Double-base deletions and signatures associated with platinum adducts in metastatic cancer

Abstract

Platinum compounds preferably create intrastrand G-G adducts to halt replication in cancer cells. When successfully bypassed, resulting somatic variants end up in the genome. Unexpectedly, current double-base substitution signatures associated with platinum lack a GG component, and we hypothesized that deletion was missing from the possible solutions to circumvent a double-base lesion. We investigated a large pan-cancer metastatic cohort and generated platinum specific de novo double-base substitution and deletion (DBSD) signatures. One of the oxaliplatin associated DBSD signatures had GG deletion as dominant characteristic and also included new and known substitutions associated with platinum. Unforeseen among the additional results, DBSD signatures with deletion of TT as singular characteristic were strongly associated with microsatellite instability in multiple cohorts and provisionally in liquid biopsies. This study showed that DBSD signatures increase our understanding and cataloging of the effects caused by platinum and, in general, may aid in the discovery of new biological insights.

Graphical Abstract

Graphical Abstract.

Introduction

Platinum compounds-cisplatin, carboplatin, and oxaliplatin-are one of the therapeutic backbones used in multiple types of cancer, and the mechanisms whereby platinum compounds work have been well studied, showing a preference to create instrastrand G-G (∼65%) and A-G (∼25%–30%) and 1%–3% interstrand adducts [1, 2]. Such lesions in principle halt the replication machinery, and in order for tumor cells to survive, the lesions must be resolved. If successfully bypassed, replication can resume, cells survive, and, when not properly repaired, the applied patch should be detectable as a somatic variant. Several pan-cancer studies using whole-genome sequencing (WGS) data have been used to study variants and their compilation into mutational signatures [3–5], obtaining recurrent patterns that occur in specific tumor types and/or are clearly driven by known carcinogenic processes, such as UV radiation or polycylic aromatic hydrocarbons from tobacco smoke, or variants that are due to intrinsic sources (e.g. cytosine deamination that are age-related or via the Apobec3 family of enzymes). The latest version of the COSMIC mutational signature database shows that the etiology of double base substitution (DBS) signature 5 is associated with prior treatment with platinum drugs (DBS5). This signature’s main characteristics are CT > AA and CT > AC mutations. Mutational signatures use a pyrimidine-oriented notation, so a 5′-CT-3′ target is equivalent to 5′-AG-3′, one of the platinum drug targets. Surprisingly, the dominant platinum target (GG, in signature notation CC) is absent. One answer to this conundrum could be that error-free copies can be generated of the strand with the GG adducts, e.g. via translesion synthesis of DNA polymerases of the Y-family [6]. Indeed, pol η has been shown to be capable of complementing the correct cytosines despite the G-G adduct [7–9]. However, platinum drugs seem quite effective in several tumor types (review [10] and references therein) showing clear benefit in, e.g. lung cancer patients treated with cisplatin [11], and colorectal patients with oxaliplatin [12]. We thus initiated this study hypothesizing that the current possible explanations to circumvent a double-base lesion lack the possibility of deletion. For example, instead of polymerases defiantly incorporating nucleotides over the lesion, be these of the wrong kind or not, the lesion may be resolved by skipping the adducted bases, leading to a deletion, as has been shown for PrimPol [13, 14]. Theoretically, platinum induced interstrand cross-links might be able to cause double-strand breaks (DSBs), the repair of which might cause a 2-bp deletion. However, a genome-wide DSB repair map showed a 2-bp deletion is not among the lesions associated with the various DSB repair pathways [15]. Current InDel (Insertion/Deletion) signatures do not capture which bases are deleted in a >1-bp context and thus do not distinguish the platinum relevant doublets (GG and AG) from platinum irrelevant ones. We used a large cohort of pan-cancer WGS data of metastatic biopsies for which the prior treatments have been documented. We aimed to investigate whether sequence specific double-base deletions occur in cancer genomes, as well as integrating double-base substitutions and deletions (DBSDs) into unified profiles for new mutational signatures. Initially, we focused on samples pretreated with platinum compounds but subsequently evaluated this in a broader sense to get a more complete picture of this unrecognized solution to double-base driven somatic variation. The work yielded 3 novel mutational signatures; one related to oxaliplatin exposure, one related to microsatellite instability (MSI), and one enriched in brain cancers.

Materials and methods

Patient cohort

Whole-genome sequencing data of in total 5822 biopsies were used of patients of diverse cancer types, available via the Center for Personalized Cancer Treatment (CPCT) consortium study (CPCT-02, NCT01855477), which was approved by the medical ethics committee of the University Medical Center Utrecht, the Netherlands. Metadata, including tumor type, pre-treatment drugs (i.e. treatments given prior to obtaining the biopsy), and MSI, were also available. Further details about the CPCT study and the technical processing of tissues and generating WGS data have been described before [5]. Samples were excluded when the biopsy site was the organ of the primary cancer type, and when multiple biopsies were available for a single patient, only the first biopsy was included.

