Common and rare mutational signatures

The national GEL sequencing endeavor delivers thousands of samples for certain tumor-types (1,009 lung, 1,355 kidney, 2,572 breast, and 1,480 bone/soft tissue cancers), an order of magnitude (or two) greater than previous WGS efforts for some organs. This permits robust detection of signatures that are rare – those occurring in 1% of the tumors or fewer. Furthermore, already-sequenced WGS cohorts such as ~3,000 primary cancers from ICGC and ~3,400 metastatic from HMF, provide a powerful means of validating findings.

We performed mutational signature extractions confined to specific tumor-types using an updated signature extraction method (Fig. 1C, fig. S1, tables S3 to S6, Materials and Methods). Briefly, for each tumor type, we clustered mutational catalogs (counts of SBS in 96-element form or DBS in 78-element form), selecting only samples with recurrent, commonly occurring profiles to perform signature extraction (fig. S1, A to C). Cases with unusual profiles and likely to have rare signatures were excluded in the first extraction. Thus, this yielded a set of highly accurate ‘common signatures’ that are prevalent for that tumor type. Next, by fitting these common signatures into all samples, cases that are likely to have additional patterns not fully explained by common signatures alone would report a high ‘error’ (or discrepancy between true sample catalog and reconstructed catalog) (fig. S1D). Potential additional signatures were then extracted from these samples to obtain a set of ‘rare signatures’ (fig. S1, E to H, Materials and Methods). Accordingly, we obtained a set of common signatures and a set of rare signatures for each tumor-type. In all, for SBS, we identified 135 common signatures and 180 rare signatures in 19 tumor types within the GEL cancer cohort.

To validate these common and rare signatures, we performed signature extractions in independent cohorts of 3,001 ICGC primary WGS cancers (19 tumor-types) and 3,417 metastatic Hartwig WGS samples (18 tumor-types). We identified 135 common signatures in ICGC, 58 rare. In Hartwig, we found 135 common signatures and 114 rare (tables S7 to S10). We performed an agnostic three-way signature comparison in 16 tissue types that were present in all three cohorts (fig. S2). We found that signatures from the same organ in different cohorts were more similar to each other than to those in other tissue type, providing reassuring evidence that mutational signatures in each organ are highly reproducible, have tissue-specificities, and were detectable regardless of sequencing platform or mutation-calling algorithms (fig. S2).

Notably, the number of common signatures in each organ is usually limited (between five and ten for SBS) and is independent of the number of samples analyzed per organ (Fig. 1, D and E, fig. S3, A and B, tables S11 and S12). By contrast, the number of rare signatures varies and is highly correlated with the number of samples analyzed (Fig. 1F and fig. S3C). This illuminates why ubiquitous, organ-specific signatures are detectable even with limited numbers of whole genomes, whereas sporadic, rare signatures are more likely to be detected with increased sample size.

Reference mutational signatures

Biologically, the same mutational processes could underpin signatures extracted from different tumor-types. Thus, we considered all common and rare GEL, ICGC, and Hartwig tissue-specific signatures together, involving 18,640 WGS cancer samples (Fig. 1G, Materials and Methods) and performed a clustering analysis to derive a set of ‘Reference Signatures’. First, we identified clusters of highly similar patterns that we termed ‘distinct patterns’ (tables S13 to S16). Each distinct pattern could be either: i) a true signature, thus observable in independent extractions of diverse organs and cohorts (recurrent pattern); ii) a mix of other signatures (mixed pattern); iii) a pattern seen in only one extraction (singleton pattern) (Fig. 1G, tables S17 and S18). Next, we determined a minimal set of ‘Reference Signatures’, which were classified as quality control (QC) green, amber or red, where green implied high-confidence signatures observed in multiple independent extractions and amber/red signatures were observed only once or were possible mathematical artefacts (Fig. 1G, tables S19 to S22). In all, we identified 82 SBS and 27 DBS high-quality signatures (figs. S4 to S6). Henceforth, we will only discuss these high-quality QC green signatures, although all signatures are available for reference in supplementary material (tables S19 and S20).

