## Main

Eukaryotic cells have many strategies to ensure the integrity of their genomes, with high-fidelity DNA replication7 and DNA-repair processes countering exogenous and endogenous DNA lesions8. The process of transcription targets DNA repair machinery to expressed genes, preferentially reducing their mutation rate after DNA damage1. Despite this targeted repair, in microorganisms, the process of transcription itself is mutagenic—a phenomenon that is referred to as transcription-associated mutagenesis (TAM)2,3. In yeast, topoisomerase 1 (Top1) activity is a major source of TAM and results in a distinctive transcription-dependent signature of 2–5 bp deletions at tandem repeat sequences9,10,11. Genome-embedded ribonucleotides have been established as a cause of Top1-TAM deletions in yeast12. Such ribonucleotides are frequently incorporated by DNA polymerases during replication, and represent the most prevalent aberrant nucleotides in the eukaryotic genome13,14. These genome-embedded ribonucleotides are normally removed by ribonucleotide excision repair (RER), a process initiated by the heterotrimeric ribonuclease H2 enzyme15. However, when Top1 cleaves at embedded ribonucleotides instead of RNase H2, this can result in small deletions16,17.

In the last decade, the widespread use of genome sequencing has enabled unbiased sampling of human mutations, substantially advancing understanding of mutagenesis in the germline18 and in neoplasia19. Multiple mutational processes act during cancer evolution, and mathematical methods decomposing tumour mutational profiles have been developed to define signatures that may correspond to individual mutagenic mechanisms19. This has successfully defined cell-intrinsic, environmental and treatment-related origins for many base-substitution signatures in cancer20,21,22. However, the origin of a substantial number of signatures remains unknown, and some may be artefactual. Recently, cancer signature analysis has been extended to indels5, small (1–49 bp) insertions and deletions. Such indels are an important class of mutations that contribute substantially to disease-causing germline variants (>20%) and human variation23.

Here we investigate an indel signature of unknown cause—ID4. We show experimentally that ID4 deletions are increased in RNase-H2-deficient cell lines and cancers and delineate a human TOP1-mediated TAM signature (ID-TOP1) that is relevant to both somatic and germline mutagenesis.

## ID4, a distinct cancer indel signature

The ID4 cancer signature, as categorized by COSMIC24, comprises 2–5 bp deletions, often with the loss of a single repeat unit in short repeat sequences5. These most commonly occur in regions in which the deleted sequence is repeated one, two or three times in tandem (Fig. 1a). Hereafter we use the term short-short tandem repeats (SSTRs) to distinguish between such short tandem repeats (STRs) with less than 5 repeats (that is, less than 6 repeat units) and microsatellite STRs with many repeats. In addition to these SSTR deletions, ID4 is characterized by small deletions at sequences with microhomology (MH), in particular, 2 bp deletions with single-nucleotide MH (SNMH). Both features are distinct from cancer deletion signatures resulting from other well-recognized mechanisms such as replication slippage and non-homologous/MH-mediated end joining (NHEJ/MMEJ) (Extended Data Fig. 1a, b). In support of a distinct aetiology, SSTR and SNMH deletions are not apparent in cancer associated with homologous recombination or mismatch repair deficiency, which are expected to have higher levels of MMEJ and replication slippage mutagenesis, respectively (Extended Data Fig. 1c, d).

## ID4 resembles a yeast mutation signature

Noting similarities to a Top1-induced TAM (Top1-TAM) in Saccharomyces cerevisiae, we reanalysed published genome-wide mutation-accumulation experiments performed with rnh201Δ pol2-M644G yeast25. This strain is particularly susceptible to Top1-TAM as it accumulates genome-embedded ribonucleotides at high levels due to RNase H2/RER deficiency and enhanced ribonucleotide incorporation by a steric-gate mutation at the catalytic site of the replicative polymerase Pol ε26. Similarities to the ID4 signature were apparent with a comparable pattern of small deletions at SSTRs, although mutational events at sites of SNMH were not evident in the yeast data (Fig. 1b). As more than 1 million ribonucleotides are incorporated by DNA polymerases per replicating mouse cell14, we reasoned that genome-embedded ribonucleotides might cause similar mutational events in mammalian cells. To experimentally assess whether TAM contributes to indel formation in human RER-deficient cells, we developed a reporter to enable sensitive and specific detection of mutational events arising from TOP1 activity in both yeast and mammals.

## Top1-dependent deletions in yeast

Mutation rates are routinely measured in S. cerevisiae using well-characterized but species-specific selectable markers (LYS2, URA3, CAN1). Thus, to establish a system that could be transferred between yeast and mammalian cells, we used an approach inspired by the Traffic Light reporter assay27, incorporating both positive and negative selection cassettes in a single transcriptional unit (Fig. 1c). The hygromycin-resistance gene (HygroR) was used both as the mutational target and negative selection marker. Indels causing a 2 bp frameshift within HygroR, including 2 bp deletions, result in translation of an otherwise out-of-frame P2A self-cleaving peptide and the neomycin-resistance (NeoR) gene, permitting positive selection of mutated colonies with neomycin (Extended Data Fig. 2a). To enrich the target for 2 bp tandem repeats, in silico redesign incorporated synonymous substitutions such that SSTRs accounted for >50% of the HygroR open reading frame.

For validation, the reporter was inserted into the S. cerevisiae genome and fluctuation assays were performed to assess the mutation rates in strains deficient for RER and/or Top1. A 37-fold increase in mutation rate was observed for the rnh201Δ (RNase H2 null) strain compared with the wild type (Fig. 1d), with a mutation rate of 6.1 × 10−9 per bp per generation (95% confidence interval = 5.4 × 10−9–6.9 × 10−9), whereas the increased mutation rate was abolished in the rnh201Δ top1Δ double-mutant strain, consistent with Top1-dependent mutagenesis at genome-embedded ribonucleotides12,28. Notably, there was a 10-fold decrease in the mutation rate for top1Δ compared with the wild-type strain, and a 35-fold decrease in 2 bp SSTR deletions (Extended Data Fig. 2b), consistent with previous reports10,11. Furthermore, the observed mutational spectrum was most similar for wild-type and rnh201Δ strains, but substantially different compared with the top1Δ and rnh201Δ top1Δ strains (Fig. 1e and Extended Data Fig. 2c–f). Taken together, we conclude that the same Top1-mediated mutations occur, albeit at different frequencies, in wild-type cells when RER is functional and in RNase-H2-deficient strains when elevated levels of ribonucleotides are present in the genome.

## TOP1-mediated mutations in human cells

Having validated the reporter in yeast, the same 2 bp repeat-enriched HygroR sequence was used to determine whether TOP1-mediated mutagenesis at embedded ribonucleotides is conserved in human cells (Fig. 2 and Extended Data Fig. 2g). NeoR was replaced by the puromycin-resistance (PuroR) gene, with reporter expression driven from the mammalian ubiquitous CAG promoter, permitting rapid antibiotic selection in mammalian cells. This modified reporter was inserted at the AAVS1 safe harbour locus in HeLa cells (Fig. 2a and Extended Data Fig. 3a–e). CRISPR–Cas9-mediated genome editing targeting the catalytic site of RNASEH2A was then used to generate two independent knockout (KO) reporter clones, alongside a control clone that had also been processed through the editing and clonal selection steps (Fig 2b, c and Extended Data Fig. 3). The control clone retained RNase H2 activity, whereas there was complete loss of cellular RNase H2 activity in the KO clones, accompanied by high levels of ribonucleotides in genomic DNA (Fig. 2b, c and Extended Data Fig. 3f, g).

In fluctuation assays, RNase-H2-null clones demonstrated a significant 3.1-fold increase in mutation rate (Fig. 2d) and 5.2-fold more 2 bp SSTR deletions (Extended Data Fig. 3h) compared with RNase-H2-proficient cells (RNASEH2A+), consistent with conservation of TOP1-directed mutagenesis in human cells. As in yeast (Fig. 1e), the overall mutational profile of reporter mutations was similar between RNase-H2-proficient and null HeLa cells (cosine similarity = 0.89, P < 10−4), predominantly comprising 2 bp SSTR deletions (Fig. 2e).