Double base events

Somatic mutations were available for each sample in a standard VCF file, and these were parsed to extract DBSs and deletions. In the existing DBS classification, in total 78 different substitutions are discerned [3] in 10 strand-agnostic unique double base targets: AC, AT, CC, CG, CT, GC, TA, TC, TG, and TT. Complementary bases are pyrimidine-catalogued, i.e. a GG > TT substitution is listed as CC > AA. This classification was assimilated and complemented with a deletion of the double base target as additional option, yielding a total of 88 possible DBSD contexts. The integration of both substitutions and deletions makes this method uniquely distinct from both existing DBS signatures as well as distinct from the InDel signatures algorithms. This resulted in a mutation profile for each sample of the number of events of each of the 88 DBSD contexts. This was used as input for SigProfilerExtractor (v1.1.23) [16], which was used to extract de novo mutational signatures. SigProflerExtractor was run with default settings except setting the minimum_signatures at 7. Default settings included reference_genome: GRCh37, exome: False, maximum_signatures: 25 and nmf_replicates: 100. The choice of the tumor types was first based on the pre-treatment metadata, including only tumor types where platinum pre-treatment was given. Next, tumor types that were used in the analyses were selected based on balancing sufficient number of samples for the analyses versus the confounding effect on extracting mutational signatures when the input data contains dozens of tumor types, each likely obfuscating the signatures by introducing tumor type specific patterns. Figure 1 shows the tumor types that were selected per platinum compound, which had both sufficient numbers for the pre-treated and untreated patients. Each platinum pre-treated (or untreated) cohort was separately used to generate de novo mutational signatures of the DBSD counts. SigProfilerExtractor automatically calculates an optimal solution for the number of signatures, and for each of the six analyses, this optimal solution was used. Sigminer (v2.3.0) [17] was used to fit the signatures to sample profiles, i.e. the mutation spectrum of a sample is associated with given signatures to calculate the relative contribution of each of the signatures. Sigminer was also used to visualize the signatures. To verify if a mutation was present on the transcribed strand, Gencode v49 was used (https://www.gencodegenes.org/). To determine differences in mutational frequencies on the leading and lagging strand, the global asymmetry of GG over CC was calculated [18] in the so-called left and right replicating regions. That regional information was obtained via https://github.com/getzlab/AsymTools2/, but since the original method only deals with single nucleotide changes, we needed to augment the provided data with the number of possible occurrences of GG or CC on the reference strand for each left or right replicating region. The number of observed GG/CC deletions was corrected for by the number of possible occurrences for this to happen, after which the ratio of leading and lagging strand mutations could be calculated (further detailed in [18]).

Figure 1.

Sample numbers included in our study by primary tumor type and pre-treatment.

Validation using publicly available data

To verify results from our metastatic cohort, several publicly available cohorts were used. First, data from a primary colorectal cancer cohort [19] of 1063 samples were downloaded from the European Nucleotide Archive (https://www.ebi.ac.uk/ena/) under accession number PRJEB61514. MSI calls for these patients were included in the supplementary table of the manuscript. Second, somatic variants of 1599 PCAWG samples (primary pan-cancer), processed via the same pipeline as the metastatic cohort [20], were used to verify the general deletion profile. Furthermore, WGS data of 24 CRC cell lines were downloaded from the National Institutes for Health Sequence Archive, Bioproject PRJNA523380 (CCLE, Cancer Cell Line Encyclopedia). Variants were called and annotated by the same algorithms used for the metastatic cohort (Sage and Pave, https://github.com/hartwigmedical/hmftools), without using a matched normal. Variants were filtered with a general population frequency > 0.00015 in either ‘EUR_AF’ (European 1000 Genomes Phase 3), ‘gnomADe_NFE_AF’ (Non-Finnish-European gnomAD exome or ‘gnomADg_NFE_AF’ (Non-Finnish-European gnomAD genome). MSI calls for these cell lines were obtained via supplementary data [21]. Both for the primary CRC as well as for the CCLE cohort, the available MSI calls were inferred using genomic characteristics.

cfDNA, ddPCR, and shallow sequencing

Samples of three metastatic cancer patients of whom the primary tumor had a MSI-High status were used; patients were part of a clinical study which was approved by the Medical Ethics Committee of the Erasmus University Medical Center (MEC-2017-238). Circulating cell-free DNA (cfDNA) was extracted from 4 ml of blood plasma with the QIAamp Circulating Nucleic Acid Kit (Qiagen), according to manufacturer’s instructions. The resulting cfDNA eluate was analyzed for MSI with the droplet digital PCR (ddPCR) MSI kit (Biorad), which uses three different assays to detect changes in five microsatellite markers; BAT-25, BAT-26, NR-21, NR-24, and Mono-27. The experiments were performed using the Naica Crystal Miner PCR system (Stilla). This system generates a maximum of 30 000 droplet crystals per sample, followed by amplification of nucleic acid molecules in each droplet. The ddPCR mixture consisted of 1× Naica multiplex PCR mix Buffer A and B (Stilla), 0.1 µM Alexa Fluor 647 Hydroxylamine (Thermo Fisher), nuclease free water, 1× Assay (Biorad), and 5.4 ng cfDNA. Twenty-six microliters was loaded onto a Sapphire Chip (Stilla). The following protocol was used for the thermal cycling conditions: Cycling started with partitioning at 40°C for 12 min, during which the pressure increased to atmospheric pressure (AP) + 1000 mbar, followed by a decrease to +950 mbar. This pressure remained until release, where pressure was regulated down to AP again. The second step in the thermal cycle started at 95°C for 10 min, followed by 40 cycles of denaturation at 94°C for 30 s and 55°C for 1 min. After these 40 cycles, release began with gradually decreasing temperature down to 25°C in 33 min. After polymerase chain reaction (PCR) amplification, the Naica Prism3 (Stilla) and accompanying software was used to analyze the individual droplets and detect fluorescent signals with a focus setting of 0.74 mm and exposure times of 65 msec blue (FAM) and 125 msec green (HEX). The mutational fractional abundance was calculated according to protocol: concentration of mutant divided by the sum of mutant and wild-type concentration.

