biological causes, tumours may be driven by mutations in cancer-associated genes that are not yet described for that tumour type. Using driver discovery algorithms on tumours with no known drivers, no individual genes reached significance for point mutations. However, we identified a recurrent CNA that spanned SETD2 in medulloblastomas that lacked known drivers (Fig. 3d), indicating that restricting hypothesis testing to missing-driver cases can improve power if undiscovered genes are enriched in such tumours. Inactivation of SETD2 in medulloblastoma significantly decreased gene expression (P = 0.002) (Extended Data Fig. 4c). Notably, SETD2 mutations occurred exclusively in medulloblastoma group-4 tumours (P < 1 × 10−4). Group-4 medulloblastomas are known for frequent mutations in other chromatin-modifying genes, and our results suggest that SETD2 loss of function is an additional driver that affects chromatin regulators in this subgroup.

37Two tumour types had a surprisingly high fraction of patients without identified driver mutations: chromophobe renal cell carcinoma (44%; 19 out of 43) and pancreatic neuroendocrine cancers (22%; 18 out of 81) (Extended Data Fig. 4a). A notable feature of the missing-driver cases in both tumour types was a remarkably consistent profile of chromosomal aneuploidy—patterns that have previously been reported, (Fig. 3e). The absence of other identified driver mutations in these patients raises the possibility that certain combinations of whole-chromosome gains and losses may be sufficient to initiate a cancer in the absence of more-targeted driver events such as point mutations or fusion genes of focal CNAs.

38Even after accounting for technical issues and novel drivers, 5.3% of PCAWG tumours still had no identifiable driver events. In a research setting, in which we are interested in drawing conclusions about populations of patients, the consequences of technical issues that affect occasional samples will be mitigated by sample size. In a clinical setting, in which we are interested in the driver mutations in a specific patient, these issues become substantially more important. Careful and critical appraisal of the whole pipeline—including sample acquisition, genome sequencing, mapping, variant calling and driver annotation, as done here—should be required for laboratories that offer clinical sequencing of cancer genomes.

39Some somatic mutational processes generate multiple mutations in a single catastrophic event, typically clustered in genomic space, leading to substantial reconfiguration of the genome. Three such processes have previously been described: (1) chromoplexy, in which repair of co-occurring double-stranded DNA breaks—typically on different chromosomes—results in shuffled chains of rearrangements, (Extended Data Fig. 5a); (2) kataegis, a focal hypermutation process that leads to locally clustered nucleotide substitutions, biased towards a single DNA strand– (Extended Data Fig. 5b); and (3) chromothripsis, in which tens to hundreds of DNA breaks occur simultaneously, clustered on one or a few chromosomes, with near-random stitching together of the resulting fragments– (Extended Data Fig. 5c). We characterized the PCAWG genomes for these three processes (Fig. 4).

40Chromoplexy events and reciprocal translocations were identified in 467 (17.8%) samples (Fig. 4a, c). Chromoplexy was prominent in prostate adenocarcinoma and lymphoid malignancies, as previously described,, and—unexpectedly—thyroid adenocarcinoma. Different genomic loci were recurrently rearranged by chromoplexy across the three tumour types, mediated by positive selection for particular fusion genes or enhancer-hijacking events. Of 13 fusion genes or enhancer hijacking events in 48 thyroid adenocarcinomas, at least 4 (31%) were caused by chromoplexy, with a further 4 (31%) part of complexes that contained chromoplexy footprints (Extended Data Fig. 5a). These events generated fusion genes that involved RET (two cases) and NTRK3 (one case), and the juxtaposition of the oncogene IGF2BP3 with regulatory elements from highly expressed genes (five cases).

41Kataegis events were found in 60.5% of all cancers, with particularly high abundance in lung squamous cell carcinoma, bladder cancer, acral melanoma and sarcomas (Fig. 4a, b). Typically, kataegis comprises C > N mutations in a TpC context, which are probably caused by APOBEC activity–, although a T > N conversion in a TpT or CpT process (the affected T is highlighted in bold) attributed to error-prone polymerases has recently been described. The APOBEC signature accounted for 81.7% of kataegis events and correlated positively with APOBEC3B expression levels, somatic SV burden and age at diagnosis (Supplementary Fig. 5). Furthermore, 5.7% of kataegis events involved the T > N error-prone polymerase signature and 2.3% of events, most notably in sarcomas, showed cytidine deamination in an alternative GpC or CpC context.

