MMEJ repair of breaks at TA repeats maintains ecDNA and cancer fitness

ecDNA comprises circular, megabase-sized DNA elements that frequently carry oncogene amplifications and are found in approximately 17% of all human cancers¹. Because ecDNA segregates randomly during cell division, it drives substantial intratumoural heterogeneity^2,3. Serving as hubs for high-level gene expression and rapid genomic evolution, ecDNAs confer an adaptive advantage to cancer cells under therapeutic and environmental pressure. Their presence is associated with aggressive tumour behaviour, resistance to targeted therapies and poor clinical outcomes⁴.

Mechanistic studies have established non-homologous end joining (NHEJ) as a central pathway driving the formation of ecDNA, particularly following chromothripsis, in which it ligates shattered chromosomal fragments into circular, oncogenic elements^7,8. CRISPR–Cas9-based screens and genetic studies have further emphasized the critical role of core NHEJ factors, including LIG4, as well as homologous recombination (HR) components such as BRCA1, in catalysing ecDNA formation^9,10. Yet how these elements are maintained remains largely unknown. ecDNAs are subject to persistent replication stress from disorganized replication and increased transcription, often exceeding that of linear chromosome amplification¹¹. This drives transcription–replication conflicts and the accumulation of DNA breaks during S phase¹². Exogenous DNA damage induced by hydroxyurea or ionizing radiation further exacerbates ecDNA instability, promoting its sequestration into micronuclei and subsequent elimination or reintegration into chromosomes^13,14. Despite this vulnerability, ecDNAs are stably maintained and expand in tumours, suggesting the engagement of DNA repair pathways that can resolve breaks and preserve ecDNA integrity under ongoing stress.

The error-prone microhomology-mediated end joining (MMEJ) pathway of DNA double-strand break (DSB) repair has emerged as a key regulator of extrachromosomal genetic elements¹⁵. MMEJ facilitates the circularization and replication of retrotransposons such as HMS-Beagle in flies and intracisternal A-particle (IAP) in mouse cells, controls the chromosomal integration of T-DNA (transferred DNA) in plants and plasmid DNA in mammalian cells, and contributes to the accumulation of microDNA in cancer cells. MMEJ has also been implicated in the oncogenic integration of the hepatitis B virus into human genomes (reviewed in ref. ¹⁵). Finally, MMEJ serves as a backup during ecDNA formation, driving circularization following excision induced by two widely spaced CRISPR–Cas9-generated breaks⁹. These roles underscore the central role of MMEJ in non-chromosomal DNA processing, stabilization and integration. In this study, we demonstrate that MMEJ components, unlike those of canonical NHEJ or HR, are essential for maintaining ecDNA levels in cancer cells. This dependency is underpinned by the accumulation of DNA breaks at TA-rich sequences in circular genomes, which are predisposed to forming stable secondary cruciform structures and may serve as hotspots for rearrangement^16,17,18. While FANCM unwinding of secondary structure suppresses TA repeat fragility, structures that escape resolution are cleaved by the ERCC1–ERCC4 endonuclease complex, generating DSBs that are repaired by Polθ-mediated MMEJ. Finally, single-cell whole-genome sequencing and pan-cancer whole-genome sequencing analysis converge on the same conclusion: TA repeat fragile sites are recurrent hotspots for structural rearrangements within ecDNA. Together, our findings reveal that TA repeat instability drives structural rearrangements on ecDNA.