The mutation rate for RNase-H2-null HeLa cells (8.0 × 10−9 per bp per generation; 95% confidence interval﻿ = 6.7–9.5 × 10−9) was similar to that observed for rnh201Δ yeast (Fig. 1d), whereas the rate was substantially higher for RNASEH2A+ control cells compared with wild-type yeast. However, the increased mutation rate in RNase-H2-null HeLa cells probably underestimates the true impact of RER deficiency in human cells as, despite the fact that the control RNASEH2A+ HeLa reporter cells retained protein expression (Fig. 2b), the clone had also acquired mutations at the CRISPR editing site that reduced enzymatic activity (Fig 2c), causing a moderate increase in genomic ribonucleotide content (Extended Data Fig. 3f, g).

To confirm these findings, we used a complementary approach to establish the relevance of such mutational events genome-wide, performing mutation-accumulation experiments using human hTERT RPE-1 (TP53−/−) diploid cell lines. Ancestral populations for RNase-H2-wild-type and RNase-H2-null cells (RNASEH2A-KO or RNASEH2B-KO; Extended Data Fig. 4a–d) were established after initial single-cell sorting, and the clones were then grown for approximately 100 generations. Single-cell sorting was performed every 25 generations, creating bottlenecks to capture accumulating mutations (Fig. 3a). Combined variant calling on whole-genome sequencing (WGS) from paired ancestral and end-point cultures identified a total of 1,698 acquired high-confidence indel mutations, captured by at least 3 out of 4 variant callers. Consistent with TOP1-mediated mutagenesis, among all indel categories, only 2–5 bp deletions were found to be substantially (7.4-fold) and significantly enriched in RNase-H2-null RPE-1 cells compared with the wild type (Fig. 3b and Extended Data Fig. 4e, f), with an estimated rate of 1.1 × 10−10 2–5 bp deletions per generation per bp for KO and 1.4 × 10−11 for the wild type. Of these deletions in RNase-H2-null cells, 82% were 2 bp deletions, of which 48% were at SSTRs (Extended Data Fig. 4g). Furthermore, signature decomposition using SigProfilerExtractor5 reported a 21% ID4 contribution in RNase-H2-null cells that increased to 61% after background mutation patterns were subtracted to identify RER-deficiency-specific mutation signatures (Fig. 3c, d and Extended Data Fig. 5). The ID4 signature was substantially enriched in transcribed genomic regions (Extended Data Fig. 5e). ID5, a clock-like signature5, was also enriched in KO cells, probably due to slower growth and the longer culture time needed to achieve the same number of doublings for RNase-H2-null cells14.

## MH deletions specific to mammals

Small deletions at sequences with MH are an additional feature of ID4 (Fig. 1a) that is not observed in rnh201Δpol2-M644G yeast (Fig. 1b). However, consistent with a ribonucleotide-induced mutational origin in mammalian cells, MH deletions are observed frequently in RNase-H2-deficient RPE-1 cells, in which SNMH sites account for 31% of 2 bp deletions, indicating that, in humans, deletions at SNMH sites share the same aetiology as deletions occurring at SSTRs. Taken together, our reporter and mutation-accumulation experiments demonstrate that genome-embedded ribonucleotides cause a similar mutational signature in yeast and mammalian cells. Thus, topoisomerase-1-mediated mutagenesis probably also occurs in humans and is associated with 2–5 bp deletions at SSTR and SNMH sequences.

## ID4 mutations in a mouse cancer model

To determine whether TOP1-induced mutations resulting in the ID4 signature can be detected in vivo, we next studied an RER-deficient mouse cancer model in which Villin-cre conditional deletion of Rnaseh2b and Tp53 results in intestinal malignancy29. WGS analysis of paired tumour–normal tissue samples from 6 mice identified a total of 989 high-confidence tumour-specific somatic indels. An analysis of the resulting mutational signature established that ID4 substantially contributed in all tumours (Fig. 4a, b and Extended Data Fig. 6a), accounting for 32% of acquired indels. Consistent with a transcription-associated process, the ID4 signature was again most evident in transcribed genomic regions (Fig. 4b). The commonly occurring cancer signatures5 ID1, ID2 and ID5 were also observed, consistent with expectations of multiple mutational processes active in neoplasia.

The observed ID4 mutation spectrum corresponded closely to that observed in the RPE-1 mutation-accumulation experiment: 28% of indels were at 2–5 bp deletions, of which the majority were again 2 bp deletions (82%) predominantly at SSTRs (51%) and sites of SNMH (34%) (Extended Data Fig. 6b, c). This is consistent with the occurrence of TOP1-induced somatic mutations at genome-embedded ribonucleotides in vivo, conserved across different tissue and cellular contexts, and shows that this process can be detected in a cancer setting.

## A sequence motif for ID4 mutations

Although COSMIC defines the ID4 signature on the basis of indel size and repeat/MH context (Fig. 1a), the number of indels in the mouse RER-deficient tumour model enabled us to further investigate the characteristics of mammalian topoisomerase-1-induced mutations. We focussed our analysis on 2 bp deletions, as such events represented 81% of >1 bp deletions in the context of tandem repeats and 85% of deletions in sequences with MH.

First, we classified all 2 bp deletions at STR/SNMH sequences into six non-redundant dinucleotide classes, grouping together complementary sequences (Fig. 4c). We noted that the deleted sequences substantially deviated from genome-wide frequencies, with a complete absence of CC/GG and CG/GC deletions, as well as an overrepresentation of the CT category (containing CT, TC, GA and AG deletions). All of the observed deletions therefore included at least one thymidine (T), which functionally could be accounted for by the very strong preference of mammalian topoisomerase 1 to cleave at a phosphodiester bond with a T immediately upstream30.

Next, to investigate the wider sequence context, we aligned sequences containing all 228 two-bp deletions (Extended Data Fig. 6d), which indicated that deletions preferentially occur when T nucleotides are spaced at a two-base interval. Indeed, this TNT motif was present in 100% of SNMH (n= 77) and STR sites (n= 124), providing a common unifying sequence context for both deletion types (Fig. 4d), a finding that was replicated in both our RPE-1 (Extended Data Fig. 6e) and yeast datasets (Extended Data Fig. 7). We found that TNT is substantially over-represented at deletion sites compared with the genome-wide null expectation. Furthermore, although the TNT motif is common at tandem repeat sequences, 2 bp deletions at this motif are still significantly enriched when considering the occurrence of 2 bp STR and SNMH sequences in mouse and human genomes (Fig. 4e and Extended Data Fig. 6f), and STR sequences in the yeast genome (Extended Data Fig. 7).

## A model for TOP1-mediated deletions

To account for thymidines spaced at a two-base interval and the occurrence of mammalian SNMH deletions, we developed a revised model based on the established strand realignment model for yeast Top1-mediated mutagenesis12,16,17. In this ‘TNT model’, TOP1 cleaves preferentially 3′ of an embedded ribouridine, with nucleophilic attack by the 2′-OH of the ribose ring resulting in TOP1 release and formation of a non-ligatable nick with a terminal 2′,3′-cyclic phosphate (Fig. 4f (i–iii)). This then provides a substrate for TOP1 cleavage 2 bp or more upstream17, preferentially at a thymidine30. When this second cleavage event happens at a base that is identical to that of the first cleaved nucleotide—an event that is more likely at STR and MH sequences—strand realignment can then occur, resulting in a nick that is permissive to religation and TOP1 cleavage complex (TOP1cc) reversal (Fig. 4f (iv–vi)). An alternative mechanism of sequential Top1 cleavage, in which double-strand breaks occur due to nicking of opposite strands31 could not be reconciled with our TNT model, but may account for deletions occurring at non-STR/SNHM sites. Within the TNT motif, deletions were most common at CT and GT dinucleotides in both mammals and yeast (Fig. 4c and Extended Data Figs. 6 and 7b, e), which may be explained, at least in part, by preferential incorporation of ribouridine at CT and GT dinucleotides32 (Extended Data Fig. 7f).