42Kataegis events were frequently associated with somatic SV breakpoints (Fig. 4a and Supplementary Fig. 6a), as previously described,. Deletions and complex rearrangements were most-strongly associated with kataegis, whereas tandem duplications and other simple SV classes were only infrequently associated (Supplementary Fig. 6b). Kataegis inducing predominantly T > N mutations in CpTpT context was enriched near deletions, specifically those in the 10–25-kilobase (kb) range (Supplementary Fig. 6c).

43Samples with extreme kataegis burden (more than 30 foci) comprise four types of focal hypermutation (Extended Data Fig. 6): (1) off-target somatic hypermutation and foci of T > N at CpTpT, found in B cell non-Hodgkin lymphoma and oesophageal adenocarcinomas, respectively; (2) APOBEC kataegis associated with complex rearrangements, notably found in sarcoma and melanoma; (3) rearrangement-independent APOBEC kataegis on the lagging strand and in early-replicating regions, mainly found in bladder and head and neck cancer; and (4) a mix of the last two types. Kataegis only occasionally led to driver mutations (Supplementary Table 5).

44We identified chromothripsis in 587 samples (22.3%), most frequently among sarcoma, glioblastoma, lung squamous cell carcinoma, melanoma and breast adenocarcinoma. Chromothripsis increased with whole-genome duplications in most cancer types (Extended Data Fig. 7a), as previously shown in medulloblastoma. The most recurrently associated driver was TP53 (pan-cancer odds ratio = 3.22; pan-cancer P = 8.3 × 10−35; q < 0.05 in breast lobular (odds ratio = 13), colorectal (odds ratio = 25), prostate (odds ratio = 2.6) and hepatocellular (odds ratio = 3.9) cancers; Fisher–Boschloo tests). In two cancer types (osteosarcoma and B cell lymphoma), women had a higher incidence of chromothripsis than men (Extended Data Fig. 7b). In prostate cancer, we observed a higher incidence of chromothripsis in patients with late-onset than early-onset disease (Extended Data Fig. 7c).

45Chromothripsis regions coincided with 3.6% of all identified drivers in PCAWG and around 7% of copy-number drivers (Fig. 4d). These proportions are considerably enriched compared to expectation if selection were not acting on these events (Extended Data Fig. 7d). The majority of coinciding driver events were amplifications (58%), followed by homozygous deletions (34%) and SVs within genes or promoter regions (8%). We frequently observed a ≥2-fold increase or decrease in expression of amplified or deleted drivers, respectively, when these loci were part of a chromothripsis event, compared with samples without chromothripsis (Extended Data Fig. 7e).

46Chromothripsis manifested in diverse patterns and frequencies across tumour types, which we categorized on the basis of five characteristics (Fig. 4a). In liposarcoma, for example, chromothripsis events often involved multiple chromosomes, with universal MDM2 amplification and co-amplification of TERT in 4 of 19 cases (Fig. 4d). By contrast, in glioblastoma the events tended to affect a smaller region on a single chromosome that was distant from the telomere, resulting in focal amplification of EGFR and MDM2 and loss of CDKN2A. Acral melanomas frequently exhibited CCND1 amplification, and lung squamous cell carcinomas SOX2 amplifications. In both cases, these drivers were more-frequently altered by chromothripsis compared with other drivers in the same cancer type and to other cancer types for the same driver (Fig. 4d and Extended Data Fig. 7f). Finally, in chromophobe renal cell carcinoma, chromothripsis nearly always affected chromosome 5 (Supplementary Fig. 7): these samples had breakpoints immediately adjacent to TERT, increasing TERT expression by 80-fold on average compared with samples without rearrangements (P = 0.0004; Mann–Whitney U-test).

47An unanswered question for clustered mutational processes is whether they occur early or late in cancer evolution. To address this, we used molecular clocks to define broad epochs in the life history of each tumour,. One transition point is between clonal and subclonal mutations: clonal mutations occurred before, and subclonal mutations after, the emergence of the most-recent common ancestor. In regions with copy-number gains, molecular time can be further divided according to whether mutations preceded the copy-number gain (and were themselves duplicated) or occurred after the gain (and therefore present on only one chromosomal copy).

