The interplay of mutagenesis and ecDNA shapes urothelial cancer evolution

Abstract

Advanced urothelial cancer is a frequently lethal disease characterized by marked genetic heterogeneity¹. In this study, we investigated the evolution of genomic signatures caused by endogenous and external mutagenic processes and their interplay with complex structural variants (SVs). We superimposed mutational signatures and phylogenetic analyses of matched serial tumours from patients with urothelial cancer to define the evolutionary dynamics of these processes. We show that APOBEC3-induced mutations are clonal and early, whereas chemotherapy induces mutational bursts of hundreds of late subclonal mutations. Using a genome graph computational tool², we observed frequent high copy-number circular amplicons characteristic of extrachromosomal DNA (ecDNA)-forming SVs. We characterized the distinct temporal patterns of APOBEC3-induced and chemotherapy-induced mutations within ecDNA-forming SVs, gaining new insights into the timing of these mutagenic processes relative to ecDNA biogenesis. We discovered that most CCND1 amplifications in urothelial cancer arise within circular ecDNA-forming SVs. ecDNA-forming SVs persisted and increased in complexity, incorporating additional DNA segments and contributing to the evolution of treatment resistance. Oxford Nanopore Technologies long-read whole-genome sequencing followed by de novo assembly mapped out CCND1 ecDNA structure. Experimental modelling of CCND1 ecDNA confirmed its role as a driver of treatment resistance. Our findings define fundamental mechanisms that drive urothelial cancer evolution and have important therapeutic implications.

Whole-genome sequencing of matched serial tumours from patients identifies two key mutagenic factors (APOBEC3 and chemotherapy) and extrachromosomal DNA-forming structural variants that drive treatment resistance in urothelial cancer.

Main

Mutagenesis and DNA repair constantly interact during cancer, resulting in distinct mutational signatures and extensive structural genomic changes. The apolipoprotein B mRNA-editing enzyme catalytic subunit 3 (APOBEC3) cytidine deaminases³ cause somatic mutational signatures characterized by cytosine to thymine or guanine substitutions in the genomes of various human cancers, especially urothelial carcinoma (UC)^3–5. In a previous study¹, we investigated the evolutionary dynamics of chemotherapy-resistant advanced UC. However, several fundamental questions remain: (1) how and whether the relative timing, clonality and velocity of endogenous mutagenic stresses (for example, APOBEC3) and extrinsic mutagenic processes (for example, chemotherapy) shape UC evolution and (2) how high-order SVs, including ecDNA, shape UC evolution and drug resistance. Our ability to predict the clinical trajectory and prevent drug resistance in patients with advanced cancer depend on answering these fundamental questions.

To address these questions, we performed whole-genome sequencing (WGS) of matched sets of primary and metastatic UC tumours and morphologically normal urothelial samples. We analysed single base substitution (SBS), doublet base substitution (DBS) and small insertion and deletion (indel; ID) mutational signatures and superimposed them on phylogenetic maps of primary and metastatic tumours collected at different time points. We compared the relative timing and clonality of these mutagenic processes over the patients’ lifespan. We then applied junction balance analysis (JaBbA)² to serially collected tumour samples to analyse the topology of complex SVs and track ecDNA evolution. We investigated the patterns and the timing of APOBEC3 and chemotherapy-induced mutations within ecDNA-forming SV events. We mapped the changes in ecDNA-forming SV events following systemic therapy to gain insights into their role as putative drivers of drug resistance. We reconstructed the genetic maps of recurrent CCND1 ecDNA-forming SVs. We mapped out the circular structure of ecDNAs using state-of-the-art Oxford Nanopore Technologies (ONT) long-read WGS. Finally, using two different experimental models, we demonstrate the fitness advantage of CCND1 ecDNA in driving resistance to therapy.

Timing mutagenic processes in UC evolution

We performed WGS of 77 urothelial tumours from 50 patients, including 28 pre-platinum and 49 post-platinum chemotherapy samples from comparable-quality freshly frozen (FF) and formalin-fixed paraffin-embedded (FFPE) tissues (Extended Data Fig. 1a and Supplementary Tables 1–3). The APOBEC3 cytidine deaminases are the dominant mutagenic enzymes in UC. We found that signatures SBS2 (refs. ^4,6), SBS13 (refs. ^4,6) and DBS11 (ref. ⁴) contributed an average of 17%, 20% and 17%, respectively, of all corresponding mutations in our cohort (Extended Data Fig. 1b,c). ID9 (ref. ⁴), a signature characterized by single cytidine deletions, correlated with APOBEC3-induced SBS signature contributions (R = 0.67) (Extended Data Fig. 1d). As expected, there was no significant correlation between our recently described APOBEC3G-induced mutational signature with APOBEC3A and APOBEC3B SBS2 and SBS13 signatures (R = −0.24, P = 0.20) or chemotherapy-associated SBS31 and SBS35 signatures (R = −0.11, P = 0.56)⁷. Contributions of endogenous mutational signatures did not differ between matched primary and metastatic tumours from the same patient (Extended Data Fig. 1e). Next, we investigated the SBS31 (refs. ^4,8), SBS35 (refs. ^4,8) and DBS5 (refs. ^4,9) signatures associated with platinum-based chemotherapy, the backbone of treatment regimens for advanced UC (Extended Data Fig. 1c). We observed contributions from SBS31 (mean 18%), SBS35 (mean 3%) and DBS5 (mean 23%) (Extended Data Fig. 1c) in the post-chemotherapy tumours.

Extended Data Fig. 1. Mutational signatures.

(a) Schematic of anatomical sites of primary and metastatic urothelial cancer samples. For each site, the number of tumours stratified by chemotherapy treatment status is listed. (b) Mutational signatures in 77 advanced urothelial carcinoma and 5 morphologically normal urothelium in the WCM-UC cohort. Two samples lacked calling of the DBS signature due to an insufficient number of DBS variants. One tumour exhibited an absence of any detected SV events. N/A: not applicable. (c) The landscape of mutational signatures induced by endogenous mutagenic processes and exogenous exposures in advanced urothelial carcinoma. From top to bottom, bar graphs represent the contribution of mutational signatures induced by APOBEC3, platinum chemotherapy, and tobacco smoking, respectively, in 77 urothelial tumours. Each panel includes associated SBSs (top bar plots) and DBSs (bottom bar plots) attributed to each mutagenic process. The samples are ordered by group (sex, chemotherapy status, and smoking status) and by the contribution of SBS signatures. (d) Pearson’s correlation of ID signatures with SBS contribution induced by APOBEC3 (top panels) and platinum chemotherapy (bottom panels). Each dot represents one tumour. Pre-chemotherapy tumours were excluded from this analysis. (e) Endogenous mutational signatures between paired primary and metastatic samples of four patients. The SBS contribution of multiple metastatic samples within a patient was collapsed to the mean. Two-sided paired Wilcoxon’s test.

Source Data

We then generated phylogenetic trees for patients with at least two tumour samples to determine the timing and clonality of mutagenic processes over the arc of tumour progression. By fitting signatures at each node, we superimposed the mutational signature composition onto the phylogenetic tree to show dynamic signature changes over time (Methods, Fig. 1a and Extended Data Fig. 2). In 11 out of 14 reconstructed evolutionary trees, SBS2 and SBS13 dominated the early truncal nodes, constituting ≥25% of all mutations and persisting throughout the lifetime of the tumour (Extended Data Fig. 2). This result underscores the role of APOBEC3 in both urothelial carcinogenesis and progression. Chemotherapy-induced mutations were more prevalent in ‘leaf’ nodes compared with APOBEC3 mutations (P = 5.9 × 10⁻⁵), matching the late timing of therapy (Fig. 1b).

Fig. 1. Timing mutagenic processes in UC evolution.

a, Phylogenetic tree depicting UC evolution for patient WCMIV063. Each node represents SNVs (centre), mean CCF across all SNVs within the node (blue numbers) and genes affected by high-impact SNVs. SBS, DBS and ID signature proportions are represented as concentric circles (periphery to the centre, respectively) within each node. The number of samples collected from different tumour sites is indicated on the human body diagram. Illustration of a human adapted from ref. ¹, Springer Nature America. b, Leaf/trunk variant fold change of platinum chemotherapy-induced and APOBEC-induced variants. n = 12 tumours, P = 5.91 × 10⁻⁵. c, Clonality fold change of APOBEC3 (SBS2 and SBS13), chemotherapy (SBS31 and SBS35) and ageing (SBS1) mutational signatures in 44 post-chemotherapy tumours. n = 28 (ageing), n = 29 (chemotherapy), n = 42 (APOBEC). Early/late APOBEC–chemotherapy: P = 2.1 × 10⁻⁷; early/late ageing–chemotherapy: P = 1.8 × 10⁻⁴; clonal/subclonal APOBEC–chemotherapy: P = 4.3 × 10⁻⁴. d, Velocity fold change of APOBEC3-induced, chemotherapy-induced and ageing-associated mutagenesis. n = 34 (APOBEC SBS), n = 30 (chemotherapy SBS), n = 28 (APOBEC DBS), n = 24 (chemotherapy DBS) patients. APOBEC SBS–chemotherapy SBS: P = 0.0139; APOBEC DBS–chemotherapy DBS: P = 7.0 × 10⁻⁴. e, Upset and alluvial plots depicting shared SBS2 and SBS13 mutations (top, vertical bars) between normal urothelium and metastatic UC tumours in patient WCMIVG010. S on the y axis of the bottom plot indicates the sample identifier. Alluvial y axes represent SNV counts. Image of human bladder created by Servier (https://smart.servier.com/) and adapted from Bioicons (https://bioicons.com/), under a CC BY 3.0 licence. f, Maximum likelihood estimates of dN/dS ratios for significant genes (q < 0.1). Circle size represents coding SNV counts across the cohort. Colour indicates the proportion of APOBEC3-induced coding mutations. TP53 had the highest dN/dS values (inset, top-right corner). Two-sided Wilcoxon rank-sum test was used in b–d. Unless otherwise specified, boxes show the median and the interquartile range (IQR); the lower whisker indicates Q1 – 1.5 × the IQR; the upper whisker indicates Q3 + 1.5 × the IQR. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, NS, not significant.

Source Data

Extended Data Fig. 2. Phylogenetic trees from individual patients.

(a–m) Superimposed mutational signatures on nodal branching points in phylogenetic trees for patients with at least two tumour samples. (Top) Timeline and clinical course (vertical lines) in the natural history of the disease. (Bottom) Each node represents the number of SNVs (centre), mean CCF across all SNVs in the node (blue numbers), and genes affected by high-impact SNVs. SBS, DBS, and ID signature proportions are represented as concentric circles (periphery to center, respectively) within each node. The number of samples collected from different tumour sites is indicated on the human body diagram.

Source Data

We classified the clonality of somatic mutations in each of the 71 tumours (Methods). Consistent with the patterns observed in the phylogenetic trees, APOBEC3-induced mutations were significantly earlier (1.85-fold compared with 0.48-fold, P = 2.1 × 10⁻⁷) and more clonal (1.49-fold compared with 0.8-fold, P = 4.3 × 10⁻⁴) than chemotherapy-induced mutations (Fig. 1c and Extended Data Fig. 3a). Comparing SBS signature clonality between post-chemotherapy and pre-chemotherapy samples revealed significantly more clonal early APOBEC3-induced mutations in chemotherapy-treated tumours (P = 0.012) but not other signatures (P > 0.05) consistent with selection of clones habouring APOBEC3-induced mutations under therapeutic pressure (Extended Data Fig. 3b).

Extended Data Fig. 3. APOBEC3 and chemotherapy-induced mutations and DNA damage footprints.

(a) Clonality fold-change of mutations induced by the APOBEC3 (SBS2/13), platinum chemotherapy (SBS31/35), and aging (SBS1) mutational signatures in 44 post-chemotherapy tumours (dots) in the WCM-UC cohort. n = 28 (aging), n = 29 (chemo), n = 42 (APOBEC). Two-sided Wilcoxon rank sum test. Each line connects samples collected from the same patient. Early/late APOBEC-Chemo: P = 2.1 × 10⁻⁷, Early/late Aging-Chemo: P = 1.8 × 10⁻⁴, Clonal/subclonal APOBEC-Chemo: P = 4.3 × 10⁻⁴. (b) The clonality of detected mutational signatures in 70 pre- and post-chemotherapy samples. Each dot represents one tumour. n = 66 (APOBEC), n = 49 (aging), n = 11 (tobacco). Two-sided Wilcoxon’s test. (c) Velocity fold change of APOBEC3-, chemotherapy-induced and aging-associated mutagenesis. n = 34 patients. Two-sided Wilcoxon rank sum test. Each connecting line represents matched samples collected from the same patient. APOBEC SBS - chemo SBS: P = 0.0139, APOBEC DBS - chemo DBS: P = 7.0 × 10⁻⁴. (d) The mutagenic velocity fold change of APOBEC3-induced mutational signatures SBS2/13 and DBS11 compared to aging signatures SBS1 in current and former smokers (n = 36) versus never smokers (n = 14). Two-sided Wilcoxon’s test. (e) OncoPrint showing mutations attributed to platinum chemotherapy-induced signatures SBS31/35 on cancer genes (rows) in UC tumours treated with chemotherapy (columns). (f) Enriched pathways affected by mutations in post-chemotherapy UC samples. Each point represents a pathway in the NCI-Nature 2016 gene set library. The pathways are plotted based on the first two UMAP dimensions. Pathways with more similar gene sets are positioned closer together. Pathways are coloured by automatically identified clusters computed with the Leiden algorithm. Circles with black outlines indicate pathways with FDR-adjusted P value ≤ 0.002. Two-sided Fisher’s exact test. (g) Upset plots showing shared APOBEC3-induced SBS2/13 mutations (top, vertical bars) between different normal urothelium, primary metastatic UC tumours from each patient (bottom, vertical connected dots). From top-bottom, left-right: Patient IDs WCMIVG001, WCMIVG065, WCMIVG091, WCMIVG010, WCMIVG013. S##: sample ID. Different tumour samples are numbered. (h) (top) Cancer cell fraction (CCF) distributions and (bottom) Clonal and subclonal mutation count for APOBEC3-induced mutations in chromatin-modifying genes compared to Cancer Gene Consensus (CGC) genes and non-cancer genes. n = 1640 mutations across 70 samples. Two-sided unpaired Wilcoxon’s test, no correction for multiple testing. Each dot represents one variant. (i) Boxplot showing the percentage of pRPA-positive nuclei and γH2AX-positive nuclei under different treatments (DMSO: n replicates = 2 and Dox: n replicates = 4) with error bars representing standard error of the mean (SEM). Two-sided Mann-Whitney test. γH2AX: P = 5.1 × 10⁻⁷, pRPA: P = 1.3 × 10⁻²¹.