Shallow whole genome sequencing (sWGS) on the three cancer samples used for ddPCR, as well as 41 anonymous healthy blood donors, was performed as follows: 5 ng of cfDNA per sample was used for input in the ThruPLEX DNA-Seq HV kit (Takara-Bio) to prepare indexed sequencing libraries. Libraries were sequenced to a median 5.0× depth of coverage on an Illumina NEXTseq2000, generating at least 70 M 150 bp paired-end reads. Raw sequencing data were processed using Fastp (v0.23.4) for adapter removal, and reads were aligned to the hg38 reference genome using bwa-mem2 (v0.7.17). Reads were sorted using samtools (v1.21). Next, duplicate reads were removed, and base recalibration was performed using GATK (v4.3.0.0). The resulting bam files were processed using mismatchfinder (v0.1.0) [22], which is a cfDNA specific variant caller. The resulting VCF files were parsed to extract the DBD counts, and the number of TT/AA deletions (and as percentage of total DBDs) were used to associate with the ddPCR scores.

Statistics

Stata (v14) and R (v4.3.2) were used for statistical analyses and visualizations. Continuous variables were analyzed using either a Mann–Whitney U test or a Kruskal–Wallis test depending on the number of categories.

Results

Of the complete pan-cancer metastatic cohort of 5822 samples, 3763 have a recorded prior-treatment status, of which samples were selected that had prior carboplatin, cisplatin, and oxaliplatin treatment. Of note, for the vast majority of cases, this was not a mono-therapy, but combined with other drugs. Figure 1 shows the numbers involved of the subgroups we analyzed. For carboplatin pre-treated cases, we selected breast, esophagus, and ovary as primary tumor types. For cisplatin, lung and urothelial tract were selected, and for oxaliplatin, only colorectal was selected. For each of the selected tumor types, cases were also included that were treatment-naïve. Next, we extracted double-base variants, using, besides the 78 existing possible substitutions, also deletion of the double base, and generated de novo double-base mutational signatures separately for each of the three cohorts pre-treated with a platinum compound, and also separately for their treatment-naïve counterparts of the same tumor type. This enterprise yielded six sets of mutational signatures.

Double-base substitution and deletion signatures

Using the WGS data of metastatic biopsies that were pre-treated with carboplatin containing therapies, we obtained eight signatures, while we obtained seven signatures for the same tumor types but without pre-treatment (Supplementary Fig. S1A). Five of the eight carboplatin signatures were also present [cosine similarity (CS) > 0.9] in the no pre-treatment signatures and are thus likely tumor type specific, not treatment related (Supplementary Table S1a). Of the three signatures that are carboplatin specific, 2 are closely related to the already described platin-associated DBS5 signature. In DBS5, the main characteristics are CT > AA and CT > AC, and these are now split into two separate signatures (Sig-3 and Sig-6, Fig. 2A/B and Supplementary Fig. S1A). The pyrimidine oriented notation in mutational signatures means that the CT target is equivalent to AG, which is a platinum target. The third carboplatin signature is noisy, in that it contains many types of characteristics, all with low percentage contributions. Thus, we did not observe a new carboplatin specific profile.

Figure 2.

Mutational signatures. A–D show the relative contribution of 88 possible DBSDs. XX indicates a deletion, and that channel is located most right in each of the targets (arrowheads). A and B are related to DBS5, which combines the CT > AA and CT > AC peaks. C/D are signatures specifically observed in metastatic CRC samples that were pre-treated with oxaliplatin containing therapy, with CT > AA and TC > AA in DBSD-3 (C), and deletion of CC (indicated by the black box) and CT > AA as the main characteristics in DBSD-2 (D). All signature weights are in Supplementary Table S2.

For the cisplatin analyses, both the pre-treated and not pre-treated cohort yielded eight de novo signatures, again showing five signatures present in both cohorts (CS > 0.9, Supplementary Fig. S1B/Supplementary Table S1a). Also in this analysis, the main peaks of DBS5 are split into two separate cisplatin specific DBSD signatures. The third cisplatin specific signature has a CS of 0.854 to a signature also observed in the not pre-treated cases (also the third signature in that cohort, Supplementary Fig. S1B). Both have CC > AA as its main characteristic, while the additional low-abundant characteristics of the cisplatin profile (TG > AT and TG > CT) are not platinum targets. Thus, also in these analyses, we observed no new cisplatin specific signatures.

