A redefined InDel taxonomy provides insights into mutational signatures

Abstract

Despite their deleterious effects, small insertions and deletions (InDels) have received far less attention than substitutions. Here we generated isogenic CRISPR-edited human cellular models of postreplicative repair dysfunction (PRRd), including individual and combined gene edits of DNA mismatch repair (MMR) and replicative polymerases (Pol ε and Pol δ). Unique, diverse InDel mutational footprints were revealed. However, the prevailing InDel classification framework was unable to discriminate these InDel signatures from background mutagenesis and from each other. To address this, we developed an alternative InDel classification system that considers flanking sequences and informative motifs (for example, longer homopolymers), enabling unambiguous InDel classification into 89 subtypes. Through focused characterization of seven tumor types from the 100,000 Genomes Project, we uncovered 37 InDel signatures; 27 were new. In addition to unveiling previously hidden biological insights, we also developed PRRDetect—a highly specific classifier of PRRd status in tumors, with potential implications for immunotherapies.

This paper describes a framework for classifying small insertions and deletions from genomic data and applies it to a large dataset comprising seven tumor types. The analysis highlights new insertion and deletion signatures and a classifier of postreplicative repair dysfunction.

Main

Small insertions and deletions (InDels; <100 bp) represent the second most prevalent form of genetic variation after substitutions¹. InDel mutagenesis is substantial and nonrandom, reflecting underlying mutational processes, be they benign by-products of normal physiology or malign consequences of exogenous exposures and/or endogenous dysfunctions.

While studies of mutational processes over the last decade have focused primarily on substitutions^2,3, recent advancements in InDel detection and annotation^4–6 have led to the identification of 18 small InDel signatures (IDS) in human cancers⁷. These signatures were defined using an 83-channel classification system (that is, 83 InDel subtypes, herein referred to as COSMIC-83), founded upon characteristics such as the InDel size, nucleotides affected, lengths of flanking mononucleotide/polynucleotide repeat and sequence homology at the InDel junctions. Subsequent re-analysis of the same dataset using a revised algorithm reported nine additional de novo IDS⁸.

Accurate InDel characterization is crucial for biological and clinical purposes. For instance, a high proportion of microhomology-mediated deletions is a key predictor of clinically actionable homologous recombination deficiency, carrying the greatest weight in homologous recombination deficiency detection and classification algorithms^9,10. Additionally, the detection of microsatellite instability (MSI) in mismatch repair (MMR)-deficient tumors relies on measuring genome-wide InDel mutagenesis and/or analyzing panels of mononucleotide/dinucleotide repeats^11–13. Recent studies have also highlighted how 2–5 bp short deletions and 2–4 bp duplications are distinct readouts of TOP1 activity (amplified in RNaseH2-null cells)¹⁴ and TOP2A dysfunction¹⁵, respectively.

Given the increasing importance of InDel mutagenesis in tumor classification and prediction of therapeutic sensitivities^16–19, we established a ‘ground truth’ set of experimental IDS focused on ‘postreplicative repair deficiency’ (PRRd), which encompasses defects in MMR and replicative polymerase proofreading—biological abnormalities that often display exquisite sensitivity to immune checkpoint inhibition (ICI)^20–22. We uncovered inherent limitations in the prevailing InDel classification schema (COSMIC-83) that hamper its ability to distinguish biologically distinct signatures. To address this, we explored whether incorporating sequences 5′ and 3′ to an InDel–features crucial for substitution classification–could contribute additional understanding and resolution power. Additionally, sequence motifs known to increase mutational vulnerability and their genome-wide prevalence were factored into our proposition. Our approach unambiguously classifies every InDel into a specific subcategory. Here we demonstrate that our alternative InDel taxonomy uncovers new etiologies of InDel mutagenesis, offering mechanistic insights and potential clinical added value.

Results

Diversity of InDel patterns in PRRd

We generated a ‘ground truth’ set of isogenic cellular models by introducing CRISPR edits to key PRRd-associated genes in an hTERT-immortalized RPE1 (TP53^−/−) cell line²³. We created four single MMR gene knockouts (ΔMLH1, ΔMSH2, ΔMSH3 and ΔSETD2 (ref. ²⁴)), two knock-in missense mutants of polymerase Pol ε (POLE exonuclease mutant p.P286R and p.L424V), two mutants of Pol δ (POLD1 exonuclease mutant p.S478N and polymerase mutant p.R689W) and three double mutants with combined polymerase proofreading mutation and MMRd (POLD1^S478N/+ΔMLH1, POLD1^S478N/+ΔMSH2 and POLE^P286RΔMSH2; Supplementary Tables 1 and 2). Successfully edited clones were propagated in culture for approximately 45–50 days to permit mutation accumulation. Subsequently, two to five daughter subclones were isolated per genotype for whole-genome sequencing (WGS) and mutational signature analyses (Fig. 1a).

Fig. 1. Isogenic PRRd human cell lines exhibit distinct InDel patterns.

a, Mutation accumulation experiment in TP53-null hTERT-immortalized retinal pigment epithelial cell (hTERT-RPE1^TP53-null, herewith referred to as the background control). b, InDel burden and average InDel fold increase of CRISPR gene edits (n = 2–5 subclones per genotype; Supplementary Tables 1–3). Red dashed line represents the mean InDel burden of control subclones. The y axis shows the InDel burden in log scale. c, Distinguishing COSMIC-83 InDel profiles of edited subclones from background control. Light blue error bars depict the mean ±3 s.d. of cosine similarities between n = 100 bootstrapped InDel profiles of unedited controls and the background profile (Extended Data Fig. 1d) aggregated from n = 7 unedited subclones. The x axis shows the InDel count in log scale. d, COSMIC-83 InDel mutational signatures associated with gene edits following background subtraction (Supplementary Table 4). e, Key features of COSMIC ID1, ID2 and ID7 (v.3.3). f, Heatmap of cosine similarities between gene-edit IDS and COSMIC IDS (v.3.3). Known, proposed etiologies are annotated above the heatmap (blue). g, Decomposed solution of gene-edit InDel signatures in d into COSMIC IDS (v.3.3).

Except for ΔSETD2, we observed elevated InDel burdens in all gene edits compared to an unedited control (background) (Fig. 1b and Supplementary Table 3). The mutation burden was approximately twofold higher in ΔMSH3 and POLD1^R689W, tenfold in POLD1^S478N, POLE^L424V and POLE^P286R, 55-fold in ΔMSH2 and ΔMLH1 and particularly significant in combined gene edits—around 200-fold in POLD1^S478N/+ΔMLH1 and POLE^P286RΔMSH2, and 300-fold in POLD1^S478N/+ΔMSH2.

All lines except ΔSETD2 showed variations in their COSMIC-83 InDel signature profiles compared to control (Fig. 1c and Supplementary Table 4). We noted discriminative characteristics between gene edits (Fig. 1d and Extended Data Fig. 1). Dominant 1 bp T deletions at homopolymers of 6 bp or more (poly-T₆₊) were observed for ΔMLH1, ΔMSH2 and ΔMSH3, while POLD1^S478N and POLE^P286R showed exclusive 1 bp T insertions at poly-T₅₊. POLD1^R689W, POLE^L424V and all three combined polymerase/MMRd edits predominantly exhibited 1 bp T insertions at long homopolymers, although not exclusively, with variations of 1 bp T deletions between different genotypes. Together, these experiments revealed unique, diverse InDel signatures among different PRRd mutants. Remarkably, mutations within the same gene but affecting different functional protein domains manifested signature variations (that is, POLD1 exonuclease p.S478N versus polymerase p.R689W).

Extended Data Fig. 1. COSMIC-83 InDel signatures of PRR deficiency in isogenic human cell line models.

a, UMAP clustering of de novo COSMIC-83 InDel profiles of all gene-edit subclones (n = 37). UMAP, uniform manifold approximation and projection for dimension reduction. b, Cosine similarities between individual gene-edit subclones and the averaged InDel profile of background controls (d). c, Relative proportion of seven InDel mutation types of each gene-edit. Ins, insertion; Del. deletion; R, repeat; M, microhomology; C, cytosine; T, thymine. d, Aggregated InDel profile of hTERT-RPE1^TP53-null background control. e, Cosine similarities among experimentally generated COSMIC-83 PRRd InDel signatures (Fig. 1d). f, Unsupervised clustering of experimentally generated COSMIC-83 PRRd InDel signatures with cosine distance. Dashed vertical line marks 0.1 cut-off.