Source Data

Next, we compared the fold change in the mutagenic velocity of APOBEC3 and chemotherapy relative to the ageing signature SBS1 per month of exposure. Chemotherapy exposure resulted in higher mutational rates than APOBEC3 for both SBS (P = 0.014) and DBS (P = 7.0 × 10⁻⁴) (Methods, Fig. 1d and Extended Data Fig. 3c). Smoking history did not affect APOBEC3-associated SBS2, SBS13 or DBS11 mutational velocities (P > 0.05) (Extended Data Fig. 3d). Notably, platinum-based chemotherapy generated hundreds of SBS and a few DBS mutations per month of exposure (Fig. 1d and Extended Data Fig. 3c). We identified deleterious SBS31-assigned C[C>T]T splice variants in ACVR2A, SLC34A2 and JAK2, but no recurrent chemotherapy-induced gene mutations were observed (Extended Data Fig. 3e). Cisplatin exposure was also associated with mutations in the putative tumour-suppressor genes DCC and ARID1B (Extended Data Fig. 3e). We examined regulatory elements and found no significant difference in their mutation rates between noncoding regions affected by platinum chemotherapy and ageing SBS1 (P > 0.05) (Supplementary Table 4). Furthermore, none of the five significantly mutated noncoding regulatory regions had SBS31 or SBS35 mutations, indicating limited impact on driver gene regulation (Methods and Supplementary Table 4). Gene-list enrichment analysis of SBS31 and SBS35 mutations exclusive to platinum chemotherapy-treated patients revealed enrichment in the IFNγ pathway (odds ratio (OR) = 34.4; adjusted P = 0.0018), ERBB4 signalling (OR = 33.46; adjusted P = 0.0022) and VEGFR3 signalling (OR = 53.27; adjusted P = 0.0014). The PDGFRβ signalling pathway (OR = 20.31; adjusted P = 5.6 × 10⁻⁵) and signalling pathways regulated by the RET tyrosine kinase (OR = 33.46; adjusted P = 0.0022) (Methods and Extended Data Fig. 3f) were also enriched. This suggested involvement of these pathways in tumour progression and drug resistance. Together, these data show that chemotherapy induces bursts of subclonal mutations that potentially affect tumour progression.

Based on the early and clonal distribution of APOBEC3-induced mutations (Fig. 1c), we proposed that APOBEC3 mutagenesis drives urothelial carcinogenesis. Thousands of APOBEC3-induced mutations, including high- to moderate-impact driver mutations in ARID1A, PARD3, CYP51A1 and SETBP1, were shared between urothelial tumours and morphologically normal urothelium (Methods, Fig. 1e and Extended Data Fig. 3g). The median cancer cell fraction (CCF) of high-impact APOBEC3-induced mutations was 1.0 in chromatin-modifying genes associated with early urothelial field defects^10–12, supporting their clonal nature (Extended Data Fig. 3h and Supplementary Table 5). Notably, three SBS13-associated high-impact ARID1A mutations were shared by one normal urothelium sample and three distinct tumours in patient WCMIVG010 (Extended Data Fig. 1e).

Furthermore, we analysed the ratio of nonsynonymous to synonymous mutations (dN/dS) and identified an enrichment of APOBEC3-induced mutations in crucial oncogenes (PIK3CA) and tumour suppressor genes (TP53, RB1, KDM6A and ARID1A), with dN/dS ratios of >10. This result suggests that APOBEC3-induced mutations in these genes have undergone intense and persistent selection, potentially providing a fitness advantage that predates tumorigenesis (Methods and Fig. 1f).

Consistent with the role of APOBEC3-induced mutagenesis as an initiating event of DNA damage in early carcinogenesis, we observed a significant increase in the percentage of cells with pRPA and γH2AX foci (P < 0.0001), markers of single-stranded DNA (ssDNA) and DNA double-strand breaks (DSB), respectively, after regulated induction of APOBEC3A in isogenic UC cell lines (Methods and Extended Data Fig. 3i). Furthermore, we identified an enrichment of SBS2 and SBS13-induced composite mutations¹³ in cancer-associated genes, including chromatin-modifying genes (KMT2D, KDM6A, KMT2A, ARID1A and ARID2) in UC tumours, consistent with the processive nature of APOBEC3 enzymes (Supplementary Table 6). These composite mutations potentially augment the selective fitness of the initial singleton mutations¹³ and promote urothelial carcinogenesis.

Our findings suggest that early APOBEC3-induced mutations drive urothelial carcinogenesis. This is significant because APOBEC3-induced mutagenesis accounts for up to 67% of all single nucleotide variants (SNVs) in UC¹⁴. Furthermore, our data show the interplay between the early clonal APOBEC3-induced mutations and the late subclonal bursts of chemotherapy-induced mutations, which cooperate to shape metastatic and treatment-resistant phenotypes.

Complex ecDNA-forming SVs in UC

Genomic instability emerges from the the interplay between mutations and complex, large-scale chromosomal variants¹⁵. To investigate these processes, we used JaBbA² to provide a comprehensive analysis of higher-order SVs within our Weill Cornell Medicine (WCM)-UC cohort (Methods, Fig. 2a and SV glossary). We identified a median of 19 SVs per sample (interquartile range (IQR) of 9–36), and 84.5% had one or more complex SVs (median of 2, IQR of 1–5) (Fig. 2a). In our WCM-UC cohort, templated-insertion chains (TICs) (49.3%), chromoplexy (33.8%) and breakage–fusion–bridges (BFBs) (31%) were the most prevalent complex SV events (Extended Data Fig. 4a). Notably, 11.3% harboured ‘tyfonas’², carrying a significantly higher SV junction burden than other complex SVs (P < 7.8 × 10⁻⁵) (Extended Data Fig. 4b). Duplications in the genome were associated with TP53 SNV and indel (SNV/indel) mutations (P = 1.0 × 10⁻⁶, false discovery rate (FDR) = 7.2 × 10⁻³) (Supplementary Table 7). Duplication, translocation and chromothripsis were associated with FGFR3 SNVs/indels (P = 3.7 × 10⁻⁵, FDR = 0.10), FBXW7 (P = 1.7 × 10⁻⁴, FDR = 0.23) and CDKN2A homozygous deletions (P = 1.7 × 10⁻⁴, FDR = 0.12) (Supplementary Table 7). We also characterized COSMIC signatures positively (for example, SBS40 and chromoplexy, FDR-adjusted P = 0.021) or negatively correlated with SV events (for example, ID11 and chromoplexy, FDR-adjusted P = 0.021) (Methods and Extended Data Fig. 4c–e).

Fig. 2. The interplay between APOBEC3 and platinum chemotherapy-induced mutagenesis and ecDNA during UC evolution.

a, Heatmap displaying junction burdens of complex SVs (y axis) in 71 UC tumours (x axis). Tyfonas, BFBs and DMs are grouped by their association with ecDNA. The heatmap is normalized to the average junction burden of our cohort and scaled in the natural log. Other complex SVs, from top to bottom, include chromothripsis, chromoplexy, TICs, quasi-reciprocal pairs, rigma and pyrgo. b, Schematic depicting the impact of APOBEC3 and chemotherapy-induced mutagenesis on ecDNA biogenesis during cancer evolution. Images of chromosomes were created by KKT Madhusanka and adapted from Adobe Stock (https://stock.adobe.com). c, Proportion of kataegic events co-localizing with SVs. DM (non-ecDNA) events were identified by JaBbA but not classified as ‘cyclic’ by AmpliconArchitect. d, Kataegic events on ecDNA-forming SVs showed a significantly shorter median distance to the nearest breakpoint compared to kataegis on other non-ecDNA SVs and kataegis without SV association. ecDNA-forming SVs median size in this cohort was 5.56 Mb, used as the upper limit of the distance to the closest breakpoints for kyklonas. The median distance of all mutations within a kataegic event was collapsed to one measurement per event. Two-sided Wilcoxon rank-sum test. e, Violin plots of VAF distributions for APOBEC3-induced kyklonic mutations (n = 147) and chemotherapy-induced SNVs (n = 716) on ecDNA-forming SVs. Plots extend between the maximum and minimum of the distribution. Two-sided Wilcoxon rank sum test. P < 2.22 × 10⁻¹⁶. f, Contributions of mutational signatures with sample/VAF tranche combinations in all ecDNA-forming SVs. Only tranches with >100 mutations are displayed. g, ecDNA-forming SVs harbouring ≥1 kyklonic event in patient WCMIVG035S01. CN track: JaBbA genome graph showing CNs for rearranged DNA segments (grey vertices) with SV junctions (aqua blue edges) within circular ecDNA events. Kyklonas track: normalized VAF of kyklonic mutations. Non-clustered track: normalized VAF of non-clustered APOBEC3-induced and platinum-induced signature mutations on ecDNA-forming SVs.

Source Data

Extended Data Fig. 4. Complex SVs.

(a) The proportion of simple and complex SV events from 71 samples from 50 patients in our UC cohort. (b) Tyfonas events have a significantly higher junction burden than other complex SVs. n = 248 SV events. Two-sided Wilcoxon rank-sum test was used in panels b, d, e, f, g, and h. Each dot represents one event. In order from left to right: P = 0.0017, P = 7.8 × 10⁻⁵, P = 0.0002, P = 6.9 × 10⁻⁶, P = 4.7 × 10⁻⁶, P = 1.4 × 10⁻⁷, P = 3.2 × 10⁻⁶, P = 1 × 10⁻¹⁵. (c) The relationship between tumour mutation burden (TMB) and SV types in WCM-UC tumours. In panels c, d, e, n = 71 samples split by the presence or absence of JaBbA events. TMB: the total number of SNVs and insertion-deletion mutations (INDELs). (d) The association between genome-wide SBS/DBS/ID signature contributions and the presence of SV types called in 71 WCM-UC tumours. A total of 8 SBSs, 8 DBSs, and 13 IDs with at least 0.05 contribution in at least 10% of samples were filtered for comparisons. FDR-adjusted P value ≤ 0.25 was used to nominate significant hits, given the limited size of our cohort. (e) The copy number variant (CNV) features associated with SV types called in WCM-UC tumours. The top 16 associations with significant FDR-adjusted P value ≤ 0.25 are presented. The y axis represents the count of CNV segments assigned to the feature by SigProfilerMatrixGenerator and formatted as copy number: loss-of-heterozygosity status: segment length. (c-e) ecDNA-forming SVs: grouped tumours having tyfonas, BFB, and DM events that overlapped with an AmpliconArchitect cyclic call. Present: at least one event of the SV type was called in the sample. Absent: no calling of the SV event in the tumour. (f) The junction burden of simple and complex SVs in UC tumours stratified by APOBEC3-induced mutational loads. n = 71 samples split by APOBEC burden. (g) Multivariable correlation of clinical characteristics with the fraction of genome altered (FGA). n = 71 samples. (h) Enrichment of structural variants in TP53 mutant urothelial cancers. (Left) The fraction of genome altered is significantly higher (P = 1.5 × 10⁻⁵) in TP53-mutant UC tumours compared to TP53-wild-type (WT) tumours. n = 71 samples split by TP53 mutational status. (Right) 42 tumours harbouring TP53 high and moderate-impact mutations exhibited a significant increase in the number of total junctions (P = 4.9 × 10⁻⁷), deletions (P = 0.0028), duplications (P = 2 × 10⁻⁷), chromoplexy (P = 0.0036), TIC (P = 0.0086), and BFB (P = 4.6 × 10⁻⁵) compared to 28 TP53-wild-type tumours. (c-h), each dot represents one sample. (i) The bar plot depicts the proportion of JaBbA events that overlapped with AmpliconArchitect’s cyclic calls. (j) Comparison of one-way overlaps between JaBbA and AmpliconArchitect. Each point is an AmpliconArchitect ecDNA-forming SV event. Dotted lines indicate nested cutoffs. (k) Individual JaBbA event (x axis) with multiple overlapping nested AmpliconArchitect ecDNA-forming SV calls (y axis). ‘Nested’ is defined as ≥ 90% one-way AmpliconArchitect overlap and ≤ 30% one-way JaBbA overlap.

Source Data

No correlation between APOBEC3-induced mutations and specific SV categories was identified (Extended Data Fig. 4f). However, patients over 60 years old had a higher fraction of the genome altered (P = 0.0023) than patients aged under 60 years (Extended Data Fig. 4g). Moreover, UC tumours with TP53 mutations exhibited a significantly higher fraction of the genome altered (P < 0.001) and number of total junctions (P = 4.9 × 10⁻⁷) and several SV classes (for example, duplications, P = 2.0 × 10⁻⁷; BFB, P = 4.6 × 10⁻⁵) (Extended Data Fig. 4h). These findings suggest that the underlying genomic instability due to mutations in genomic integrity guardians like TP53 is crucial to the biogenesis of complex SVs.

Cancer cell ecDNA is a megabase-scale circular DNA subclass that amplifies oncogenes¹⁶. To detect ecDNA-forming SVs, we focused on two defining characteristics: (1) amplicons with high copy numbers (CNs) and (2) circular structure. Within the JaBbA framework, tyfonas, BFBs or double minutes (DMs)² are amplicons with CNs twofold higher than the genomic ploidy, thereby satisfying the first criterion. Further analysis revealed that tyfonas (100%), BFBs (81.5%) and DMs (63.2%) were the most common JaBbA-defined events that overlapped with the circular event annotations by AmpliconArchitect¹⁷, which implicated their association with ecDNAs^2,18,19 (Extended Data Fig. 4i). JaBbA identified larger amplification events encompassing those identified by AmpliconArchitect (Extended Data Fig. 4j,k). We determined that 35% of samples in our cohort harboured one or more events that could generate ecDNA-forming SVs, highlighting their high prevalence in UC.