48Chromothripsis tended to have greater relative odds of being clonal than subclonal, suggesting that it occurs early in cancer evolution, especially in liposarcomas, prostate adenocarcinoma and squamous cell lung cancer (Fig. 5a). As previously reported, chromothripsis was especially common in melanomas. We identified 89 separate chromothripsis events that affected 66 melanomas (61%); 47 out of 89 events affected genes known to be recurrently altered in melanoma (Supplementary Table 6). Involvement of a region on chromosome 11 that includes the cell-cycle regulator CCND1 occurred in 21 cases (10 out of 86 cutaneous, and 11 out of 21 acral or mucosal melanomas), typically combining chromothripsis with amplification (19 out of 21 cases) (Extended Data Fig. 8). Co-involvement of other cancer-associated genes in the same chromothripsis event was also frequent, including TERT (five cases), CDKN2A (three cases), TP53 (two cases) and MYC (two cases) (Fig. 5b). In these co-amplifications, a chromothripsis event involving multiple chromosomes initiated the process, creating a derivative chromosome in which hundreds of fragments were stitched together in a near-random order (Fig. 5b). This derivative then rearranged further, leading to massive co-amplification of the multiple target oncogenes together with regions located nearby on the derivative chromosome.

49In these cases of amplified chromothripsis, we can use the inferred number of copies bearing each SNV to time the amplification process. SNVs present on the chromosome before amplification will themselves be amplified and are therefore reported in a high fraction of sequence reads (Fig. 5b and Extended Data Fig. 8). By contrast, late SNVs that occur after the amplification has concluded will be present on only one chromosome copy out of many, and thus have a low variant allele fraction. Regions of CCND1 amplification had few—sometimes zero—mutations at high variant allele fraction in acral melanomas, in contrast to later CCND1 amplifications in cutaneous melanomas, in which hundreds to thousands of mutations typically predated amplification (Fig. 5b and Extended Data Fig. 9a, b). Thus, both chromothripsis and the subsequent amplification generally occurred very early during the evolution of acral melanoma. By comparison, in lung squamous cell carcinomas, similar patterns of chromothripsis followed by SOX2 amplification are characterized by many amplified SNVs, suggesting a later event in the evolution of these cancers (Extended Data Fig. 9c).

50Notably, in cancer types in which the mutational load was sufficiently high, we could detect a larger-than-expected number of SNVs on an intermediate number of DNA copies, suggesting that they appeared during the amplification process (Supplementary Fig. 8).

51We integrated the set of 88 million germline genetic variant calls with somatic mutations in PCAWG, to study germline determinants of somatic mutation rates and patterns. First, we performed a genome-wide association study of somatic mutational processes with common germline variants (minor allele frequency (MAF) > 5%) in individuals with inferred European ancestry. An independent genome-wide association study was performed in East Asian individuals from Asian cancer genome projects. We focused on two prevalent endogenous mutational processes: spontaneous deamination of 5-methylcytosine at CpG dinucleotides (signature 1) and activity of the APOBEC3 family of cytidine deaminases (signatures 2 and 13). No locus reached genome-wide significance (P < 5 × 10−8) for signature 1 (Extended Data Fig. 10a, b). However, a locus at 22q13.1 predicted an APOBEC3B-like mutagenesis at the pan-cancer level (Fig. 6a). The strongest signal at 22q13.1 was driven by rs12628403, and the minor (non-reference) allele was protective against APOBEC3B-like mutagenesis (β = −0.43, P = 5.6 × 10−9, MAF = 8.2%, n = 1,201 donors) (Extended Data Fig. 10c). This variant tags a common, approximately 30-kb germline SV that deletes the APOBEC3B coding sequence and fuses the APOBEC3B 3′ untranslated region with the coding sequence of APOBEC3A. The deletion is known to increase breast cancer risk and APOBEC mutagenesis in breast cancer genomes,. Here, we found that rs12628403 reduces APOBEC3B-like mutagenesis specifically in cancer types with low levels of APOBEC mutagenesis (βlow = −0.50, Plow = 1 × 10−8; βhigh = 0.17, Phigh = 0.2), and increases APOBEC3A-like mutagenesis in cancer types with high levels of APOBEC mutagenesis (βhigh = 0.44, Phigh = 8 × 10−4; βlow = −0.21, Plow = 0.02). Moreover, we identified a second, novel locus at 22q13.1 that was associated with APOBEC3B-like mutagenesis across cancer types (rs2142833, β = 0.23, P = 1.3 × 10−8). We independently validated the association between both loci and APOBEC3B-like mutagenesis using East Asian individuals from Asian cancer genome projects (βrs12628403 = 0.57, Prs12628403 = 4.2 × 10−12; βrs2142833 = 0.58, Prs2142833 = 8 × 10−15) (Extended Data Fig. 10d). Notably, in a conditional analysis that accounted for rs12628403, we found that rs2142833 and rs12628403 are inherited independently in Europeans (r2<0.1), and rs2142833 remained significantly associated with APOBEC3B-like mutagenesis in Europeans (βEUR = 0.17, PEUR = 3 × 10−5) and East Asians (βASN = 0.25, PASN = 2 × 10−3) (Extended Data Fig. 10e, f). Analysis of donor-matched expression data further suggests that rs2142833 is a cis-expression quantitative trait locus (eQTL) for APOBEC3B at the pan-cancer level (β = 0.19, P = 2 × 10−6) (Extended Data Fig. 10g, h), consistent with cis-eQTL studies in normal cells,.