Reference Signatures were compared and matched with COSMIC mutational signatures (14), confirming 42 previously described COSMIC SBS signatures and 9 COSMIC DBS signatures (Fig. 2, A and B, fig. S3, D to G, fig. S4, fig. S6A and table S19). We found 40 previously unreported high-confidence SBS signatures and 18 previously unreported DBS signatures in this analysis (fig. S5 and fig. S6B). Respecting prior nomenclature (14), these SBS signatures have been numbered from 95 onwards, and DBS signatures from 12 onwards (table S19). Note that COSMIC and/or Reference Signatures are a simplified means of discussing signatures that are mutually present across tissues. However, they are purely mathematical constructs - an averaged result across different organs - thus organ-specific signatures are more likely to be accurate biological representations of the mutational processes that occur within a tissue. We also provide the numbers of mutations associated with each reference signature per sample (tables S23 and S24), and matrices to map each reference signature to organ-specific signatures (tables S25 and S26), for SBS and DBS respectively.

Open in a separate window

Fig. 2

SBS signatures across 18,640 WGS cancers.

(A) Frequency of SBS signatures in the present study. Orange bars highlight 42 signatures reported in this study and present in COSMIC v3.2. Blue bars highlight 40 previously unreported signatures found in the present analysis. (B) Same information as A with log scale on y-axis. (C) Four previously unreported (SBS107, SBS100, SBS110 and SBS121) and five recently reported (SBS92, SBS93, SBS94,SBS125, and SBS127) common SBS signatures found in many organ systems. (D) Previously unreported common SBS signatures found in single organs only. (E) A few examples of previously unreported rare signatures. TSB, transcriptional strand bias; RSB, replication strand bias. SBS signatures can be browsed here: https://signal.mutationalsignatures.com/explore/study/6?mutationType=1.

Previously unreported mutational signatures

Double (DBS) and triple (TBS) base substitution signatures

We adopted similar principles to identify 39 DBSs, including 27 high-confidence ones (Materials and Methods, table S20 and fig. S6). We performed three additional evaluations. First, we curated dinucleotides for each DBS signature in the GEL dataset to check that they were in cis. Second, for a DBS signature that was correlated with an SBS signature, an in-silico analysis assessing whether the DBS pattern could be expected given the SBS pattern was performed (Materials and Methods). Third, we investigated up to 10 nt of mutational context of relevant dinucleotides for each DBS signature. These assessments were critical in refuting several DBS signatures as being simply due to chance, described below.

Of eleven previously described COSMIC DBS signatures (14), we identified nine and were unable to extract DBS6 or DBS9 (Fig. 3A, fig. S3, F and G, and fig. S6A). Of our 27 high-confidence DBS signatures, 17 had bona fide dinucleotides in cis. We confirmed previously reported signatures and their associated etiologies: DBS1 (UV light), DBS2 (smoking), DBS5 (platinum therapy), and DBS11, associated with APOBEC, here verified as APOBEC-induced given the 10 nt sequence context analysis showing a TpCC preponderance (Fig. 3B). DBS7 was previously reported as associated with MMR defects (14), while we find associations with SBS17 instead (fig. S7A). DBS8, mostly in colorectal cancer, showed dinucleotide variants often preceded by a Cytosine and followed by an Adenine (fig. S7, B and C).

Open in a separate window

Fig. 3

DBS signatures across the cohort.