APOBEC3 and chemotherapy mutate ecDNA

By juxtaposing APOBEC3 and chemotherapy-induced mutations on ecDNA-forming SVs, we aimed to uncover  the chronology and mechanisms of ecDNA biogenesis (Fig. 2b). APOBEC3-induced clustered mutations in ecDNA (kyklonas)²⁰ were found in 69% of 42 ecDNA-forming SV events (Fig. 2c and Extended Data Fig. 5). Kyklonas occurred near SV breakpoints with a median distance of 16.2 kb, with 48.6% within 10 kb, and were significantly closer to SV junctions than chromosomal kataegic events (P < 0.001) (Fig. 2d). The topographical proximity of APOBEC3-induced kyklonic events and SV breakpoints suggests APOBEC3-induced DSB as an intermediary step in ecDNA biogenesis.

Extended Data Fig. 5. APOBEC3 and platinum chemotherapy-induced mutagenesis in ecDNA-forming SVs.

The graphs depict 28 ecDNA-forming SV events having at least one kyklonic event in addition to the events depicted in Fig. 2g. CN track: JaBbA genome graph showing CN for rearranged DNA segments (grey vertices) with SV junctions (aqua blue edges) within circular ecDNA-forming SV events. Kyklonas track: normalized VAF of APOBEC3-induced kyklonas. Non-clustered track: normalized VAF of non-clustered mutations assigned to APOBEC3-induced and platinum chemotherapy-induced signatures on ecDNA-forming SV. Chr: chromosome.

Source Data

ecDNA-forming SVs mainly harboured APOBEC3-induced SBS2 and SBS13, followed by chemotherapy-associated SBS31 and SBS35, with minimal contributions from tobacco smoking-associated SBS4, SBS29 and SBS92 (Extended Data Fig. 6a). To investigate the relative timing of APOBEC3 and chemotherapy-induced mutagenesis in relation to ecDNA biogenesis, we examined the variant allele frequency (VAF) of APOBEC3 kyklonic clusters and non-clustered chemotherapy mutations on ecDNA-forming SVs. Kyklonas events had a significantly higher VAF than chemotherapy mutations in ecDNA (P < 2.22 × 10⁻¹⁶) (Fig. 2e). Overall, 10.2% of kyklonic mutations occurred early in the evolution of UC (VAF > 0.67), whereas 93.4% (669 out of 716) of chemotherapy-induced mutations on ecDNA-forming SVs had low VAF < 0.33 (Fig. 2f,g and Extended Data Fig. 5). Further analysis showed that the APOBEC3-induced mutational burden was significantly higher in ecDNA-forming SVs than in non-ecDNA regions (P = 1.3 × 10⁻⁵) (Extended Data Fig. 6b). These findings suggest that APOBEC3-induced mutagenesis and ecDNA biogenesis overlap topographically and chronologically.

Extended Data Fig. 6. Mutational profiles of ecDNA-forming SVs.

(a) Mutational profiles of all mutations on ecDNA-forming SVs. (Top) SNV count and (Bottom) Mutational signature contributions to each sample with ecDNA-forming SV/VAF tranche combination. Only sample/VAF tranche combinations >100 mutations are displayed. Signatures with less than 10% contribution for a sample are marked as “Other”. (b) Comparison of APOBEC mutational burden between ecDNA-forming SV and non-ecDNA-forming SV regions within 27 samples harbouring at least one ecDNA-forming SV event. Two-sided paired Wilcoxon test.

Source Data

ecDNA-driven oncogene amplification in UC

To identify recurrent ecDNA-driven oncogene amplification, we used FishHook to search for regions with recurrent structural variation (Methods). FishHook identified 31 loci, 17 (55%) of which were within 500 kb of a cancer-related gene (CCND1, AXIN2, CHEK2, ERBB2, FHIT, KRAS, MDM2 or NUMA1) (Extended Data Fig. 7a). We identified 4 out of 71 samples showing ERBB2 amplification in ecDNA-forming SVs, including DM, BFB or tyfonas events (Extended Data Fig. 7b). All four JaBbA ERBB2 SV events were part of AmpliconArchitect cyclic ecDNA calls and showed extremely high ERBB2 CNs (median of 30, range of 11–285) (Extended Data Fig. 7c). Notably, the 11q13.3 locus containing CCND1 was involved in most SVs (13 samples), 92% of which were ecDNA-forming SVs (Extended Data Fig. 7b).

Extended Data Fig. 7. CCND1 is commonly involved in recurrent putative ecDNA-forming complex SVs in urothelial carcinoma.

(a) Manhattan plot showing the significantly recurrent breakpoints (SRBs) identified by FishHook and their distance to the nearest Cancer Gene Consensus (CGC) genes in UC whole-genome sequences. Each dot represents an FDR-adjusted (Benjamini-Hochberg) P value, and a cutoff of 0.25 (horizontal solid line) was used to nominate significant hits. (b) JaBbA-classified SV events were overlapped with FishHook SRB hits to identify the frequency and class of SV events occurring in significantly recurrent breakpoint regions with the nearest CGC genes (left y axis) in each tumour (x axis). The panel was arranged by decreasing the total number of SV events in a particular chromosomal region. (c) Combined genome graphs of AmpliconArchitect (top track) with JaBbA (CN and read depth tracks). AmpliconArchitect classifies cyclic amplicons. JaBbA tracks show the chromosomal locations and copy-number alteration for DNA segments (black vertices) and the corresponding JaBbA SV events (coloured edges) in 4 samples with ERBB2 ecDNA amplification. CN: copy number, Chr: chromosome, BFB: breakage-fusion-bridge, TIC: templated-insertion chain. (d) Kaplan–Meier survival curves of overall survival (OS) (left) and disease-free survival (DFS) (right) in the TCGA Pan-Cancer Atlas stratified by CCND1 alteration and log-rank test P values showing worse OS and DFS in patients harbouring CCND1 alterations (gain and amplification).

Source Data

Dynamics of CCND1 ecDNA-forming SVs under therapy

Using AmpliconArchitect and JaBbA, we identified ecDNA-forming SV events in 67% (10 out of 15) of WCM-UC tumours with CCND1 amplification (Fig. 3a). These events had a mean CN of 47, 3.6-fold higher than other non-cyclic CCND1 amplifications (Extended Data Fig. 3a). To validate these findings, we examined CCND1 ecDNA amplification in the dataset previously analysed by AmpliconArchitect^21,22, which includes a subset of tumours from The Cancer Genome Atlas (TCGA) and Pan-Cancer Analysis of Whole Genomes (PCAWG) pan-cancer cohorts (n = 3,731 and n = 1,291, respectively). Amplification through ecDNA-forming SVs was found in 70% of TCGA/PCAWG bladder cancer cases with CCND1 amplification. This is the second-highest following pancreatic adenocarcinoma (75%) and higher than ovarian (67%) and oesophageal (59%) cancers (Fig. 3a). Pan-Cancer Atlas analysis of patients with CCND1 alterations in 32 cancer types showed significantly worse overall survival (P < 0.0001) and disease-free survival (P = 0.002), which highlighted the correlation between CCND1 high CN and poor clinical outcomes (Extended Data Fig. 7d). Cyclin D1 forms a complex with CDK4 and CDK6 (CDK4/6) to phosphorylate RB1. This releases the E2F transcription factors, which orchestrate cell cycle progression through the G1/S checkpoint²³. CDKN2A encodes p16 (p16^INK4a), a tumour suppressor that inhibits CDK4/6 (Fig. 3b). In our cohort, homozygous CDKN2A loss was significantly more common in tumours with CCND1 amplification, including those with CCND1 ecDNA-forming SVs (P = 0.041), than in tumours with other amplifications (Extended Data Fig. 3a). In samples with CCND1 ecDNA amplification, 87% had CDKN2A deletions (mean CN fold change relative to ploidy of 0.34), whereas the downstream signalling pathway genes CDK4, CDK6 and RB1 were predominantly wild type (Fig. 3b). RNA sequencing of a subset of the tumours in our cohort confirmed that tumours with CCND1 gain had significantly higher normalized CCND1 mRNA expression (adjusted P = 0.038, log₂(fold change) = 1.26) (Fig. 3c). Our results suggest that UC tumours with ecDNA-mediated CCND1 amplification activate the p16–cyclin D1–CDK4/6–RB1 pathway.

Fig. 3. Increased complexity and CN of CCND1 ecDNA amplicons following systemic therapy.

a, Distribution of circular, heavily rearranged, linear or non-cyclic amplicons in samples with CCND1 amplification from our WCM bladder cancer (BLCA-WCM) cohort, BLCA-TCGA cohort and TCGA/PCAWG pan-cancer subset cohorts analysed in a previous study²². b, Top, schematic of the p16–cyclin D1–CDK4/6–Rb pathway. Bottom, normalized CN alteration heatmap of chromosomal regions 9p21.3 (CDKN2A), 11q13.3 (CCND1) and genes CDK4 and CDK6 and RB1 in 71 UC tumours. Chromosomal regions and genes are not drawn to scale. c, Normalized CCND1 and LTO1 mRNA expression. d, ecDNA-forming SV events (up to 3 MB) in post-chemotherapy (post-chemo) tumours had significantly higher mean JCN (P = 0.016) and maximum JCN (P = 0.032) than ecDNA-forming SV events in pre-chemotherapy (pre-chemo) tumours. n = 15 events in 12 tumours. Two-sided Wilcoxon rank-sum test. Each dot represents one ecDNA-forming SV event. e–h, Combined genome graphs of AmpliconArchitect (top track) with JaBbA (CN and read depth tracks). JaBbA tracks show the chromosomal locations and CN alteration for DNA segments (black vertices) and the corresponding JaBbA SV events (coloured edges) in patient WCMIV091 (e,f) and in patient WCMIV076 (g,h) before (e,g) and after (f,h) systemic therapy for primary bladder tumour. Patient WCMIV091 was treated with neoadjuvant abemaciclib, a CDK4/6 inhibitor, for 4 weeks as part of a clinical trial. Patient WCMIV076 received a neoadjuvant combination of gemcitabine (Gem) and cisplatin (Cis). Top right schematics show that in patient WCMIV091, FGFR1 was rearranged with CCND1 in the same ecDNA, whereas in patient WCMIV076, FGF19 and CCND1 were co-amplified in ecDNA. The chromosomal locations of CCND1, FGFR1 and FGF19 are highlighted in red.

Source Data

To understand the role of ecDNA in therapy resistance, we investigated the dynamics of ecDNA-forming SV events in matched tumours from patients undergoing systemic therapy. For ecDNA-forming SV events up to 3 Mb, post-chemotherapy tumours had significantly higher mean and maximum junction copy number (JCN) (P = 0.016 and P = 0.032, respectively) than chemotherapy-naive tumours, suggesting increased genetic complexity following treatment (Fig. 3d). After systemic treatment, CCND1 ecDNA-forming SV events increased their integer CN, which reflects the mean CN of ecDNAs per cell, from 17 to 26 in patient WCMIV091 and from 12 to 24 in patient WCMIV076 (Fig. 3e–h). This increase is consistent with mathematical models of ecDNA selection under positive selection²⁴. AmpliconArchitect reconstruction of ecDNA structures showed co-amplification of FGFR1 and FGF19 with CCND1 in patient WCMIV091 and patient WCMIV076, respectively, and confirmed their CN increases following systemic therapy (Fig. 3e–h). Collectively, these data show the high prevalence of ecDNA-mediated CCND1 amplification in UC and suggest it has a role in drug resistance.

Mapping CCND1 ecDNA-forming SVs

We constructed genomic maps of CCND1 ecDNA-forming SV amplicons and discovered that each amplicon is unique in structure, acquiring promoters, enhancers and super-enhancers of distinct oncogenes such as ERBB4 and LRP5 (Extended Data Figs. 8a and 9). Our maps revealed the common amplification of oncogenes FGF3, FGF4 and FGF19 and the preservation of CCND1 local promoter and regulatory elements in the shared CCND1 ecDNA-forming SV regions (Extended Data Fig. 8b).

Extended Data Fig. 8. Genetic maps of CCND1 ecDNA-forming structural variants.

(a) The putative reconstruction of representative ecDNA-forming SV amplicons harbouring CCND1 across the cohort. From the innermost to outermost ring: FANTOM5 promoters⁸⁶ (promoters located on the positive strand in red, and minus strand in blue), FANTOM5 enhancers⁸⁷, super-enhancers⁸⁸, coverage, and oncogene models. Coverage is normalized by segment dosage. Gene directionality reflects transcription orientation relative to segment orientation. WCM: WCMIVG. (b) Visualization of the genic and regulatory elements contained within the commonly amplified region shared between CCND1 ecDNA-forming SVs. From top to bottom, tracks include GRCh38 coordinates, the 204 Kb common ecDNA region, NCBI RefSeq gene models, FANTOM5 transcription start site (TSS) peak, ENCODE H3K27Ac, H3K4Me1, and H3K4Me3 histone marks. Chromosome 11 regions involved in ecDNA amplicons are coloured purple. (c) Schematic of Oxford Nanopore Technologies (ONT) long-read WGS of a patient-derived bladder cancer tumour (WCMIVG101), followed by Flye read assembly to reconstruct ecDNA amplicons. (d) Circos plot reconstruction of a CCND1-ecDNA from ONT long-read sequencing. Annotation from inside to outside: assembled contiguous segment size, FANTOM5 promoters, FANTOM5 enhancers, super-enhancers, selected genes, coverage, and circular assembly aligned to GRCh38. (e) Read length and count of split-read alignments supporting both ends of the circularizing junctions of the assembled CCND1-ecDNA.

Source Data

Extended Data Fig. 9. Putative reconstruction of ecDNA-forming SV amplicons across 50 bladder cancer patients.

Starting from the innermost ring, annotations include FANTOM5 promoters⁸⁶ (promoters located on the positive strand are coloured red, and those on the minus strand are coloured blue), FANTOM5 enhancers⁸⁷, super-enhancers⁸⁸, coverage, and oncogene models. Coverage is normalized by segment dosage (e.g., the inclusion of two of the same genomic segments will halve the displayed coverage). The directionality of the genes corresponds to the direction in which they are transcribed relative to the orientation of the segment.

Source Data

To further characterize the genetic map of CCND1 ecDNA, we performed ONT long-read WGS of a bladder tumour from patient WCMIVG101, followed by de novo assembly of the putative ecDNA using Flye (Methods and Extended Data Fig. 8c). De novo assembly produced an 840 kb circular contig (Extended Data Fig. 8d). We queried both ends of the circularizing junction for split-read alignments and found 1,028 reads that precisely supported the junction (with 768 reads anchored on ≥1 kb on each side, 100 reads anchored ≥5 kb on each side, and one read spanning ≥10 kb on each side) (Extended Data Fig. 8e). Interphase fluorescence in situ hybridization (FISH) confirmed significant CCND1 amplification (Methods and Extended Data Fig. 10a).

