Cancer 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 decades19,20. ‘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 surveillance21. 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.

This 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.

Initial studies using massively parallel sequencing demonstrated the feasibility of identifying every somatic point mutation, copy-number change and structural variant (SV) in a given cancer1,2,3. 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 types22.

The pan-cancer analysis of whole genomes

The 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 Atlas23,24,25, scientific working groups concentrated their efforts on analyses best-informed by whole-genome sequencing data.

We 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).

To 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).

The 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 computing26,27 to distribute alignment and variant calling across 13 data centres on 3 continents (Supplementary Table 3). Core pipelines were packaged into Docker containers28 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).

Benchmarking of genetic variant calls

To 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 sequencing29. 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.

Fig. 1: Validation of variant-calling pipelines in PCAWG.
figure 1

a, Scatter plot of estimated sensitivity and precision for somatic SNVs across individual algorithms assessed in the validation exercise across n = 63 PCAWG samples. Core algorithms included in the final PCAWG call set are shown in blue. b, Sensitivity and precision estimates across individual algorithms for somatic indels. c, Accuracy (precision, sensitivity and F1 score, defined as 2 × sensitivity × precision/(sensitivity + precision)) of somatic SNV calls across variant allele fractions (VAFs) for the core algorithms. The accuracy of two methods of combining variant calls (two-plus, which was used in the final dataset, and logistic regression) is also shown. d, Accuracy of indel calls across variant allele fractions.

Next, 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).

Analysis of PCAWG data

The 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 paper4,5,6,7,8,9,10,11,12,13,14,15,16,17,18 (Extended Data Table 3).

Pan-cancer burden of somatic mutations

Across 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 mutation30, and a single carcinogen can cause a range of DNA lesions31.

Panorama of driver mutations in cancer

We 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 drivers32. 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.

We 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 knowledge33, 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.

This 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 approaches32.

Fig. 2: Panorama of driver mutations in PCAWG.
figure 2

a, Top, putative driver mutations in PCAWG, represented as a circos plot. Each sector represents a tumour in the cohort. From the periphery to the centre of the plot the concentric rings represent: (1) the total number of driver alterations; (2) the presence of whole-genome (WG) duplication; (3) the tumour type; (4) the number of driver CNAs; (5) the number of driver genomic rearrangements; (6) driver coding point mutations; (7) driver non-coding point mutations; and (8) pathogenic germline variants. Bottom, snapshots of the panorama of driver mutations. The horizontal bar plot (left) represents the proportion of patients with different types of drivers. The dot plot (right) represents the mean number of each type of driver mutation across tumours with at least one event (the square dot) and the standard deviation (grey whiskers), based on n = 2,583 patients. b, Genomic elements targeted by different types of mutations in the cohort altered in more than 65 tumours. Both germline and somatic variants are included. Left, the heat map shows the recurrence of alterations across cancer types. The colour indicates the proportion of mutated tumours and the number indicates the absolute count of mutated tumours. Right, the proportion of each type of alteration that affects each genomic element. c, Tumour-suppressor genes with biallelic inactivation in 10 or more patients. The values included under the gene labels represent the proportions of patients who have biallelic mutations in the gene out of all patients with a somatic mutation in that gene. GR, genomic rearrangement; SCNA, somatic copy-number alteration; SGR, somatic genome rearrangement; TSG, tumour suppressor gene; UTR, untranslated region.

To address the frequency of non-coding driver point mutations, we combined promoters and enhancers that are known targets of non-coding drivers34,35,36,37 with those newly discovered in PCAWG data; this is reported in a companion paper4. 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 mutations4.

Across 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 shown38. Across tumour types, there are differences in which classes of mutation affect a given genomic element (Fig. 2b).

We 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 genes39, DNA-damage response genes40 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).

PCAWG tumours with no apparent drivers

Although 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.

Technical 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 contamination41, 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 sequencing42.