To evaluate the role of MMEJ in ecDNA maintenance, we compared a pair of isogenic cell lines derived from the same patient: COLO320DM cells, which carry MYC-ecDNA (also known as double minutes (DM)), and the control COLO320HSR cells, which contain a chromosomally integrated MYC amplification within a homogeneously staining region (HSR)¹⁹. The cell lines were exposed to two selective small-molecule inhibitors of Polθ, a key effector of the MMEJ pathway^20,21. RP-6685 (Polθi-Pol) targets the polymerase domain of Polθ and blocks MMEJ at micromolar concentrations²². RP-2119 (Polθi-Hel) is a more potent inhibitor, targeting the helicase domain of Polθ and disrupting its ATPase activity with nanomolar efficacy²³ (Extended Data Fig. 1a–d). Following a 7-day Polθ inhibitor treatment, metaphase spreads and fluorescence in situ hybridization (FISH) staining revealed a significant reduction in MYC-ecDNA in COLO320DM cells (Fig. 1a,b), whereas chromosomally amplified MYC in COLO320HSR cells remained unaffected (Extended Data Fig. 1e). These findings were replicated across multiple ecDNA-positive cancer cell lines, including PC3-DM prostate cancer (Extended Data Fig. 1f), SNU-16 gastric cancer (Extended Data Fig. 1g) and NCI-H716 colorectal cancer cells (Extended Data Fig. 1h), all of which showed significant reductions in ecDNA copy number upon Polθ inhibition and are in agreement with a previous report implicating MMEJ in ecDNA stability²⁴. To isolate the role of MMEJ in ecDNA maintenance from its potential role in formation, we used an established neural stem cell line in which transient Cre expression generates MYC ecDNA while permanently excising the chromosomal MYC locus, precluding any further de novo ecDNA formation²⁵. Polθ inhibitors were applied only after Cre-mediated recombination was complete, ensuring that all ecDNA present was pre-existing and no new copies could arise. This resulted in a marked and progressive depletion of ecDNA over time (Extended Data Fig. 1i–l), establishing that MMEJ is required to sustain the existing ecDNA pool independently of its formation.

To assess the contribution of other DSB repair pathways in maintaining ecDNA, we inhibited NHEJ using a DNA-PKcs inhibitor (1 μM NU7441) and suppressed HR by short hairpin RNA (shRNA)-mediated depletion of BRCA2 (Extended Data Fig. 1m–o). Analysis of MYC-FISH staining on metaphase spreads revealed that neither treatment significantly reduced ecDNA levels (Fig. 1a,c,d and Extended Data Fig. 1f,g,o), supporting a specific and non-redundant role for MMEJ in maintaining ecDNA levels. Given the well-established synthetic lethality between MMEJ and the two canonical DSB repair pathways^20,26, we could not determine whether the residual ecDNA observed following Polθ inhibition was maintained through compensatory repair by NHEJ or HR. To corroborate these findings, we deleted RHNO1 (coding for RHINO), which promotes Polθ recruitment to DNA lesions²⁷. Deletion of RHNO1 in COLO320DM cells resulted in a significant reduction in ecDNA copy number (Fig. 1e and Extended Data Fig. 1p), consistent with a role for MMEJ in maintaining ecDNA.

Polθ inhibition also led to the accumulation of DNA damage in ecDNA-containing cells. COLO320DM cells exhibited higher baseline levels of DNA damage compared with COLO320HSR cells, as evidenced by increased accumulation of 53BP1 and γH2AX foci¹², and the DNA damage was further increased by Polθ inhibition. We observed similar results in PC3-DM cells, which showed increased DNA damage foci upon Polθ inhibition compared with PC3-HSR cells (Fig. 1f and Extended Data Fig. 2a,b). ecDNAs are known to exhibit high transcriptional activity, leading to frequent replication–transcription conflicts that can generate DNA breaks¹². Pharmacological inhibition of transcription using triptolide or a CDK9 inhibitor rescued the elevated DNA damage observed upon Polθ inhibition, implicating MMEJ in resolving damage inflicted by transcription (Fig. 1g and Extended Data Fig. 2c–f). To confirm that the detected DNA damage occurred specifically at ecDNA loci, we performed combined 53BP1 immunofluorescence and MYC-FISH staining, which revealed significant colocalization of DNA damage foci with ecDNA (Extended Data Fig. 2g,h), indicating that Polθ inhibition induces DNA damage specifically at extrachromosomal elements.

Faithful inheritance of ecDNA during cell division is driven by its tethering to chromosomal DNA during mitosis, a mechanism that ensures co-segregation and inheritance in daughter cells. Specifically, inhibition of BRD4 with JQ1 or loss of the long non-coding RNA PVT1 triggered extensive, acute ecDNA untethering within 6–24 h of perturbation³. More broadly, CpG-rich promoters have been shown to act as retention elements that facilitate ecDNA tethering²⁸. To investigate whether MMEJ inhibition disrupts ecDNA tethering, we performed live-cell imaging in COLO320DM cells with TetO-tagged ecDNA and co-expressing TetR–eGFP and H2B–mCherry^3,29. Unlike BRD4 inhibition, which causes rapid and severe ecDNA untethering^3,29, Polθ inhibition had no detectable effect at 48 h and resulted in only a modest increase in ecDNA detachment after 72 h (Extended Data Fig. 3a,b). This suggests that any effect of MMEJ loss on chromosomal tethering is indirect, and likely to be a secondary consequence of accumulated DNA damage and ecDNA instability rather than direct regulation of tethering.