Extended Data Fig. 10. CCND1 amplification in UC tumour and UMUC3 cell line.

(a) Interphase FISH demonstrating CCND1 amplification (ratio >2) (left). n = 1 independent experiment. JaBbA and AmpliconArchitect reconstruction of a CCND1 ecDNA-forming SV event in WCMIVG101 UC tumour (right). (b) Metaphase FISH showing CCND1 chromosomal gain in UMUC3 bladder cancer cell line. Centromeric Ch11 (red) and CCND1 signals (green). n = 1 independent experiment. (c) Composite bar chart illustrating the mean cell cycle proportions of UMUC3 bladder cancer cells with episomal CCND1 compared to control cells. The cell cycle phases (G1, S, and G2/M) were quantified using flow cytometry of cells stained with DAPI and compared between cells with episomal CCND1 and controls under no-drug and cisplatin treatment conditions. Statistical significance was determined using a two-sided unpaired student’s t-test with FDR adjustment from six replicates.

Source Data

CCND1 ecDNA drives chemotherapy resistance

We experimentally modelled the extrachromosomal configuration of CCND1 and examined its selective fitness relative to chromosomal amplification under therapeutic pressure. To that end, we transduced the UMUC3 bladder cancer cell line, which harbours CCND1 chromosomal gain, with a non-integrating lentiviral episomal vector harbouring CCND1–mCherry or a control-GFP vector (Methods and Extended Data Fig. 10b). Equal numbers of isogenic UMUC3 cells with GFP-tagged empty vector and mCherry-tagged CCDN1 episomal ecDNA were co-cultured in the presence of cisplatin or vehicle in a competition assay. This strategy enabled us to model the heterogeneity of ecDNA in UC tumour cells and to longitudinally monitor adaptive fitness to chemotherapy by measuring the ratio of GFP/mCherry fluorescence signals using the Incucyte live cell imaging system (Fig. 4a). UMUC3 cells with mCherry-tagged CCND1 episomal ecDNA displayed a significantly higher ratio to the GFP-empty vector control cells under cisplatin treatment than with vehicle (2.171 compared with 1.172, FDR-adjusted P = 1.3 × 10⁻⁵) (Fig. 4b). This suggests positive selection of clones with CCND1 ecDNA. Co-cultured mCherry and GFP-expressing cells were then separated after 96 hours of treatment using fluorescence-activated flow sorting and underwent single-cell 10x RNA sequencing to determine how CCND1 ecDNA alters their transcriptomic profiles under selective pressure. Cells harbouring CCND1 episomal ecDNA significantly activated downstream E2F signalling pathways compared with empty GFP vector cells under vehicle (normalized enrichment score (NES) = 3.45, FDR q < 0.0001) and cisplatin treatment (NES = 3.37, FDR q < 0.0001) (Fig. 4c). DNA repair, unfolded protein response and mTORC1 signalling pathways were also differentially upregulated in cells with CCDN1 episomal ecDNA compared with empty controls in vehicle and cisplatin conditions, suggesting that these pathways potentially drive CCND1-mediated adaptive fitness (Fig. 4d). Cell cycle analysis showed that CCND1 episomal vector-transduced cells exhibited faster G1-to-S transition, with significant decreases in G1 phase proportion in both normal culture and cisplatin treatment conditions (FDR-adjusted P < 0.0037). Under cisplatin treatment, there was also a significant increase in cells in the S phase (FDR-adjusted P = 0.022) (Extended Data Fig. 10c and Supplementary Figs. 1–5). These data demonstrate that CCND1 ecDNA increases fitness under selective therapeutic pressure by driving cell cycle progression.

Fig. 4. CCND1 ecDNA drives adaptive fitness to cisplatin chemotherapy.

a, Schematic of experiments using non-integrating CCND1 episomal lentiviral vector transduction in the UMUC3 bladder cancer cell line and the competition assay, monitored by Incucyte live-cell imaging and single-cell 10x RNA sequencing. b, Ratio of mCherry CCND1 episomal ecDNA to GFP empty cells over 96 h, demonstrating a selective advantage for CCND1 ecDNA. n = 25 areas per well. Two-sided unpaired t-test. Data are mean ± s.d. At each 6-h point from 0–96 h, FDR-adjusted P = 0.02, 0.11, 0.12, 0.047, 0.022, 0.0083, 0.0013, 0.00031, 0.00021, 7.1 × 10⁻⁵, 8.0 × 10⁻⁶, 7.0 × 10⁻⁶, 7.0 × 10⁻⁶, 7.0 × 10⁻⁶, 7.0 × 10⁻⁶ and 1.3 × 10⁻⁵. c, Uniform manifold approximation and projection (UMAP) (left) and boxplot (right) comparing E2F target gene expression scores in single cells with GFP empty control (vehicle n = 7,602, cisplatin n = 8,177) and mCherry CCND1 ecDNA (vehicle n = 8,336, cisplatin n = 4,320) under cisplatin (P = 5.3 × 10⁻⁹¹) and vehicle (P = 8.3 × 10⁻⁴⁹) treatment. Two-sided Mann–Whitney test. d, Bubble plot showing enriched gene sets for differentially expressed genes between GFP empty control and mCherry CCND1 ecDNA samples under vehicle and cisplatin treatments. Bubble sizes represent NES values. Black borders denote significance (FDR q < 0.01). e, JaBbA detection and AmpliconArchitect confirmation of cyclic ecDNA-forming DM events in SF295 cell lines. f, Metaphase FISH showing CCND1 amplification on ecDNA. n = 1 independent experiment. Scale bar, 10 µm. g, SDS–PAGE western blot validating CCND1 knockdown. n = 3 independent experiments. For gel source data, see Supplementary Fig. 1. h, Schematic of CCND1 shRNA and scramble shRNA control isogenic cell lines in a competition assay under cisplatin selection and monitoring by Incucyte live-cell imaging. i, Cell ratio after 96-h treatment of mCherry-positive cells with shRNA-mediated CCND1 knockdown versus GFP-positive cells with scramble shRNA control, showing reduced fitness of the CCND1 knockdown cells (P = 0.034). n = 25 areas per well. One-sided unpaired t-test. Data are mean ± s.d. Image of six-well plate created by KeHan and adapted from Bioicons (https://bioicons.com/), under a CC BY 1.0 licence.

Source Data

To complement the episomal overexpression experiments, we performed shRNA-mediated knockdown of CCND1 encoded by CCND1-amplifying ecDNA in the SF295 cell line. JaBbA and AmpliconArchitect analyses identified DM events generating more than 50 copies of CCND1 ecDNA (Fig. 4e) in this cell line. Metaphase FISH confirmed the presence of these CCND1-amplifying ecDNAs (Fig. 4f). We then established isogenic SF295 shRNA-CCND1–mCherry and scramble shRNA-control–GFP cells. SDS–PAGE western blotting confirmed CCND1 knockdown (Fig. 4g and Supplementary Figs. 1–5). Equal numbers of isogenic SF295 cells with shRNA-CCND1–mCherry and scramble shRNA-control–GFP were competitively co-cultured for 96 hours with and without cisplatin (Fig. 4h). The reduction in fitness in CCND1 knockdown cells was more marked under cisplatin treatment compared to controls (P = 0.034) (Fig. 4i). Collectively, our experiments confirm that the acquisition of CCND1 ecDNA significantly enhances the adaptive fitness of tumour cells against chemotherapy.

Discussion

In this study, we performed WGS of matched samples collected from distinct anatomical regions and at consecutive time points from patients with advanced UC. Our analyses revealed a previously uncharted timeline of the activity of two key mutagenic factors: APOBEC3 cytidine deaminases and platinum-based chemotherapy. By comprehensively mapping complex SVs in advanced UC, we identified a high prevalence of ecDNA-forming SVs during UC progression. Notably, ecDNA-forming SVs were mutagenized by APOBEC3 and chemotherapy at distinct time points. Furthermore, we discovered that SV-forming events undergo dynamic evolution incorporating additional DNA segments in the process. This suggests that ecDNA-forming SVs serve as adaptive reservoirs for oncogene amplification^25,26 and contribute to therapy resistance.

Previous pan-cancer studies^12,21,27 have evaluated the timing and clonality of mutagenic processes of primary tumours but were limited by the paucity of longitudinal samples. Our study analysed serial samples from the same patient using metastatic biopsies and rapid autopsies to reconstruct phylogenetic trees that demonstrated that APOBEC3-induced mutagenesis occurs early in the natural history of UC. Building on our previous work using whole-exome sequencing¹, this study used WGS to identify early clonal APOBEC3 mutations in ecDNAs and in benign urothelium, providing support of APOBEC3-induced mutagenesis in the early stages of bladder carcinogenesis. Moreover, the observed early clonal expansion of mutations in chromatin-modifying enzymes within microscopically normal urothelium underscores the key role of APOBEC3-induced mutations in creating early urothelial field defects that are primed for carcinogenesis^10,11,28.

Our results are aligned with observations of relatively early timing of SBS2 and SBS13 mutations in UC in the PCAWG and TCGA datasets^12,14,29, and previous findings of APOBEC3-induced mutations in histologically normal urothelium^10,11. These data are also consistent with our previous work showing that the transgenic expression of human APOBEC3G promotes mutagenesis and shorter survival in a murine bladder cancer model⁷. Previous studies¹⁴ described the prognostic link between high APOBEC3-induced mutational signatures in primary bladder cancers and improved patient outcomes. Other studies³⁰ showed inferior overall survival for patients treated with chemotherapy whose tumours harbour higher APOBEC3-induced mutations.

Our data demonstrated that APOBEC3A upregulation induces a significant increase in pRPA foci likely marking ssDNA overhangs at DSB sites marked by γH2AX foci. These data suggest that APOBEC3-induced deamination of small quantities of ssDNA arising under physiological conditions such as DNA replication³¹ can lead to a forward feedback loop that generates additional ssDNA substrates and subsequent episodic bursts of APOBEC3-induced mutagenesis³².

Together, these findings highlight the emerging role of the APOBEC3 enzymes as putative drivers of tumorigenesis. These observations present opportunities for cancer control through early detection of actionable APOBEC3-induced driver mutations in the normal urothelium in patients at a higher risk of developing UC and for intercepting this process at the pre-cancerous stages.

We quantified the rate of chemotherapy-induced mutations, which generated hundreds of mutations per month of exposure, a finding consistent with previous studies⁸. These mutations are potentially functionally significant, converging on the same genes in our post-chemotherapy tumours and in cisplatin-resistant bladder cancer cell lines³³. Our results align with previous studies implicating VEGFR3, ERBB4, PDGFRβ, RET tyrosine kinase and IFNγ-induced JAK2–STAT3 in cisplatin resistance^34–39. Targeting these pathways offers a potential avenue for preventing or reversing cisplatin resistance. Notably, cisplatin induces genome-wide mutations rather than specific recurrent hotspot mutations. The acquisition of these ‘off-target’ mutations in pathways not directly related to the primary mechanism of action of the drug potentially contributes to resistance in a pattern similar to that previously implicated in the evolution of microbial resistance⁴⁰. It is plausible that the accumulation of these ‘off-target’ cisplatin-induced mutations can result in a similar fitness advantage in chemotherapy-treated cancers. These findings add to our understanding of the role of DNA-damaging chemotherapy as a significant mutagenic force that shapes cancer evolution.

We identified a significant burden of complex SVs in our UC cohort. We focused on ecDNA-forming SVs that drive tumour evolution and introduce an additional layer of genetic heterogeneity, thus posing additional challenges due to the emergence of therapeutic resistance^22,24,41. Using JaBbA² and AmpliconArchitect¹⁷, we identified a high prevalence of tyfonas^2,19, BFBs^16,42 and DMs^16,43 as complex ecDNA-forming SVs. We found that ERBB2 and CCND1 were involved in these ecDNA-forming SVs often resulting in high CN amplification. These findings are important in light of the recent accelerated approval by the Food and Drug Administration of the antibody–drug conjugate fam-trastuzumab deruxtecan-nxki for metastatic HER2-positive solid tumours, including UC⁴⁴.

The mechanisms of ecDNA biogenesis are not fully understood⁴⁵, but the occurrence of DNA DSBs followed by DSB repair⁴⁶ is thought to be crucial. Chromothripsis or chromosomal shattering was previously identified as a mechanism driving ecDNA evolution⁴⁷. The co-occurrence of chromothripsis and kataegis, putatively caused by APOBEC3 enzyme activity^48,49, suggest a potential role for APOBEC3-induced DSBs in ecDNA biogenesis. Our findings support this hypothesis by showing the topographical overlap between APOBEC3-induced mutations and ecDNA-forming SV junctions, and the early timing of APOBEC3-induced mutagenesis within these events.

Our data suggest that the interaction between ecDNA and chemotherapy-induced mutagenesis compounds subclonal diversification and cancer heterogeneity. Emerging techniques such as single-cell Circle-seq⁵⁰ can characterize the evolution of ecDNA mutagenized by APOBEC3 and chemotherapy in tumour subclones driving aggressive cancer phenotypes.

We discovered that the formation of ecDNA-forming SVs is a prevalent mechanism for generating CCND1 CN amplification in UC⁵¹. Our analysis shows that UC has the second-highest percentage of CCND1 ecDNA-forming SVs among tumour types in the TCGA/PCAWG pan-cancer cohort^21,22. Previous studies have shown that CCND1 is enriched in specific molecular subtypes of bladder cancer, including the Lund Uro subtype⁵². A detailed understanding of the association of CCND1 ecDNA with molecular subtypes warrants future investigation. As CCND1 ecDNA amplification events were found to co-occur with CDKN2A^del RB1^WT tumours, we predict that the CCND1^ecDNA-amp CDKN2A^del RB1^WT genotypic configuration promotes cell cycle progression. Notably, we discovered that CCND1 ecDNA-forming-SVs evolve following systemic therapy, increasing their complexity by integrating additional DNA segments. This is consistent with the emerging literature suggesting that ecDNA acts a dynamic pool that is constantly integrating new DNA segments, as a mechanism of adaptive oncogene amplification^25,26. New strategies, including the use of DNA-dependent protein kinase inhibitors to disrupt the non-homologous DNA end joining pathway, can limit ecDNA size and inhibit the evolution of resistance to targeted therapy in melanomas²⁵. Similar therapeutic strategies are potentially applicable for intercepting ecDNA-driven resistance in UC.