Implicating TOP1-TAM as the cause of the ID4 signature enables us to include additional features in the definition of this COSMIC signature, namely preference for a TNT sequence motif at 2 bp deletion sites and enrichment in transcribed genes. Hereafter, we refer to this extended definition as ID-TOP1. To establish the relevance of the ID-TOP1 signature for human disease and genetic variation, we next examined publicly available datasets.

## ID-TOP1 in human cancer

RNASEH2B is frequently deleted in human cancer, in particular, in chronic lymphocytic leukaemia (CLL), given its proximity to a tumour suppressor locus, the DLEU2-miR-15a/miR-16-1 microRNA cluster33. Such RNase-H2-deficient human cancers should therefore be enriched for the ID4/ID-TOP1 signature. We analysed WGS data for 348 patients with CLL from two independent cohorts34,35, stratified on the basis of RNASEH2B-deletion status. Somatic variant calling identified a significant increase in 2–5 bp deletions in RNase-H2-null tumours (Fig. 5a), whereas other indels were equally represented across wild-type, heterozygous and null categories (Extended Data Fig. 8a). Of the 2–5 bp deletions in tumours with biallelic RNASEH2B loss, more than half (57%) were 2 bp deletions, which were predominantly at STR and SNMH sequences and substantially enriched for the TNT motif (Extended Data Fig. 8b, c), consistent with the ID-TOP1 mutational signature. Furthermore, mutational signature decomposition for RNase-H2-null CLL cases confirmed the presence of the ID4 signature, most apparent in genic regions (Extended Data Fig. 8d). We therefore conclude that the ID-TOP1 signature is present in human cancer and enriched in tumours that are RNase H2 deficient.

Topoisomerase 1 also causes mutations in RER-proficient cells10,11 (Fig. 1d–f) and is therefore likely to cause mutations in other cancers, with deletions expected to occur most frequently in highly transcribed genes4. Accordingly, analysis of WGS data across cancer types (International Cancer Genome Consortium (ICGC)/Pan-Cancer Analysis of Whole Genomes (PCAWG)) demonstrated that the 2–5 bp deletion rate correlates with expression levels of ubiquitously expressed genes (Pearson’s r = 0.86, P = 0.0014), with deletions markedly elevated in the most highly expressed genes (Fig. 5b), consistent with previous reports of such deletions in certain cancer genes36,37. Examination of 2 bp deletions (42% of 2–5 bp deletions) across cancer types also demonstrated them to be predominantly in STR and SNMH contexts (Extended Data Fig. 8f) and enriched for the TNT sequence motif (Fig. 5c). Furthermore, using a dataset of TOP1 cleavage events captured by TOP1-seq38, we found that 2–5 bp deletions increase in frequency with TOP1 enzymatic activity, with such deletions more prevalent in regions of high TOP1 activity (Fig. 5d). Similarly, TOP1-ID deletion rates also corresponded to TOP1 activity and transcription level, in contrast to all other deletions (Extended Data Fig. 8g, h). Taken together, this establishes a substantial role for TOP1-mediated mutagenesis in the generation of somatic deletions.

To further examine the role of transcription in deletion mutagenesis of cancer genomes, we identified genes that are highly expressed, but only in certain tissues. For prostate adenocarcinoma, highly expressed prostate-restricted genes were significantly enriched for 2–5 bp deletion mutations compared with other genes in this cancer type, as well as the same genes in other cancers (two-tailed Fisher’s exact test, odds ratio = 3.5, P = 2.5 × 10−8 after Bonferroni correction; Extended Data Fig. 8i). Importantly, this analysis considers the same sets of genes between cancer types and therefore rules out sequence composition biases as a confounding effect for elevated ID-TOP1 mutagenesis in highly expressed genes. Extending this approach in an all-versus-all comparison between 8 cancer types and 17 tissues demonstrated specificity between high expression in a tissue of origin and enrichment for 2–5 bp deletions (Fig. 5e). These results extend the relevance of TOP1-mediated mutagenesis to other cancers, confirm the ID-TOP1 mutational signature to be transcription-associated and support the occurrence of TAM in humans.

## TOP1-mediated deletions in the germline

TOP1 is ubiquitously expressed, so we reasoned that it could cause germline as well as somatic mutations. To investigate this possibility, we examined mutations from parent–child trio WGS studies in the Gene4Denovo database39. De novo mutations identified in such datasets represent germline events, as they occur in germ cells or during early embryonic cell divisions. Strikingly, 2–5 bp deletions were the largest category identified, accounting for 33% of the 40,936 de novo indels (Fig. 5f), and the majority of these were compatible with the ID-TOP1 signature. Analysis of 2 bp deletions (41% of 2–5 bp deletions) demonstrated that most occur at SSTR or MH sites (Extended Data Fig. 9a, b), with enrichment of the TNT sequence motif both genome wide and in the context of STR/SNMH sites (Fig. 5g and Extended Data Fig. 9c). Similarly, for 3 and 4 bp deletions, respectively, TNNT and TNNNT motifs were significantly over-represented compared with genome-wide expectation (Extended Data Fig. 9d), supporting sequential TOP1 cleavage and strand realignment as the underlying cause. Consistent with TOP1-TAM aetiology, 2–5 bp deletion and ID-TOP1 deletion frequency correlated with transcript expression in male germ cells (Fig. 5h and Extended Data Fig. 9e). We therefore conclude that the ID-TOP1 mutational signature also occurs in the human germline, implicating TOP1-induced strand realignment mutagenesis as an important mutational process in mammalian cells.

## Discussion

Here we establish a biological basis for the ID4 cancer signature5, experimentally demonstrating that it occurs in RNase-H2-deficient cells both in vitro and in vivo. This implicates TOP1-mediated cleavage at genome-embedded ribonucleotides as its cause. TOP1 is cell-essential in mammals and it is therefore not possible to similarly confirm a genetic dependency on TOP1 in human cells, as has been done in yeast12. However, conservation of this mechanism across eukaryotes is supported by us finding a topoisomerase-1-dependent TNT deletion motif that is present in both yeast and humans, and demonstrating that deletion frequency is dependent on human TOP1 activity levels. Previously published research also provides evidence for TOP1 mutagenesis at ribonucleotide sites in humans. The reversible transesterification reaction of type 1 topoisomerases is conserved from yeast to humans6, and human TOP1 has site-specific activity for ribonucleotides40, causing DNA breaks in mammalian RNase-H2-deficient cells33. Furthermore, the generation of 2 bp deletions through sequential TOP1 cleavage at embedded ribonucleotides has been biochemically reconstituted with both human and yeast enzymes17,31.

We define additional features of this ID-TOP1 mutational signature, with deletions strongly enriched at TNT motifs in both yeast and mammals, a sequence context that is specific to topoisomerase 1 (Extended Data Fig. 7g, h) and deletions that are most frequent in highly transcribed regions. As a consequence, we show that a TAM process that was first identified in yeast10,12,41 is relevant to higher eukaryotes, establishing TOP1-induced mutagenesis as an important process for human variation and disease. Additional signatures associated with topoisomerases or indeed RNase H2 may be identified in the future, particularly given that ID17 has recently been linked to TOP2AK743N cancers42.