52Second, we performed a rare-variant association study (MAF <0.5%) to investigate the relationship between germline PTVs and somatic DNA rearrangements in individuals with European ancestry (Extended Data Fig. 11a–c). Germline BRCA2 and BRCA1 PTVs were associated with an increased burden of small (less than 10 kb) somatic SV deletions (P = 1 × 10−8) and tandem duplications (P = 6 × 10−13), respectively, corroborating recent studies in breast and ovarian cancer,. In PCAWG data, this pattern also extends to other tumour types, including adenocarcinomas of the prostate and pancreas, typically in the setting of biallelic inactivation. In addition, tumours with high levels of small SV tandem duplications frequently exhibited a novel and distinct class of SVs termed ‘cycles of templated insertions’. These complex SV events consist of DNA templates that are copied from across the genome, joined into one contiguous sequence and inserted into a single derivative chromosome. We found a significant association between germline BRCA1 PTVs and templated insertions at the pan-cancer level (P = 4 × 10−15) (Extended Data Fig. 11d, e). Whole-genome long-read sequencing data generated for a BRCA1-deficient PCAWG prostate tumour verified the small tandem-duplication and templated-insertion SV phenotypes (Fig. 6b). Almost all (20 out of 21) of BRCA1-associated tumours with a templated-insertion SV phenotype displayed combined germline and somatic hits in the gene. Together, these data suggest that biallelic inactivation of BRCA1 is a driver of the templated-insertion SV phenotype.

53Third, rare-variant association analysis revealed that patients with germline MBD4 PTVs had increased rates of somatic C > T mutation rates at CpG dinucleotides (P < 2.5 × 10−6) (Fig. 6c and Extended Data Fig. 11f, g). Analysis of previously published whole-exome sequencing samples from the TCGA (n = 8,134) replicated the association between germline MBD4 PTVs and increased somatic CpG mutagenesis at the pan-cancer level (P = 7.1 × 10−4) (Extended Data Fig. 11h). Moreover, gene-expression profiling revealed a significant but modest correlation between MBD4 expression and somatic CpG mutation rates between and within PCAWG tumour types (Extended Data Fig. 11i–k). MBD4 encodes a DNA-repair gene that removes thymidines from T:G mismatches within methylated CpG sites, a functionality that would be consistent with a CpG mutational signature in cancer.