Cyclin D1 has a direct role in DNA damage repair independent of its canonical functions in cell cycle regulation⁵³. It is recruited to DNA damage sites in a BRCA2-dependent manner to promote RAD51 localization and stabilization, thereby playing a direct part in homologous recombination-mediated DSB repair⁵³. This suggests that CCND1 ecDNA promotes therapeutic resistance through multiple mechanisms that work in tandem to enhance proliferative capacity and DNA damage repair. We acknowledge that our data do not directly address the relative contributions of increased cell proliferation versus resistance to cell death in CCND1 ecDNA-driven drug resistance. However, our data suggest that inhibiting the CDK4/6 pathway in patients with UC and CCND1 ecDNA amplification could lead to significant advances in treating UC. Our group is currently leading the CLONEVO window-of-opportunity clinical trial (ClinicalTrials.gov identifier NCT03837821) of the CDK4/6 inhibitor abemaciclib in patients with UC, which will shed light on the feasibility of this approach.

A previous study⁵⁴ showed that high nuclear CCND1 expression in metastases is associated with increased responsiveness to chemotherapy but did not translate to improved cancer-specific survival when compared with patients with low nuclear CCND1 expression⁵⁴. This indicates that although high CCND1 expression is a potential biomarker for chemotherapy responsiveness, it might also confer enhanced adaptive fitness to tumour cells during therapy, which in turn leads to poorer survival outcomes. Supporting this hypothesis, our findings showed increased CCND1 ecDNA CN after chemotherapy compared with paired treatment-naive tumours. Furthermore, this is consistent with our experimental data demonstrating that CCND1 ecDNA significantly enhances the adaptive fitness of bladder cancer cells to cisplatin treatment.

We used a non-integrating lentiviral episomal vector to model the extrachromosomal configuration of CCND1 ecDNA. We observed a marked selection advantage in UMUC3 bladder cancer cells with CCND1 ecDNA under selective pressure from cisplatin treatment through E2F-dependent mechanisms. Similarly, knockdown of CCND1 in SF259 cells addicted to high CCND1 CN decreased cellular fitness under cisplatin treatment. These findings highlight the crucial role of CCND1 ecDNA in chemoresistance and the promise of therapeutic strategies to reverse this process. Collectively, our findings show complex interactions between therapeutic selective pressures and ecDNA dynamics revealing the functional consequences of ecDNA as a driver of therapy resistance in cancer cells. Our study is underpowered to draw conclusive correlations between mutational and SV events and clinical outcomes. Future WGS studies of larger cohorts of patients with advanced UC tumours with prospective clinical annotation are needed.

In conclusion, our findings define the fundamental mechanisms that drive UC evolution. These new insights hold promise for future research and the development of innovative therapeutic strategies aimed at intercepting urothelial carcinogenesis and treatment resistance.

Methods

Patient enrolment and tissue acquisition

All experimental procedures followed approved guidelines and were approved by the Institutional Review Boards (IRBs) at WCM. Recruited patients signed informed consent under IRB-approved protocols: WCM/New York-Presbyterian (NYP) IRB protocols for Tumour Biobanking—0201005295; GU Tumour Biobanking—1008011210; Urothelial Cancer Sequencing—1011011386; and Comprehensive Cancer Characterization (Genomic and Transcriptomic Profiling—1007011157, and Precision Medicine—1305013903). The study was open to all patients with advanced UC at WCM/NYP. Fresh-frozen and FFPE tissues from biopsies, cystectomy and nephroureterectomy specimens from patients with high-grade UC were collected. All pathology specimens were reviewed by board-certified genitourinary pathologists (J.M.M.) at WCM/NYP. Clinical charts were reviewed to record patient demographics, tobacco use, family history of cancer, concurrent cancer, treatment history, anatomical site, pathologic grade and stage using the tumour, node, metastasis (TNM) system.

Rapid autopsy procedures

The Englander Institute for Precision Medicine research protocol at WCM/NYP has been established to promote personalized medicine focused on molecular diagnostics and therapeutics. Patients were given the option to be enrolled in the IRB-approved rapid autopsy programme. In addition, patients’ next-of-kin provided written consent before autopsy. A systematic autopsy protocol is followed whereby normal and malignant fresh tissues are collected, allocating samples to be snap-frozen or formalin-fixed. The goal is to maximize the amount of tissue collected for research purposes. Once tissue collection is complete, the autopsy proceeds per the WCM Autopsy Service protocol. In this study, tissues from multiple sites were collected, with DNA extracted for WGS following haematoxylin and eosin evaluation and frozen slide annotation.

Whole-genome library preparation and sequencing

WGS libraries were prepared using a KAPA Hyper PCR+ library preparation kit following the manufacturer’s instructions. Before library preparation, FFPE samples were repaired using PreCR Repair Mix (NEB) followed by additional DNA quantification. For library preparation, DNA was sheared using a Covaris LE220. DNA fragments were end-repaired, adenylated and ligated to Illumina sequencing adapters. Libraries went through two post-ligation bead clean-ups, PCR amplification and a final post-PCR bead clean-up. Final library quality was verified using a KAPA qPCR Library Quantification kit (Roche) and a Fragment Analyzer (Agilent). Libraries were normalized, pooled and sequenced on an Illumina NovaSeq 6000 sequencer using 2 × 150 bp cycles.

WGS data pre-processing, variant calling and annotation

The New York Genome Center (v.6) somatic pipeline⁵⁵ was used to align the data and call variants. In brief, sequencing reads were aligned to GRCh38 with BWA-MEM (v.0.7.15)⁵⁶. Short alignments were removed with NYGC ShortAlignmentMarking (v.2.1) (https://github.com/nygenome/nygc-short-alignment-marking), and mate-pair information was added with GATK FixMateInformation (v.4.1.0)⁵⁷. Individual lane BAMs were merged and sorted simultaneously with Novosort markDuplicates (v.1.03.01), followed by GATK BQSR. SNVs, multi-nucleotide variants and indels were called using MuTect2 (GATK v.4.0.5.1)⁵⁸, Strelka2 (v.2.9.3)⁵⁹ and Lancet (v.1.0.7)⁶⁰. SVs were called using Svaba (v.0.2.1)⁶¹, Manta (v.1.4.0)⁶² and Lumpy (v.0.2.13)⁶³, with Svaba used for both indels and SVs. Split-read support for SVs was quantified using SplazerS (v.1.1)⁶⁴. Germline variants were called on the matched normal samples with GATK HaplotypeCaller (v.3.5) and filtered with GATK VQSR at tranche 99.6%. The positions of heterozygous germline variants were used to compute B allele frequencies in the tumour samples. Variants were merged across callers and annotated using Ensembl (v.93)⁶⁵, COSMIC (v.86)⁶⁶, 1000Genomes (Phase3)⁶⁷, ClinVar (201706)⁶⁸, PolyPhen (v.2.2.2)⁶⁹, SIFT (v.5.2.2)⁷⁰, FATHMM (v.2.1)⁷¹, gnomAD (r.2.0.1)⁷² and dbSNP (v.150)⁷³ using Variant Effect Predictor (v.93.2)⁷⁴. Somatic variants that occurred in two or more individuals in our in-house panel of normals were removed, as well as SNV/indels that had minor allele frequency ≥1% in 1000Genomes or gnomAD, and SVs overlapping with DGV (2020-02-25 release)⁷⁵, 1000Genomes or gnomAD-SV (v.2.0.1)⁷⁶. SNV/indels with tumour VAF <0.0001, normal VAF >0.2, depth <2 in either the tumour or normal sample, and normal VAF greater than tumour VAF were filtered from the final callset. SNV/indels with support from two or more callers were marked as high confidence. SVs with support from two or more callers or one caller with split-read or CNV changepoint support were marked as high confidence. Variants detected by the somatic calling pipeline in the FFPE normal tissues were removed from the tumour samples. None of the 77 remaining tumour samples was excluded from the analyses due to the absence of ABOPEC mutational signatures.

Purity and ploidy estimation

Purity and ploidy were estimated for each sample using AscatNGS (v.4.2.1)⁷⁷ and Sequenza (v.3.0.0)⁷⁸. Estimates were manually reviewed and chosen based on VAF, B-allele frequency and read depth fit. Owing to low purity, six tumours were excluded from downstream JaBbA analysis.

Study sample size definition

Our study included 83 samples: 77 histologically proven UCs and 5 morphologically normal urothelial samples from 50 patients. Five normal urothelium samples were only included in the comparisons between primary and metastatic tumours and normal urothelial samples. The bladder cancer (primary bladder adenocarcinoma) sample from patient WCMIVG101 was only used for ONT long-read WGS. This was a cross-sectional genomic cohort study of patients with advanced UC. Basic covariates (including sex, age at diagnosis, smoking status, highest stage, primary or metastatic tumour, site of tumour collection and platinum-based chemotherapy status) are listed in Supplementary Table 1.

Union of somatic variants for patients with multiple samples

For patients with multiple samples, a union of ‘HighConfidence’ somatic SNVs and indels across all the patient’s samples was generated. Pileup (v.0.15.0) (https://github.com/pysam-developers/pysam) was then run on tumour and normal BAM files to compute the counts for variants present in the union variant call format (VCF) file that were missing from each sample’s VCF. Variants that had a VAF > 0 were then rescued and added to the sample VCF. The resulting ‘union VCF’ was used for further post-processing.

Mutational signature fitting and assignment

Mutational signature fitting was performed using the R package deconstructSigs (v.1.9)⁷⁹. HighConfidence variants were input, and COSMIC (v.3.2) signatures (https://cancer.sanger.ac.uk/signatures/)⁶⁶ was used as a reference. Arguments ‘contexts.needed=TRUE’ and ‘signature.cutoff=0’ were used for SBS, DBS and ID signatures. Additionally, the experimentally derived SBS.A3G from our APOBEC3G transgenic mouse model⁷ was included to fit the signatures in patients’ tumours. Following signature fitting, for each SNV, we computed a channel of 78 posterior probabilities (corresponding to each of the SBS reference signatures) that the mutation was caused by a given COSMIC signature (similar to ref. ⁸⁰). After discarding posterior probabilities below 0.5, the most likely signature was chosen. The assignment was discarded if the trinucleotide context was not one of the top 5 reference signature peaks. APOBEC (SBS2 and SBS13) and platinum chemotherapy (SBS31 and SBS35) were treated separately based on previous knowledge (SBS2: C>T; SBS13: C>A, C>G; SBS31: C>T, T>A; SBS35: C>A, C>G, C>T, T>A). We then refined platinum chemotherapy assignments using patient clinical data. For samples from patients who did not receive platinum chemotherapy, SBS31 and SBS35 mutation assignments were removed. For patients with multiple samples, SBS31 and SBS35 mutation assignments were removed for mutations occurring in both pre-chemotherapy and post-chemotherapy samples. However, if a mutation was unassigned in one post-chemotherapy sample but was assigned to SBS31 or SBS35 in another post-chemotherapy sample from the same patient, the mutation in the first sample was reassigned.

Mutational signature assignment in matched urothelium samples

Mutational signature assignments of variants in the normal urothelium were determined using matched tumours from the same patient. Pileup (v.0.15.0) (https://github.com/pysam-developers/pysam) was run on the normal urothelium, using positions with a SBS2 or SBS13 assignment in the matched tumours. Variants with VAF > 0 were counted towards the number of variants assigned to a given mutational signature in the normal urothelium.

Signature clonality fold change analysis

To investigate the timing of ageing (SBS1), APOBEC (SBS2 and SBS13) and chemotherapy (SBS31 and SBS35) mutational processes, we computed a signature clonality fold change as previously described⁸. To compute the clonal versus subclonal fold change for each post-chemotherapy sample, we pooled all MutationTimer¹² clonal categories (early clonal, late clonal, clonal (NA), subclonal) and divided the proportion of clonal mutations by the proportion of subclonal mutations assigned to the same signature. The early versus late fold change was similarly calculated. Mutation counts were pooled for their respective COSMIC signatures. Only samples with sufficient purity for upstream JaBbA analysis were included.

For Fig. 1b, we included n = 12 patients with multiple samples who received platinum chemotherapy. Truncal mutations were node alterations shared by all samples in a tree, whereas leaf mutations were node alterations unique to each sample. To statistically substantiate the observation that APOBEC signatures dominated early truncal nodes whereas chemotherapy-associated signatures dominated later leaf nodes in phylogenetic trees, we calculated the fold change of their variants for each patient by dividing the fraction of variants associated with a mutational signature in leaf nodes by the fraction in truncal nodes for each patient. To avoid dividing by zero, a pseudo-count of 1 variant was added before the fold change calculation.

Signature clonality fold change analysis of chromatin-modifying genes

The list of chromatin-modifying genes was curated from REACTOME⁸¹ and from previous publications^14,82,83. Duplicate genes were removed. The CGC gene list is described in the section ‘Oncoprint of chemotherapy-induced mutations’. Notably, among the 743 CGC genes examined, 47 genes overlapped with the 259 chromatin-modifying genes, aligning with the frequent observations of mutations in chromatin-remodelling genes in cancer⁸⁴. Control genes were randomly sampled from 500 genes not in the list of CGC genes and chromatin-modifying genes (Supplementary Table 5). The Ensembl variant effect predictor (VEP)⁷⁴ was used to assign functional impact predictions to the detected variants, whereas the R package MutationTimeR (v.1.00.2)¹² was used to compute the clonality and CCF for each variant (see the section ‘Mutation timing and CCF calculation’). Moderate-impact and high-impact variants predicted by the VEP were analysed to prioritize early urothelial mutations with functional consequences.

Estimating the relative velocity fold change of mutagenic processes

Mutagenic velocity (rate of signature accumulation per month of exposure) was calculated for ageing (SBS1), APOBEC (SBS2, SBS13 and DBS11) and platinum chemotherapy (SBS31, SBS35 and DBS5) signatures. Mutation counts from signatures fitted by deconstructSigs⁷⁹ was divided by the exposure time to mutagenic processes (months). The estimated exposure time for ageing and chemotherapy was between the date of sample collection and the date of birth or the date of the initial chemotherapy treatment, respectively. For patients with an unknown initial diagnosis date, it was assumed to be 1 year before the earliest date of tissue collection. Duration from birth and date of diagnosis were retroactively calculated from age at diagnosis and sample collection dates. For APOBEC, the exposure time is assumed to be from the sample collection dating back to 10 years before the UC diagnosis date, as we expected that APOBEC mutagenesis occurs long before the diagnosis of UC. All samples were given one pseudo count of the ageing signature to prevent division by zero in samples with no ageing signature contribution. Samples with zero APOBEC signature contribution and post-chemotherapy samples with zero chemotherapy contribution were excluded from the calculation. The velocity induced by each mutational process within patients with multiple samples was averaged. Fold change was calculated by dividing APOBEC and chemotherapy velocity by ageing velocity.