Oxaliplatin-pre-treated cases show eight de novo DBSD and seven signatures in the not pre-treated colorectal cases (Supplementary Fig. S1C/Supplementary Table S1). Here, only three signatures are overlapping in pre-treated/not pre-treated samples (CS > 0.9), with one signature (DBSD-4) with a CS of 0.881 (a general deletion profile, except deletions of CC). Again, the main peaks of DBS5 are split into two separate signatures, but the CT > AA peak of DBS5 is now joined by TC > AA in DBSD-3 (Fig. 2C). Both CT and TC are the targets for platinum drugs in their equivalent AG and GA. Another oxaliplatin specific signature (DBSD-7, Supplementary Fig. S1C) has a main peak of TT > AA at 15.6% contribution and many other minor contributing characteristics, without any clear association with the platinum working mechanism. Remarkably, there was a clear new profile (DBSD-2), which has deletion of CC (CC > XX, deletions are indicated as a separate channel next to the options for substitution, labeled as “XX”) and CT > AA as main peaks (Fig. 2D). This is the first time the dinucleotide CC, the equivalent of GG that is the main target of creating an adduct by platinum drugs, appears in a platinum pre-treatment specific signature. And not only as target for substitution—there is some contribution of CC > AA/AG/AT, but the main effect is deletion of CC (CC > XX, indicated by the black box in Fig. 2D). To a lesser degree, CT deletion is also among the characteristics (CT > XX).

Since a CC-centered profile did not materialize in the carboplatin and cisplatin analyses, we studied the new oxaliplatin DBSD-2 signature in a more fine-grained overview of the pre-treatments. Where oxaliplatin is predominantly accompanied with a pyrimidine-antagonist (PYR) like capecitabine or fluorouracil in our colorectal cohort (76% and 13%, respectively), in the other cancer types, the platinum compound is often combined with other PYRs like gemcitabine or pemetrexed. We calculated the relative contribution of the oxaliplatin DBSD signatures in all samples of our cohort and associated the relative contribution of oxaliplatin DBSD-2 with detailed pre-treatment groups (Fig. 3A). In the carboplatin pre-treated cohort, only 38 (19.8%) samples included a PYR in the same line as carboplatin (30 gemcitabine, 8 capecitabine). Of the cisplatin pre-treated cohort, 73% of samples (n = 105) included a PYR in the same line as the cisplatin, mostly gemcitabine (n = 61) and pemetrexed (n = 39). The group with the highest median contribution DBSD-2 was single oxaliplatin (19.3%), followed by oxaliplatin + capecitabine (15.1%). Of the other compounds, highest median contribution was seen in the combination with gemcitabine (12.1% and 7.1% for carboplatin and cisplatin, respectively). PYRs might contribute to signatures other than oxaliplatin DBSD-2; however, when combining the relative contribution of DBSD-2 with the total mutational DBSD load, the patterns for DBSD-2 did not change (e.g. single oxaliplatin still had the highest median, Supplementary Fig. S2). Evaluating oxaliplatin DBSD-2 in all platinum pre-treated patients, 33% (33 of 99), 21% (27 of 131), and 56% (152 of 270) of carboplatin, cisplatin, and oxaliplatin pre-treated samples, respectively, showed at least 10% contribution of this signature. Next, we used WGS data of 1063 primary colorectal cancers [19] as “reverse” validation; the oxaliplatin DBSD-2 signature should have a low contribution in this primary CRC cohort where 88% of cases were untreated (n = 937). We extracted the DBSD events and again calculated the relative contribution of the oxaliplatin DBSD signatures for each sample. The median contribution of oxaliplatin DBSD-2 was significantly lower in primary untreated CRC than in oxaliplatin pre-treated metastatic CRC (Mann–Whitney P = 2.1e-24, Fig. 3B).

Figure 3.

Oxaliplatin DBSD-2 signature contribution by pre-treatment type and origin. (A) Samples from our cohort were grouped according to their pre-treatment and pyrimidine antagonist (PYR), including capecitabine, cytarabine, fluorouracil, gemcitabine, pemetrexed, and methotrexate. Other indicate a combination with other drugs, but excluding PYRs. (B) Oxaliplatin DBSD-2 signature contribution in untreated primary CRC and oxaliplatin pre-treated metastatic patients. P-value is derived from a Mann–Whitney U test.

Platinum-acquired double deletion

We next investigated if the deletion of CC or the equivalent GG in platinum pre-treated samples was limited to two bases. Furthermore, we studied the double-base deletions individually, i.e. outside of the mutational signatures whereby combinations of events are considered. To make it so, we counted the number of 1, 2, 3, 4, 5, and >5 base deletions for each sample, and calculated the odds ratio (OR) of samples with platinum pre-treatment versus those of the same cancer-type without any pre-treatment (Fig. 4A).

Figure 4.