54Fourth, we assessed long interspersed nuclear elements (LINE-1; L1 hereafter) that mediate somatic retrotransposition events–. We identified 114 germline source L1 elements capable of active somatic retrotransposition, including 70 that represent insertions with respect to the human reference genome (Fig. 6d and Supplementary Table 7), and 53 that were tagged by single-nucleotide polymorphisms in strong linkage disequilibrium (Supplementary Table 7). Only 16 germline L1 elements accounted for 67% (2,440 out of 3,669) of all L1-mediated transductions detected in the PCAWG dataset (Extended Data Fig. 12a). These 16 hot-L1 elements followed two broad patterns of somatic activity (8 of each), which we term Strombolian and Plinian in analogy to patterns of volcanic activity. Strombolian L1s are frequently active in cancer, but mediate only small-to-modest eruptions of somatic L1 activity in cancer samples (Extended Data Fig. 12b). By contrast, Plinian L1s are more rarely seen, but display aggressive somatic activity. Whereas Strombolian elements are typically relatively common (MAF > 2%) and sometimes even fixed in the human population, all Plinian elements were infrequent (MAF ≤ 2%) in PCAWG donors (Extended Data Fig. 12c; P = 0.001, Mann–Whitney U-test). This dichotomous pattern of activity and allele frequency may reflect differences in age and selective pressures, with Plinian elements potentially inserted into the human germline more recently. PCAWG donors bear on average between 50 and 60 L1 source elements and between 5 and 7 elements with hot activity (Extended Data Fig. 12d), but only 38% (1,075 out of 2,814) of PCAWG donors carried ≥1 Plinian element. Some L1 germline source loci caused somatic loss of tumour-suppressor genes (Extended Data Fig. 12e). Many are restricted to individual continental population ancestries (Extended Data Fig. 12f–j).

55One of the hallmarks of cancer is the ability of cancer to evade cellular senescence. Normal somatic cells typically have finite cell division potential; telomere attrition is one mechanism to limit numbers of mitoses. Cancers enlist multiple strategies to achieve replicative immortality. Overexpression of the telomerase gene, TERT, which maintains telomere lengths, is especially prevalent. This can be achieved through point mutations in the promoter that lead to de novo transcription factor binding,; hitching TERT to highly active regulatory elements elsewhere in the genome,; insertions of viral enhancers upstream of the gene,; and increased dosage through chromosomal amplification, as we have seen in melanoma (Fig. 5b). In addition, there is an ‘alternative lengthening of telomeres’ (ALT) pathway, in which telomeres are lengthened through homologous recombination, mediated by loss-of-function mutations in the ATRX and DAXX genes.

56As reported in a companion paper, 16% of tumours in the PCAWG dataset exhibited somatic mutations in at least one of ATRX, DAXX and TERT. TERT alterations were detected in 270 samples, whereas 128 tumours had alterations in ATRX or DAXX, of which 71 were protein-truncating. In the companion paper, which focused on describing patterns of ALT and TERT-mediated telomere maintenance, 12 features of telomeric sequence were measured in the PCAWG cohort. These included counts of nine variants of the core hexameric sequence, the number of ectopic telomere-like insertions within the genome, the number of genomic breakpoints and telomere length as a ratio between tumour and normal. Here we used the 12 features as an overview of telomere integrity across all tumours in the PCAWG dataset.

57On the basis of these 12 features, tumour samples formed 4 distinct subclusters (Fig. 7a and Extended Data Fig. 13a), suggesting that telomere-maintenance mechanisms are more diverse than the well-established TERT and ALT dichotomy. Clusters C1 (47 tumours) and C2 (42 tumours) were enriched for traits of the ALT pathway—having longer telomeres, more genomic breakpoints, more ectopic telomere insertions and variant telomere sequence motifs (Supplementary Fig. 9). C1 and C2 were distinguished from one another by the latter having a considerable increase in the number of TTCGGG and TGAGGG variant motifs among the telomeric hexamers. Thyroid adenocarcinomas were markedly enriched among C3 samples (26 out of 33 C3 samples; P < 10−16); the C1 cluster (ALT subtype 1) was common among sarcomas; and both pancreatic endocrine neoplasms and low-grade gliomas had a high proportion of samples in the C2 cluster (ALT subtype 2) (Fig. 7b). Notably, some of the thyroid adenocarcinomas and pancreatic neuroendocrine tumours that cluster together (cluster C3) had matched normal samples that also cluster together (normal cluster N3) (Extended Data Fig. 13a) and which share common properties. For example, the GTAGGG repeat was overrepresented among samples in this group (Supplementary Fig. 10).

58Somatic driver mutations were also unevenly distributed across the four clusters (Fig. 7c). C1 tumours were enriched for RB1 mutations or SVs (P = 3 × 10−5), as well as frequent SVs that affected ATRX (P = 6 × 10−14), but not DAXX. RB1 and ATRX mutations were largely mutually exclusive (Extended Data Fig. 13b). By contrast, C2 tumours were enriched for somatic point mutations in ATRX and DAXX (P = 6 × 10−5), but not RB1. The enrichment of RB1 mutations in C1 remained significant when only leiomyosarcomas and osteosarcomas were considered, confirming that this enrichment is not merely a consequence of the different distribution of tumour types across clusters. C3 samples had frequent TERT promoter mutations (30%; P = 2 × 10−6).