Fig. 3: Analysis of patients with no detected driver mutations.
figure 3

a, Individual estimates of the percentage of tumour-in-normal contamination across patients with no driver mutations in PCAWG (n = 181). No data were available for myelodysplastic syndromes and acute myeloid leukaemia. Points represent estimates for individual patients, and the coloured areas are estimated density distributions (violin plots). Abbreviations of the tumour types are defined in Extended Data Table 1. b, Average detection sensitivity by tumour type for tumours without known drivers (n = 181). Each dot represents a given sample and is the average sensitivity of detecting clonal substitutions across the genome, taking into account purity and ploidy. Coloured areas are estimated density distributions, shown for cohorts with at least five cases. c, Detection sensitivity for TERT promoter hotspots in tumour types in which TERT is frequently mutated. Coloured areas are estimated density distributions. d, Significant copy-number losses identified by two-sided hypothesis testing using GISTIC2.0, corrected for multiple-hypothesis testing. Numbers in parentheses indicate the number of genes in significant regions when analysing medulloblastomas without known drivers (n = 42). Significant regions with known cancer-associated genes are labelled with the representative cancer-associated gene. e, Aneuploidy in chromophobe renal cell carcinomas and pancreatic neuroendocrine tumours without known drivers. Patients are ordered on the y axis by tumour type and then by presence of whole-genome duplication (bottom) or not (top).

Finally, 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 clones43, recurrent driver mutations in these clones can enter panels of normals.

With 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 genes44, and our results suggest that SETD2 loss of function is an additional driver that affects chromatin regulators in this subgroup.

Two 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 reported45,46 (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.

Even 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.

Patterns of clustered mutations and SVs

Some 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 rearrangements47,48 (Extended Data Fig. 5a); (2) kataegis, a focal hypermutation process that leads to locally clustered nucleotide substitutions, biased towards a single DNA strand49,50,51 (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 fragments52,53,54,55 (Extended Data Fig. 5c). We characterized the PCAWG genomes for these three processes (Fig. 4).

Fig. 4: Patterns of clustered mutational processes in PCAWG.
figure 4

a, Kataegis. Top, prevalence of different types of kataegis and their association with SVs (≤1 kb from the focus). Bottom, the distribution of the number of foci of kataegis per sample. Chromoplexy. Prevalence of chromoplexy across cancer types, subdivided into balanced translocations and more complex events. Chromothripsis. Top, frequency of chromothripsis across cancer types. Bottom, for each cancer type a column is shown, in which each row is a chromothripsis region represented by five coloured rectangles relating to its categorization. b, Circos rainfall plot showing the distances between consecutive kataegis events across PCAWG compared with their genomic position. Lymphoid tumours (khaki, B cell non-Hodgkin’s lymphoma; orange, chronic lymphocytic leukaemia) have hypermutation hot spots (≥3 foci with distance ≤1 kb; pale red zone), many of which are near known cancer-associated genes (red annotations) and have associated SVs (≤10 kb from the focus; shown as arcs in the centre). c, Circos rainfall plot as in b that shows the distance versus the position of consecutive chromoplexy and reciprocal translocation footprints across PCAWG. Lymphoid, prostate and thyroid cancers exhibit recurrent events (≥2 footprints with distance ≤10 kb; pale red zone) that are likely to be driver SVs and are annotated with nearby genes and associated SVs, which are shown as bold and thin arcs for chromoplexy and reciprocal translocations, respectively (colours as in a). d, Effect of chromothripsis along the genome and involvement of PCAWG driver genes. Top, number of chromothripsis-induced gains or losses (grey) and amplifications (blue) or deletions (red). Within the identified chromothripsis regions, selected recurrently rearranged (light grey), amplified (blue) and homozygously deleted (magenta)) driver genes are indicated. Bottom, inter breakpoint distance between all subsequent breakpoints within chromothripsis regions across cancer types, coloured by cancer type. Regions with an average inter breakpoint distance <10 kb are highlighted. C[T>N]T, kataegis with a pattern of thymine mutations in a Cp TpT context.

Chromoplexy 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 described47,48, 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)56, and the juxtaposition of the oncogene IGF2BP3 with regulatory elements from highly expressed genes (five cases).

Kataegis 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 activity49,50,51, 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 described57. 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.