We also examined the substitution profiles of all gene edits (Extended Data Fig. 2 and Supplementary Table 5). Intriguingly, MMRd lines showed lower substitution-to-InDel ratios compared to control, while polymerase-dysfunction (Pol-dys) lines exhibited markedly increased ratios (Extended Data Fig. 2h). This suggests that genome instability is predominantly driven by an excess of InDel mutagenesis in MMRd, whereas substitution mutagenesis plays a more significant role in polymerase proofreading dysfunction. Furthermore, mutational asymmetry analyses revealed enrichment of both substitutions and InDels on the leading strand for POLE mutants while POLD1 mutants exhibited lagging strand bias, specifically T insertions at homopolymeric tracts of 5–7 nts (Supplementary Fig. 1 and Supplementary Table 6). This is in keeping with the hypothesized preferential activity of Pol ε and Pol δ in leading and lagging strand synthesis, respectively^25,26, suggesting that POLE/POLD1 mutants tend to accumulate 1 bp A insertions on the nascent strand while replicating through 5–7 nts poly-T-tracts^27,28. This lends support to the proposition that polymerase ε and δ are more proficient at detecting incorrectly paired bases at template adenines²⁷.

Extended Data Fig. 2. Substitution mutational signatures associated with PRR deficiency in isogenic human cell line models.

a, De novo substitution burden, and average substitution fold-increase of CRISPR gene edits and unedited control (n = 2–5 subclones per gene edit; Supplementary Table 3). Red dashed line represents the mean substitution burden of control. b, Distinguishing substitution profiles of edited subclones from background controls. Light green error bars depict the mean ±3 s.d. of cosine similarities between n = 100 bootstrapped substitution profiles of unedited controls and the background profile aggregated from n = 7 unedited subclones. The x axis shows the substitution count in log scale. c, UMAP clustering of de novo substitution profiles of all subclones (n = 37). UMAP, uniform manifold approximation and projection for dimension reduction. d, Relative proportion of six substitution mutation types of each edit genotype. e, Substitution signatures associated with PRRd gene edits following background subtraction. f, Cosine similarities among experimentally generated PRRd substitution signatures in e. g, Cosine similarities between experimentally generated PRRd substitution signatures and cancer-derived SBS reference signatures (RefSig)³. h, Averaged SNV to InDel ratio of gene edits. Bar represents the averaged ratio for each genotype.

Nevertheless, while the diversity of experimental InDel profiles was appreciable among PRRd genotypes, it was difficult to disambiguate gene-edit signatures from background mutagenesis. Clustering analyses and direct comparisons of gene edit and control InDel profiles showed extremely high similarity (cosine similarity > 0.9; Extended Data Fig. 1a,b,d). Discrimination between MMRd and Pol-dys signatures was also limited (Extended Data Fig. 1e). Unsupervised clustering using cosine distance revealed mainly two groups of signatures—deletion-driven MMRd signatures and insertion-driven polymerase mutant signatures (Extended Data Fig. 1f). We thus investigated the sufficiency of COSMIC-83 taxonomy, given that signal variation among the ten gene edits was primarily observed in two channels—1 bp T insertions at poly-T₅₊ and 1 bp T deletions at poly-T₆₊.

Limitations of current InDel taxonomy

We compared experimental gene-edit InDel signatures with COSMIC IDS⁷. InDel signatures of ΔMSH2 and ΔMLH1 showed no similarity to the purported MMRd-associated ID7 (Fig. 1f). Instead, ΔMSH2 and ΔMLH1 signatures most resembled ID1 and ID2, ascribed to normal replication errors associated with nascent and template strand slippage, respectively (Fig. 1d–g).

The COSMIC-83 taxonomy aggregates 1 bp InDels at homopolymers >5 bp into single channels (that is, T₆₊ for deletions and T₅₊ for insertions, respectively; Fig. 1e and Extended Data Fig. 3a–d). Yet, the probability of microInDel formation increasing with the length of simple nucleotide repeats in MMRd²⁹ is a recognized hallmark of MSI³⁰. We surmised that the conflation of discriminatory signals within longer homopolymers into single ‘insertion at T₅₊’ or ‘deletion at T₆₊’ channel likely reduces the separative capacity for signature extraction. Hence, MMRd signatures cannot be distinguished from signatures of normal replication errors. This contrasts with corresponding PRRd-associated substitution signatures, which manifest as distinct and diverse patterns amongst MMRd and/or Pol-dys cancers^3,7,31 (Extended Data Fig. 2b–g).

Extended Data Fig. 3. InDel frequency distribution across pan-cancer datasets and the consolidation of 476 full InDel channels into a redefined 89-channel InDel classification framework.

a, Aggregated COSMIC-83 InDel profile of ICGC samples, with (left, n = 3,323) and without (right, n = 3,287) hypermutator samples. b, Aggregated 89-channel InDel profile of ICGC samples, with (left, n = 3,323) and without (right, n = 3,287) hypermutator samples. c, Percentage of total InDels per channel for ICGC cohort, with COSMIC-83 (left) or 89-channel (right) format. d, Percentage of total InDels per channel for Hartwig cohort, with COSMIC-83 (left) or 89-channel (right) format. e, Aggregated 476-full-channel InDel profiles of ICGC, Hartwig and GEL cohorts. f, Final, consolidated 89-channel InDel profiles of aggregated samples from GEL cohort, with (left, n = 11,585) and without (right, n = 10,792) hypermutator samples.

Notably, ID7 lacks signal within reputedly the most informative homopolymer channel (>5 bp). Instead, signals are only present in channels associated with ID1 and ID2 (Fig. 1e), resulting in systematic misattribution of all MMRd gene-edit signatures to ID1 and ID2 (Fig. 1f,g). Moreover, InDel signatures of POLE, POLD1 mutants and all combined polymerase/MMRd edits were indistinguishable from ID1 using COSMIC-83 taxonomy and sometimes indistinguishable from each other (Extended Data Fig. 1e). The signature of polymerase mutant POLD1^R689W did not resemble any reported signatures. Because InDel mutagenesis of gene edits occurred predominantly at longer homopolymers and was erroneously assigned to ID1 and/or ID2 (Fig. 1g), we explored whether expanding on the long homopolymer channels and modifying the information presented in individual InDel channels could improve the resolution to disentangle the ostensibly alike but distinct biological signatures without compromising the power for signature extraction.

A new framework for classifying InDels

As with substitutions, incorporating surrounding sequence characteristics may enhance the discriminatory capacity of InDel catalogs for signature analyses³². We first classified InDels according to whether they were insertions, deletions or complex InDels (simultaneous insertions and deletions; Fig. 2a). Within insertions and deletions, InDels were subclassified by motif size (1 bp versus ≥2 bp). For 1 bp InDels, we considered the nucleotide content (C/G versus A/T motifs), the 5′ and 3′ flanking bases and the length of homopolymeric tracts. For InDels ≥2 bp, we identified the maximally repetitive motif within the InDel and accounted for its repeat length in the 3′ sequence (Supplementary Note 1). For deletions occurring with microhomology at the InDel junction, we considered the deletion motif length (L) and the microhomology length (M). This comprehensive taxonomy yielded 476 nonoverlapping InDel subcategories (channels; Supplementary Table 7 and Supplementary Note).

Fig. 2. Redefined InDel taxonomy improves discriminatory power and reveals differential InDel patterns associated with PRR gene edits.

a, Proposed InDel classification schema and an example 89-channel InDel profile of an ΔMSH2 subclone. b, Distinguishing 89-channel InDel profiles of edited subclones from background control. Light blue error bars depict the mean ± 3 s.d. of cosine similarities between n = 100 bootstrapped InDel profiles of unedited controls and the background profile (Extended Data Fig. 5b) aggregated from n = 7 unedited subclones. The x axis shows the InDel count in log scale. c, Cosine similarities of edited subclones and bootstrapped controls in COSMIC-83 InDel profiles against 89-channel InDel profiles. Two-tailed Wilcoxon signed-rank test, P = 1.917 × 10⁻⁷). d, The 89-channel InDel mutational signatures associated with PRRd gene edits following background subtraction (Supplementary Table 4; https://signal.mutationalsignatures.com/explore/main/experimental/experiments?study=7). Ins, insertion; Del, deletion.