The substantial contribution of ID-TOP1 deletions to germline mutagenesis has particular importance given that such deletions will be disproportionately disruptive, particularly in transcribed regions. Notably, such deletions occur in the context of normal RER function, consistent with the mutagenic potential of topoisomerase 1 in physiological wild-type settings10,11 (Fig. 1d). Given that genome-embedded ribonucleotides are the most common endogenous lesion in replicating mammalian cells14, they are the most likely sites of TOP1-TAM mutagenesis, where TOP1 could cleave before their removal by RNase H2-dependent RER. Processing of TOP1cc may be an alternative, less frequent source of 2–5 bp deletions41, but we did not detect ID4 in cancers treated with topoisomerase 1 inhibitor (Extended Data Fig. 8j). The canonical function of TOP1 is to relieve DNA topological stress, arising during both transcription and replication6 (Extended Data Fig. 10). Thus, TOP1-mediated deletions are not restricted to transcribed regions of the genome, with deletions also evident in non-genic regions with high TOP1 activity (Extended Data Fig. 8k). However, overall, enhanced TOP1 activity associated with transcription accounts for more frequent mutagenesis within genes.

Given the essential nature of topoisomerase activity across tissues and cell states, TOP1-mediated mutagenesis probably occurs in many contexts. The frequent TOP1-mediated human germline mutations (Fig. 5i–k) and the identification of ID4 at early embryonic stages43 suggest developmental vulnerability to TOP1-TAM. Moreover, 2–5 bp somatic deletions at SSTRs are also observed at high frequency in non-dividing neurons36, and ID4 has been identified in multiple tumour types5. As such, this mutational process is likely to be important not only in cancers with RER deficiency, but also those with high TOP1 activity and tumours with defects in relevant repair mechanisms, such as enzymes that process TOP1cc6 or non-ligatable TOP1-induced nicks44,45,46. Furthermore, alternative RER pathways may exist47 that could reduce TOP1 mutagenesis. The ID-TOP1 signature may provide a useful biomarker with potential future diagnostic and therapeutic use48, for example, as an indicator of TOP1-induced genome instability targetable by PARP or ATR inhibitors33,49.

In conclusion, alongside its essential role in relieving DNA torsional stress, TOP1 also drives mutagenesis in somatic and germline contexts, relevant to neoplasia, inherited disease and human variation.

## Methods

### Plasmids

A description of all plasmids used in this work is provided in Supplementary Table 1. The S. cerevisiae reporter was generated by DNA synthesis (GeneArt Gene Synthesis, Thermo Fisher Scientific; gBlocks Gene Fragments, IDT) and conventional cloning (restriction, ligation and Quikchange site-directed mutagenesis). The final construct (pTCW12) was used for S. cerevisiae reporter strain construction and fluctuation assays. A Gateway compatible reporter construct for mammalian cells (pTCW14) was generated in a similar manner using a combination of DNA synthesis and conventional cloning strategies. Gateway cloning was then used to move the reporter cassette into pAAVS-Nst-CAG-Dest (a gift from K. Woltjen; Addgene plasmid, 80489)51 to generate pTCW15 for targetting it to the human AAVS1 locus.

### In silico redesign of the hygromycin-resistance gene

To increase the frequency of 2 bp tandem repeats, synonymous substitutions were introduced in the 1 kb hph coding sequence, the Klebsiella pneumoniae hygromycin-resistance gene (HygroR)52. Using Python, a 5-codon (15-base) sliding window was moved one codon at a time, to identify all possible synonymous permutations. Permutations were ranked on the basis of tandem dinucleotide repeat sequence length, with the highest ranking sequences used to replace whole codons, prioritising dinucleotide repeats over mononucleotide repeats. Edited codons were then censored from subsequent permutation. Subsequently, to eliminate stop codons that would arise after a 2 bp deletion or equivalent frameshift mutations, further synonymous changes were made, where possible preserving tandem repeat sequences.

### Yeast strains and growth conditions

All S. cerevisiae strains used in this research (Supplementary Table 2) are isogenic with BY4741 (ref. 53) and were grown at 30 °C. TOP1 and RNH201 open reading frames were deleted using one-step allele replacement using PCR products generated from plasmid templates with selection cassettes (Supplementary Table 2) and primers containing 60-nucleotide homology directly upstream and downstream of the open reading frame. Gene deletions were confirmed by PCR. The 2 bp deletion reporter was inserted at the AGP1 locus using a PCR product amplified from pTCW12 using primers AGP1-MX6-F and AGP1-MX6-R (Supplementary Table 3). The correct reporter insertion was confirmed by PCR and Sanger sequencing. Growth under selection was on YPD medium (10 g l−1 yeast extract, 20 g l−1 bactopeptone, 20 g l−1 dextrose, 20 g l−1 agar) supplemented with hygromycin B (300 mg l−1), nourseothricin (100 mg l−1) and/or G418 (1 g l−1), or on synthetic defined medium (6.7 g l−1 yeast nitrogen base without amino acids, complete supplement single dropout mixture (Formedium), 20 g l−1 dextrose, 20 g l−1 agar).

### Fluctuation assays in yeast

Fluctuation assays were performed as previously described54. Yeast was grown overnight in YPD medium with hygromycin B (300 mg l−1), plated on YPD and grown at 30 °C to obtain individual colonies derived from a single cell without HygroR mutations. For each strain, 16 independent colonies were then used to inoculate 5 ml YPD, and grown for 3 days at 30 °C with shaking at 250 rpm. Cells were pelleted by centrifugation and resuspended in 1 ml of H2O. Undiluted suspensions for each culture were plated (100 μl per plate) on 2 YPD plates supplemented with 1 g l−1 G418, with the exception of rnh201Δ for which a 10−2 dilution was used. Furthermore, each suspension was serially diluted to 10−6 of which 100 μl per plate was spread on 2 YPD plates to estimate the total number of viable cells per culture. Plates were incubated at 30 °C for 2–3 days, and colonies counted. Mutation rates were determined in Microsoft Excel 2016 for each individual culture, and an overall rate for each strain calculated using the Lea Coulson method of the median55. The number of mutants for each culture was ranked, and those ranked 4th and 13th were used to calculate the rates that define the lower and upper limits of the 95% CI56. A single G418-resistant colony for each independent culture was used to determine the spectrum of frame-shift mutations. A 1.3 kb region including HygroR was amplified in two overlapping amplicons (primers S297F and S1113R; S752 and S1658R) using FastStart PCR Master Mix (Roche) and direct colony PCR (5 min at 95 °C; then 35 cycles of 30 s at 95 °C, 30 s at 58 °C and 45 s at 72 °C; and then 45 s at 72 °C). Each amplicon was Sanger sequenced using the primers described in Supplementary Table 3, and analysed using Sequencher v.5.4.6 (Gene Codes Corporation) and/or Mutation Surveyor v.3.30 (SoftGenetics). Mutation rates (per bp) were calculated for 1,032 bp of sequence in which productive frameshift mutations can occur.