Oncoprint of chemotherapy-induced mutations

To address the biological significance of the mutations induced by cisplatin treatment, we examined mutations in coding regions assigned to SBS31 and SBS35, which are induced by previous treatment with platinum-based chemotherapy. CGC genes were curated from Cancer Gene Consensus (CGC v.95 Tier 1 and 2 genes) and TCGA 2017 BLCA driver genes, excluding FLAGS (frequently mutated genes)⁸⁵, with BRCA2, KMT2C and KMT2D rescued back into the gene list.

Association of alterations in driver genes with the presence of SV types

We investigated the association between alterations in driver genes and the presence of SV types using a predefined gene list (Supplementary Table 5). For each combination of gene and SV type, we conducted association tests using the Fisher’s test with a FDR threshold of <0.25. The tested associations included moderate to high impact SNVs/indels, amplifications, a combination of amplifications and SNVs/indels, homozygous deletions and a combination of homozygous deletions and SNVs/indels.

Platinum chemotherapy mutations in regulatory regions and recurrence FishHook analysis

We used three databases for analysis: FANTOM5 for promoters⁸⁶ (209,911 promoters, 4.2 MB), FANTOM5 for enhancers⁸⁷ (63,285 enhancers, 18.6 MB) and previously described super-enhancers⁸⁸ (8,589 super-enhancers, 468.5 MB, with the union taken across 86 cell types). The initial step involved overlaying all SNVs onto these features, disregarding mutational signature assignment, for a comprehensive starting point in our analysis. We then conducted FishHook analysis to identify any regulatory elements that are mutated more than expected based on background. We re-used covariates from FishHook SV analysis (see the section ‘Regions of recurrent structural variation’), and used the union (ignoring strand) of promoters, enhancers and super-enhancers as the list of hypotheses to test. Each patient could only contribute one mutation per hypothesis. We included local mutation density as a covariate and kept the FDR cut-off at 0.25.

SNV enrichment analysis

Enrichr refers to an integrative web-based search engine of various gene set libraries and methods to compute gene set enrichment with interactive visualization of the enrichment results^89–91. In this study, we interrogated the NCI-Nature 2016 pathway database to search for pathways showing a significant increase in the number of mutations assigned to platinum chemotherapy-induced SBS31 and SBS35. The P value was computed from a two-sided Fisher’s exact test, and pathways with adjusted P ≤ 0.2 are reported. The enriched pathways were visualized by Appyter available at Enrichr.

Mutation timing and CCF calculation

The R package MutationTimeR (v.1.00.2)¹² was run using the union of somatic variants, allele-specific CN output from JaBbA, patient sex information and previously estimated purity values. Parameter n.boot was set to 200. MutationTimeR infers a multiplicity for each mutation and assigns a timing based on the multiplicity and the allele-specific CN configuration at that locus. Using MutationTimeR multiplicities, CCF was computed as previously described⁹².

Mutational signature analysis of phylogenetic trees

Phylogenetic trees were generated using LICHeE (v.1.0)⁹³ for each patient with multiple samples. LICHeE was run with a variant-by-CFF matrix as input, the ‘maxVAFAbsent’ argument of 0.0, the ‘minVAFPresent’ argument of 0.5, the ‘maxClusterDist’ of 0.2, and in cell prevalence mode (-cp). The top-scoring tree was selected. The variants in each of the resulting nodes of the phylogenetic tree output were then fed through deconstructSigs as described above to estimate a set of mutational signature proportions for each node.

Calculating the ratio of dN/dS

Genes with evidence of positive selection were detected using the R package dndscv^94–96 using default parameters and GRCh38-specific reference data as supplied by the developers. As per the developer’s instructions, mutations shared across multiple samples from the same patient were only listed once. This method of calculating the dN/dS ratio adjusts for other confounders of clonal selection, such as large gene sizes and a high regional mutation rate⁹⁷. The following classification was used: dN/dS = 1, neutral selection; dN/dS > 1, positive selection; dN/dS < 1, negative selection. Higher missense and nonsense dN/dS ratios indicate a clonal selection of oncogenes and tumour suppressor genes, respectively.

Detection of complex structural variants with JaBbA

Read counts were corrected for GC percentage and mappability in 1-kb bins using fragCounter (https://github.com/mskilab/fragCounter) for all tumour and normal samples. A coverage panel of normal was built from the normal samples and used to denoise the tumour coverage data using dryclean (commit bda8065)⁹⁸. Denoised tumour coverage profiles, B allele frequencies and high-confidence SVs were used as input to JaBbA (v.1.1)², an algorithm that integrates CNV and SVs into a junction-balanced genome graph, computing integer copy numbers for both. Default parameters were used, except for the slack penalty, which was increased to 1,000. Simple inversions, translocations, duplications and deletions, as well as TICs, quasi-reciprocal pairs, rigma, pyrgo, tyfonas, BFB cycles and DMs, were called on the junction-balanced genome graph using the JaBbA companion R package gGnome (commit c390d80)². Using the integer CN as output by JaBbA, we computed the fraction of genome altered, defined here as the proportion of autosomes not in a neutral copy state, as defined by sample ploidy. For samples with an intermediate average ploidy (fractional value of 0.4–0.6, for example, 3.5), the copy-neutral state was set as the closest two integer values (for example, for a ploidy of 3.5, the copy-neutral states would be 3 and 4). Otherwise, the copy-neutral state was set as the rounded ploidy.

Regions of recurrent structural variation

A Gamma–Poisson model, as implemented in the R package FishHook (commit 06e3927)⁹⁹, was used to discover regions of recurrent SVs. The genome was partitioned into 100-kb non-overlapping bins, and the union of breakpoints from each patient was used as input as previously described¹⁰⁰. Regions overlapping intervals of mappability <1 and centromeres >25% were excluded from the analysis. Covariates were added to model the background mutation rate, including the following:

• Nucleotide frequency, dinucleotide frequency and trinucleotide frequency

• H3K4me3 marks (ENCODE accession: ENCFF191IBA), H3K27ac marks (ENCFF208GHP), H3K4me1 marks (ENCFF759BRD), H3K3me3 marks (ENCFF983DSU)

• DNase hypersensitivity sites (ENCFF823HYK)

• Replication timing (https://github.com/skandlab/MutSpot/tree/master/features/Ch38), fragile sites (HGNC 2021) and

• RepeatMasker long interspersed nuclear element, short interspersed nuclear element, long terminal repeat, simple repeat and DNA transposon annotations from UCSC¹⁰¹.

A FDR-adjusted (Benjamini–Hochberg) P value cut-off of 0.25 was used to nominate significant breakpoint hits. Cancer-related genes were from the COSMIC Cancer Gene Consensus¹⁰².

Detection of ecDNA-forming structural variants

We noted that JaBbA does not impose any kind of cyclic constraint when calling high-level amplifications (for example, DMs, tyfonas and BFBs). To determine whether these events were ecDNA-forming SVs, AmpliconArchitect (v.1.2)¹⁷ was run using default parameters on tumour BAMs downsampled to 10x, considering only intervals with integer CNs greater than 4 (inferred by JaBbA) and longer than 10 kb. For two runs on which these parameters could not be applied, we used 10x downsample, CN > 5 and 100-kb intervals. Any high-level amplifications detected by JaBbA overlapped with an AmpliconArchitect-detected cyclic amplicon were nominated as ecDNA-forming SVs.

Kataegis and kyklonas identification

We ran the SigProfilerClusters (v.1.1.2)¹⁰³ software with default parameters to identify kataegis loci, computing a sample-dependent inter-substitution distance of clustered mutations and requiring a kataegis event to have a consistent VAF. Kataegis events contained completely within the footprint of an ecDNA-forming SV are classified as kyklonas.

Relative timing of APOBEC and chemotherapy mutations on ecDNA-forming SVs

As a secondary method for determining the relative contributions and timing of mutational processes acting on ecDNA, we partitioned SNVs in ecDNA-forming SVs by VAF into three groups (VAF ≤ 0.333, 0.333 < VAF ≤ 0.667 and VAF > 0.667) for each sample. We then ran deconstructSigs as described above and VAF combination, requiring that a particular combination have at least 50 SNVs. The relative contribution of APOBEC and chemotherapy mutageneses on ecDNA-forming SVs are presented in Fig. 2f, and all mutational processes in Extended Data Fig. 6a.

Analysis of CCND1 ecDNA amplification in TCGA/PCAWG pan-cancer cohorts

The AmpliconArchitect analyses from a previous study²² were accessed from GitHub. Samples were from both TCGA and PCAWG datasets (based on their identifiers, either SA-XXXX or TCGA-XXXX). Sample barcodes from all specimens were analysed for amplicon intervals corresponding to the genomic location of CCND1 (chromosome 11: 69455855–69469242). The resulting amplicon intervals were then matched to the type of amplification event (heavily rearranged, circular-BFB or linear) as determined by AmpliconArchitect. Only one CCND1 amplification event was identified in cervical cancer and none in B cell lymphoma, glioblastoma, sarcoma, renal or colorectal cancer. The CDKN2A and CCND1 mutation and CN status of bladder cases having CCND1 amplicons were retrieved from cBioPortal. The study abbreviations for different cancer types are listed on the National Cancer Institute website (https://gdc.cancer.gov/resources-tcga-users/tcga-code-tables/tcga-study-abbreviations).

RNA sequencing analysis

STAR aligner (v.2.7.3a)¹⁰⁴ was run in two-pass mode (--twoPassMode Basic¹⁰⁵) versus GRCh38. The STAR index was created using Gencode (v.39) for splice junction annotation, with overhang=99, and including only canonical chromosomes (chromosomes 1–22, X, Y and M). Gencode (v.28) was used for gene quantification. The GTF was filtered to include only genes from the following types (‘gene_type’ field): protein_coding, lincRNA, antisense, IG_LV_gene, IG_V_gene, IG_V_pseudogene, IG_D_gene, IG_J_gene, IG_J_pseudogene, IG_C_gene, IG_C_pseudogene, TR_V_gene, TR_V_pseudogene, TR_D_gene, TR_J_gene, TR_J_pseudogene, TR_C_gene. This filtering was performed using the ‘mkgtf’ utility from CellRanger (https://support.10xgenomics.com/single-cell-gene-expression/software/pipelines/latest/advanced/references#mkgtf). Transcripts annotated as ‘retained_intron’ were also excluded, using the script collapse_annotation.py from the GTex pipeline (https://github.com/broadinstitute/gtex-pipeline/tree/master/gene_model) with argum–t --collapse_only. This filtering was performed to avoid under-counting due to overlap for instances in which genes from these categories overlapped other gene types (like pseudogene versions of the same gene) or in which genes would otherwise be non-overlapping except for ‘retained_intron’ isoforms. The featureCounts¹⁰⁶ module from the subread package (v.1.4.3-p1) was used for read quantification, with the argument ‘-s 2’ for strand-specific samples and argument ‘-s 0’ for non-strand-specific samples. The module ComBat_seq¹⁰⁷ from the sva package (R v.4.1.3, sva v.3.42.0) was used for batch correction. The samples were treated as four batches based on prep type, with a 2 × 2 design for ‘totalRNA’ or physical rRNA removal versus ‘mRNA’ or poly-A selection, and strand-specific versus non-strand-specific. The 39 samples with corresponding DNA were included. DNA and RNA were extracted from the same tumour sample within the same patient except when tumour tissues were not sufficient for both. DESeq2 (R v.3.6.1, DESeq2 v.1.24.0)¹⁰⁸ was used for library size normalization after batch correction. DESeq2 was also used to perform differential expression between samples with versus without gain of CCND1 CN. Gain is defined as CN gain relative to ploidy.

ecDNA enrichment statistical analysis in the WCM-UC cohort

The co-occurence of CCND1 amplification and homozygous loss of CDKN2A was determined using Fisher’s exact test (fisher.test in R v.4.0.0). CCND1 amplification was defined as having a gene modal copy number (CN) greater than two times the sample ploidy or greater than 8. Homozygous loss of CDKN2A was defined as having a gene modal CN of zero. Separate Fisher’s exact tests were performed within CCND1 amplifications, splitting by ecDNA-forming SV status, and across the entire cohort, comparing CCND1 ecDNA-forming SV⁺ to CCND1 ecDNA-forming SV^– samples.

Overall survival analysis

To determine the impact of CCND1 alterations on clinical outcomes, we queried data from 10,953 patients (10,967 samples) across 32 cancer types included in the Pan-Cancer Atlas¹⁰⁹ from cBioPortal (https://bit.ly/4cjAYof). Out of 10,622 patients with overall survival data available, 2,128 patients had CCND1 alterations, including CN gains and CN amplifications. We performed survival analysis using the built-in survival function of cBioPortal.

UCSC Genome browser view of shared CCND1 ecDNA-forming SV regions

The JaBbA amplicon footprints corresponding to each of the putative CCND1 ecDNA-forming SVs were intersected, which produced a single 204 kb genomic interval (chromosome 11: 69614476–69818943). This interval was uploaded as a custom track in UCSC Genome Browser¹⁰¹. Additional tracks selected included NCBI RefSeq gene models (release 2023-03-21), FANTOM5 transcription start site peaks⁸⁶, and H3K27Ac, H3K4Me1 and H3K4Me3 marks from ENCODE¹¹⁰. A tutorial on how to operate the interactive browser interface is available at GitHub (https://github.com/mskilab-org/gGnome.js/blob/master/README.md#the-ggnomejs-interface).

Visualization of ecDNA-forming SV reconstructions

For each putative ecDNA-forming SV (for example, JaBbA-called tyfonas, DMs or BFBs overlapping an AmpliconArchitect cyclic call), the amplicon-associated subgraph was extracted from the full JaBbA genome graph. Each amplicon was then decomposed into constituent walks using the ‘peel’ function in the R package gGnome (commit c390d80), using junction CN to rank walks. In brief, ‘peel’ attempts to iteratively decompose a subgraph by finding walks through the graph that maximize some junction feature (here, JCN). For each iteration, the top-ranked walk was ‘peeled’ off of the subgraph, and the procedure was repeated.