We examined whether all 476 channels were informative. By analyzing the InDel distribution across all channels in 18,522 tumors covering most cancer types from the International Cancer Genome Consortium (ICGC)/The Cancer Genome Atlas (TCGA)³³, Hartwig³⁴ and the Genomics England (GEL) 100,000 Genomes Project³⁵ (Extended Data Fig. 3e), we identified noninformative channels (that is, channels with no signal) and consolidated those with low signal to reduce the total number of InDel channels to 89 (Fig. 2a, Extended Data Fig. 3f and Supplementary Table 8). Overall, compared to COSMIC-83, the 89-channel taxonomy expands upon channels that had most of the signals, here 1 bp A/T InDels, into a larger array of channels, and condenses longer InDels and/or genome motifs infrequent in the genome (where signals were scant or nonexistent) into fewer InDel subcategories (Extended Data Fig. 4). Although the final numbers are not vastly different between the two classification systems, our data-driven approach, incorporating sequence contexts and enhancing signal distribution of mononucleotide/polynucleotide repeat tracts into additional channels, provides alternative information to the mutational signature extraction and assignment process, potentially increasing the likelihood of detecting new biologically meaningful signatures.

Extended Data Fig. 4. Channel mapping between COSMIC-83 and 89-channel InDel classification methods using InDel data from the ICGC cohort (n = 3,323).

There is no simple, direct 1-to-1 mapping between the methods. In general, the 89-channel taxonomy expands upon channels that had most of the signals, here 1 bp A/T InDels, into a larger array of channels, and condenses longer InDels and/or genome motifs infrequent in the genome (where signals were scant or nonexistent) into fewer InDel subcategories. NM, not mapped.

To test this, we applied the new 89-channel InDel taxonomy to our ground truth gene-edit dataset (Supplementary Table 4). Cosine similarities between experimental InDel profiles and control were much lower with the 89-channel format than with COSMIC-83 (Fig. 2b,c and Extended Data Fig. 5a,b), indicating that the new classification improved separation of gene edits from the background (mean cosine similarity, 0.68 ± 0.08 for 89-channel versus 0.89 ± 0.11 for COSMIC-83; two-tailed Wilcoxon signed-rank test, P = 1.917 × 10⁻⁷). We subsequently determined signatures associated with each gene edit using the 89-channel format. The resultant signatures showed more evenly distributed signals across the entire 89-channel profile (Figs. 1d, 2d and Extended Data Fig. 5c). Gene-edit signatures were also more readily discernible from one another (mean signature pairwise cosine similarity, 0.57 ± 0.25 for 89-channel versus 0.64 ± 0.3 for COSMIC-83; two-tailed Wilcoxon signed-rank test, P = 1.483 × 10⁻⁵; Extended Data Fig. 5d,e). Notably, InDel signatures of combined MMRd/polymerase mutants were not simply the sum of the individual mutational processes, likely reflecting the biological interactions of Pol ε and Pol δ with MMR in suppressing InDel formation during the replication of repetitive DNA.

Extended Data Fig. 5. The 89-channel InDel signatures of post-replicative repair (PRR)-deficient isogenic human cell line models.

a, UMAP clustering of de novo 89-channel InDel profiles of gene-edit subclones (n = 37). UMAP, uniform manifold approximation and projection for dimension reduction. b, Aggregated 89-channel background InDel profile of hTERT-RPE1^TP53-null control subclones (n = 7). c, Gini index of experimentally generated InDel signatures (n = 10) in COSMIC-83 versus 89-channel format. Two-tailed Wilcoxon signed-rank test, P = 0.001953. d, Cosine similarities among experimentally generated 89-channel PRRd InDel signatures (Fig. 2d). e, Unsupervised clustering of experimentally generated 89-channel PRRd InDel signatures using cosine distance. Dashed vertical line marks 0.1 cut-off.

Interestingly, we noted that while MMRd deletions are particularly amplified at longer homopolymers (8–9 bp > 5–7 bp > 0–4 bp), polymerase mutants displayed strikingly different distribution of excess insertion mutagenesis at shorter homopolymers (5–7 bp > 8–9 bp > 0–4 bp; Fig. 2d). The higher InDel rates in shorter homopolymers conferred by defective proofreading of Pol ε and Pol δ likely reflect the distance over which they interact with duplex DNA upstream of the polymerase active site²⁹. Indeed, crystal structures of Pol ε and Pol δ have shown numerous contacts made within 5–7 bp of the polymerase active sites with duplex DNA^36,37, with experimental model reinforcing this optimal distance²⁹, explaining how proofreading may offer reduced protection against InDels outside of this ‘footprint’ (that is, unpaired bases further upstream of the active site³⁸; longer runs where MMR plays a more crucial role²⁹). These unique insights were only appreciable due to the new 89-channel format, offering enhanced capturing of biological variation.

To compare the discriminatory capacity of both classification systems, we also performed de novo signature extraction on our ground truth experimental dataset (n = 37; Extended Data Fig. 6a). With COSMIC-83, only two de novo signatures were extracted—one dominated by T insertions at poly-T₅₊ (ID83A) and the other by T deletions at poly-T₆₊ (ID83B; Extended Data Fig. 6b). In contrast, the 89-channel format yielded four signatures, matching our expectation of a predominantly deletion-driven MMRd signature (InD89B), a predominantly insertion-driven polymerase signature (InD89D) and two distinct signatures with differing proportions of InDels (InD89A and InD89C), likely reflecting the combined polymerase/MMRd phenotypes (Extended Data Fig. 6c).

Extended Data Fig. 6. De novo signature extraction of COSMIC-83 versus 89-channel InDel catalogs demonstrates differential signature recovery.

a, Signature selection plot of de novo extraction by SigProfilerExtractor³⁹ of n = 37 experimental samples in COSMIC-83 or 89-channel InDel catalogs. Suggested solutions are shaded in orange or purple for COSMIC-83 and 89-channel catalogs, respectively. b, Suggested de novo solution (two signatures) by SigProfilerExtractor for COSMIC-83 experimental cohort (n = 37). c, Suggested de novo solution (four signatures) by SigProfilerExtractor for 89-channel experimental cohort (n = 37). d, Signature selection plots of de novo extraction by signature.tools.lib³, SigProfilerExractor³⁹ and MuSiCal⁸ of n = 52 ICGC colorectal cancers in COSMIC-83 versus 89-channel InDel catalogs. Suggested solutions are shaded in orange or purple for COSMIC-83 and 89-channel InDel catalogs, respectively. e, Selected de novo solution (five signatures) from signature.tools.lib for ICGC colorectal cancer COSMIC-83 catalogs. f, Selected de novo solution (eight signatures) from signature.tools.lib for ICGC colorectal cancer 89-channel catalogs. g, Suggested de novo solution (five signatures) by SigProfilerExtractor for ICGC colorectal cancer COSMIC-83 catalogs. h, Suggested de novo solution (six signatures) by SigProfilerExtractor for ICGC colorectal cancer 89-channel catalogs. Signatures from MuSiCal extraction are not shown.

Finally, to determine whether this observed relationship between channel information content and signature extraction extended to other datasets and workflows, we applied three different algorithms^3,8,39 to an unrelated cohort of 52 colorectal WGS from ICGC³³ (Extended Data Fig. 6d). All three algorithms failed to discern all available signatures using COSMIC-83, reaching a discrimination limit of five, yielding sparse signatures with signal density highly concentrated in two channels (Extended Data Fig. 6d,e,g). Contrarily, the 89-channel format consistently enabled the detection of more de novo signatures across all algorithms used (Extended Data Fig. 6f,h). The extracted signatures also displayed signals across more channels, highlighting the superior performance of the 89-channel classification over COSMIC-83 in uncovering additional, true mutational processes.