### Cell lines

A summary of the human cell lines used in this research is provided in Supplementary Table 4. All cells were grown at 37 °C and 5% CO2, authenticated using STR DNA profiling in the laboratories of origin and shown to be mycoplasma negative through routine testing. HeLa cells (a gift from G. Stewart, University of Birmingham, UK; originally purchased from ATCC) were grown in Dulbecco’s Modified Eagle Medium (DMEM; Gibco/Thermo Fisher Scientific) supplemented with 10% fetal bovine serum (FBS), 50 U ml−1 penicillin and 50 μg ml−1 streptomycin. hTERT RPE-1 cells (a gift from D. Durocher, University of Toronto, Canada; originally purchased from ATCC) were grown in DMEM/F12 medium mixture (Gibco/Thermo Fisher Scientific) supplemented with 10% FBS, 50 U ml−1 penicillin and 50 μg ml−1 streptomycin. The 2 bp deletion reporter was integrated at the AAVS1 safe harbour locus in HeLa cells using a published CRISPR–Cas9 targeting protocol51. HeLa cells were transfected with pXAT2 and pTCW15 in Opti-MEM reduced-serum medium using Invitrogen Lipofectamine 2000 (Thermo Fisher Scientific). After 48 h cells were replated in medium containing 500 µg ml−1 G418 and, after another 48 h and a second round of replating in selective medium, single cells were sorted into 96-well plates using a BD FACSJazz instrument (BD Biosciences). The resulting G418-resistant clones were screened by PCR for reporter integration at the correct locus, retention of integration-free AAVS1 and Sanger sequencing of resulting PCR products. Single-locus integration was confirmed by fluorescence in situ hybridization (FISH) as previously described57, using pTCW16 to generate a fluorescently labelled probe. The full reporter sequence of selected clones was checked, with amplification of a 1.9 kb fragment using Prime Star Max PCR Master Mix (Takara Bio) with the primers HygroR_up and PuroR_rev (40 cycles of 10 s at 98 °C, 15 s at 70 °C and 2 min at 72 °C), followed by Sanger sequencing with additional primers (Supplementary Table 3). To generate RNASEH2A-KO reporter cells, the selected parental HeLa reporter clone was transfected with pMAR526 and pMAR527 (Supplementary Table 1), using Lipofectamine 2000. Then, 48 h after transfection, single EGFP-expressing cells were sorted into 96-well plates and grown until colonies formed. Initial screening was based on PCR amplification (primers RNASEH2A-ex1F and RNASEH2A-ex1R) of the CRISPR–Cas9-targeted region of RNASEH2A with mutations present in selected clones determined by Sanger sequencing. The cellular RNase H2 status was then confirmed by immunoblotting, RNase H2 enzymatic activity assay and alkaline gel electrophoresis to determine ribonucleotide content of genomic DNA (detailed methods are provided below).

### Fluctuation assays in human

Hygromycin resistant HeLa reporter cells (400 µg ml−1 hygromycin B) were recovered from frozen stocks in the absence of selection. The next day, 10 wells of a 96-well plate were seeded with 2,000 cells per well for each line. The experiment was performed with the operator blinded to the identity of the cell lines. Cells were cultured under non-selective conditions and replated subsequently in 24-well, 6-well plates and ultimately T75 flasks, in which they were grown to confluence. Cells were then dissociated using Gibco TrypLE (Thermo Fisher Scientific) and the cells were counted using a Moxi Z automated cell counter. After serial dilution, 1,000 cells were plated into two 10 cm plates for each culture and grown for 14 days to determine plating efficiency. All other cells were plated into two 10 cm plates, and 0.5 µg ml−1 puromycin was added after 4 h, with the medium subsequently changed every 2–3 days for 14 days to remove dead cells and maintain a puromycin concentration of 0.5 µg ml−1.

To establish mutation spectra, colonies were removed by scraping and then cultured in a 96-well plate. When confluent, cells were lysed with 75 μl DirectPCR Lysis Reagent (Viagen Biotech) and 0.4 mg ml−1 PCR-grade Proteinase K (Roche), heating overnight at 55 °C followed by 45 min at 85 °C. Only one sample per independent culture was used for PCR amplification and Sanger sequencing to determine the nature of mutations in the HygroR coding sequence. A 1.24 kb region including HygroR was amplified with Prime Star Max PCR Master Mix (Takara Bio), HygroR_up and H1327R primers (40 cycles of 10 s at 98 °C, 15 s at 70 °C, 2 min at 72 °C). Sanger sequencing was then performed with additional primers (Supplementary Table 3) and mutations identified using Mutation Surveyor v.3.30 (SoftGenetics). All mutants showed double traces of equal height from the point of indel mutations, consistent with the presence of two copies of the reporter in all reporter lines. As FISH indicated the presence of the reporter at a single AAVS1 locus, we inferred that two copies of the reporter were inserted in tandem at this locus. As a 2 bp deletion or equivalent frameshift mutation in either HygroR copy would bring the associated PuroR coding sequence into the translated reading frame, we corrected mutation rate calculations (per bp) for the presence of two copies.

To determine colony numbers, plates were washed with PBS, fixed with 2% formaldehyde in PBS for 10 min, rinsed with water and colonies were stained with 0.1% crystal violet solution for 10 min. Plates were then washed with water and left to dry before counting colonies. After counting, the experiment was performed in an unblinded manner. Mutation rates were determined for each individual culture in Microsoft Excel 2016, and an overall rate for wild-type and KO strains was calculated using the Lea Coulson method of the median. The number of mutants for each culture was ranked, and appropriate ranks56 were used to calculate the rates that define the lower and upper limits of the 95% CI.

### Immunoblotting

Whole-cell extracts (WCE) to determine protein levels of RNase H2 subunits by immunoblotting and for RNase H2 activity assays were prepared as previously described58. Equal amounts of protein from WCE were separated by SDS–PAGE on 4–12% NuPAGE gels and transferred to PVDF. Membranes were probed in 5% milk (w/v; Marvel Original Dried Skimmed), TBS + 0.2% Tween-20 (v/v) with the following antibodies: sheep anti-RNase H2 (raised against human recombinant RNase H2, 1:1,000)14; mouse anti-RNASEH2A G-10 (Santa Cruz Biotechnologies, sc-515475, A1416, 1:1,000); rabbit anti-GAPDH (Abcam, ab9485, 1:2,000, GR3380498-1). For detection, we used rabbit anti-sheep immunoglobulins/HRP (Dako, P04163, 00047199, 1:2,000); goat anti-mouse immunoglobulins/HRP (Dako, P0447, 20039214, 1:10,000); anti-rabbit IgG, HRP-linked antibodies (Cell Signaling Technologies, 7074S, 29, 1:10,000); Amersham ECL Prime Western Blotting Detection Reagent (GE Healthcare Life Sciences) and an ImageQuantLAS4000 device, or IRDye secondary antibodies and an Odyssey CLx Imaging System (LI-COR Biosciences). The uncropped immunoblots are presented in Supplementary Fig. 1.

### RNase H2 activity assays

To assess cellular RNase H2 activity, a FRET-based fluorescence substrate release assay was performed as previously described14. In brief, RNase-H2-specific activity was determined by measuring the cleavage of double-stranded DNA substrate containing a single embedded ribonucleotide. Activity against a DNA-only substrate of the same sequence was used to correct for background activity. Substrates were formed by annealing a 3′-fluorescein-labelled oligonucleotide (GATCTGAGCCTGGGaGCT or GATCTGAGCCTGGGAGCT; uppercase DNA, lowercase RNA) to a complementary 5′-DABCYL-labelled DNA oligonucleotide (Eurogentec). The reactions were performed in 100 μl reaction buffer (60 mM KCl, 50 mM Tris–HCl pH 8.0, 10 mM MgCl2, 0.01% BSA, 0.01% Triton X-100) with 250 nM substrate in black 96-well flat-bottomed plates (Costar) at 24 °C. WCE was prepared as described above, protein concentrations determined using the Bio-Rad Bradford Protein Assay and the final protein concentration per reaction was 50 ng μl−1. Fluorescence was read (100 ms) every 5 min for up to 90 min using the VICTOR2 1420 multilabel counter (Perkin Elmer), with a 480 nm excitation filter and a 535 nm emission filter. Initial substrate conversion after background subtraction was used to calculate RNase H2 enzyme activity.

### Alkaline gel electrophoresis

To determine the presence of excess genome-embedded ribonucleotides in nuclear DNA, alkaline gel electrophoresis of RNase-H2-treated genomic DNA was performed as previously described58. In brief, total nucleic acids were isolated from pellets from around 1 million cells by incubation in ice-cold buffer (20 mM Tris-HCl pH 7.5, 75 mM NaCl, 50 mM EDTA) with 200 µg ml−1 proteinase K (Roche) for 10 min on ice, followed by addition of N-lauroylsarcosine sodium salt (Sigma-Aldrich) to a final concentration of 1%. Nucleic acids were extracted using phenol–chloroform, then isopropanol-precipitated and dissolved in nuclease-free water. For alkaline gel electrophoresis, 500 ng of total nucleic acids was incubated with 1 pmol of purified recombinant human RNase H2 (isolated as previously described59) and 0.25 µg of DNase-free RNase (Roche) for 30 min at 37 °C in 100 µl reaction buffer (60 mM KCl, 50 mM Tris–HCl pH 8.0, 10 mM MgCl2, 0.01% Triton X-100). Nucleic acids were ethanol-precipitated, dissolved in nuclease-free water and 250 ng was separated on 0.7% agarose gels in 50 mM NaOH, 1 mM EDTA. After overnight electrophoresis, the gel was neutralised in 0.7 M Tris–HCl pH 8.0, 1.5 M NaCl and stained with SYBR Gold (Invitrogen). Imaging was performed on a FLA-5100 imaging system (Fujifilm), and densitometry plots were generated using AIDA Image Analyzer v.3.44.035 (Raytest).