We next tested whether MMEJ prevents ecDNA elimination via micronuclear sequestration, a process previously linked to DNA damage in ecDNA^12,13,30. Interphase FISH revealed that Polθ inhibition substantially increased MYC-positive micronuclei in COLO320DM cells but not in COLO320HSR cells (Fig. 1h,i). Inhibition of DNA-PKcs did not induce micronucleation, consistent with NHEJ being dispensable for ecDNA maintenance (Fig. 1i). By contrast, co-treatment of cells with hydroxyurea, which induces replication stress, and Polθ inhibitors further increased MYC-positive micronuclei (Extended Data Fig. 3c,d). Notably, whereas micronuclei in Polθ-inhibited COLO320DM cells were enriched for MYC-positive signal, they showed no enrichment for centromeric or telomeric sequences (Fig. 1h–j and Extended Data Fig. 3e,f). These results suggest that ecDNA accumulation in micronuclei is not driven by piggybacking on missegregated chromosomes due to Polθ inhibition but rather reflects selective sequestration of ecDNA.

Having established that Polθ inhibition reduces ecDNA copy number and leads to the accumulation of DNA damage factors at extrachromosomal elements, we next investigated whether these effects translate into a fitness cost. Despite reducing ecDNA levels and increasing the accumulation of DNA damage, Polθ inhibition alone did not impair COLO320DM proliferation (Fig. 2a), alter cell cycle (Fig. 2b and Extended Data Fig. 4a–d) or confer a competitive disadvantage in co-culture assays (Fig. 2c and Extended Data Fig. 4e,f). These observations suggest that, in the absence of strong selective pressure for high levels of oncogenes or drug-resistance genes encoded on ecDNA, cells can tolerate a reduction in ecDNA abundance without apparent fitness cost. Moreover, the lack of growth inhibition indicates that, although Polθ inhibition elicits detectable DNA damage, this burden does not reach the threshold required to compromise cellular proliferation.

We therefore explored whether increasing the overall DNA damage load might reveal a proliferative defect in cells carrying ecDNA. It has been established that treatment with Chk1 inhibitors leads to increased DNA damage in ecDNA-containing tumour cells and compromised growth¹². We therefore treated COLO320DM and COLO320HSR cells with Chk1 inhibitors alone or in combination with Polθi-Hel. Whereas Chk1 inhibition alone had a small but notable effect on cellular growth, co-treatment with Polθi-Hel produced a marked increase in DNA damage and a cellular growth defect in COLO320DM but not COLO320HSR cells (Extended Data Fig. 5a–f).

We next tested whether Polθ inhibition could selectively compromise cells whose survival depends on ecDNA-driven gene amplification. To address this, we first utilized a methotrexate (MTX) selection system in which engineered HeLa cells rely on high-copy DHFR amplification on ecDNA (HeLa-DM) or HSRs (HeLa-HSR) to withstand MTX-induced cytotoxicity⁷ (Fig. 2d). Cells were pre-treated with Polθi-Hel for one week, followed by combined MTX and Polθi treatment for an additional 2 weeks. Metaphase FISH confirmed a marked reduction in DHFR ecDNA in HeLa-DM cells (Fig. 2e and Extended Data Fig. 5g). Notably, whereas HeLa-HSR cells continued to proliferate, HeLa-DM cells exhibited a significant growth defect upon combination treatment (Fig. 2f,g), suggesting that MMEJ inhibition selectively compromises cells that depend on ecDNA-driven oncogene amplification for survival under selective pressure. To corroborate these findings in a clinically relevant context of oncogene amplification, we examined SNU-16 gastric cancer cells, which exhibit FGFR2 amplification on ecDNA. We treated cells with a sub-saturating dose of the FGFR inhibitor (6.25 nM) that had no discernible effect on proliferation as a single agent. Notably, combining this low dose with Polθ inhibition was sufficient to suppress cell growth (Fig. 2h and Extended Data Fig. 5h). Together, these results demonstrate that MMEJ inhibition does not broadly impair proliferation but instead selectively exposes vulnerabilities in cells carrying ecDNA, both when DNA damage is exacerbated and when survival depends on the dosage of ecDNA-encoded oncogenes or drug resistance genes.