New InDel signatures (InDs) in seven cancer types

To explore the impact of our new InDel taxonomy on signature discovery beyond PRRd phenotypes in human cancers, we analyzed seven tumor types (n = 4,775) known to display clinically relevant high tumor mutational burden (TMB) due to a range of abnormalities (for example, MMRd, environmental ultraviolet (UV) radiation, APOBEC-related mutagenesis)—bladder (n = 347), brain (CNS, n = 392), colorectal (n = 2,146), endometrial (n = 695), lung (n = 958), stomach (n = 181) and skin (n = 56) cancers from the GEL 100,000 Genomes Project³⁵ (Fig. 3a).

Fig. 3. De novo signature extraction using redefined InDel taxonomy uncovers 37 InDS in seven cancer types in the GEL cohort.

a, InDel burden across seven cancer types (n = 4,775; left) and the number of mutations contributed by each InD to the GEL tumors. The size of each dot represents the proportion of samples of each tumor type that shows the mutational signature. The color of each dot represents the median mutation burden (per Mb) of the signature in samples that show the signature. b, Profiles of 37 consensus InDel mutational signatures (InDS) extracted and curated from seven GEL cancer cohorts (Supplementary Table 10; https://signal.mutationalsignatures.com/explore/main/cancer/signatures?mutationType=3&study=7). Putative etiologies are provided in the top-left squircles. N-Slip, nascent strand slippage; T-Slip, template strand slippage; NHEJ, nonhomologous end joining.

We performed mutational signature analysis per tumor type as previously described³ (Fig. 3a, Extended Data Fig. 7, and Supplementary Tables 9–11; Methods). We identified 37 consensus InDel signatures, referred to as InDS (to distinguish from COSMIC IDS; Fig. 3b). Ten signatures shared characteristics mappable to known IDS (InD1, InD2a, InD3a/InD3b, InD4a, InD6, InD8, InD9a, InD13 and InD18)⁷. The remaining 27 were new.

Extended Data Fig. 7. Hierarchical clustering and dendrogram partitioning of n = 111 tissue-specific signatures into n = 40 distinct patterns.

Hierarchical clustering and dendrogram partitioning of n = 111 tissue-specific signatures into n = 40 distinct patterns using cosine distance (1 − cosine similarity) as distance metric, with a cut-off of 0.15. The 40 distinct patterns were manually reviewed, revised and inspected for mixed patterns to produce the final 37 consensus InDS (Fig. 3 and Supplementary Tables 9–11; Methods).

Exogenous exposures underlie five InDS. InD3a and InD3b often co-occurred in lung cancers with tobacco exposure. InD3a/InD3b clustered with experimental signatures induced by benzo(a)pyrene and its metabolite benzo(a)pyrene diol epoxide (Extended Data Figs. 8 and 9), supporting the notion that they represent modulated versions of tobacco-related DNA damage. InD13, characterized by T deletions at TT dinucleotides, is linked to UV damage, and InD18, found exclusively in colorectal samples, is due to colibactin exposure⁴⁰. InD32 was identified in samples with prior exposure to platinum and was associated with a new platinum-associated signature, SBS112 (ref. ³).

Extended Data Fig. 8. InDel signatures of mutagen exposures and DNA repair/replication gene knockouts in human induced pluripotent stem cells, as well as proposed etiology for InD9a.

a, InDel signatures of exposures to environmental mutagens in human induced pluripotent stem cells (hiPSC). Data reanalyzed from previous study⁵². b, InDel signatures of DNA repair/replication gene knockouts in human induced pluripotent stem cells. Data reanalyzed from previous study⁴⁶. c, Extended sequence context of 1 bp C deletions of InD9a showed TTCT/TTCA enrichment for APOBEC deamination. APOBEC, apolipoprotein B mRNA-editing enzyme, catalytic polypeptide. d, Proposed etiology for APOBEC-associated InD9a. APOBEC, apolipoprotein B mRNA-editing enzyme, catalytic polypeptide; TLS, translesion synthesis; UNG, Uracil-N-glycosylase.

Source data

Extended Data Fig. 9. Heatmap clustering of n = 26 experimental InDel signatures and n = 37 cancer-derived InDS.

Heatmap clustering of n = 26 experimentally generated signatures and 37 cancer-derived InDS. Experimentally generated signatures include n = 10 from the current study, n = 10 from the mutagen study⁵² (Extended Data Fig. 8a) and n = 6 from iPSC knockout study⁴⁶ (Extended Data Fig. 8b). Putative sources (left) and etiologies (right) of the signatures are annotated if known. HRd, homologous recombination deficiency; PAH, polycyclic aromatic hydrocarbons; N-Slip, nascent strand slippage; T-Slip, template strand slippage; MMRd, mismatch repair deficiency.

Twenty InDS had probable endogenous origins (Extended Data Fig. 9). Several have been described, including InD1 and InD2a, errors associated with nascent and template strand slippage during normal DNA replication, respectively⁷. InD1 and InD2a were seen universally across all tumor types except CNS and skin cancers, which showed a tissue-specific variant, InD2b (Fig. 3a). InD4a is attributed to TOP1 transcription-associated mutagenesis¹⁴. InD6, marked by microhomology-mediated deletions, is associated with deficiency in HR repair⁷. InD8, which had deletions with little to no microhomology at deletion junctions, likely reflects the footprint of nonhomologous end-joining activity and/or radiotherapy⁴¹.

InD9a, correlated with SBS2 and SBS13 hypermutation, featured 1 bp C deletions at TCT and TCA (mutated base underlined), identical to mutable motifs characteristic of SBS2/SBS13, particularly at short poly-T tracts. It was presumptively induced by APOBEC (Extended Data Fig. 8c), corroborated by experimental evidence from an APOBEC overexpression DT40 model⁴². We proposed a mutagenesis mechanism wherein following C-to-U deamination at TCT by APOBEC, uracil removal by UNG leaves an uninformative abasic site. Template strand slippage can then occur at this short repetitive T tract, leading to a C deletion (Extended Data Fig. 8d). For reasons currently unclear, we also found similar C-deletion-dominated InD9b/InD9c, which, although resembled InD9a, lacked the predilection for a preceding T, and was possibly caused by an alternative mechanism.

Interestingly, we extracted eight gene-specific MMRd and Pol-dys InDS. MMRd-InD7 contrasts with COSMIC ID7. InD7 is characterized by the expected excess of 1 bp and 2 bp deletions, particularly at longer mononucleotide/dinucleotide repeat tracts. InD7 clustered with experimental signatures of ΔMLH1, ΔMSH2 and ΔMSH6 (Extended Data Fig. 9). We also identified InD19 (due to PMS2 deficiency), InD14 (associated with POLD1 exonuclease mutations), InD15 (associated with POLE exonuclease mutations), InD16a and 16b (resulting from concurrent loss of POLE proofreading and MMR), InD21 (associated with combined POLD1 proofreading defect and MMRd) and InD20, which we found through experimental investigations, was due to MMRd occurring on a POLE dysfunction background.

The remaining 12 signatures were of uncertain etiology. Five were probably artifacts—InD27 and InD28 often co-occurred, incurring thousands of InDels, and were related to SBS57, potentially an amplification or a sequencing artifact⁷. InD28m was likely a mixed signature of InD28 and InD4, remaining to be resolved with larger cohorts. InD5 and InD10 were ubiquitous and possibly artifacts.

While C insertions dominated both InD26 and InD30 at poly-C tracts followed by a 3′A, InD30 C insertions induced thousands of insertions at homopolymers CCC and CCCC, whereas InD26 C insertions mainly occurred at longer CCCCC and were not associated with hypermutation.

Three InDS (InD31, InD24 and InD12) showed striking correlations with signatures of other classes. InD31 displayed distinct C deletions at short homopolymers (<5 bp) followed by 3′G and T deletions at short homopolymers (<5 bp) followed by 3′A. It was only reported in samples with novel rare SBS105 (ref. ³), and often co-occurred with InD8. InD24 deletions peaked prominently at GTA and GTG and were strongly correlated with DBS8, which shows double substitutions at the same motifs (TGTG > TAGG/TTGG). InD12 exhibited C deletions between dinucleotides AA and AT and was associated with DBS25 featuring a tall peak at TT dinucleotide. Despite clear co-occurrence, the causes for these signatures remain cryptic.