59There was a marked predominance of RB1 mutations in C1. Nearly a third of the samples in C1 contained an RB1 alteration, which were evenly distributed across truncating SNVs, SVs and shallow deletions (Extended Data Fig. 13c). Previous research has shown that RB1 mutations are associated with long telomeres in the absence of TERT mutations and ATRX inactivation, and studies using mouse models have shown that knockout of Rb-family proteins causes elongated telomeres. The association with the C1 cluster here suggests that RB1 mutations can represent another route to activating the ALT pathway, which has subtly different properties of telomeric sequence compared with the inactivation of DAXX—these fall almost exclusively in cluster C2.

60Tumour types with the highest rates of abnormal telomere maintenance mechanisms often originate in tissues that have low endogenous replicative activity (Fig. 7d). In support of this, we found an inverse correlation between previously estimated rates of stem cell division across tissues and the frequency of telomere maintenance abnormalities (P = 0.01, Poisson regression) (Extended Data Fig. 13d). This suggests that restriction of telomere maintenance is an important tumour-suppression mechanism, particularly in tissues with low steady-state cellular proliferation, in which a clone must overcome this constraint to achieve replicative immortality.

61The resource reported in this paper and its companion papers has yielded insights into the nature and timing of the many mutational processes that shape large- and small-scale somatic variation in the cancer genome; the patterns of selection that act on these variations; the widespread effect of somatic variants on transcription; the complementary roles of the coding and non-coding genome for both germline and somatic mutations; the ubiquity of intratumoral heterogeneity; and the distinctive evolutionary trajectory of each cancer type. Many of these insights can be obtained only from an integrated analysis of all classes of somatic mutation on a whole-genome scale, and would not be accessible with, for example, targeted exome sequencing.

62The promise of precision medicine is to match patients to targeted therapies using genomics. A major barrier to its evidence-based implementation is the daunting heterogeneity of cancer chronicled in these papers, from tumour type to tumour type, from patient to patient, from clone to clone and from cell to cell. Building meaningful clinical predictors from genomic data can be achieved, but will require knowledge banks comprising tens of thousands of patients with comprehensive clinical characterization. As these sample sizes will be too large for any single funding agency, pharmaceutical company or health system, international collaboration and data sharing will be required. The next phase of ICGC, ICGC-ARGO (https:// www.icgc-argo.org/), will bring the cancer genomics community together with healthcare providers, pharmaceutical companies, data science and clinical trials groups to build comprehensive knowledge banks of clinical outcome and treatment data from patients with a wide variety of cancers, matched with detailed molecular profiling.

63Extending the story begun by TCGA, ICGC and other cancer genomics projects, the PCAWG has brought us closer to a comprehensive narrative of the causal biological changes that drive cancer phenotypes. We must now translate this knowledge into sustainable, meaningful clinical treatments.

64We compiled an inventory of matched tumour–normal whole-cancer genomes in the ICGC Data Coordinating Centre. Most samples came from treatment-naive, primary cancers, although a small number of donors had multiple samples of primary, metastatic and/or recurrent tumours. Our inclusion criteria were: (1) matched tumour and normal specimen pair; (2) a minimal set of clinical fields; and (3) characterization of tumour and normal whole genomes using Illumina HiSeq paired-end sequencing reads.