Odds Ratios of pre-treated versus not pre-treated cases, by treatment type. (A) The number of deleted bases by pre-treatment type and (B) the number of two-base deletions, by nucleotide composition of the deletion and pre-treatment. Of note, we excluded MSI cases, since further described results showed extremely high number of delTT events, obscuring the pre-treatment status.

A 2 base deletion and >5 base deletion were significantly higher only in oxaliplatin pre-treated colorectal cancer patients, with an OR of 23.2 (P < .0001) and 2.6 (P = .017), respectively. Thus, a 2 base deletion is clearly the dominant effect, and to obtain further details on the double-base deletions, we stratified them by type (Fig. 4B). Without the restriction of co-occurring deletion patterns which are used in mutational signatures, deletions of CC are enriched in cases pre-treated with oxaliplatin and carboplatin, GC deletions in cisplatin and carboplatin pre-treated cases, and CG deletions in oxaliplatin pre-treated cases (Fig. 4B). Again, deletion of CC shows the largest effect in oxaliplatin pre-treated cases. Still, there are samples without oxaliplatin pre-treatment that have deletions of CC, and we next investigated the sequence context of these deletions. We used the variant annotation to retrieve the positions of deletions, and where we previously unified CC and GG events as CC for the mutation signatures, we now chose to depict the results using GG, which is the platinum target. For reported CC deletions, we used the complementary sequence, included 10 nt before and after the deleted bases, and stratified by pre-treatment status (Fig. 5A and Supplementary Fig. S3 for carboplatin and cisplatin contexts).

Figure 5.

Details on deletion of GG. (A) Sequence logos of deletion of GG (indicated by the box) with the relative composition of preceding and subsequent nucleotides surrounding the deletion. Of note, in the case a GG is deleted in a repeat of G-nucleotides, a variant caller cannot know which of the two bases are deleted and defaults to positioning that variant at the beginning of the repeat. This is why after the deleted GG, the next Gs of the repeat are all to the right. (B) The number of deletions of GG in just the 2 Gs or in a repeat > 5 Gs, by pre-treatment type.

Deletions of GG in samples not pre-treated with oxaliplatin are virtually all found as part of a homopolymer of G-nucleotides, while GG deletions observed in oxaliplatin pre-treated samples are most often singular, i.e. not observed in a repeat of G-nucleotides. GG deletions in carboplatin or cisplatin pre-treated samples have a very similar sequence context to that of oxaliplatin. In contrast, only in the untreated samples of the oxaliplatin cohort (colorectal tumor type) are the GG deletions observed in a G-repeat, while in the other tumor types this is not the case (Supplementary Fig. S3). We further quantified these results by enumerating the number of consecutive Gs (Supplementary Fig. S3) after the deletion, by platinum pre-treated cases and not pre-treated cases. Figure 5B shows the number of events of just 2 Gs or deleted in 6 or more consecutive Gs. First, the number of GG deletions is, in absolute numbers, much higher in oxaliplatin pre-treated cases compared to carboplatin or cisplatin pre-treated cases. Most striking is that in the CRC subgroup, 95.1% of the events (sample size corrected) where only the 2 Gs are deleted are found in oxaliplatin pre-treated cases (67.0% and 62.1% in carboplatin and cisplatin pre-treated cases, respectively). In contrast, when GG is deleted as part of a repeat of Gs, the percentage found in the pre-treated cases are 27.6%, 66.3%, and 48.3% for oxaliplatin, carboplatin, and cisplatin, respectively. Thus, the effect of deletion of just GG appears strongest in oxaliplatin pre-treated colorectal cases. We evaluated if this effect was also present in non-colorectal cases that were pre-treated by oxaliplatin (46 oxaliplatin pretreated non-colorectal samples versus 139 untreated samples of the same tumor types). Here, of all just GG deletions, 74.4% (sample size corrected) occur in the oxaliplatin pre-treated samples. This is significantly lower (Poisson regression, Incidence Rate Ratio 0.15, P = 1.18e-44) than the colorectal oxaliplatin pre-treated cases (95.1%), so the effect appears particularly evident in the colorectal cancers. Lastly, since a large portion of oxaliplatin associated deletions occur in just the 2 Gs and less so in G repeats, we checked if this was also the case for substitutions. In colorectal samples pre-treated with oxaliplatin, in total 25 185 substitutions of GG were observed, of which the vast majority occurred in just the 2 Gs (n = 15 912, 63.2%) or 3 Gs (n = 7513, 29.8%). Substitutions of GG where the GG is part of a repeat of 6 or more Gs account for only 0.3% (n = 77).