(A) Frequency of DBS signatures in the present study. (B) Flanking sequence context surrounding mutated dinucleotides of DBS11, which is correlated with APOBEC-related SBS2, to demonstrate a preference for TpCCpN context similar to the TpCpN sequence predilection of APOBECs. (C) Correlation of DBS with SBS exposures across cohorts. Numbers in each column report the number of organs implicated in the correlative analyses. A correlation is computed independently for each organ and the correlations are displayed as a boxplot. Boxplots denote median (horizontal line) and 25th to 75th percentiles (boxes). The lower and upper whiskers extend 1.5×IQR (IQR: inter-quartile range). (D) Examples of previously unreported DBS signatures. (E) Samples with TBS1. The total number of samples and total number of TNVs are too low to perform a formalmutational signature analysis. All DBS signatures identified in the present study can be browsed here: https://signal.mutationalsignatures.com/explore/study/6?mutationType=2.

We confirm DBS5 and DBS18 are associated with prior platinum exposure (18). Mutational context analysis indicates that these are distinct signatures: DBS5 has the tallest peak of CT>AA mutations without preference in flanking sequences, while DBS18 has the tallest peak of CT>AC mutations, where the dinucleotide is always preceded by a Cytosine (fig. S7D). Both signatures have a TG>GT peak most frequently followed by a Guanine (fig. S7, E and F).

DBS13 and DBS20 were low-burden signatures that appear to correlate with each other and SBS8 (Fig. 3, C and D). DBS16 was associated with SBS10d (Fig. 3C), a hypermutator signature recently reported as due to polymerase δ (POLD) dysfunction (19). DBS22 is not associated with very prominent peaks (highest probabilities only 7%). However, it appears to be correlated with SBS9 and is only seen in lymphoid cancers (Fig. 3, C and D). DBS26 is similar to DBS7 and correlates with SBS17 in esophageal and stomach cancers (Fig. 3, C and D). DBS30 was observed in one lymphoid cancer sample and may be related to treatment (fig. S6B).

DBS25 is characterized by an excess of TT>AA that, on inspection, reveals a triple base substitution signature (TBS). Exploring all triple base possibilities, we obtain an 864-channel profile that systematically reports an excess of TTT>AAA followed by TTT>GAA, TTT>CAA, and minor contributions of TTG>AAA, and TTC>AAA. We propose that this is called TBS1 (Fig. 3E and table S27). However, the number of mutations and implicated samples are too low to perform a formal mutational signature analysis.

Our curation steps uncovered several DBS signatures, including previously reported ones, that comprise adjacent substitutions that are not in cis and are simply the mathematical outcome of an associated SBS hypermutator (fig. S8). For example, DBS3 and DBS10 were similar and correlated with polymerase ε (POLE)-attributed SBS10a (fig. S8, A to C). In silico analysis showed that a DBS pattern that recapitulates DBS3/DBS10 could be reproduced from hypermutated samples of SBS10a (fig. S8D). The alleged double substitutions were not, in fact, in cis. Similarly, DBS12 (associated with SBS105), DBS14 (associated with SBS14), DBS29 (associated with SBS20), and DBS37 (associated with SBS26) could all be generated mathematically from their associated SBS signatures (fig. S8, E to H), indicating that these were not true dinucleotides, but simply single nucleotide variants occurring next to each other by chance. One exception, DBS24 – associated with SBS90, attributed to duocarmycin exposure – has a pattern that can be mostly recapitulated by simulation of SBS90, apart from the CT>AA component (fig. S8I). Three signatures were not in the GEL cohort and could not be curated (DBS23, DBS32, DBS35) due to lack of access to sequencing data.

Contrasting previously unreported signatures with previously reported endogenous processes

Deamination and amplified deamination

Pervasive patterns of deamination are widely observed in malignant and non-malignant tissues. SBS1 characterized by C>T mutations at CpG is due to deamination of methyl-cytosine, while SBS2 and SBS13 are due to APOBEC-related deamination. Both are likely physiological: SBS1 occurs by natural hydrolytic processes, while SBS2 and SBS13 arise through transient single-stranded DNA availability (20).