InD4b and InD29 shared features with InD4a and InD8, respectively. Whether they represented tissue-specific variants, were mixed or caused by different mechanisms requires further investigation. InD11 appeared related to InD1 and might be an oversplit signature frequently enriched in high InDel burden samples, such as those with MMRd and Pol-dys. Seen in bladder and colorectal cancers, InD23 showed a striking pattern of longer insertions (≥5 bp) at nonrepeats. These insertions were almost exclusively tandemly duplicated from immediate neighboring sequences. InD33 was most prominent in one CNS tumor treated with temozolomide; however, its etiology remains unknown.

In summary, 5 InDS were of likely exogenous origins (InD3a, InD3b, InD13, InD18 and InD32), 20 were endogenous (InD1, InD2a, InD2b, InD4a, InD4b, InD6, InD7, InD8, InD9a, InD9b, InD9c, InD11, InD14, InD15, InD16a, InD16b, InD19, InD20, InD21 and InD29) and 12 had uncertain sources (InD5, InD10, InD12, InD23, InD24, InD26, InD27, InD28, InD28m, InD30, InD31 and InD33).

A signature-based classifier of PRR dysfunction

PRRd subtypes, typified by MSI, are clinically actionable with potential selective sensitivity to immunotherapies^20–22. Current methods of detecting PRRd mainly rely on immunohistochemistry (IHC) staining of MMR proteins (but not for polymerase mutants) and/or PCR-based assays to determine MSI at selected genomic loci. These assays are not sensitive or robust enough, especially in nonepithelial tissues¹⁶. Using insights from this study, we therefore explored constructing a classifier for tumor PRRd stratification, reporting MMRd, Pol-dys and mixed MMRd/Pol-dys as distinct classes versus PRR proficiency.

We used 571 GEL cancers assigned as MMRd (n = 214), Pol-dys (n = 36), mixed MMRd/Pol-dys (n = 41) or PRR-proficient (controls, n = 280) based on confirmed causal genotypes and allelic status, and/or supporting IHC staining (Fig. 4a and Supplementary Table 12). Samples treated as controls had neither MMRd and/or Pol-dys confirmed through the lack of driver mutations in key MMR genes (that is, MLH1, MSH2, MSH6, and PMS2), POLE, POLD1, and displayed no evidence of MSI associated with these abnormalities⁴³. We trained multiple multinomial elastic net regression models applying 7:3 partitioning iteratively across the dataset. Through exploring all possible features/models (Supplementary Table 13), we identified exposures of SBS and InDS associated with MMRd, Pol-dys and mixed MMRd/Pol-dys, as well as the ratio of total InDels to substitutions as the most predictive features (Fig. 4b and Supplementary Table 14; Methods). The final model, termed PRRDetect (postreplicative repair detect), was retrained on the entire dataset (n = 571). Then, in an independent validation cohort of 504 ICGC breast cancers^44,45 and 847 GEL cancers, for which the true labels of PRRd were known, PRRDetect achieved an AUROC (area under the ROC curve) of 1 and an AUPRC (precision–recall curve) of 0.99 at distinguishing PRR-dysfunctional from PRR-proficient samples, performing superiorly to other MSI/MMRd detection tools, including MSIseq⁴³, MMRDetect⁴⁶ and TMB—an approved biomarker for immunotherapies^{20,21,47–49} (Fig. 4c,d and Supplementary Tables 12, 15 and 16).

Fig. 4. PRRDetect improves the detection of tumors with PRR dysfunction.

a, Simplified workflow of the development of PRRDetect classifier. (1) Initial exploratory training using 571 ground truth samples. (2) Final retraining to produce the PRRDetect classifier. -ve ctrl, negative control. b, Distribution of coefficients across seven genomic features contributing to the final PRRDetect classifier. Green error bars depict the mean ± s.d. from ten replicates of training in cross-validation. Red dots indicate the final coefficients chosen for each class prediction (Supplementary Table 14). c, Validation and application of PRRDetect on independent cancer cohorts. d, ROC curves demonstrating the superior performances of PRRDetect on independent cancer cohort (n = 1,351) against alternative biomarker strategies. P values were calculated using two-sided nonparametric test based on the bootstrap distribution (10,000) of the difference in AUCs⁵³. MMRDetect, P < 2.2 × 10⁻¹⁶; MSIseq, P = 6.617 × 10⁻¹⁵; TMB, P < 2.2 × 10⁻¹⁶. e, PRRDetect results of n = 1,335 ICGC and Hartwig cancers, ordered from the lowest to the highest prediction probability across the x axis (left to right) for MMRd (purple), combined MMRd/Pol-dys (blue) and Pol-dys samples (orange). Negative samples were ordered by TMB in increasing order from left to right. Results of MSIseq, MMRDetect, cancer gene driver annotation and cancer tissue origin are labeled at the bottom tracks. Dashed rectangle highlights the extent of false positive overcalling if using TMB > 10 mutations per Mb as a cutoff. f, Concordance of calls among TMB-high (>10 mutations per Mb), positive exposure to SBS signatures that impart hypermutation and PRRDetect prediction across n = 1,335 ICGC and Hartwig cancers. g, Concordance of calls among TMB-high (>10 mutations per Mb), positive exposure to SBS signatures that impart hypermutation and PRRDetect prediction across n = 4,775 GEL tumors. muts, mutations.

Next, to survey the prevalence of PRRd across alternative cancer cohorts, we applied PRRDetect on seven cancer types commonly enriched with hypermutator samples from ICGC³³ and Hartwig³⁴ (Fig. 4c,e,f, Extended Data Fig. 10a,b and Supplementary Table 17). PRRDetect predicted 3.7% (50/1,335) samples as PRR-dysfunctional, correctly identifying all Pol-dys, MMRd/Pol-dys samples and missing two subclonal MMRd samples (based on available published driver information for PRRd status). MSIseq missed 6 of 43 PRRDetect-predicted MMRd, 2 of 4 mixed MMRd/Pol-dys cases while displaying poor concordance for detecting pure Pol-dys cases (that is, missed all 7 cases). Unsurprisingly, PRRDetect captured all MMRDetect-positive cases. However, MMRDetect failed to identify all PRRd cases as it was not designed to detect Pol-dys/mixed phenotypes and missed seven MMRd samples. Crucially, we noted that many PRRDetect-positive cases did not have an associated driver mutation identified (33/50). This is clinically significant. Of 50 PRRDetect-positive cases, 39 were MMRd (only 8 had an associated driver mutation), 7 were Pol-dys (all had driver mutations in polymerase proofreading domains) and 4 were predicted as mixed MMRd/Pol-dys (2 had POLE exonuclease mutations and none had MMR drivers). If PRRDetect predictions were all true and sequencing approaches focused exclusively on identifying driver events associated with these deficiencies were used, a significant proportion of cases (66%) could be missed.

Extended Data Fig. 10. Prediction results of PRRDetect on n = 1,335 cancers from ICGC/Hartwig selected cohorts and n = 4,775 GEL cancers, stratified by tissue types.

a, Venn diagrams showing the concordance and discordance between different predictors in selected ICGC and Hartwig cohort (n = 1,335). b, PRRDetect prediction of ICGC and Hartwig seven cancer types (n = 1,335; Supplementary Table 17). c, PRRDetect prediction of GEL seven cancer types (n = 4,775; Supplementary Table 12).

Given that PRRd cancers often present with high TMB, and TMB is used as a biomarker for immunotherapies, we explored the limits of TMB-based patient stratification. With an FDA-approved TMB cutoff of 10 mutations per Mb⁴⁹, just over a tenth of 459 cases classified as TMB-high (50/459, 10.9%) had predicted PRR dysfunction (Fig. 4f, Extended Data Fig. 10b and Supplementary Table 17). The majority of other cases (353/459, 76.9%) had high TMB from tobacco, UV and APOBEC exposure; 56 (12.2%) were due to alternative causes. Thus, across independent cancer cohorts where MMRd and Pol-dys are known to occur at higher frequencies, ~89% of the samples classified as TMB-high may not have the intrinsic biological underpinnings associated with response to immunotherapies, with implications for the use of TMB as a selective biomarker for ICI^50,51.