We also evaluated if transcription-dependent regulation of replication has a role by checking if the deleted GG occurred on the transcribed strand. In all samples 34% (28 464 of 83 828), the GG was deleted on the transcribed strand. However, this was unrelated to the pre-treatment, i.e. GG deletions were equally often observed on the transcribed strand of oxaliplatin pretreated patients (33.3%, 2311 of 6932) as in not pre-treated colorectal patients (33.1%, 1081 of 3269). Similar percentages were observed in the cisplatin and carboplatin groups (Supplementary Fig. S4A), the uniformity in observed transcription strand asymmetry percentages indicating no obvious role for transcription coupled repair, but likely reflect differences in DNA composition of the transcribed/untranscribed strand. Finally, we analyzed if the deletion of GG specifically was located on the leading or lagging strand during replication to evaluate if specific repair mechanisms might be involved. Global asymmetry [18] analysis showed a ratio of 1.081 of GG deletion on leading over lagging strand when including all samples. Restricting to only GG deletions in oxaliplatin pre-treated CRC patients, the ratio of leading over lagging was just 1.016, indicating no preference where deletions occurred (Supplementary Fig. S4B).

A simplified MSI signature and a general deletion profile

Evaluating the oxaliplatin DBSD signatures, several signatures were clearly associated with the already known DBS signatures, like DBS2, DBS4, DBS5, DBS7, DBS8, and DBS11 (https://cancer.sanger.ac.uk/signatures/, indicated in Supplementary Table S1a). We also compared the relative contribution of the oxaliplatin DBSD signatures to the existing insertion/deletion (InDel) signatures [3, 23] (Supplementary Table S3). Only two comparisons showed a CS > 0.9: oxaliplatin DBSD-1 with ID12 and oxaliplatin DBSD-5 with ID3. The latter association is not unexpected: oxaliplatin DBSD-5 (CC > AA peak) is highly similar to DBS2 (also a CC > AA peak), which was already known to be associated with ID3 (https://cancer.sanger.ac.uk/signatures/dbs/dbs2/).

Overall, besides the platinum specific signatures, two novel signatures emerged. First, defiant of different tumor types and pre-treatment status, in all six de novo DBSD analyses, the first deconvoluted signature is very specific, showing a singular peak of deletion of TT (TT > XX) with on average 82.9% contribution (range 57.0%–91.2%). We calculated the relative contribution of the oxaliplatin DBSD signatures in all 736 colorectal samples. Associating the relative contribution of oxaliplatin DBSD-1 with the available metadata showed a clear association with MSI, as previously estimated using WGS characteristics (MSIseq, [5, 24]). In the group that was used to derive the DBSD signatures (colorectal samples pre-treated with oxaliplatin), only four samples were labeled as MSI, but TT deletion appeared sufficiently dominant to surface as signature. In total, 35 of 736 colorectal samples were labeled as MSI (4.8%), of which nine samples were used in the de novo signature analysis (four samples with pre-treatment, five samples without). The median contribution of oxaliplatin DBSD-1 in these nine samples was 100% [95% confidence interval (CI) 90.5%–100%], with 7 of 9 showing >91% contribution. For the colorectal samples with MSI that were not used in generating the signatures, the median contribution was also 100% (95% CI 98.7%–100%). Both were significantly higher compared to the MSS cases (Fig. 6A). Of note, the aforementioned similarity of oxaliplatin DBSD-1 with InDel signature ID12 (Supplementary Table S3) was indeed based on consistent characteristics; our DBSD-1 is deletion of TT, and the main characteristic of ID12 is a 2 bp deletion (unspecified base) in a 6+ homopolymer. The relative contribution of ID12 is significantly higher in MSI cases (41.9% MSI versus 3.7% MSS, P = 1.41e-12, Supplementary Fig. S5A). Currently ID12 has an unknown etiology, and we hereby submit that ID12’s etiology is MMR deficiency related.

Figure 6.

MSI-related analyses and details on a general deletion signature. (A, B) Relative contribution of DBSD-1 with MSI status of metastatic and primary CRC. (C, D) The number of TT-deletions (log₁₀ scale) by MSI status in primary CRC and CRC cell lines (CCLE, Cancer Cell Line Encyclopedia). The red dashed line indicates a level of 1000 delTTs. All P-values are from the Mann–Whitney U test.

To further validate the association of oxaliplatin DBSD-1 with MSI, we again used the 1063 primary colorectal cancers [19], of which 223 were reported as MSI (inferred using MSISensor2). Figure 6B shows a significantly higher relative contribution of oxaliplatin DBSD-1 in MSI (median 73.3% in MSI versus 1.2% in MSS, P = 5.62e-117).