### Mutation-accumulation experiment

TP53-KO hTERT RPE-1 cells without and with loss-of-function mutations in RNASEH2A or RNASEH2B, introduced by CRISPR–Cas9 genome editing, a gift from D. Durocher, have previously been described33. RNase-H2-proficient (wild type), RNASEH2A-KO and RNASEH2B-KO cells were single-cell sorted into 96-well plates using the BD FACSJazz instrument (BD Biosciences). Multiple individual clones for each were expanded to confluent T75 flasks for cryopreservation and genomic DNA isolation of these ancestral populations. Moreover, lines were again single-cell sorted into 96-well plates to start the mutation-accumulation experiment. Cultures were expanded by subsequent growth in 24-well, 6-well plates and T75 flasks until confluent (approximately 25 population doublings), and this process of single-cell sorting and expansion was repeated four more times, providing bottlenecks to capture mutations that occurred since the previous sort. From the first to the last single-cell sort, a total of approximately 100 population doublings occurred and the final culture was expanded for cryopreservation and genomic DNA isolation of these end-point populations.

Genomic DNA was isolated using phenol extraction as previously described58, for alkaline gel electrophoresis and WGS. Library preparations and sequencing were performed by Edinburgh Genomics. Libraries were prepared using Illumina SeqLab specific TruSeq PCRFree High Throughput library preparation kits according to the manufacturer’s instructions, with DNA samples sheared to a mean insert size of 450 bp. Libraries were sequenced using paired-end reads on the Illumina HiSeqX instrument using v2.5 chemistry to achieve minimum mean genome-wide sequencing depth of 30× per sample.

### Mouse WGS analysis

Villin-cre+Trp53fl/flRnaseh2bfl/fl mice with epithelial-specific deletion of Trp53 and Rnaseh2b on a C57Bl/6J background have been described previously29. Animal experiments were conducted with appropriate permission, in accordance with guidelines for animal care of the Christian-Albrechts-University, in agreement with national and international laws and policies. No randomization or blinding was performed. Paired tumour–normal DNA was isolated from small intestinal tumours (Trp53−/−Rnaseh2b−/−) and liver tissue (Trp53+/+Rnaseh2b+/+) from female mice (aged 52 weeks), using the Qiagen DNeasy Blood & Tissue Kit. Library preparations and sequencing were performed by Edinburgh Genomics using Illumina DNA PCR-Free Library Prep according to the manufacturer’s instructions. Paired-end sequencing was performed by Edinburgh Genomics on a NovaSeq 6000 using v1.5 chemistry. Mean genome-wide sequencing depth of at least 30× for liver samples and 60× for tumour samples was obtained.

### S. cerevisiae WGS analysis

WGS SRA files for rnh201Δ pol2-M644G S. cerevisiae25 from the NCBI Sequence Read Archive (SRA) were converted to FASTQ files using SRA Toolkit v.2.5.4-1 (SRA Toolkit Development Team; http://ncbi.github.io/sra-tools/). FASTQ reads were aligned to the GSE56939_L03_ref_v2 reference genome60 (Supplementary Table 5) and sorted BAM files were created using BWA-MEM (v.0.7.12)61, and deduplicated with SAMBLASTER (v.0.1.22)62. To select high-quality indel variants, GATK (v.3.6-0) Haplotype Caller (without base quality score recalibration)63 variant calling was performed with ‘Hard Filters’ (--filterExpression “QD < 2.0 || FS > 200.0 || ReadPosRankSum < −20.0”) . Filtering for strain-specific variants was performed as previously described60, with minor modifications. The filters were as follows: (1) eliminate variants shared with an ancestral clone; (2) required ≥20 reads for variant allele in descendent; (3) exclusion of repetitive sequences as defined in ref. 60; and (4) reference/variant depth ratio 0.4–0.6, <0.4 if homozygous variant allele.

### RPE-1 WGS analysis

FASTQ files were converted to unaligned BAM format and Illumina adaptors were marked using GATK (v.4.1.9.0) FastqToSam and MarkIlluminaAdapters tools64. Reads were aligned to the human genome (hg38, including alt, decoy and HLA sequences) using BWA-MEM (v.0.7.16)61 and read metadata were merged using GATK’s MergeBamAlignment tool. PCR and optical duplicate marking and base quality score recalibration were performed using GATK. Variants from NCBI dbSNP build 151 were used as known sites for base quality score recalibration. Post-processed alignments were genotyped using Mutect2, Strelka2, Platypus and SvABA using somatic calling models for each pair of ancestral and end-point cultures, as described below.

### Mouse WGS analysis

FASTQ processing and alignment were performed as for RPE-1 WGS analysis, using the GRCm38 mouse genome reference and known variant sites from the Mouse Genomes Project65 (REL-1807-SNPs_Indels) for base quality score recalibration. Somatic variant calling of post-processed alignments was performed using Mutect2, Strelka2, Platypus and SvABA for each tumour-liver pair, as described below. Somalier v.0.2.12 (https://github.com/brentp/somalier) was used to confirm each paired tumour and liver sample originated from the same animal.

### Human ethics approval

Data generated from Genomics England 100,000 genomes and ICGC-CLL studies were analysed. In these respective studies, informed consent for participation was obtained. Ethical approval for Genomics England 100,000 genomes project: East of England and South Cambridge Research Ethics Committee; CLL-ICGC: International Cancer Genome Consortium (ICGC) guidelines from the ICGC Ethics and Policy committee were followed and the study was approved by the Research Ethics Committee of the Hospital Clínic of Barcelona.

### CLL WGS analysis

#### Genomics England

CLL tumour–normal pairs (n = 198) were processed as part of the 100,000 Genomes Project (pilot and main programme v8). Samples were sequenced using the Illumina HiSeq X System with 150 bp paired-end reads at a minimum of 75× coverage for tumours and 30× coverage for germline samples. Reads were mapped to GRCh38 using ISAAC aligner (v.03.16.02.19)66. Single-nucleotide variants (SNVs) and indels were called using Strelka v.2.4.7 using somatic calling mode. Structural and copy number variants were called using Manta (v.0.28.0) and Canvas (v.1.3.1)67, respectively. Samples with a tumour purity estimate from Canvas of less than 50% were excluded from analysis. RNASEH2B copy number was determined using a combination of Canvas, Manta, read depth counts with samtools (v.1.9) and confirmed by manual inspection using IGV (v.2.5.0)68.

#### ICGC

WGS from the ICGC-CLL cohort35 (n = 150) was reanalysed. Raw reads were mapped to the human reference genome (GRCh37) using BWA-MEM (v.0.7.15)61. BAM files were generated, sorted and indexed, and optical or PCR duplicates were flagged using biobambam2 (https://gitlab.com/german.tischler/biobambam2, v.2.0.65). Copy-number alterations were called from WGS data using Battenberg (cgpBattenberg, v.3.2.2)69, ASCAT (ascatNgs, v.4.1.0)70, and Genome-wide Human SNP Array 6.0 (Thermo Fisher Scientific) data35 reanalysed using Nexus 9.0 Biodiscovery software (Biodiscovery). RNASEH2B copy number was established by combining the three analyses and manual review with IGV.