We asked whether this trend extended to the larger GEL cohort (n = 4,775). Among the 1,371 TMB-high cases, nearly half (677, 49.4%) were predicted as having MMRd and/or Pol-dys (Fig. 4g), of which, only ~50% of them had an identified driver. The remaining 564 (41.1%) had high TMB due to alternative mutagenic exposures; 130 (9.5%) were due to other undetermined causes. Furthermore, beyond revealing PRR dysfunction in typical tumor types such as colorectal cancers (19%, 400/2,146) and uterine cancers (37%, 255/695), PRRDetect predicted PRRd in a small but notable proportion of stomach (11/181, 6%), bladder (3/347, 1%), CNS (3/392, 1%) and lung cancers (8/958, 1%; Extended Data Fig. 10c and Supplementary Table 12). This reinforces two important clinical points—first, PRRd is not restricted to colorectal and uterine cancers despite being more prevalent in these tumor types; second, WGS can serve as a tumor-agnostic assay uncovering PRRd and any other actionable biological abnormalities in the future.

Discussion

The ability to distill biologically relevant signatures is heavily influenced by how mutations are represented or classified, more so than the underlying algorithms used for signature extraction. Here we showed that a classification schema that aggregates potentially discriminatory signals into only a few channels and/or does not take surrounding sequence context into account is limited in its ability to discern biologically insightful InDel patterns, irrespective of the extraction algorithms used. Consequently, some of the currently reported InDel signatures may correspond to multiple mutational processes, affecting the specificity of their assignments. To overcome this limitation, we proposed an alternative InDel taxonomy that incorporates flanking sequence context and distributes signals across a broader set of channels, offering increased discriminative capability without sacrificing power for signature extraction. Using this framework, we captured the distinct MSI phenotypes and true biological diversity of PRRd-associated InDel patterns, evident in both isogenic cellular models and patient tumors. Indeed, these InDel signatures mirror the diverse PRRd-associated SBS signatures in cancers^3,7,31.

Furthermore, we deciphered 37 consensus InDS from seven cancer types. We confirmed ten previously described IDS^7,14,42, including those associated with tobacco use, UV exposure and APOBEC activity and reported eight new InDS of MMRd and polymerase proofreading dysfunction. While we have offered putative causes and associations for several new signatures, our current understanding of InDel mutagenesis remains incomplete. Future studies incorporating more cancer types and/or larger sample cohorts will help uncover additional signatures and illuminate new etiologies. The possibility of adapting the taxonomy in the future, to include features currently not explorable due to the limitations incurred by technological error rates of calling InDels using short-read WGS (that is, at longer simple repeats), could also be revealing.

Our classifier, PRRDetect, is highly sensitive and specific. It utilizes both SBS and InD signatures to stratify tumors by PRRd subtypes and, to our knowledge, is the only tool with this capability. Importantly, we found a lack of concordance between the current biomarkers of MSI/MMRd and the true biological state. Particularly, TMB, despite being FDA approved, is nonspecific^50,51. This has profound clinical implications as more than 50% of TMB-high (>10 mutations per Mb) cancers arise from biological abnormalities and environmental exposures that have no substantiated biological basis for immunotherapies, potentially impacting patient outcomes. PRRDetect can also detect samples that have signatures of PRR deficiency but for which no drivers can be detected (nearly 50% of all PRRd cases). A limited sequencing assay would have simply missed a substantial portion of these tumors. Finally, our classifier does not distinguish between MMRd genotypes despite clear differences between, for example, MLH1, MSH2, MSH6 and PMS2; currently, there is no clinical indication to do so. However, should it become clinically important to distinguish between these genes, it shall be possible to do so.

In summary, our study highlights how mutation classification directly impacts the accuracy of signature analysis. Our decision to leverage the surrounding sequence context for classifying InDels stems from mechanistic work demonstrating the relationship between InDel formation and flanking 3′ and 5′ sequences^32,52. Nevertheless, optimal classification remains an active research area. Alternative schema could unveil additional mutational processes in the future. Unraveling the landscape of InDel mutagenesis through the refined framework described here will hopefully translate into meaningful benefits for cancer patients.

Methods

The experiments described herein did not require approval from a specific ethics board.

Cell lines

All cell line models generated and used in this study are provided in Supplementary Table 1. All cells were grown in DMEM/F12 medium (Gibco/Thermo Fisher Scientific) supplemented with 10% FBS, at 37 °C and 5% CO₂ in a humidified incubator. The original hTERT-RPE1 ΔTP53 cell was generated from a previous study²³. To generate the remaining isogenic CRISPR-edited cell lines, 200k RPE1 ΔTP53 cells per each edit were electroporated with preformed ribonucleoprotein complex (RNP with final concentrations of 120 pmol gRNAs and 100 pmol Alt-R Cas9) in supplemented SE buffer, using nucleocuvette and AMAXA 4D-Nucleofector (Lonza) on program EH-158 according to the manufacturer’s instructions. Following electroporation, cells were replated into fully supplemented DMEM/F12 medium to recover for 48 h. For knock-ins, homology-directed repair (HDR) donor oligos were supplied with the RNP for electroporation, and cells were replated for recovery in medium spiked in with final concentrations of 2 μM M3814 (Selleckchem) and 0.5 μM Alt-R HDR Enhancer (Integrated DNA Technologies) for the first 24 h. All cells were then cultured for additional 2 to 4 days to allow for gene editing and eventually subjected to limiting dilution on 96-w plates to isolate single-cell clones. Targeted clones were screened for successful frameshift edits for knockouts or incorporation of intended missense mutations for knock-ins via Sanger sequencing. All gRNAs, sequencing primers and genotypes of cell lines generated in this study are summarized in Supplementary Tables 1 and 2.

Mutation accumulation experiment and sequencing

Double mutants were harder to establish but had similar doubling time to single mutants. Edited clones were cultured for 45–55 days (~40 to 50 doublings) to allow for mutation accumulation before a second round of single-cell limiting dilution was performed to isolate two to three daughter subclones per edit genotype, providing a bottleneck to capture mutations that had occurred since the isolation of the initial edited parental clones.

Genomic DNA was isolated from all samples using Quick-DNA Miniprep Plus Kit (ZymoResearch) following the manufacturer’s protocol. WGS libraries were prepared and sequenced (150 bp, paired end, 25×) on the Illumina NovaSeq 6000 platform by Novogene.

Somatic variant calling

Short reads were aligned to GRCh38/hg38 using BWA-MEM 0.7.17-r1188. Substitutions and InDels in genetic reference clones and daughter subclones were called as previously described⁴⁶. Postprocessing filters were applied to improve the specificity of mutation-calling. Specifically, for single-nucleotide variant (SNV) calls by CaVEMan⁵⁴ (v.1.13.15), we used CLPM = = 0 and ASMD ≥ 140. To reduce false positive calls by Pindel⁵⁵ (v.3.2.0), we used QUAL ≥ 250 and REP < 10. Cell clones with an average variant allele frequency of <0.4 were designated as polyclonal and excluded from all subsequent quantitative analyses (that is, estimation of mutation burden). A variant allele frequency filter of 0.2 was applied to substitutions and InDels. De novo substitutions and InDels in subclones were obtained by subtracting from respective parental clones whenever available, or by removing mutations shared among subclones. De novo mutation counts are provided in Supplementary Table 3. Rearrangements were not analyzed as they were too few to be informative.

Mutational signature analysis of experimental samples

The derivation of gene edit-associated mutational signatures with cosine similarity, bootstrapping and background subtraction was performed using published framework (https://github.com/xqzou/COMSIG_KO)⁴⁶. In short, we first (1) determined the background mutational signature in background control by aggregating the unedited and untreated subclone mutational profiles, (2) determined the difference between the mutational profiles of the edited clones and background mutation profile and if an edit generates a signature, we (3) removed background mutation profile from the mutation profile of the edited subclone (Supplementary Tables 4 and 5).

We also used Uniform Manifold Approximation and Projection (UMAP)⁵⁶ to cluster the InDel profiles. Experimentally derived signatures were compared to published reference signatures^3,7 using signature.tools.lib (v.2.4.4) from https://rdrr.io/github/Nik-Zainal-Group/signature.tools.lib/.