To understand the mechanistic basis for the observed dependency of ecDNA maintenance on MMEJ, we considered three distinct but potentially overlapping scenarios: (1) a global rewiring in DNA repair pathway preference favouring MMEJ over HR or NHEJ; (2) locus-specific alterations in repair pathway usage at ecDNA amplifications compared with their chromosomal counterparts; and (3) accumulation of endogenous DNA lesions uniquely within ecDNA that necessitate MMEJ for resolution (Fig. 3a).

To assess whether global DNA repair pathway utilization is altered in ecDNA-positive cells, we applied AAVS1-Repair-seq. This sequencing-based assay quantitatively profiles the relative contributions of HR, NHEJ and MMEJ following CRISPR–Cas9-mediated cleavage at the AAVS1 safe harbour locus (Fig. 3a). In this system, HR events are identified by incorporating a co-transfected donor oligonucleotide bearing a traceable sequence marker. NHEJ is characterized by small insertion–deletion mutations (indels) or blunt-end repair events. Finally, MMEJ is distinguished by specific repair events that exhibit microhomology at junctions, are characterized by deletions or insertions, and depend on Polθ^22,23. Comparing the spectrum of repair events between COLO320DM and COLO320HSR cells revealed no significant differences in the relative contributions of each repair pathway upon DNA break induction at the AAVS1 locus (Fig. 3b,c). More specifically, quantification of MMEJ-dependent repair junctions (Extended Data Fig. 6a–c) revealed comparable frequencies in ecDNA-positive and ecDNA-negative cell lines, suggesting that the accumulation of ecDNA does not broadly alter the landscape of DSB repair pathway usage in cells (Fig. 3b,c). Similar results were obtained when a Repair-seq assay was used to compare repair events at the MYC locus in COLO320DM and COLO320HSR cells (Fig. 3b,c and Extended Data Fig. 6d). Together, these data argue against both a global shift in repair preference and a locus-specific change in pathway choice as the primary explanation for heightened MMEJ dependence in ecDNA-positive cells.

We therefore considered the third possibility: that ecDNA itself is an endogenous source of DNA lesions dependent on MMEJ for resolution. To test this, we performed END-seq, a high-resolution technique for genome-wide mapping of DSBs³¹ in COLO320DM and COLO320HSR cells treated with Polθ inhibitors. Both cell types showed similar read coverage across the regions within the ecDNA locus, including on chromosome 8, a 1.58-Mb region encompassing the amplified MYC locus, chromosome 6 (126 kb) and chromosome 13 (168 kb) (Fig. 3d and Extended Data Fig. 6e). Furthermore, CRISPR–Cas9-induced breaks at the MYC locus were detected in both lines, validating assay sensitivity (Extended Data Fig. 6f). Of note, cells carrying ecDNA exhibited more breaks at the MYC locus following CRISPR-mediated cleavage, possibly reflecting increased chromatin accessibility on the circular DNA, which has been shown to lack canonical chromatin features and higher-order organization^30,32.

Analysis of END-seq reads revealed recurrent endogenous DSBs at multiple discrete loci within the MYC ecDNA in COLO320DM cells, whereas the corresponding regions in COLO320HSR cells were less prone to breakage (Fig. 3e and Extended Data Fig. 6g). These breaks were not randomly distributed; motif analysis revealed a pronounced enrichment for TA dinucleotide repeats at ecDNA break sites, a pattern that was absent at breaks mapping to linear chromosomal DNA (Fig. 3e–g and Extended Data Fig. 6h,i). Specifically, 11 of the 20 END-seq peaks within the chromosome 8 amplicon of COLO320DM cells overlapped with TA repeat tracts spanning at least 6 repeat units, a proportion well above that seen at other genomic loci or in COLO320HSR cells (Fig. 3e–h and Extended Data Fig. 6j). These TA-rich break sites were observed across independent Polθ inhibitor conditions and single guide RNA (sgRNA) treatments. They were consistently detected in COLO320DM but not COLO320HSR cells, indicating that they reflect an intrinsic susceptibility to breakage on MYC-containing ecDNA rather than a treatment artefact (Fig. 3e,h and Extended Data Figs. 6j–l and 7). Furthermore, stratification of END-seq signal by TA repeat length revealed a positive correlation between repeat length and DSB burden in COLO320DM cells (Extended Data Fig. 6l). Together, these findings establish TA repeats as hotspots of intrinsic DNA fragility on ecDNA.