### Colorectal cancer WGS analysis

Irinotecan-treated (n = 39) and irinotecan-untreated (n= 78) colorectal cancers from the 100,000 Genomes Project Colorectal Cancer Domain were 1:2 matched using a multivariate greedy matching algorithm without replacement, implemented in the Matching R-package71. Matching was conducted considering sex, age at sampling, whether a primary tumour or metastasis had been sequenced, microsatellite instability status, and whether the individual had previously received radiotherapy, oxaliplatin, capecitabine or fluorouracil treatment.

### Somatic variant calling

Somatic variant calling was performed in parallel using four distinct methods: Mutect2 (as part of GATK v.4.1.9.0)72,73, Strelka2 (v.2.1.9.10)74, SvABA (v.1.1.3)75 and Platypus (v.0.8.1)76. High-confidence indel calls were defined as the intersected output of these four tools, where variants passed all filters for ≥3 of 4 callers. The intersection was performed using the bcftools (v.1.10.2)77 isec function after normalizing variant calls and left-aligning ambiguous alignment gaps using the bcftools norm function. For Platypus (v.0.8.1)76, joint calling all samples in each cohort was performed before filtering for somatic variants; the other variant callers were run in paired tumour–normal mode. For the RPE-1 mutation-accumulation experiment the end-point and ancestral cultures were defined as ‘tumour’ and ‘normal’ samples, respectively. Variant filtering strategies were optimized to both available information on segregating genetic variation for humans and mice, and the functionality of each calling method as detailed below.

#### Mutect2

Unfiltered genotypes for all normal samples were combined to filter germline variants. Somatic calls were obtained using GATK’s FilterMutectCalls command. Human polymorphism data and allele frequencies from gnomAD78 were provided to Mutect2 for the filtering of germline variants.

#### SvABA

Germline indel and structural variants were filtered using the --dbsnp-vcf and --germline-sv-database options. Mouse indels were obtained from Mouse Genomes Project version 5 SNP (ftp://ftp-mouse.sanger.ac.uk/REL-1505-SNPs_Indels/mgp.v5.merged.indels.dbSNP142.normed.vcf.gz); structural variants from SV release version 5 (ftp://ftp-mouse.sanger.ac.uk/REL-1606-SV/mgpv5.SV_insertions.bed.gz and ftp://ftp-mouse.sanger.ac.uk/REL-1606-SV/mgpv5.SV_deletions.bed.gz). Human indels were extracted from NCBI dbSNP build 151 and common structural variants from dbVAR (https://hgdownload.soe.ucsc.edu/gbdb/hg38/bbi/dbVar/).

#### Strelka2

Candidate small indels for each pair were first generated by Manta (v.1.6.0)79 in somatic calling mode. Strelka2 was then executed in somatic calling mode for each pair with Manta’s candidate small indels output provided to the --indelCandidates option.

#### Platypus

Germline variants were filtered on the basis of any normal sample with ≥2 variant allele reads. Somatic variant calls for each sample pair were retained if tumour/end-point sample > 2 variant reads; site depth > 9; and normal sample read depth ≥ 20, <2 variant reads. Moreover, a >10× ratio of tumour to normal for variant/total depth was required.

For Genomics England CLL tumour–normal pairs, pre-existing Strelka2 calls from the 100,000 Genomes Project pipeline were used, while variant calling with Mutect2, Platypus and SvABA was performed as above. Colorectal cancer tumour–normal pairs from Genomics England were processed as for Genomics England CLL but without Mutect2 analysis. For ICGC CLL, somatic indels were called using Mutect2 (GATK v.4.0.2.0)72,73, Strelka2 (v.2.8.2)74, SvABA (v.1.1.0)75 and Platypus (v.0.8.1)76. Candidate small indels generated by Manta (v.1.2)79 were used as input for Strelka2. Mutect2, Strelka2 and SvABA were run in paired tumour–normal mode. somaticMutationDetector.py (https://github.com/andyrimmer/Platypus/tree/master/extensions/Cancer) was used to identify somatic indels called by Platypus with a minimum posterior of 1. SNVs called by Platypus were considered to be somatic if they had at least 2 alternative reads in the tumour, fewer than 2 alternative reads in the normal, a minimum tumour VAF of 10× the control VAF, and a minimum depth of 10.

### Germline mutation analysis

De novo WGS variants were downloaded from the Gene4Denovo database (Supplementary Table 5). Reference assembly conversion errors were removed by discarding variants for which the reference allele did not match the genome reference at the given position or for which the variant position was greater than the length of the reference chromosome. Furthermore, individuals with total de novo variants below the 10th (n = 33) or above the 90th (n = 140) percentile were excluded. For germline gene expression we used predefined expression groups80 based on Ensembl release 90 annotation (ftp://ftp.ensembl.org/pub/release-90/gtf/homo_sapiens/Homo_sapiens.GRCh38.90.gtf.gz). Initially stratified as nine expression groups from 1 (unexpressed) to 9 (high), we collapsed them into a smaller set of unexpressed (1), low (2, 3, 4), mid (5, 6, 7) and high (8, 9). The annotations were converted to GRCh37 coordinates using liftover (kent source v.417). Genomic segments overlapping multiple distinct expression groups, due to overlapping genes, were assigned to the higher of those expression groups. For each expression group, we summed the count (c) of de novo indels contained within the genomic span of those genes. This was converted to rate estimates by dividing by the union genomic span (g nucleotides) of genes in that expression group, and adjusting for the number of mutated genomes considered (n); rate = c/(gn). To obtain 95% CIs, gene selection was bootstrapped (sampled to an identical number with replacement) 100 times and the 0.025 and 0.975 quantiles of the bootstrapped rate calculation taken as the 95% CI.

### ICGC pan-cancer expression analysis

The ICGC PCAWG somatic mutations50 (https://dcc.icgc.org/api/v1/download?fn=/PCAWG/consensus_snv_indel/final_consensus_passonly.snv_mnv_indel.icgc.public.maf.gz) and ICGC PCAWG ‘baseline’ gene expression data50 were obtained (ArrayExpress, https://www.ebi.ac.uk/arrayexpress/experiments/E-MTAB-5200/). Genomic annotation of gene extents on the GRCh37 reference genome match the Ensembl version 75 annotation (http://ftp.ensembl.org/pub/release-75/gtf/homo_sapiens/Homo_sapiens.GRCh37.75.gtf.gz) of the ICGC gene expression calls. Mean, median and maximal gene expression (transcripts per million (TPM)) were calculated for each gene across the 76 ICGC baseline gene expression tissues/samples. Only genes annotated on the main autosomal chromosomes, 1 to 22, and the X chromosome were considered. Overlapping genes were removed, retaining only the most abundantly (highest median, then mean in the case of ties) expressed genes from overlapping pairs. This filtering was applied hierarchically, starting with the most abundant. Following ref. 81, genes with housekeeping-like expression were defined as those with maximal expression of less than ten times median expression. Housekeeping-like genes were decile-binned into expression groups on the basis of median expression. Mutations were stratified by type (1 bp deletion, 2–5 bp deletion) or by the ‘TN*T’ motif defined below and counted by intersection with the annotated genomic extents of genes in each expression group.