65We collected genome data from 2,834 donors, representing all ICGC and TCGA donors that met these criteria at the time of the final data freeze in autumn 2014 (Extended Data Table 1). After quality assurance (Supplementary Methods 2.5), data from 176 donors were excluded as unusable, 75 had minor issues that could affect some analyses (grey-listed donors) and 2,583 had data of optimal quality (white-listed donors) (Supplementary Table 1). Across the 2,658 white- and grey-listed donors, whole-genome sequences were available from 2,605 primary tumours and 173 metastases or local recurrences. Matching normal samples were obtained from blood (2,064 donors), tissue adjacent to the primary tumour (87 donors) or from distant sites (507 donors). Whole-genome sequencing data were available for tumour and normal DNA for the entire cohort. The mean read coverage was 39× for normal samples, whereas tumours had a bimodal coverage distribution with modes at 38× and 60× (Supplementary Fig. 1). The majority of specimens (65.3%) were sequenced using 101-bp paired-end reads. An additional 28% were sequenced with 100-bp paired-end reads. Of the remaining specimens, 4.7% were sequenced with read lengths longer than 101 bp, and 1.9% with read lengths shorter than 100 bp. The distribution of read lengths by tumour cohort is shown in Supplementary Fig. 11. Median read length for whole-genome sequencing paired-end reads was 101 bp (mean = 106.2, s.d. = 16.7; minimum–maximum = 50–151). RNA-sequencing data were collected and re-analysed centrally for 1,222 donors, including 1,178 primary tumours, 67 metastases or local recurrences and 153 matched normal tissue samples adjacent to the primary tumour.

66Demographically, the cohort included 1,469 men (55%) and 1,189 women (45%), with a mean age of 56 years (range, 1–90 years) (Supplementary Table 1). Using population ancestry-differentiated single nucleotide polymorphisms, the ancestry distribution was heavily weighted towards donors of European descent (77% of total) followed by East Asians (16%), as expected for large contributions from European, North American and Australian projects (Supplementary Table 1).

67We consolidated histopathology descriptions of the tumour samples, using the ICD-0-3 tumour site controlled vocabulary. Overall, the PCAWG dataset comprises 38 distinct tumour types (Extended Data Table 1 and Supplementary Table 1). Although the most common tumour types are included in the dataset, their distribution does not match the relative population incidences, largely owing to differences among contributing ICGC/TCGA groups in the numbers of sequenced samples.

68To generate a consistent set of somatic mutation calls that could be used for cross-tumour analyses, we analysed all 6,835 samples using a uniform set of algorithms for alignment, variant calling and quality control (Extended Data Fig. 1, Supplementary Fig. 2, Supplementary Table 3 and Supplementary Methods 2). We used the BWA-MEM algorithm to align each tumour and normal sample to human reference build hs37d5 (as used in the 1000 Genomes Project). Somatic mutations were identified in the aligned data using three established pipelines, which were run independently on each tumour–normal pair. Each of the three pipelines—labelled ‘Sanger’–, ‘EMBL/DKFZ’, and ‘Broad’– after the computational biology groups that created or assembled them—consisted of multiple software packages for calling somatic SNVs, small indels, CNAs and somatic SVs (with intrachromosomal SVs defined as those >100 bp). Two additional variant algorithms, were included to further improve accuracy across a broad range of clonal and subclonal mutations. We tested different merging strategies using validation data, and choses the optimal method for each variant type to generate a final consensus set of mutation calls (Supplementary Methods S2.4).

69Somatic retrotransposition events, including Alu and LINE-1 insertions, L1-mediated transductions and pseudogene formation, were called using a dedicated pipeline. We removed these retrotransposition events from the somatic SV call-set. Mitochondrial DNA mutations were called using a published algorithm. RNA-sequencing data were uniformly processed to quantify normalized gene-level expression, splicing variation and allele-specific expression, and to identify fusion transcripts, alternative promoter usage and sites of RNA editing.

70Calls of common (≥1% frequency in PCAWG) and rare (<1%) germline variants including single-nucleotide polymorphisms, indels, SVs and mobile-element insertions (MEIs) were generated using a population-scale genetic polymorphism-detection approach,. The uniform germline data-processing workflow comprised variant identification using six different variant-calling algorithms,, and was orchestrated using the Butler workflow system.

71We performed call-set benchmarking, merging, variant genotyping and statistical haplotype-block phasing (Supplementary Methods 3.4). Using this strategy, we identified 80.1 million germline single-nucleotide polymorphisms, 5.9 million germline indels, 1.8 million multi-allelic short (<50 bp) germline variants, as well as germline SVs ≥ 50 bp in size including 29,492 biallelic deletions and 27,254 MEIs (Supplementary Table 2). We statistically phased this germline variant set using haplotypes from the 1000 Genomes Project as a reference panel, yielding an N50-phased block length of 265 kb based on haploid chromosomes from donor-matched tumour genomes. Precision estimates for germline SNVs and indels were >99% for the phased merged call-set, and sensitivity estimates ranged from 92% to 98%.