Replicative strand asymmetry

Replicative strand bias analysis was performed for 1 bp InDels only. InDels were mapped to leading or lagging strand using Repli-seq data (MCF-7) of the ENCODE project⁵⁷. All 1 bp InDels were aligned with pyrimidines as the mutated base, following the convention in the field. This facilitated the identification of the InDel’s corresponding strand. IntersectBed⁵⁸ in BEDTools (v.2.26.0-114-g4c407ce) was utilized to identify mutations overlapping specific genomic features. To assess a specific mutational signature, the ‘expected’ ratio of InDels between lagging and leading strands was calculated according to the distribution of the repeats in these regions. The ‘observed’ ratio of InDels among different strands was determined by mapping InDels to genomic coordinates of all leading/lagging regions. The asymmetry between different strands was quantified by calculating the odds ratio of InDels occurring on one strand (for example, leading) versus the other strand (for example, lagging). P values were computed using binomial test or Χ² test and corrected for multiple testing using the Benjamini–Hochberg method.

De novo extraction of InDel signatures from hTERT-RPE1 samples

De novo signature extraction and decomposition of mutational signatures were performed using SigProfilerExtractor³⁹ (v.1.1.18), along with SigProfilerMatrixGenerator⁵⁹ (v.1.2.4). The recommended default settings (including 500 NMF replicates) were applied (https://github.com/AlexandrovLab/SigProfilerExtractor). Signatures were also extracted using Indel.signature.tools³ (v.2.4.4) with default settings (20 bootstraps, 500 repeats per bootstrap, matched clustering). The number of InDel signatures selected per channel set was determined by maximal drop in average silhouette width of clustered signatures. For extraction using MuSiCal⁸ (https://github.com/parklab/MuSiCal, v.1.0.0), the default hyperparameters included random initialization (init hyperparameter), the minimum volume nonnegative matrix factorization (MVNMF) algorithm (method hyperparameter), 20 MVNMF replicate runs (n_replicates hyperparameter) and between 10,000 and 1,000,000 MVNMF iterations per replicate run (min_iter and max_iter hyperparameters, respectively).

GEL cohort

WGS data (v.8) for colorectal (n = 2,146), endometrial (n = 695), stomach (n = 181), bladder (n = 347), brain (CNS, n = 392), lung (n = 958) and skin cancers (n = 56) were considered in this study, subset from the curated 12,222 sample cohort previously described³. Indels were called with Strelka⁶⁰ (v.2.4.7) using somatic calling mode.

InDel segmentation

We exploited the relationship between InDels and their associated 3′ sequence context to determine, for each InDel, the minimal InDel prefix that is maximally repetitive in the 3′ sequence context and within the InDel itself. In brief, for each prefix of the InDel, we identified the maximally repetitive substring in the associated suffix and 3′ context. This applies to all InDel variants that are left-aligned and parsimonious. We termed these partitions of the InDel and 3′ context ‘segmentations’ and selected the segmentation that contains the smallest prefix that has maximum repetition in the 3′ sequence context and the InDel. Segmentation produced multiple values quantifying the repetitiveness of the prefix sequence (termed, ‘unit’). These values, along with the sequence context, may be used to group InDels into biologically relevant, nonoverlapping InDel subcategories or ‘channels’. See Supplementary Note for details.

InDel channel construction

Using the segmentation values for each InDel, we constructed a set of 476 nonoverlapping InDel channels. By surveying the InDel frequency distribution across all 476 channels in 17,253 tumors covering most cancer types from ICGC/TCGA, Hartwig and the GEL 100,000 Genomes Project, we discarded channels with no signal and consolidated those with low signal to reduce the total number of InDel channels to 89. The final 89-channel set was used for the experimental gene-edit dataset and the cancer WGS from the GEL 100,000 Genome Project considered in this study. Generally, channels may be grouped into six broad divisions—1 bp insertions, 1 bp deletions, ≥2 bp insertions, ≥2 bp deletions, ≥ 2 bp deletions with evidence of microhomology and complex InDels. Channels were constructed such that each InDel could only be unambiguously assigned to a single channel. A complete description of each channel and exemplar InDels are included in Supplementary Table 8. A description of the reasoning behind channel construction is presented in Supplementary Note.

InDel signature extraction of GEL cohort

Our approach to signature extraction was motivated by previous study³. We observed that hypermutator samples strongly influenced signature extraction using the standard β-divergence NMF model. We sought to filter out hypermutator samples from our initial extraction. For each tissue, we first removed samples with a total InDel burden <100, and then clustered sample profiles according to their cosine similarity (cosine distance, 1 − cosine similarity) using hierarchical clustering with complete linkage. Clusters of similar samples were determined by thresholding the resulting dendrogram such that the average silhouette width was maximized, and within-cluster variation was minimized.

To determine cluster-specific hypermutators, for each cluster, we fit the total burden per sample using a two-component Gaussian mixture model (mixtools⁶¹) compared to a one-component mixture model, using the Bayesian information criteria for model selection. Hypermutators were defined as the union of samples with a total burden more than third quartile + 1.5 × IQR, where quartiles were calculated over the total dataset, and samples with a greater than 50% probability of being generated by the higher burden Gaussian distribution. Normal and hypermutator clusters were then manually reviewed per tissue, and only normal clusters were used in primary extraction.

Signature extraction was performed per tissue as described, with an increased number of bootstraps (40) and repetitions per bootstrap (1,000), to increase final solution stability.

Signature refitting and reliably determining excess variation

We sought to determine whether excess variation indicative of rare or unextracted signatures was present in sample catalogs using a published framework³. In refitting tissue-specific catalogs, hypermutator samples included, with signatures extracted from nonhypermutator samples using FitMS³, we observed profiles with lower InDel counts displayed higher degrees of error (as measured by total residual normalized by sample burden). Generally, error decreased logarithmically as InDel burden increased. Therefore, using a single threshold on fitting error to define samples with excess variation excluded a large proportion of samples.

To more accurately calibrate our expectation of fitting error, and therefore, our threshold for detecting excess variation, we performed a parametric bootstrapping procedure to generate a sample-specific expected error distribution. For each sample catalog, we constructed a multinomial distribution using the per-channel density from the normalized reconstructed profile produced by FitMS. Using this distribution, we simulated 10,000 sample profiles with a total burden equal to the sample burden, fit these profiles with FitMS and calculated the resulting error distribution.

Comparing the experimentally derived error distribution to the resulting null distribution allowed us to estimate an empirical P value. This procedure was repeated for all samples in the GEL cohort, and P values were corrected for multiple testing. To control the false discovery rate at 5%, samples with an adjusted P value less than 0.05 were selected for further analysis to determine rare signatures.

Rare signature extraction and refinement

For each tissue, samples with excess variation were clustered using the residual signal after subtracting out FitMS reconstructed profiles. Hierarchical clustering with average linkage and Euclidean distance resulted in multiple clusters per tissue. A secondary round of signature extraction per cluster was performed using the primary tissue-specific signature set to initialize the signature matrix. An additional one to five rare signatures were determined per cluster, and the number of rare signatures was determined as the minimal value of n, such that rare signatures were not found to perfectly recapitulate cluster members or match common signatures. All extracted rare signatures across a tissue were subject to manual curation to identify recurrent patterns, and the rare signature exemplars that displayed minimal mixing with common signatures were selected. Using this consensus set of common and rare signatures per tissue, all samples in each catalog were refitted using FitMS to determine per-sample signature exposures.

Deriving consensus IDS

We clustered InDel signatures from seven GEL cancer cohorts (111 tissue-specific InDS) using hierarchical clustering with average linkage and cosine distance and derived a set of consensus signatures following a published framework³. The associated tissue-specific signature to reference signature conversion matrix (Supplementary Tables 9–11) was then used to transform tissue-specific signature exposures to consensus signature (final InD) exposures for downstream analyses.

Annotating InD reference signatures using gene-edit and mutagen exposure signatures

Variants in samples generated in previous studies^46,52 were reclassified and analyzed using our refined InDel classification scheme. Experimental signatures were obtained via background subtraction and determined using a bootstrapping and cosine similarity-based framework previously described⁴⁶. Consensus InDel signatures and all experimentally derived signatures were clustered using hierarchical clustering and cosine distance.