For the analysis of tissue-biased gene expression, the 76 ICGC baseline samples were grouped by annotated tissue (such as breast, prostate, kidneys, liver) and matched where possible to the tissue of origin for ICGC cancer types. For each tissue, the median expression (in TPM) of each gene was calculated for (1) within-tissue samples and (2) for all other samples. The 90th quantile of gene expression (q90, top 10%) within a tissue was set as a threshold for high level expression. Genes with high expression in a tissue (1) but a median expression of less than q90*0.1 in the other tissues (2) were considered to be highly expressed but tissue restricted (HETR). For the set of HETR genes from a tissue, we counted the number of 2–5 bp deletions within the annotated genomic extent of the HETR genes in a cancer type of interest. We similarly counted 2–5 bp deletions in all other genes for that cancer type, and counted both the HETR and non-HETR 2–5 bp deletions from all other cancer types within the ICGC cohort. For each cancer–tissue pair, this provided four sets of counts, analysed using two-tailed Fisher’s exact tests using the R function fisher .test. A positive odds ratio indicated enrichment of 2–5 bp deletions in the HETR genes, compared with a background of the remainder of the ICGC cohort in which HETR genes are not highly expressed. For each cancer type considered, this test was repeated for each tissue type (n = 17). Analyses were carried out for eight of the ICGC cohort cancer types that met the combined criteria of having a well-matched and known tissue of origin among the ICGC baseline samples, and requiring the cancer type cohort to have at least n= 2,500 2–5 bp deletions in aggregate. This represents n= 17 × 8 = 136 statistical tests, adjusted by Bonferroni correction. Odds ratios (r) for mutation depletion were transformed to their reciprocal (1/r) for display purposes.

### ICGC pan-cancer TOP1-seq analysis

Data corresponding to two replicates of TOP1-seq, a modified ChIP–seq technique to immunoprecipitate only catalytically engaged TOP1 (ref. 38), were downloaded from the NCBI Gene Expression Omnibus database (accession code GSE57628, samples GSM1385717 and GSM1385718). Autosomal chromosomes 1 to 22 and the X chromosome were divided into 1 kb bins and, for each bin, the amount of mappable sequence was determined using Umap’s regions mappable using 36-mers82 to approximate the read length of the TOP1-seq data. For each 1 kb window, the TOP1-seq signal within mappable regions was summed for each replicate and the mean signal was calculated. This mean was divided by the amount of mappable sequence to calculate the TOP1-seq signal per bp and each 1 kb window was then assigned to decile bins using this value.

Somatic deletion calls from ID4-positive PCAWG samples (as defined in https://dcc.icgc.org/api/v1/download?fn=/PCAWG/mutational_signatures/Signatures_in_Samples/SP_Signatures_in_Samples/PCAWG_SigProfiler_ID_signatures_in_samples.csv) were counted within the same 36-mer mappable regions for each 1 kb window and either stratified by type (1 bp deletion, 2–5 bp deletion) or by the TN*T motif defined below. Relative rates of deletions in each category were calculated relative to the first TOP1-seq signal decile.

### Mutational signatures

De novo extraction and decomposition of mutational signatures was performed in Python v.3.8.5 using SigProfilerExtractor (v.1.1.0)5, along with SigprofilerMatrixGenerator (v.1.1.14/1.1.15)83 and SigprofilerPlotting (v.1.1.27). The recommended default settings (including 500 NMF replicates) were applied (https://github.com/AlexandrovLab/SigProfilerExtractor). Subtraction of mutations in RPE-1 wild-type cells from those detected in RNase H2 null cells was performed as follows. The average number of indels per line for each of the 83 categories was determined for the three wild-type lines. Counts per category for AKO and BKO lines were subtracted using these averages, with negative values set to 0. SigProfilerExtractor was then performed on the resulting WT-subtracted AKO and BKO ID-83 matrices for both de novo signature detection and decomposition analysis.

### Indel sequence context analysis

WGS indels were categorized on the basis of repeat sequence context. Genome-wide occurrence of short repeats and regions of MH were identified and filtered to include only the mappable genome, defined by Umap’s regions mappable using 100-mers82. For both WGS-identified indel variants and genome-wide occurrence, scoring of 2 bp deletions compliant with the TNT motif at MH/SSTR sites required the deleted bases to match the sequence NT with a T immediately 5′ of the deleted dinucleotides. More generally, for varying sized deletions these were considered to fit a TN*T motif if the deletion lay within an SSTR or region of MH containing the motif TN(d – 1)T where d is the length of the deletion. Genome-wide occurrences were estimated from 100,000 randomly generated deletions of given lengths within the mappable genome. For SSTRs and MH regions, all regions containing the respective motifs (TN(r – 1))n or TN(r – 1)T were identified (where r is the length of the repeat unit and n > 1), and the fraction of SSTR/MH sequence containing TNT motifs was determined against total SSTR/MH sequence in the mappable genome.

To derive a null expectation for de novo deletions matching the TNT, TNNT and TNNNT motif for 2, 3 and 4 bp deletions, respectively, deletions at repeats from the Gene4Denovo database were first classified by deletion length, repeat type (STR or MH) and repeat length. Bootstrap samples of corresponding repeats from the genome were generated with 1,000 replicates. That is, for each deletion category an equal number of repeats of matching repeat type, repeat unit length and total repeat length were randomly drawn from the genome for each bootstrap sample.

### Sequence logos

Genomic sequences containing 2 bp deletions were reversed and complemented when the deleted dinucleotide contained an adenosine (A), except when the dinucleotide was AT or TA. For SNMH and STR deletions, the position of the deleted dinucleotide cannot be unequivocally assigned and, therefore, the deleted sequence was right-aligned in the repeat/MH region, either to the most 3′ T, where present, or otherwise to the limit of the repeat/MH region. Sequences were converted to bit score matrices and logos were drawn using Logomaker (v.0.8)84.

### Embedded ribonucleotide sequence context analysis

EmRiboSeq data from rnh201Δ yeast prepared during mid-log phase growth85 were obtained (Supplementary Table 5) and aligned to the sacCer3 reference genome as previously described to identify the genomic coordinates of genome-embedded ribonucleotides86. The Bedtools (v.2.30.0)87 utilities groupby, slop and getfasta were used to extract and count the sequence context of genome-embedded ribonucleotides with downstream analysis and plotting implemented in R (v.4.0.5). Genome sequence composition-adjusted relative rates were calculated as previously described32 such that, under the null expectation of no sequence bias in ribonucleotide incorporation, all sequence contexts have an expected relative rate of 1/n where n is the number of contexts considered.

### Statistical methods

Statistical testing was performed using GraphPad Prism v.9.1.1, Python v.3.8.5 or R v.3.3.1. Two-sided non-parametric Mann–Whitney U-tests were performed for quantitative measurements; multiple testing correction, FDR set at 0.05; and, for categorical data, Fisher’s exact tests were performed in Python using stats.fisher_exact from scipy v.1.6.3. Calculation of cosine similarities was performed as follows. Mutations for each strain were converted into a vector, with ordered values representing different mutation categories as a proportion of total mutations. These were then compared in a pairwise manner. Given two vectors A and B, the cosine similarity (cos(θ)) was calculated as:

$$\cos (\theta )=\frac{{\sum }_{i=1}^{n}{{\bf{A}}}_{i}{{\bf{B}}}_{i}}{\sqrt{{\sum }_{i=1}^{n}{({{\bf{A}}}_{i})}^{2}}\sqrt{{\sum }_{i=1}^{n}{({{\bf{B}}}_{i})}^{2}}}$$

Hierachical clustering used the hclust function of R (v.4.1.0) with complete linkage clustering of pairwise cosine distances (1 − cosine similarity) between ID-83 mutation spectra, with 41 categories of productive reporter frameshift mutations. For bootstrap support, n = 1,000 bootstrap datasets were generated by sampling with replacement the mutations observed with a strain, for each strain, and then calculating the cosine distance and hierarchical clustering for each bootstrap dataset. Reported bootstrap scores are the percentage of bootstrap replicates hierarchical clustering of which supports the clustering to the right of the indicated position.

To test the significance of cosine similarities, we used a null model based on the Dirichlet-multinomial distribution. In brief, when comparing two mutation count vectors, with total mutations m1 and m2, over n mutation classes, we constructed a distribution of cosine values by comparing 10,000 simulated pairs of random vectors generated as follows. For each simulated pair, we sampled from a Dirichlet-multinomial distribution with the concentration parameters as a vector of ones of dimension n, and number of trials as m1 for the first vector in the pair, and m2 for the second vector. The null distribution was obtained by computing the cosine similarity of the 10,000 pairs of mutation count vectors.

### Reporting summary

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