After obtaining a set of walks for each amplicon, the non-cyclic walks and walks containing genomic intervals with CNs fewer than eight were removed. Each walk that passed the aforementioned filters was then compared against the associated AmpliconArchitect cyclic event. The best overlapping walk was then kept, and the others were discarded. Finally, walks with reciprocal overlap ≥80% were selected for visualization with CycleViz (v.0.1.5)¹¹¹ (https://github.com/AmpliconSuite/CycleViz). The preset BUSHMAN was used for displaying oncogenes contained within ecDNA reconstructions.

ONT library preparation and sequencing

Samples were fragmented to 10 kb using Covaris g-TUBE (COVARIS, 520079). DNA repair and end-prep were performed using the NEBNext Companion Module for ONT (NEB, E7180S), incubating at 20 °C for 5 min, followed by 5 min at 65 °C. The sample was cleaned up using 1× Ampure XP beads (Beckman Coulter, A63881) and washed twice on a magnetic rack with 80% ethanol. Sequencing adaptors and ligation buffer from an Oxford Nanopore Ligation Sequencing kit (ONT, LSK1114) were ligated to DNA ends using Quick T4 DNA ligase (NEB, E6056) for 45 min at room temperature. The sample was cleaned up using 0.45× Ampure XP beads (Beckman Coulter, A63881), washing twice on a magnetic rack with the long-fragment buffer (ONT, LSK114) before eluting in 32 µl of elution buffer (ONT, LSK114). Sequencing libraries were prepared by adding the following to the eluate: 100 µl sequencing buffer (ONT, LSK114), 68 µl loading solution (ONT, LSK114) and 0.5 µl sequencing tether (ONT, LSK114). Samples were run on a PromethION (R10.4.1) flow cell using the PromethION sequencer. Sequencing runs were operated using MinKNOW software (v.23.07.12). Live base-calling was performed in high accuracy (HAC) mode. Read N50 was 12,186 bp, and mean coverage was 58×.

Assembly of circular ecDNA contigs

ONT reads were aligned to GRCh38 with minimap2 (v.2.26-r1175)¹¹² with flags -a–L -–D --cs -x map-ont, and coordinate-sorted with samtools (v.1.18)¹¹³. Reads overlapping the amplified regions of interest were extracted to FASTQ using samtools and used as input into assembly. A total of 196,924 long reads overlapped CCND1 ecDNA reference coordinates with N50 10,087 and mean coverage 1,403×. GRCh38 ALT contig chr11_KI270827v1_alt overlapped one of the regions of interest and had many reads aligned to it instead of the corresponding locus on chromosome 11, so these were included in the assembly for the relevant amplicon. Reads were assembled with Flye (v.2.9.2-b1786)¹¹⁴ in metagenomics mode, with the –nano-hq setting and –read-error set to 0.03 as recommended for ONT Q20+ chemistry. A single round of polishing was used. Following assembly, the Flye assembly summary was reviewed to confirm that a circular contig was assembled. The assembled contig was mapped back to GRCh38 using minimap2 with assembly alignment preset -ax asm5 for confirmation that the assembly footprint on the reference genome matched the amplicon detected using Illumina sequencing.

Cell culture

Bladder cancer cell lines UMUC3 (CRL-1749) and 5637 (HTB-6), and HEK293T (CRL-3216) were purchased from the American Type Culture Collection. The SF295 cell line was purchased from Neuromics (SF-001). Cell lines UMUC3, 5637 and SF295 were validated by STR testing and/or morphology by the manufacturer. The UMUC3, SF295, 5637 and HEK293T cells were incubated in a humidified constant 37 °C and 5% CO₂ incubator in EMEM (ATCC), RPMI-1640 (Gibco) (for SF295 and 5637) and DMEM (Gibco) medium, respectively, with 100 U ml^–1 penicillin–streptomycin (Gibco) and 10% tetracycline-free FBS (Omega Scientific). The cells were tested negative for Mycoplasma contamination with a PCR detection kit (ABM) according to the manufacturer’s instructions.

Plasmids for episomal ecDNA and shRNA knockdown

The non-integrating lentiviral episomal vector we used in this study has homogeneous long terminal repeat regions that promote the circularization of the transduced CCND1 DNA. As this vector is integrase-deficient, it models the extrachromosomal configuration of CCND1 ecDNA. Finally, the scaffold/matrix attachment region (S/MAR) sequence, which interacts with the nuclear matrix through scaffold attachment factor A (SAF-A), ensures the self-replication and partitioning of the episome during mitosis¹¹⁵. The pBMN (CMV-copGFP-Puro-SMAR) plasmid, a gift from M. Essand (Addgene, 80391)¹¹⁶, was modified to include a mCherry-T2A-5′HA-tagged CCND1 coding sequence, synthesized with XbaI and AgeI sites by Integrated DNA Technologies. Following digestion with XbaI and AgeI, the fragment was cloned into the pBMN vector between the CMV promoter and Puro-SMAR sequence using T4 DNA ligase (New England BioLabs) (Supplementary Figs. 1–5). This plasmid was transformed into One Shot Stbl3 Chemically Competent Escherichia coli (Thermo Fisher Scientific) and collected using a PureLink HiPure Plasmid Filter Maxiprep kit (Thermo Fisher Scientific), according to the manufacturer’s instructions. The shRNA CCND1 knockdown plasmid (pLV-mCherry-T2A-Puro-U6-shRNA) and shRNA control plasmid (pLV-EGFP-T2A-Puro-U6-scramble control) were generated by VectorBuilder. The target CCND1 sequence was TGTTGTAAGAATAGGCATTAA. The plasmids were validated by Sanger sequencing.

Cell line generation for episomal ecDNA and shRNA knockdown

HEK293T cells were transfected with the interested plasmid in combination with pPAX2 and pVSV-G plasmids (Addgene). The virus-containing medium was collected every 24 h for 3 days post-transfection, filtered using a 0.22-µm sterile 50 ml tube top filter (Corning) and concentrated using Lenti-X Concentrator (Takara). The concentrated lentivirus was used for transduction with 10 µg ml^–1 polybrene (Millipore Sigma). Transduced cells were selected using 2 µg ml^–1 puromycin (Gibco). The expression level was validated by SDS–PAGE western blotting of cyclin D1.

Competition assay tracking using the Incucyte living-image system

The GFP-tagged pBMN empty vector UMUC3 cells were co-cultured with mCherry-tagged pBMN CCND1 episomal ecDNA UMUC3 cells in a 1:1 ratio, initially seeded into a 6-well plate. Likewise, scramble shRNA SF295 cells were co-cultured with CCND1 knockdown SF295 cells in a 1:1 ratio, initially seeded into a 6-well plate. After 24 h of seeding, the co-cultured cells were treated with 1.5 µM cisplatin or vehicle control. Cisplatin (at the stock concentration of 1 mg ml^–1) was freshly dissolved in medium instead of DMSO to prevent loss of activity¹¹⁷ and was protected from light. Each 6-well plate was then placed in the Incucyte system (Sartorius) for live tracking of GFP and mCherry signals every 6 h. To sustain the selective pressure of cisplatin, the medium containing cisplatin was refreshed after 48 h. For replicates, 25 spots within each well were imaged. After 96 h of monitoring, GFP and mCherry signals were quantified using Incucyte software (Incucyte, 2022B, Rev2). Subsequently, the ratio of mCherry and GFP signal area was calculated and normalized to the baseline (time 0) measurements.

Single-cell RNA sequencing

The competitive assay experiment was repeated in T75 flasks for single-cell RNA sequencing to identify the impact of CCDN1 episomal ecDNA on the transcriptional profiles under selective pressure. Following a 96-h exposure to either vehicle or cisplatin (1.5 µM), live cells tagged with GFP and mCherry were isolated using fluorescence-activated cell sorting (BD FACSAria II) and subsequently processed by the WCM genomic core facility. The 10x Genomics chromium single-cell 3′ library RNA sequencing kit was used to construct libraries, and sequencing was performed using Illumina NovaSeqXplus with paired-end sequencing, producing read 1 of 28 bp and read 2 of 90 bp.

Single-cell RNA sequencing analysis

Alignment

The obtained sequencing data underwent processing through 10x cellranger (v.7.1.0) with the ‘ --include-introns’ option set to false. Specifically, cellranger mkfastq was used to generate Fastq files. For the alignment, we used the 10x prebuilt GRCh38-2020-A human reference.

Preprocessing

Raw single-cell RNA sequencing results of four samples (vehicle–GFP, vehicle–mCherry, cisplatin–GFP and cisplatin–mCherry) were processed using CellRanger (v.7.1.0). Matrix files for each pair of samples (GFP and mCherry cells under the same treatment) were read with Scanpy (v.1.9.6)¹¹⁸ and concatenated in the same AnnData object (v.0.10.3)¹¹⁹. Cells were filtered to have at least 200 genes and a mitochondrial count below 20%. Genes were filtered to those present in at least three cells. Transcriptomes were normalized by their total size and scaled to 10,000 UMI counts, and highly variable genes were selected using Scanpy pp.highly_variable_genes function with batch_key=‘sample’. The data matrix underwent logarithmic transformation using the natural logarithm using Scanpy pp.log1p function. We scaled the gene expression per gene by subtracting the mean and dividing by the standard deviation using the Scanpy pp.scale function. Then we performed principal component analysis using the Scanpy tl.pca function with svd_solver = ‘arpack’ and performed batch correction using Harmony¹²⁰.

E2F score calculation

The E2F score was calculated in each single cell by calculating the average expression of all 200 genes in the E2F target pathway from the Molecular Signatures Database (MSigDB)¹²¹ subtracted the average expression of a reference set of genes using the Scanpy tl.score_score_genes function. We constructed a k-nearest neighbour graph using the pp.neighbors function from Scanpy, setting the number of neighbours to 25 and the number of principal components to 40. We then applied UMAP for dimensionality reduction using the Scanpy tl.umap function for each anndata pair (GFP and mCherry cells under the same treatment). For visualizing E2F transcription factor activity across individual cells, we generated a UMAP colour-coded by the pre-calculated E2F score to differentiate between low and high E2F activity states. We then plotted E2F scores as boxplots and statistically compared the E2F score between each pair of samples using the Mann–Whitney test.

Differential gene expression and gene set enrichment analysis

To compare the transcriptomic profiles between each pair of samples (GFP and mCherry cells under the same treatment), we performed differential expression of genes on pre-processed anndata between GFP and mCherry samples under each treatment using t-test as implemented in Scanpy tl.rank_genes_groups function. Gene enrichment analysis of significantly upregulated and downregulated genes was done using the GSEApy package (v.1.1.1) with the GSEApy prerank function. Differentially regulated pathways were identified using gene set enrichment analysis from the GSEA (Broad Institute) package based on signatures from MSigDB (MSigDB_Hallmark_2020)¹²¹. Bubble plots of enriched gene sets were generated based on NESs and associated FDR q value.

Cell cycle analysis

The UMUC3 bladder cancer cells were seeded in a 6-well plate and cultured for 24 h, followed by overnight synchronization in serum-free medium. Cells were then treated with either EMEM with 10% FBS or EMEM with 10% FBS with 1.5 µM cisplatin, freshly prepared without DMSO. After 24 h, cells were collected, resuspended in 200 µl of ice-cold PBS, and dropwise fixed with 3 ml of 80% ice-cold ethanol while vortexing to minimize clumping, then stored overnight at −30 °C. The fixed cells were washed with PBS to remove the ethanol and stained with a DAPI solution (1 mg ml^–1, BD Pharmingen, diluted 1:1,000 in 0.1% Triton X-100 PBS) for 30 min at room temperature in the dark. Stained cells were analysed on a BD LSRFortessa flow cytometer (BD Biosciences) using a 405 nm laser with a 450/50 nm emission filter. The DNA content was analysed using FlowJo (BD Biosciences), for which cell gating was based on forward and side scatter to isolate single cells, and cell cycle proportions were calculated using the Dean–Jett–Fox model¹²².

FISH analysis

CCND1 (labelled green) and centromeric chromosome 11 (labelled orange) probes were used (Empire Genomics). Before use, all probes were validated on metaphase spreads.

To assess extrachromosomal CCND1 localization, the bladder cancer cell line UMUC3 and the human glioblastoma cell line SF295 were first treated with colcemid (0.1 µg ml^–1) for 1 h to obtain a metaphase preparation. A hypotonic solution was added, then cells were fixed in methanol and acetic acid (3:1). The fixed cells were placed on microscopy glass slides and dried for 1 week at room temperature. Prepared slides were dehydrated in a series of ethanol. Probes were then added to the slides, which were denatured and hybridized, washed and then counterstained with DAPI before visualization. CCND1 was considered present in ecDNA when the green signal was observed outside the chromosome spread, and the centromeric chromosome 11 served as control².

To assess gene amplification, CCND1 enumeration from interphase FISH was carried out on FFPE tissue slides (5 µm thick). CCND1 amplification was determined by a ratio of CCND1 gene signal (green) to centromere signal (orange) higher than two. A minimum of 100 nuclei per slide were scored. Analysis was carried out using a fluorescence microscope (Olympus BX51; Olympus Optical). CytoVision (v.7.3.1) software (Leica Biosystems) was used for imaging^123,124.

Preparation of metaphase spreads

Cells were grown to 80% confluency in a 15-cm dish and metaphase-arrested by adding KaryoMAX colcemid (10 µl ml^–1, Gibco) for 1–2 h. Cells were washed with PBS, trypsinized (Gibco) and centrifuged at 200g for 10 min. We added 10 ml of 0.075 M KCl preheated to 37 °C, 1 ml at a time, vortexing at maximum speed in between. Afterwards, cells were incubated for 20 min at 37 °C. Then, 5 ml of ice-cold 3:1 methanol and acetic acid (kept at −20 °C) was added, 1 ml at a time, followed by resuspension of the cells by flicking the tube. The sample was centrifuged at 200g for 5 min. Addition of the fixative followed by centrifugation was repeated four times. Two drops of cells within 200 µl of methanol and acetic acid were dropped onto prewarmed slides from a height of 15 cm. Slides were incubated overnight.