PRRDetect model

We trained PRRDetect, a multinomial elastic net regression model, on a subset of 571 GEL cancers confidently assigned as MMRd (n = 214), Pol-dys (n = 36), mixed MMRd/Pol-dys (n = 41) or PRR-proficient (negative controls, n = 280) based on manual curation of relevant driver mutations and/or supporting immunohistochemistry where possible. Samples with neither Pol-dys and/or MMRd were confirmed to lack driver mutations in key MMR genes (that is, MLH1, MSH2, MSH6 and PMS2), POLE, POLD1, and displayed no evidence of MSI associated with these abnormalities⁴³.

To create our classifier, we explored a range of feature combinations as model inputs, including (1) summed exposures of SBS and InDel signatures related to PRR deficiency (MMRd, MMRd/Poly-dys, Poly-dys); (2) feature set (1) combined with TMB; (3) feature set (1) combined with total InDel/SNV ratio; (4) summed exposure of SBS signatures related to PRRd; (5) summed exposure of InD signatures related to PRRd. For each feature set, we constructed the model using either proportion (that is, normalized signature exposure) or the absolute values of the features (that is, raw mutation count contributed by each signature). In total, ten model structures were attempted (five sets of features × two normalizations; Supplementary Table 13).

For all models, the feature values were first log₂ transformed, then z-score normalized using the formula $x^{\prime }=\frac{x-\mu }{\sigma }$. We used the implementation of multinomial elastic net regression (glmnet) in caret (https://topepo.github.io/caret/). In each training iteration, we first partitioned the cohort into 70% for training and 30% for testing, retaining relative proportions of MMRd, Pol-dys, mixed and negative categories across the training and test datasets. A ten-repeat tenfold cross-validation strategy was adopted within the 70% training group. A grid search approach was used to determine the best combination of two hyperparameters (that is, α, which acts as a balancing factor between a lasso and a ridge penalty; and λ, which defines the strength of the penalty), aiming to minimize the log loss. The model was then tested on the held-out 30% test set. We performed the 70:30 repartitioning ten times to assess the stability/robustness of the model (Fig. 4a). We then retrained the model on the entire cohort of 571 samples using a repeated tenfold cross-validation strategy as in the first round to obtain the final feature coefficients for individual models.

To identify the best-performing model, seven measures were defined—training accuracy of the final model, median accuracy on the test sets, s.d. of α, s.d. of λ, median coefficient s.d., median multiclass area under the curve (AUC) on the test set and training multiclass AUC of the final model (Supplementary Table 13). Eventually, PRRDetect was selected as the one having the following input variables: (1) summed exposures to MMRd-associated SBS 6, 15, 26, 44, 97; (2) summed exposures to Pol-dys-associated SBS 10a, 10d; (3) summed exposures to combined MMRd/Pol-dys SBS 14, 20; (4) summed exposures to MMRd-associated InD7, InD19; (5) summed exposures to Pol-dys InD14, InD15; (6) summed exposures to combined MMRd/Pol-dys InD16a, InD16b, InD20, InD21; and (7) total InDel to total SNV ratio, with proportional normalization of the first six features (Supplementary Table 14). The final PRRDetect was retrained using the abovementioned features on the entire curated cohort of 571 samples.

The model outputs a categorical distribution across the four PRRd subclasses (that is, MMRd, Pol-dys, MMRd/Pol-dys or Neg). A sample is PRR-proficient if the probability for the ‘negative’ class is >0.5; otherwise, the positive class (MMRd, MMRd/Pol-dys or Pol-dys) with the highest probability is assigned.

PRRDetect in additional cohorts

We validated PRRDetect in an independent ICGC breast cohort (n = 504)^44,45 and a subset of held-out samples from GEL (n = 847), for which the true PRRd labels were established based on immunohistochemistry staining of four MMR proteins (PMS2, MLH1, MSH2 and MSH6) and driver mutations. The final cohort consists of 1,351 samples, for which we also computed the MSIseq and MMRDetect prediction results. The ROC curves and their relative AUC values were calculated using R package ‘pROC’⁵³. P values were calculated using the roc.test() function with bootstrap method with 10,000 resamples.

To survey the prevalence of PRRd in other cancer cohorts, we applied PRRDetect to two additional datasets not included in InD signature extraction and PRRDetect training—the ICGC/TGCA pan-cancer dataset³³ and Hartwig Medical Foundation metastatic cancer cohort³⁴ (n = 1,335), focusing on seven cancer types commonly enriched with samples with high InDel burdens. InDels from individual samples in these cohorts were processed to 89-channel profiles as was done for GEL cohort samples. For these datasets, we used published driver annotations as PRRd labels.

Statistics and reproducibility

All comparisons were between biologically independent samples. No statistical method was used to predetermine sample size. No data were excluded from the analyses. The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment. Further details are provided in the Reporting Summary.

Reporting summary

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

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41588-025-02152-y.

Author contributions

G.C.C.K., A.S.N., X.Z. and S.N.-Z. conceived the project and designed the experiments. G.C.C.K., S.B., J.Y., C.B. and S.J.Z. performed gene editing and mutation accumulation experiments. G.C.C.K., A.S.N., G.R. and X.Z. designed and implemented computational analyses, with input from A.D., A.M.P., L.C., D.B., L.H., J.D., Y.M., S.S., J.C., M.B. and H.R.D. S.N.-Z. supervised the work. Data interpretation and write-up were provided by G.C.C.K., A.S.N., X.Z., G.R. and S.N.-Z., with inputs from all the other authors. All authors had the opportunity to edit the manuscript and approve the final manuscript.

Peer review

Peer review information

Nature Genetics thanks Geoff Macintyre and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Data availability

Raw sequence files from the hTERT-RPE1 mutation accumulation experiment are deposited at the European Genome-Phenome Archive with accession EGAD50000000209. Mutation calls have been deposited at Mendeley (10.17632/3k2tpx9ssr.2). RPE1 cells can be obtained directly from the authors. The curated data are available for general browsing from our reference mutational signatures website, Signal (https://signal.mutationalsignatures.com). All other data were previously published^46,52.

Primary data from the 100,000 Genomes Project, which are held in a secure research environment, are available to registered users. See https://www.genomicsengland.co.uk/research for further information or contact research-network@genomicsengland.co.uk. ICGC/TCGA WGS data can be obtained from https://dcc.icgc.org/releases/PCAWG. Hartwig metastasis WGS data can be obtained from Hartwig Medical Foundation through standardized procedures and request forms that can be found at https://www.hartwigmedicalfoundation.nl/en/appyling-for-data/.

Mutagen signatures from human induced pluripotent stem cells (iPS)⁵² can be accessed via https://data.mendeley.com/datasets/m7r4msjb4c/2. Human iPS knockout signatures can be obtained directly from 10.1038/s43018-021-00200-0 (ref. ⁴⁶). The results of RPE1 experimental signatures can be browsed at https://signal.mutationalsignatures.com/explore/main/experimental/experiments?study=7. InD signatures of the seven cancer types are accessible at https://signal.mutationalsignatures.com/explore/main/cancer/signatures?mutationType=3&study=7. Source data are provided with this paper.

Code availability

The R source code of PRRDetect is available via GitHub at https://github.com/Nik-Zainal-Group/PRRDetect and Zenodo at 10.5281/zenodo.14906103 (ref. ⁶²). InDel segmentation and signature classification script can be accessed via GitHub at https://github.com/Nik-Zainal-Group/indelsig.tools.lib and Zenodo at 10.5281/zenodo.14906117 (ref. ⁶³).

Competing interests

A.D., A.S.N., G.C.C.K., G.R., H.R.D., X.Z. and S.N.-Z. hold patents or have submitted applications on clinical algorithms of mutational signatures: MMRDetect (PCT/EP2022/057387), HRDetect (PCT/EP2017/060294), clinical use of signatures (PCT/EP2017/060289), rearrangement signature methods (PCT/EP2017/060279), clinical predictor (PCT/EP2017/060298), hotspots for chromosomal rearrangements (PCT/EP2017/060298), InDel signature methods (PCT/EP2024/077959) and PRRDetect (PCT/EP2024/078030). All other authors declare no competing interests.

Extended data

is available for this paper at 10.1038/s41588-025-02152-y.

Supplementary information

The online version contains supplementary material available at 10.1038/s41588-025-02152-y.