TA repeats are intrinsically predisposed to adopt non-B DNA conformations, such as cruciform, making them well-established hotspots of genome instability⁵. TA repeats can function as DNA unwinding elements to promote double helix melting during replication⁶. Studies in Escherichia coli and yeast have shown that transcription-induced negative supercoiling is sufficient to drive cruciform extrusion at TA-rich sequences^33,34,35, supporting the notion that the transcriptional hyperactivity of ecDNA may potentiate the structural vulnerability of its TA-rich loci. Consistent with this, END-seq peaks at TA loci 1–3 overlap with the highly transcribed MYC–PVT1 region of the chromosome 8 ecDNA amplicon (Fig. 3e), possibly linking break formation to active transcription. Mammalian cells have evolved opposing enzymatic activities to manage secondary structures forming at TA-rich regions: translocases such as FANCM and SMARCAL1, and the helicase WRN, act to suppress cruciform accumulation, while the structure-specific endonucleases ERCC1–ERCC4 and MUS81 resolve them through cleavage, generating DSBs in the process^5,36,37,38. The balance between these activities, therefore, determines whether cruciform structures are safely resolved or converted into DNA breaks that must be repaired.

To dissect the relative contributions of these factors, we targeted each genetically in COLO320DM and COLO320HSR cells. Loss of FANCM or SMARCAL1, but not WRN, elevated DNA damage, an effect further compounded by Polθ inhibition, with a similar trend observed upon co-depletion of RHINO and FANCM. Conversely, depletion of ERCC1 or ERCC4 rescued the DNA damage induced by Polθ inhibition (Fig. 4a–c and Extended Data Fig. 8a–d). Our data also showed that ecDNA-positive micronucleation induced by Polθ inhibition was further increased by FANCM depletion and suppressed by ERCC1 or ERCC4 depletion (Fig. 4d,e). Consistent with damage occurring specifically on ecDNA, FANCM depletion increased colocalization of 53BP1 with MYC (Extended Data Fig. 8e) and reduced ecDNA expression (Extended Data Fig. 8f), supporting a model in which nucleolytic cleavage at TA cruciform drives ecDNA destabilization. Critically, transcriptional inhibition with triptolide or CDK9i rescued the DNA damage caused by FANCM loss (Fig. 4f), linking the transcriptional hyperactivity of ecDNA to its structural vulnerability at TA repeat fragile sites.

To confirm that these perturbations converge specifically on TA repeat-containing regions at ecDNA, we performed END-seq in COLO320DM and COLO320HSR cells depleted of FANCM or treated with Polθ inhibitors and normalized the TA repeat-associated END-seq signal to total read coverage across the amplified MYC locus. Both FANCM depletion and Polθ inhibition independently increased DSBs specifically at TA repeat-containing loci within ecDNA, and their combination produced an additive increase in break accumulation at these sites. By contrast, no such enrichment was observed in COLO320HSR cells (Fig. 4g). Together, our findings posit a two-tiered mechanism that protects the highly transcribed ecDNA from the inherent fragility of TA repeat-containing sequences: FANCM resolves transcription-induced DNA secondary structures, likely to be cruciforms, before they can be cleaved by ERCC1–ERCC4, while Polθ-mediated MMEJ repairs the DSBs that escape this first line of defence (Fig. 4h). The functional importance of both tiers is underscored by the observation that, whereas FANCM loss and Polθ inhibition were individually tolerated, the combination selectively impaired proliferation in COLO320DM, but not COLO320HSR, cells (Extended Data Fig. 9).