Kataegis events were frequently associated with somatic SV breakpoints (Fig. 4a and Supplementary Fig. 6a), as previously described50,51. 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).

Samples 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).

We identified chromothripsis in 587 samples (22.3%), most frequently among sarcoma, glioblastoma, lung squamous cell carcinoma, melanoma and breast adenocarcinoma18. Chromothripsis increased with whole-genome duplications in most cancer types (Extended Data Fig. 7a), as previously shown in medulloblastoma58. The most recurrently associated driver was TP5352 (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 disease59 (Extended Data Fig. 7c).

Chromothripsis 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).

Chromothripsis 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 amplification60 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).

Timing clustered mutations in evolution

An 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 tumour49,61. 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)7.

Chromothripsis 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 melanomas62. We identified 89 separate chromothripsis events that affected 66 melanomas (61%); 47 out of 89 events affected genes known to be recurrently altered in melanoma63 (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.

Fig. 5: Timing of clustered events in PCAWG.
figure 5

a, Extent and timing of chromothripsis, kataegis and chromoplexy across PCAWG. Top, stacked bar charts illustrate co-occurrence of chromothripsis, kataegis and chromoplexy in the samples. Middle, relative odds of clustered events being clonal or subclonal are shown with bootstrapped 95% confidence intervals. Point estimates are highlighted when they do not overlap odds of 1:1. Bottom, relative odds of the events being early or late clonal are shown as above. Sample sizes (number of patients) are shown across the top. b, Three representative patients with acral melanoma and chromothripsis-induced amplification that simultaneously affects TERT and CCND1. The black points (top) represent sequence coverage from individual genomic bins, with SVs shown as coloured arcs (translocation in black, deletion in purple, duplication in brown, tail-to-tail inversion in cyan and head-to-head inversion in green). Bottom, the variant allele fractions of somatic point mutations.

In 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).

Notably, 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).

Germline effects on somatic mutations

We 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 dinucleotides5 (signature 1) and activity of the APOBEC3 family of cytidine deaminases64 (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 level65 (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 genomes66,67. 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 cells68,69.

Fig. 6: Germline determinants of the somatic mutation landscape.
figure 6

a, Association between common (MAF > 5%) germline variants and somatic APOBEC3B-like mutagenesis in individuals of European ancestry (n = 1,201). Two-sided hypothesis testing was performed with PLINK v.1.9. To mitigate multiple-hypothesis testing, the significance threshold was set to genome-wide significance (P < 5 × 10−8). b, Templated insertion SVs in a BRCA1-associated prostate cancer. Left, chromosome bands (1); SVs ≤ 10 megabases (Mb) (2); 1 kb read depth corrected to copy number 0–6 (3); inter- and intra chromosomal SVs > 10 Mb (4). Right, a complex somatic SV composed of a 2.2-kb tandem duplication on chromosome 2 together with a 232-base-pair (bp) inverted templated insertion SV that is derived from chromosome 5 and inserted inbetween the tandem duplication (bottom). Consensus sequence alignment of locally assembled Oxford Nanopore Technologies long sequencing reads to chromosomes 2 and 5 of the human reference genome (top). Breakpoints are circled and marked as 1 (beginning of tandem duplication), 2 (end of tandem duplication) or 3 (inverted templated insertion). For each breakpoint, the middle panel shows Illumina short reads at SV breakpoints. c, Association between rare germline PTVs (MAF < 0.5%) and somatic CpG mutagenesis (approximately with signature 1) in individuals of European ancestry (n = 1,201). Genes highlighted in blue or red were associated with lower or higher somatic mutation rates. Two-sided hypothesis testing was performed using linear-regression models with sex, age at diagnosis and cancer project as variables. To mitigate multiple-hypothesis testing, the significance threshold was set to exome-wide significance (P < 2.5 × 10−6). The black line represents the identity line that would be followed if the observed P values followed the null expectation; the shaded area shows the 95% confidence intervals. d, Catalogue of polymorphic germline L1 source elements that are active in cancer. The chromosomal map shows germline source L1 elements as volcano symbols. Each volcano is colour-coded according to the type of source L1 activity. The contribution of each source locus (expressed as a percentage) to the total number of transductions identified in PCAWG tumours is represented as a gradient of volcano size, with top contributing elements exhibiting larger sizes.