Two rare signatures also characterized by C>T transitions at CpG are SBS96 and SBS95, differing by their ability to demonstrate marked hypermutator phenotypes and relative C>T peak heights (Fig. 4, A and B). SBS96, present in 18 of 12,222 GEL samples (0.15%, table S23, reported as due to inherited and/or acquired mutations in MBD4 (21), has C>T at ACG as its tallest peak. We identified germline truncating MBD4 mutations with loss of heterozygosity (LOH) of the alternative allele to explain 12 of 18 samples (6/10 patients) with SBS96 (table S28 and Fig. 4C). MBD4 germline variants were also seen in 35 other GEL patients, yet SBS96 was not observed in their tumors because the wild-type parental allele was intact in all assessable cases. We note that SBS96 was observed in extremely rare cancers such as myxofibrosarcomas and uveal melanoma. SBS95 is distinguishable from SBS96 by having its tallest peak at CCG and by exhibiting transcriptional strand bias. SBS95 occurred in a lymphoid and a stomach cancer in GEL and one head and neck cancer in the ICGC cohort (table S23). None had MBD4 mutations. The cause for SBS95 remains unclear.

Open in a separate window

Fig. 4

Signatures associated with endogenous mutational processes.

(A) Five signatures characterized by substitutions at NCG nucleotides are contrasted to each other. SBS profiles are shown, along with prevalence of signatures across three cohorts (GEL, ICGC and Hartwig), transcriptional strand bias (TSB) and replication strand bias (RSB). (B) Mutationburdens associated with each signature in all tumor-types. SBS1 is common and seen in all cohorts in many tumor-types. SBS95, SBS96, SBS87, and SBS105 are rare and are associated with a higher mutation burden when compared to SBS1. Y-axis shows mutation count on log-scale. Summaries comparing signatures in etiological groupings can be found here: https://signal.mutationalsignatures.com/explore/study/6. (C) SBS96 signature and prevalence among samples in the three cohorts. The Venn diagram illustrates the number of patients in GEL with SBS96 and biallelic loss ofMBD4. (D) SBS108 signature and prevalence among samples in the three cohorts. The Venn diagram illustrates the number of patients in GEL with SBS108 and biallelic OGG1 G308E. (E) SBS30 signature and prevalence among samples in the three cohorts. The Venn diagram illustrates the number of patients in GEL with SBS30 and biallelic loss of NTHL1. Only samples with SBS30 were inspected for NTHL1 mutations.(F) SBSs associated with MMR and POLE/POLD gene defects. (G) Proportion of samples across GEL organs with MMR or POLE/POLD defects related signatures or with high HRDetect score. (H) Proportion of samples with MMR biallelic loss and/or POLE/POLD dysregulation in each group of samples presenting an SBS mutational signature in (F). (I) Proportion of HRDetect-high samples with biallelic loss in genes linked to homologous recombination repair, number of samples in parenthesis.

Two signatures were characterized by C>N at CpG (Fig. 4A). SBS87 (22), with its tallest peak at CCG, was observed in one breast cancer. A related signature with C>N at all CpGs, SBS105, was reported in one breast and one bladder cancer in GEL. Although we have not found a cause for SBS105, it is associated with DBS12, a mathematical outcome of a high rate of SBS105 (fig. S8E), and does not exhibit transcriptional strand bias. Mechanistically, SBS105 would require deamination at CpGs followed by generic misincorporation during DNA replication and/or repair, not limited to the A-rule (23), to generate this pattern.

Despite all occurring at CpGs, these signatures have distinguishing characteristics. Discriminating MBD4-related SBS96 is particularly important given reports that such tumors have sensitivities to checkpoint therapies (24).

Environmental sources of mutational signatures

UV-like C>N signatures at CCN and TCN

We reinforce SBS7a (defined by C>T at CCN and TCN) in skin tumors with associated DBS1 characterized by CC>TT dinucleotides (33) (Fig. 5, A and B). However, we highlight three signatures that occurred at similar trinucleotides CCN/TCN and that could be miscalled as UV-related but may be due to alternative etiologies.