SDS–PAGE western blotting

Cells were lysed with RIPA buffer with added protease inhibitor (Thermo Fisher Scientific) to produce whole protein lysate. Protein concentration was determined using a Pierce BCA assay (Thermo Fisher Scientific). The lysate was run on a 10% SDS–PAGE gel with MOPS buffer and transferred using an iBlot system (Thermo Fisher Scientific). Rabbit anti-cyclin D1 (Abcam, ab134175, 1:2,000), mouse anti-α-tubulin (Millipore, 05-829, 1:1,000), mouse anti-Rb (Cell Signaling Technology, 9309S, 1:1,000), rabbit anti-phospho-Rb (Ser807/811) (Cell Signaling Technology, 8516S, 1:2,000) and rabbit anti-p16^INK4A (Cell Signaling Technologies, 80772, 1:1,000) antibodies were separately used as primary antibodies at 4 °C overnight. HRP-conjugated secondary antibodies (goat anti-rabbit, 32260, and goat anti-mouse, 32230, Invitrogen, 1:1,000) were incubated for 1 h at room temperature. Blots were developed using Luminata Forte Poly HRP substrate (Millipore) reagent, imaged on a ChemiDoc imager (Bio-Rad) and analysed using Image Lab (Bio-Rad v.6.1.0) software.

Establishing stable 5637 cell line with doxycycline-inducible APOBEC3A expression

All-in-one doxycycline-inducible lentivirus backbone included APOBEC3A cDNA with HA-tag and Venus with continuously expressing promoter was a gift from M. D. Weitzmann^125,126. The lentivirus was transduced into the 5637 bladder cancer cell line, followed by the fluorescence-activated cell sorting of Venus-positive cells. APOBEC3A expression level was validated by SDS–PAGE western blotting using a HA-tag antibody.

Immunofluorescence microscopy

Cells were plated on Chamber slides (ThermoFisher Scientific) and treated with vehicle (DMSO) or doxycycline for 24 h. Cells were fixed in 4% paraformaldehyde at room temperature for 10 min. The fixed nuclei were treated with 0.5% Triton X-100 in PBS and blocked by 2.5% normal goat serum in PBS. Mouse anti-phospho-histone H2A.X (Ser139) antibody (Cell Signaling Technology, 80312, 1:1,000) and rabbit anti-phospho RPA32 (S4/S8) antibody (Bethyl Fortis Life Sciences, A300-245A, 1:1,000) were used as the primary antibodies and incubated at 4 °C overnight. Then, AlexaFluor 488-conjugated AffiniPure donkey anti-rabbit antibody (Jackson ImmunoResearch Laboratories, 711-545-152, 1:1,000) and AlexaFluor 594-conjugated AffiniPure donkey anti-mouse antibody (Jackson ImmunoResearch Laboratories, 715-585-150, 1:1,000) were used as secondary antibodies and incubated for 2 h at room temperature. Negative blank controls were included by omitting primary antibodies in the protocol. After that, the slide was sealed with ProLong Diamond Antifade mountant with DAPI (Invitrogen, p36962). High-resolution imaging was performed using a fluorescent microscope (Axioscop 2 M2 with Plan Apochromat 633/1.4 NA oil differential interference contrast objective; Carl Zeiss) with a camera (CoolSNAP HQ; Photometrics) using the AF488 filter (38 HE GFP) and AF594 filter on the Axio Imager M2. Multiple fields of view for each condition were randomly picked for imaging. Z stacks (0.25 µm) were collected and subjected to constrained iterative deconvolution using Zeiss deconvolution software (Zen desk v.3.7). Images were processed using Fiji ImageJ (v.154f) software¹²⁷. Each image was stacked using the Z-stacked function of the software with the ‘Max intensity’ property. The number of nuclei in each image was manually counted in the blue DAPI channel, discarding cells with areas less than 80% present in the image. Cells were considered pRPA-positive if they formed ≥5 discrete foci or γH2AX-positive if they formed ≥5 discrete foci within nuclei¹²⁸. Mann–Whitney test was used to statistically compare percentages of positive nuclei between the two treatments using GraphPad Prism (v.10.2.0).

Statistical tests

Two-sided Wilcoxon rank-sum test, two-sided t-test or Pearson test for continuous variables was performed using R (v.4.0.0) software. Unless specified otherwise, P < 0.05 was considered significant.

Ethical approval

The studies involving human participants were reviewed and approved by the WCM IRB. Patients and participants provided their written informed consent to participate in this study. Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

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/s41586-024-07955-3.

Author contributions

Initiation and design of the study: D.D.N., W.F.H., O.E., N.R. and B.M.F. Enrolment of participants, sample collection and preparation, and clinical data collection: D.D.N., W.F.H., Z.R.G., L.W., M. Sigouros, J. Manohar, J. Moyer, D.W., S.T.T., D.M.N., J.T.N., C.N.S., A.M.M., D.S., G.I., J.M.M., O.E., N.R. and B.M.F. Experimental work: W.L., D.D.N., M.O., R.R.S., A.H., J. Moyer, M.A.A., A. Semaan, S.C., F.M.R., D.W., H.L.V. and P.A. Statistical and bioinformatics analyses: D.D.N., W.F.H., W.L., T.R.C., H.G., J.M.S., M. Shah, M.O., R.R.S., A. Sboner, O.E., N.R. and B.M.F. Supervision of research: O.E., N.R. and B.M.F. Writing of the first draft of the manuscript: D.D.N. and B.M.F. All authors contributed to the writing and editing of the revised manuscript and approved the manuscript.

Peer review

Peer review information

Nature thanks Rouf Banday, David McConkey and Roel Verhaak for their contribution to the peer review of this work.

Data availability

All BAM files and associated sample information are deposited in dbGaP under accession phs001087.v4.p1. Single-cell RNA sequencing datasets from this study have been deposited in the Sequence Read Archive with the accession number SUB14118668 (BioProject PRJNA1061081). The analysis files from single-cell RNA sequencing, ecDNA amplicon reconstructions, Incucyte live-cell images, immunofluorescence pRPA and γH2AX foci images, and the according analysis files have been deposited into Zenodo 10.5281/zenodo.11121869¹²⁹. The TCGA/PCAWG pan-cancer human cancer data²² used for CCND1 amplification analysis was obtained and modified from the supplementary information of that article²². Data for the CCND1 pan-cancer survival analysis was obtained from cBioPortal (https://bit.ly/4cjAYof). Source data are provided with this paper.

Code availability

The following open-source code and databases were used in this article: JaBbA (v.1.1) (https://github.com/mskilab/JaBbA), gGnome (commit c390d80) (https://github.com/mskilab/gGnome), AmpliconArchitect (https://github.com/virajbdeshpande/AmpliconArchitect), FishHook (commit 06e3927) (https://github.com/mskilab/fishHook), MutationTimeR (v.1.00.2) (https://github.com/gerstung-lab/MutationTimeR), deconstructSigs (v.1.9) (https://github.com/raerose01/deconstructSigs), SigProfilerClusters (v.1.1.2) (https://github.com/AlexandrovLab/SigProfilerClusters), Pileup (v.0.15.0) (https://github.com/pysam-developers/pysam), ShortAlignmentMarking (v.2.1) (https://github.com/nygenome/nygc-short-alignment-marking), BWA-MEM (v.0.7.15) (https://github.com/lh3/bwa), GATK (v.4.1.0) (https://github.com/broadinstitute/gatk), MuTect2 (v.4.0.5.1) (https://github.com/broadinstitute/gatk), Strelka2 (v.2.9.3) (https://github.com/Illumina/strelka), Lancet (v.1.0.7) (https://github.com/nygenome/lancet), Svaba (v.0.2.1) (https://github.com/walaj/svaba), Manta (v1.4.0) (https://github.com/Illumina/manta), Lumpy (v.0.2.13) (https://github.com/arq5x/lumpy-sv), SplazerS (v.1.1) (https://github.com/seqan/seqan/tree/master/apps/splazers), Ensembl (v.93) (https://www.ensembl.org), COSMIC (v.86) (https://cancer.sanger.ac.uk), COSMIC Cancer Gene Consensus (v.95) (https://cancer.sanger.ac.uk/census), ClinVar (201706) (https://www.ncbi.nlm.nih.gov/clinvar/), PolyPhen (v.2.2.2) (http://genetics.bwh.harvard.edu/pph2/index.shtml), SIFT (v.5.2.2) (http://sift-dna.org/sift4g), FATHMM (v.2.1) (http://fathmm.biocompute.org.uk), gnomAD (r.2.0.1) (https://gnomad.broadinstitute.org/), gnomAD-SV (v2.0.1) (https://gnomad.broadinstitute.org/, https://github.com/talkowski-lab/gnomad-sv-pipeline), dbSNP (v.150) (https://www.ncbi.nlm.nih.gov/snp/), Variant Effect Predictor (VEP) (v.93.2) (http://www.ensembl.org/vep), Database of Genomic Variants (DGV) (2020-02-25 release) (http://dgv.tcag.ca/), AscatNGS (v.4.2.1) (https://github.com/cancerit/ascatNgs), Sequenza (v.3.0.0) (http://www.cbs.dtu.dk/biotools/sequenza), LICHeE (v1.0) (https://github.com/viq854/lichee), fragCounter (https://github.com/mskilab/fragCounter), dryclean (commit bda8065) (https://github.com/mskilab/dryclean), RepeatMasker (created in 2010 with the original RepBase library from 2010-03-02 and RepeatMasker 3.0.1) (https://www.repeatmasker.org/species/hg.html). Scanpy (v.1.9.6) (https://github.com/scverse/scanpy), GSEApy (v.1.1.1) (https://github.com/zqfang/GSEApy), CycleViz (v.0.1.5) (https://github.com/AmpliconSuite/CycleViz) and CellRanger (v.7.1.0) (https://github.com/10XGenomics/cellranger). Custom analysis scripts and scripts to reproduce figures are available at GitHub (https://github.com/nygenome/UrothelialCancer_WGS_paper_figures). The JaBbA SV browser includes detailed interactive maps of our structure variant calls (https://urothelial-cancer-wcm-2023.nygenome.org/). Image Lab (Bio-Rad v6.1.0) (https://www.bio-rad.com/) was used for western blot image processing and analysis. CytoVision (v.7.3.1) (https://www.leicabiosystems.com/) was used for FISH imaging. Zeiss deconvolution software (Zen desk v.3.7) (https://www.zeiss.com/microscopy/en/products/software/zeiss-zen-desk.html), Fiji ImageJ (v.154f) (https://imagej.net/software/fiji/) and GraphPad Prism (v.10.2.0) (https://www.graphpad.com/) were used for immunofluorescence image processing and analysis. Incucyte software (2022B, Rev2) (https://www.sartorius.com) was used for competitive assays. FlowJo (v.10.10.0) (https://www.flowjo.com/) was used for the analysis of FACS data. R (v.4.0.0) software was used for statistical tests.

Competing interests

B.M.F.: consulting or advisory role for QED therapeutics, Boston Gene, Astrin Biosciences Merck, Immunomedics/Gilead, QED therapeutics, Guardant and Janssen; patent royalties from Immunomedics/Gilead; research support from Eli Lilly; and honoraria from Urotoday. O.E.: stock and other ownership interests from Freenome, OneThree Biotech, Owkin and Volastra Therapeutics, and personal fees from Pionyr Immunotherapeutics and Champions Oncology. S.T.T.: consulting or advisory role for 4D Pharma, Abbvie, AIkido Pharma, Amgen, Astellas Pharma, Bayer, Blue Earth Diagnostics, Clarity Pharmaceuticals, Clovis Oncology, Convergent Therapeutics, Dendreon, Endocyte, Genentech, Genomic Health, Gilead Sciences, Immunomedics, Janssen, Karyopharm Therapeutics, Medivation, Myovant Sciences, Novartis, Pfizer, POINT Biopharma, QED Therapeutics, Sanofi, Seagen, Telix Pharmaceuticals and Tolmar; research funding from Abbvie, Amgen, Astellas Pharma, AstraZeneca, AVEO, Bayer, Boehringer Ingelheim, Bristol-Myers Squibb, Clovis Oncology, Dendreon, Endocyte, Exelixis, Genentech, Immunomedics, Inovio Pharmaceuticals, Janssen, Karyopharm Therapeutics, Lilly, Medivation, Merck, Millennium, Newlink Genetics, Novartis, POINT Biopharma, Progenics, Rexahn Pharmaceuticals, Sanofi and Stem CentRx; patents, royalties, other intellectual property or patent royalty from Immunomedics/Gilead; travel, accommodation and expenses from Amgen, Immunomedics and Sanofi; and uncompensated relationships from ATLAB Pharma and Phosplatin Therapeutics. D.M.N.: consulting or advisory role for AstraZeneca, and research funding from AstraZeneca, Boehringer Ingelheim, Clovis Oncology, Exelixis, Immumedics, Janssen, Novartis, Pfizer and Zenith Epigenetics. J.T.N.: consulting or advisory role for AIQ Solutions, and travel, accommodation or expenses from Digital Science Press. C.N.S.: consulting or advisory role for Astellas Pharma, AstraZeneca, Bayer, Bristol-Myers Squibb/Medarex, Foundation Medicine, Genzyme, Immunomedics, IMPAC Medical Systems, Incyte, Medscape, Merck, MSD, Pfizer, Roche and UroToday. A.M.M.: consulting or advisory role for Eisai, Exelixis and Janssen. J.M.M.: research funding from Personal Genome Diagnostics, and travel, accommodation or expenses from Personal Genome Diagnostics. D.S.: research funding from Urogen Pharma, Cepheid, Anchiano and CryoLife. P.A.: stock and other ownership interests with Abyost Pharmaceuticals; consulting or advisory role for ArTara Therapeutics; research funding from Adaptive Biotechnologies, Janssen Oncology and natera; and patents, royalties or other intellectual property from intravesical imidazolium compounds and urine biomarkers patent application. D.D.N., W.F.H., T.R.C., H.G., J.M.S., M. Shah, Z.R.G., L.W., A.H., M. Sigouros, J. Manohar, J. Moyer, M.A.A., A. Semaan, S.C., F.M.R., D.W., M.O., R.R.S., W.L., H.L.V., A. Sboner, G.I. and N.R. declare no competing interests.

Extended data

is available for this paper at 10.1038/s41586-024-07955-3.

Supplementary information

The online version contains supplementary material available at 10.1038/s41586-024-07955-3.