Considering the strong distinction between MSS and MSI, it may be sufficient to just count the number of deletions of TT per sample to determine MSI status of a given sample. In the primary CRC cohort when using 1000 TT > XX as threshold, just 2 of 223 MSIs are missed (below the threshold), while 6 of 840 MSS are false positive (above threshold) which is a >99% sensitivity and specificity of association with the reported MSI status (Fig. 6C). Next, we evaluated the sequence context of the TT deletions and observed that the vast majority of TT deletions are found in a >6 homopolymer T-repeat (98.3%, Supplementary Fig. S5B). Since a double A or T deletion is much less prevalent in the germline of the non-Finnish European population (only 0.7% of polymorphisms of non-cancer controls with an allele frequency >0.05 are deletions of AA/TT in gnomad v3), it should be feasible to obtain tumor specific double deletion events without the need to sequence the germline DNA of a cancer patient. To test this, we used WGS data of 24 colorectal cell lines (Cancer Cell Line Encyclopedia) and called the variants without matching normal. DBSD events were extracted and we associated the number of TT deletions with the reported MSI calls of these cell lines. Figure 6D shows a perfect distinction between the number of TT deletions and MSI call. Finally, we performed a small pilot using circulating free DNA (cfDNA) of three metastatic cancer patients (colorectal, endometrium, and melanoma) whose primary tumor tissue was MSI. The cfDNA was used to measure five mononucleotide microsatellite markers using ddPCR. Table 1 shows one cfDNA was clearly positive for all five markers (high percentages of altered copies), while the other 2 had much lower altered fractions to the point that it is doubtful that there was tumor DNA in the cfDNA fraction (sample p3). This low ctDNA content was confirmed since we were unable to detect a BRAF V600E mutation in cfDNA (with arguably sufficient concentration of 18.7 ng/ul), which was known from the primary tissue of this patient. Subsequently, we shallow sequenced (∼5× coverage) this plasma cfDNA and extracted the variants without using matched normal. In addition, we also sequenced cfDNA from 41 healthy blood donors and analyzed the number of TT > XX and also as percentage of total double base deletions for all samples. Table 1 clearly shows that both the absolute number of TT deletions as well as the % of TT deletions of all DBDs is substantially higher in the sample with a high altered fraction. The average number of TT deletions in HBDs was 97.5 (95% CI 90.9–104.0). Interestingly, both patients with low altered fractions in ddPCR have TT deletions above the 95% CI of the healthy controls, potentially indicating higher sensitivity when evaluating genome-wide double base deletions. This shows it is possible to measure MSI in clinical samples without using a matched normal and on cfDNA, making it an attractive clinical diagnostic to measure residual disease or monitor therapy response.

Table 1.

cfDNA results from three cancer patients and 41 HBDs

Marker	p1	p2	p3	Mean HBD
BAT26	0.9%	57.0%	0.0%
BAT25	0.7%	65.7%	0.0%
NR24	1.0%	56.6%	0.0%
NR21	0.9%	59.6%	0.1%
Mono27	1.4%	49.0%	1.3%
nr of delTT	134	6846	110	97.5
% delTT of DBD	25.0%	65.0%	15.3%	18.4%

ddPCR results of five MSI markers and the number and percentage of TT deletions via sWGS.

The second novel and recurrent signature was a general deletion profile, where several double bases are deleted concurrently, with, in most signatures CT > XX as the most prominent contributor (signature 4 and 2 in carboplatin pre-treated/not pre-treated, signature 7 and 5 in cisplatin, and DBSD-4 and -2 in oxaliplatin, see Supplementary Fig. S1). Remarkably, CC > XX was rarely contributing to this signature. Associating the relative contribution of the oxaliplatin DBSD-4 in the complete cohort with metadata, we noticed a significantly higher contribution in metastatic samples of “Nervous system” primary tumor type (Supplementary Fig. S6A). The median contribution in nervous system samples was 75.4%, 2.1-fold higher than the overall median (Mann–Whitney P = 1.8e-31, testing nervous system versus all others). Also in primary tumors oxaliplatin DBSD-4 was significantly higher (Supplementary Fig. S6B). When including pre-treatment data of the metastatic nervous system samples, 16 of 72 (22%) had received temozolomide, an alkylating agent. Remarkably, patients pre-treated with temozolomide had a slightly lower median contribution (62.4% versus 78.0% in not pre-treated cases, Mann–Whitney P = .0003, Supplementary Fig. S6C).

Discussion

In this study, by adding deletion to substitution as an alternative way to deal with a double base lesion, we obtained several novel insights. One prime observation was that a DBSD signature emerged when evaluating metastatic biopsies of colorectal cancer patients that were pre-treated with oxaliplatin. Other results highlight a general deletion profile which seems particularly prevalent in metastatic nervous system cancer, while the use of TT deletions as an MSI biomarker enabled us to verify MSI status in both tissues and liquid biopsies.

One of the questions that initiated this study was the seemingly unnoticed lack of a GG substitution in the signature that is platinum pre-treatment associated (DBS5), while GG is the preferred target of platinum compounds [1, 2]. DBS5 has two main characteristics CT > AA and CT > AC, and in our analyses these two peaks are split over separate DBSD signatures. This split of CT > AA and CT > AC peaks in separate signatures was also previously observed [4] in a very large (>30 000 tumors) pan-cancer cohort where deletions were not included in the double base signature extraction. But even in that large cohort, like DBS5, there was no CC peak in the platinum-associated signatures (signature notation is pyrimidine-orientated, so a GG event is listed as CC). In our study, we obtained an oxaliplatin specific signature (DBSD-2) that shows CC > XX and CT > AA as the main traits, together with additional CC > AA/AG/AT substitution components. To the best of our knowledge, this is the first observation of GG specific variants in a mutational signature associated with platinum. Considering the existence of other mutagens that also cause double base lesions, e.g. UV-exposure (DBS1), tobacco smoking (DBS2), or defective mismatch repair (DBS7, DBS10), we advocate for integrating deletions alongside substitutions as possible variants when analyzing double base signatures. The current DBSD setup may be even augmented by including sequence contexts surrounding the variants (as illustrated by Fig. 5), effectively incorporating double base events with the way InDels are categorized (i.e. by using homopolymers or microhomology information).