Computational modelling and reconstruction of ecDNA structures from long-read sequencing have revealed that ecDNAs are highly complex and dynamically evolving, often comprising heterogeneous and co-existing species with extensive rearrangements and overlapping genomic footprints^39,40. To functionally assess whether TA repeat-associated fragility contributed to structural rearrangements in ecDNA, as has been suggested in linear DNA¹⁸, we performed single-cell whole-genome sequencing using direct library preparation (DLP+) in COLO320DM cells following long-term depletion with CRISPR-mediated knockout (sgFANCM) or treatment with Polθi-Hel⁴⁰. Structural variants on ecDNA, including deletions, duplications and inversions, were called from single-cell DLP+ sequencing data, and sample-specific structural variants were compared across parental (T = 0), control (DMSO; T = 2 months), sgFANCM (T = 1 month) and Polθi-Hel (T = 2 months) conditions (Fig. 5a). Both sgFANCM and Polθi-Hel treatments shifted the deletion length distribution toward larger deletions in ecDNA relative to parental and control cells. At the same time, sgFANCM additionally drove the emergence of small duplications (less than 5 kb in length) (Fig. 5b and Extended Data Fig. 10a). To determine whether these structural variants arose at sites of recurrent ecDNA breakage, we examined their distribution relative to four regions with the highest END-seq signal on the chromosome 8 ecDNA amplicon, in which the strongest peaks map to TA repeats (Extended Data Fig. 10b). In both sgFANCM and Polθi-Hel-treated cells, large deletions showed significant overlap with these four END-seq-rich regions (Fig. 5c,d and Extended Data Fig. 10c). Similarly, small duplications observed in sgFANCM-treated cells were enriched within these fragile regions (Fig. 5e,f). The near absence of small duplications following Polθ inhibition alone suggests that when FANCM surveillance fails, MMEJ resolves the resulting breaks through a repair mode that inherently favours the templated insertion of short duplications.

To investigate whether these observations extend to human tumours more broadly, we analysed the breakpoint junctions of focal amplifications from publicly available ecDNA reconstructions in the Amplicon Repository, which integrates AmpliconArchitect and AmpliconClassifier data across diverse cancer genome datasets^41,42. Focal amplification breakpoint junctions from both ecDNA and non-ecDNA amplicons were annotated for overlap with TA repeats. This analysis uncovered a statistically significant enrichment of TA repeats at ecDNA amplification breakpoints relative to linear focal amplification breakpoints (Fig. 5g–i and Extended Data Fig. 10d–f), supporting the model whereby TA-rich sequences are more prone to structural fragility in the context of circular DNA.

ecDNA has emerged as a powerful engine of tumour evolution⁴³, yet the mechanisms that enable its persistence despite its unique structural vulnerabilities remain incompletely understood. Here, we demonstrate that ecDNA maintenance reflects an active and specialized dependence on MMEJ. Although canonical repair by NHEJ and HR is essential for ecDNA formation^7,10 and for resolving Cas9-induced breaks on ecDNA, our findings show that sustaining ecDNA copy number depends critically on MMEJ activity. This reframes ecDNA as a distinct genomic entity whose structural and transcriptional challenges diverge fundamentally from those of linear chromosomal DNA. Our data support a model in which rampant transcriptional activity in the context of ecDNA generates unwinding at TA-rich loci and the formation of secondary DNA structures, such as cruciforms, that serve as recurrent sites of breakage and rely on MMEJ for repair (Fig. 4h). While transcription-driven cruciform extrusion at TA repeats is well-documented in plasmids, the organization of ecDNA into topological domains and the accumulation of R-loops at its highly transcribed loci can trap negative supercoiling locally⁴⁴, even within these megabase-scale molecules, providing sufficient torsional stress to drive cruciform extrusion and subsequent cleavage at TA-rich sequences.

HR and NHEJ do not resolve breaks at TA-rich regions in ecDNA, leaving MMEJ as the sole competent pathway. The cleavage of a cruciform might generate end structures that are refractory to Ku loading on DNA ends⁴⁵, or prevent the long-range resection necessary for HR, while the TA repeat itself provides the microhomology that licenses MMEJ. An alternative model, based on the central role of RHINO in ecDNA maintenance and the suppression of HR and NHEJ during mitosis^27,46, is that these breaks might remain unresolved until mitosis. We speculate that TA-rich breaks could persist into the M phase, when MMEJ becomes the only available means of restoring ecDNA integrity, although we cannot rule out a contribution of MMEJ during S–G₂.