Open in a separate window

Fig. 5

Signatures associated with environmental mutational processes.

(A) SBS signatures occurring at CCN and TCN, with similarities to UV-related SBS7a. (B) Correlations of the signatures in (A) with DBS signatures. (C) SBS signatures presenting T>A, with similarities to AAI-associated SBS22. (D) SBS signatures indicating prior platinum exposure. (E) Correlations of the signatures in (D) with DBS signatures. (F) SBS signatures presenting C>A with similarities to tobacco-associated SBS4. (G) Correlation of the signatures in (F) with DBS signatures. Numbers in each column report the number of organs implicated in the correlative analyses. A correlation is computed independently for each organ and the correlations are displayed as a boxplot. Boxplots denote median (horizontal line) and 25th to 75th percentiles (boxes). The lower and upper whiskers extend 1.5×IQR.

SBS129, observed once in a nodular malignant melanoma (GEL-2501934-11) and once in a leiomyosarcoma (GEL-2300438-11), is characterized by C>T transitions at CCN, particularly CCA and CCT, but not TCN trinucleotides (Fig. 5A). It is distinguishable from SBS7a by its rarity and lack of CC>TT dinucleotides. However, SBS129 presents a transcriptional strand asymmetry with excess C>T mutations on the non-transcribed strand, the same as SBS7a. Apart from somatic TP53 mutations, no other potential genetic associations have been identified.

SBS38 is identical in its trinucleotide preponderance to SBS129, except it results in C>A transversions instead (Fig. 5A). Although reported before (14), it is rare, and its etiology is unknown. Here, we identify it in 30 cancers (29 skin, one lung, table S23) in GEL and note that it can either be a dominating phenotype or occur in combination with SBS7a, SBS17, and SBS18. Notably, among the samples affected by SBS38, we found all three anorectal mucosal cancers in the GEL cohort, an aggressive, unusual mucosal melanocytic cancer. This uncommon signature occurring in a very rare tumor-type hints at a germline genetic predisposition. Yet, we have not been able to identify a causative gene. Minor transcriptional strand bias is noted with more mutations on the transcribed strand for C>A mutations.

Lastly, SBS137 was observed twice in GEL brain cancers (table S23) and would superficially seem highly similar to UV (Fig. 5A). Critically, affected tumors do not have a CC>TT DBS signature (Fig. 5B) and demonstrate transcriptional strand bias in the opposite direction to UV (table S32), with an excess of C>T mutations on the transcribed strand (likely representing an excess of G>A on the non-transcribed strand). Its tallest peak is at CCC, dissimilar to the SBS7a peak at TCC. By contrast, in a metastatic CNS lesion derived from a cutaneous primary (GEL-2906789-11), the classic appearance of SBS7a and DBS1 is observed. This suggests that SBS137 is a distinct signature with currently uncertain cause.

Mutational Signature Extraction

For each tumor sample, we counted the number of somatic mutations and constructed SBS (96 channel) and DBS (78 channel) mutational catalogs (tables S7 and S8). Mutational signatures were analyzed independently for each tumor type in each of the three cohorts (Fig. 1C). First, we clustered mutational catalogs and excluded samples with unusual profiles (hierarchical clustering using 1 – cosine similarity as distance) (fig. S1, A to C), aiming at reducing the number of rare, complicating signatures and obtaining fewer, more accurate signatures. Second, we used non-negative matrix factorization (NMF) with Kullback-Leibler divergence (KLD) optimization, repeated bootstrapping (at least 300 bootstraps), and removed poor local minima (17). We identified a set of ‘common’ mutational signatures that were organ- and cohort-specific. Third, we fitted the common signatures into all samples of a given cohort and tissue type, and identified samples with high reconstruction error to identify unexplained processes or ‘rare’ mutational signatures (details in supplementary materials) (fig. S1, D to H) (tables S9 to S12).