Further detailed analyses into deletions showed a clear preference for a 2 base deletion in oxaliplatin pre-treated samples, not of other lengths, and when analyzing the composition, a deletion of GG was very clearly associated with oxaliplatin pre-treatment and to a lesser extent carboplatin pre-treatment. These differences may again reflect previously reported differences between the effect of oxaliplatin and the other platinum compounds [25–29]. One additional characteristic of oxaliplatin here reported is that while platinum compounds are all capable of generating an adduct in adjacent Gs that are not flanked by additional G nucleotides, an oxaliplatin GG adduct is much more frequently resolved by deleting the GG bases than for carboplatin or cisplatin adducts. Thus, an oxaliplatin adduct somehow makes it easier for the molecular machinery to skip the adducted bases and create a deletion. Differences in conformational dynamics and hydrogen bond formation [28] might provide clues as to the root cause of the protein complexes that recognize and deal with oxaliplatin adducts. In contrast, an oxaliplatin GG adduct which is part of a homopolymer of Gs is most likely successfully solved since substitutions of such GG lesions in a repeat are rarely found. This suggests that the microhomology surrounding the adduct is sufficient to engage accurate translesion synthesis.

A general deletion profile (oxaliplatin DBSD-4) where several double-base targets are deleted (except CC and CG deletion) showed a significantly higher contribution in primary and metastatic nervous system malignancies. However, it does seem to be a property that is present in many cancer types (with a median of ~35% contribution). In a recent study [23] where InDel signatures were redefined on >18 000 cancers, these signatures were tested on (among others) 392 CNS tumors. This showed only signature InD9c involved in CNS as a deletion-driven profile. However, in the redefined InDel signatures, deletions of length 2 are combined with deletions of length 2–5 in a separate category, and additionally, this category did not have a high contribution in this signature. So, as of yet the biological relevance of this profile remains unclear. Its appearance in several tumor types may be attributed to different exogenous mutagenic processes. Further analyses are needed, and these could benefit from a CNS specific approach, preferably in a cohort where associations with clinical metadata might shed light on the relevance. Also, incorporating the sequence context will likely help in further delineating the relevant patterns in the different tumor types.

Lastly, including deletions as a possible variant in double base signatures, we stumbled across a serendipitous observation. In all of our de novo signature extractions, we noticed a very prominent MSI signature, with a single dominant characteristic of TT > XX where the TT deletions virtually all arise in a T-homopolymer. Validation in several independent cohorts showed its merit and robustness, and by associating with InDel signatures, vey likely explains the etiology of ID12 as MMR deficiency. Although we processed only a few samples, shallow sequencing (without matched normal) of cfDNA samples of cancer patients showed that TT > XX was concordant with the MSI status as determined by a ddPCR score of a 5-gene MSI marker panel. Although the primary tumors of all three patients displayed MSI, their circulating tumor DNA fractions varied quite a bit. The colorectal patient with very high altered fraction according to ddPCR showed almost 7000 TT > XX events, indicating that when circulating tumor DNA fragments with MSI are sufficiently present, the TT deletions are readily detected. A much larger cohort is required for exact thresholds of sensitivity and specificity to call MSI using TT > XX. Using shWGS on cfDNA is a quite cost-effective method for disease monitoring, especially when sequencing matched normal is not required (due to the low TT > XX population frequency). shWGS is a relatively rich data source and besides, e.g. enabling copy number analyses [30, 31] or fragmentomics [32], we show here it is applicable and practical to call MSI. Thus, although in its infancy, if proven consistent in a larger cohort, TT > XX may provide a clinically useful marker to monitor disease progression by tracking tumor content in plasma cfDNA in MSI-positive cases. Ultimately, this type of tests could improve current burdensome colonoscopy-based surveillance programs in genetically predisposed individuals at high risk for developing MSI cancers (i.e. Lynch syndrome carriers).

One drawback in our analyses in establishing DBSD signatures for platinum compounds is that these were predominantly based on the clinical practice of combining platinum compounds with other drugs in the treatment of patients. In theory, would the treatment have consisted of platinum mono-therapy it would have improved the specificity of the signatures.

To conclude, this study was focused on the effects caused by platinum compounds, and the results increase our knowledge in that dominion. In a more general sense, DBSD signatures are a valuable addition to the toolkit in the understanding mutagenesis in cancer.

Supplementary data

Supplementary data is available at NAR Online.

Conflict of interest

None declared.

Funding

This work was partly supported by the Dutch Cancer Society (grant KWF14976).

Data availability

WGS based somatic data and metadata used in this study were made available by the Hartwig Medical Foundation (our data request code DR-339) under a license stating data cannot be made publicly available via third party organizations. Data can thus be requested via the Hartwig Medical Foundation (https://www.hartwigmedicalfoundation.nl/en/data/data-access-request/).