Second, 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 cancer30,70. In PCAWG data, this pattern also extends to other tumour types, including adenocarcinomas of the prostate and pancreas6, 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’6. 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.

Third, 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 sites71, a functionality that would be consistent with a CpG mutational signature in cancer.

Fourth, we assessed long interspersed nuclear elements (LINE-1; L1 hereafter) that mediate somatic retrotransposition events72,73,74. 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 transductions10 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).

Replicative immortality

One of the hallmarks of cancer is the ability of cancer to evade cellular senescence21. Normal somatic cells typically have finite cell division potential; telomere attrition is one mechanism to limit numbers of mitoses75. 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 binding34,37; hitching TERT to highly active regulatory elements elsewhere in the genome46,76; insertions of viral enhancers upstream of the gene77,78; 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 genes79.

As reported in a companion paper13, 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 maintenance13, 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.

On the basis of these 12 features, tumour samples formed 4 distinct sub-clusters (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).

Fig. 7: Telomere sequence patterns across PCAWG.
figure 7

a, Scatter plot of the clusters of telomere patterns identified across PCAWG using t-distributed stochastic neighbour embedding (t-SNE), based on n = 2,518 tumour samples and their matched normal samples. Axes have arbitrary dimensions such that samples with similar telomere profiles are clustered together and samples with dissimilar telomere profiles are far apart with high probability. b, Distribution of the four tumour-specific clusters of telomere patterns in selected tumour types from PCAWG. c, Distribution of relevant driver mutations associated with alternative lengthening of telomere and normal telomere maintenance across the four clusters. d, Distribution of telomere maintenance abnormalities across tumour types with more than 40 patients in PCAWG. Samples were classified as tumour clusters 1–3 if they fell into a relevant cluster without mutations in TERT, ATRX or DAXX and had no ALT phenotype. TMM, telomere maintenance mechanisms.

Somatic 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).

There 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 inactivation80, and studies using mouse models have shown that knockout of Rb-family proteins causes elongated telomeres81. 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.

Tumour 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 tissues82 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.

Conclusions and future perspectives

The 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.

The 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 characterization83. 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 (, 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.

Extending 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.



We 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.

We 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.

Demographically, 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).

We consolidated histopathology descriptions of the tumour samples, using the ICD-0-3 tumour site controlled vocabulary89. 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.

Uniform processing and somatic variant calling

To 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 algorithm90 to align each tumour and normal sample to human reference build hs37d5 (as used in the 1000 Genomes Project91). 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’92,93,94,95, ‘EMBL/DKFZ’96,97 and ‘Broad’98,99,100,101 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 algorithms102,103 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).

Somatic retrotransposition events, including Alu and LINE-1 insertions72, L1-mediated transductions73 and pseudogene formation104, were called using a dedicated pipeline73. We removed these retrotransposition events from the somatic SV call-set. Mitochondrial DNA mutations were called using a published algorithm105. 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 editing8.

Integration, phasing and validation of germline variant call-sets

Calls 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 approach91,106. The uniform germline data-processing workflow comprised variant identification using six different variant-calling algorithms96,107,108 and was orchestrated using the Butler workflow system109.

We performed call-set benchmarking, merging, variant genotyping and statistical haplotype-block phasing91 (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 Project91 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%.

Core alignment and variant calling by cloud computing

The 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 architecture26 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.

To generate reproducible variant calling across the 13 data centres, we built the core pipelines into Docker containers28, 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).

Validation, benchmarking and merging of somatic variant calls

To 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 described29. 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.

The 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.

Next, 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 regression110,111 to integrate the calls. The merged SV set includes all calls made by two or more of the four primary SV-calling algorithms96,100,112,113. 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 multi tier system was developed for consensus copy-number calls (Supplementary Methods 2.4.3, and described in detail elsewhere7).

Overall, 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.