To define signature exposures in each sample, we used a signature ‘fit’ procedure. Briefly, the number of mutations attributed to each signature in each sample were estimated using organ-specific signatures detected in their originating cohort utilizing KLD optimization (NNLM R package) and bootstrapping (200 bootstraps) (17). Point estimates of exposures were the median of the exposures obtained from bootstrapping. Exposures below 5% of the total SBS burden or below 25% of DBS burden per sample were set to zero because of the risk of over-fitting.

Reference signatures

To permit comparability across cohorts and organs, we defined ‘reference signatures’ (Fig. 1G). In brief, we clustered all common and rare mutational signatures (757 SBS or 301 DBS signatures) (tables S13 and S14) and obtained clusters of highly similar signatures (187 SBS and 60 DBS clusters). Cluster averages were termed ‘distinct patterns’ (tables S15 and S16). We assigned each distinct pattern to one of three groups: i) a reliably recurrent distinct pattern observed in multiple independent extractions; ii) a mix of two or more distinct patterns; iii) a singleton pattern found in one organ in one cohort (tables S17 and S18). Recurrent distinct patterns were additionally clustered to remove patterns that may simply be a variant of another pattern. Mixed distinct patterns that could be estimated as a combination of two distinct patterns using non-negative least squares were dismissed. Singleton distinct patterns were also curated and dismissed if they could simply be variants of other reference signatures. If they had been reported in other studies, they were retained as reference signatures. A total of 120 SBS and 39 DBS reference signatures were identified.

A QC status was assigned to each of the reference signatures: green, amber or red. QC green signatures were those extracted independently multiple times and/or reported in orthogonal studies. QC amber status was given to signatures with limited supporting evidence, such as signatures identified in only one extraction and not reported previously. QC red status was assigned to signatures that were mathematical or alignment artefacts. After QC, 82/120 SBS and 27/39 DBS reference signatures remained QC green (tables S19 and S20, SBS/DBS final reference signatures (tables S21 and S22), exposures (tables S23 and S24)). Conversion matrices that map reference signatures to organ-specific signatures of each cohort are in tables S25 and S26).

Additional analytics relating to correlations with germline and somatic driver events can be found in supplementary materials and tables S27 to S30.

Replication and transcription strand bias were calculated as in previous work (42). Briefly, we counted classes of single nucleotide variants (C>A, C>G, C>T, T>A, T>C, T>G) taking into account whether they appeared on the lagging or leading strand (according to MCF-7 reference repliseq data), or on the transcribed or non-transcribed strand (according to gene orientation) (42). A paired two-tailed Student’s t-test was used to determine the significant deviation from the ‘natural’ bias given by the regions base content. The log2 ratio was used to determine the size of the asymmetry between the two strands (table S32).

HRDetect scores were computed as previously described (17, 30). HRDetect input features are exposures of SBS3 and SBS8, proportions of short deletions at microhomology, HRD-LOH index, and exposures of rearrangement signatures 3 and 5. Rearrangement signature exposures were estimated by using KLD optimization, bootstrapping, and previously published rearrangement signatures (17). HRDetect scores were computed both as point estimates and also as a distribution obtained from 1000 bootstrapped scores, as previously described (17) (table S31).

This work was enabled by access to data and findings generated by the 100,000 Genomes Project, under the auspices of the Pan-Cancer GeCIP (project RR239). The 100,000 Genomes Project is managed by Genomics England Limited (a wholly owned company of the Department of Health and Social Care) funded by the National Institute for Health Research and NHS England. The Wellcome Trust, Cancer Research UK and the Medical Research Council have also funded research infrastructure. The 100,000 Genomes Project uses data provided by patients and collected by the National Health Service as part of their care and support. This publication and the underlying research are facilitated by data that were generated by the Hartwig Medical Foundation (HMF) and the Center for Personalized Cancer Treatment (CPCT) in the Netherlands, and the International Cancer Genome Consortium.