72The requirement to uniformly realign and call variants on nearly 5,800 whole genomes (tumour plus normal) presented considerable computational challenges, and raised ethical issues owing to the use of data from different jurisdictions (Extended Data Table 2). To process the data, we adopted a cloud-computing architecture in which the alignment and variant calling was spread across 13 data centres on 3 continents, representing a mixture of commercial, infrastructure-as-a-service, academic cloud compute and traditional academic high-performance computer clusters (Supplementary Table 3). Together, the effort used 10 million CPU-core hours.

73To generate reproducible variant calling across the 13 data centres, we built the core pipelines into Docker containers, in which the workflow description, required code and all associated dependencies were packaged together in stand-alone packages. These heavily tested, extensively validated workflows are available for download (Box 1).

74To evaluate the performance of each of the mutation-calling pipelines and determine an integration strategy, we performed a large-scale deep-sequencing validation experiment (Supplementary Notes 1). We selected a pilot set of 63 representative tumour–normal pairs, on which we ran the 3 core pipelines, together with a set of 10 additional somatic variant-calling pipelines contributed by members of the PCAWG SNV Calling Methods Working Group. Sufficient DNA remained for 50 of the 63 cases for validation, which was performed by hybridization of tumour and matched normal DNA to a custom RNA bait set, followed by deep sequencing, as previously described. Although performed using the same sequencing chemistry as the original whole-genome sequencing analyses, the considerably greater depth achieved in the validation experiment enabled accurate assessment of sensitivity and precision of variant calls. Variant calls in repeat-masked regions were not tested, owing to the challenge of designing reliable validation probes in these areas.

75The 3 core pipelines had individual estimates of sensitivity of 80–90% to detect a true somatic SNV called by any of the 13 pipelines; with >95% of SNV calls made by each of the core pipelines being genuine somatic variants (Fig. 1a). For indels—a more-challenging class of variants to identify in short-read sequencing data—the 3 core algorithms had individual sensitivity estimates in the range of 40–50%, with precision 70–95% (Fig. 1b). Validation of SV calls is inherently more difficult, as methods based on PCR or hybridization to RNA baits often fail to isolate DNA that spans the breakpoint. To assess the accuracy of SV calls, we therefore used the property that an SV must either generate a copy-number change or be balanced, whereas artefactual calls will not respect this property. For individual SV-calling algorithms, we estimated precision to be in the range of 80–95% for samples in the 63-sample pilot dataset.

76Next, we examined multiple methods for merging calls made by several algorithms into a single definitive call-set to be used for downstream analysis. The final consensus calls for SNVs were based on a simple approach that required two or more methods to agree on a call. For indels, because methods were less concordant, we used stacked logistic regression, to integrate the calls. The merged SV set includes all calls made by two or more of the four primary SV-calling algorithms,,,. Consensus CNA calls were obtained by joining the outputs of six individual CNA-calling algorithms with SV consensus breakpoints to obtain base-pair resolution CNAs (Supplementary Methods 2.4.3). Consensus purity and ploidy were derived, and a multitier system was developed for consensus copy-number calls (Supplementary Methods 2.4.3, and described in detail elsewhere).

77Overall, the sensitivity and precision of the consensus somatic variant calls were 95% (90% confidence interval, 88–98%) and 95% (90% confidence interval, 71–99%), respectively, for SNVs (Extended Data Fig. 2). For somatic indels, sensitivity and precision were 60% (90% confidence interval, 34–72%) and 91% (90% confidence interval, 73–96%), respectively. Regarding SVs, we estimate the sensitivity of the merging algorithm to be 90% for true calls generated by any one calling pipeline; precision was estimated to be 97.5%. That is, 97.5% of SVs in the merged SV call-set had an associated copy-number change or balanced partner rearrangement. The improvement in calling accuracy from combining different pipelines was most noticeable in variants that had low variant allele fractions, which are likely to originate from subclonal populations of the tumour (Fig. 1c, d). There remains much work to be done to improve indel calling software; we still lack sensitivity for calling even fully clonal complex indels from short-read sequencing data.

78Further information on research design is available in the Nature Research Reporting Summary linked to this paper.

79Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41586-020-1969-6.