The central role of MMEJ in ecDNA maintenance extends its emerging function in stabilizing and propagating extrachromosomal genomes more broadly, including viral episomes and transposable elements¹⁵. Although the specific mechanisms of propagation may differ across these distinct DNA species, a consistent principle appears to hold: circular or episomal DNA elements rely on MMEJ to persist within the host genome. Our findings further implicate DNA secondary structure as a central driver of ecDNA fragility. While TA repeats are recognized as sources of translocations on linear chromosomes¹⁸, their instability is profoundly amplified on ecDNA, where the combination of exceptionally high transcription and closed-circular topology creates conditions uniquely permissive for cruciform extrusion, converting otherwise stable loci into recurrent hotspots for DNA breakage and rearrangement. We did not observe increased breakage at TA repeats on HSR amplifications, which have also been observed to have high levels of transcription, suggesting that the circular topology of ecDNA amplifications contributes to the TA repeat-associated fragility. A recent study showed that ecDNA replication is associated with ATM-mediated DNA damage responses involving TOP1 and TOP2B, with MMEJ contributing to subsequent repair²⁴. Together with our findings, these observations support a model in which ecDNA is subject to multiple, potentially complementary sources of DNA damage that converge on a shared reliance on MMEJ for resolution. Here we identify TA repeat-associated secondary structures as an intrinsic, sequence-encoded source of fragility and delineate the mechanism by which the resulting lesions are repaired.

Specifically, we uncover a two-tiered protection mechanism that buffers this instability. FANCM, known for its ability to remodel branched DNA structures⁴⁷, acts upstream to suppress or resolve transcription-induced cruciform structures at TA-rich sequences, thereby preventing their cleavage by the ERCC1–ERCC4 endonuclease complex and limiting DNA damage (Fig. 4). Breaks that escape this surveillance are subsequently channelled into Polθ-mediated MMEJ for repair. Consistent with this model, analysis of single-cell whole-genome sequencing data alongside large-scale whole-genome sequencing datasets from human tumours reveals that rearrangement breakpoints on ecDNA are significantly enriched at TA dinucleotide repeats (Fig. 5). This enrichment suggests that these sequence elements not only represent intrinsic structural vulnerabilities but also serve as recurrent substrates for the ongoing rearrangements that drive ecDNA remodelling and diversification in vivo.

From an evolutionary perspective, the TA fragility that we describe may not simply be a liability; it may instead facilitate ecDNA plasticity. Sequence analyses of ecDNA suggest an accumulation of DNA repair signatures after ecDNA formation, particularly SBS3, which has been associated with HR deficiency⁴⁸, suggesting that ecDNA exists in a state of ongoing repair activity. The TA repeat-associated breaks that necessitate MMEJ repair may also provide hotspots for ongoing structural rearrangement, enabling rapid remodelling of ecDNA architecture. In this light, MMEJ may have a dual role: preserving ecDNA integrity in the short term while enabling structural diversification over longer evolutionary timescales. Finally, our observations carry direct therapeutic implications. Polθ inhibition reduces ecDNA levels, but its effect on cell fitness is selectively evident when cells depend on ecDNA-encoded oncogenes, under targeted therapy, or in the context of acquired drug resistance. This context dependence suggests a rational combination strategy: oncogene-directed therapies that heighten reliance on ecDNA for survival may synergize with MMEJ inhibition to destabilize these elements and overcome resistance.

Cell culture

COLO320DM, COLO320HSR and SNU-16 cells (ATCC CCL-220, CCL-220.1 and CRL-5974, respectively) were maintained in RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS), 100 U ml⁻¹ penicillin-streptomycin, 1% non-essential amino acids, and 2 mM l-glutamine. HEK293T cells (ATCC CRL-3216), PC3-DM, and PC3-HSR (gifts from the P. Mischel laboratory) were cultured in Dulbecco’s Modified Eagle Medium (DMEM) supplemented with 10% bovine calf serum, 100 U ml⁻¹ penicillin-streptomycin, 1% non-essential amino acids, and 2 mM l-glutamine. NCI-H716 cells (ATCC CCL-251) were cultured in RPMI-1640 with the same supplements. All cells were incubated at 37 °C in a humidified atmosphere containing 5% CO₂. All cells were tested for Mycoplasma by PCR.

HeLa cells were cultured as previously described⁷. In brief, cells were grown in DMEM supplemented with 10% FBS (Omega), 100 U ml⁻¹ penicillin, 100 U ml⁻¹ streptomycin, and 2 mM l-glutamine. For MTX experiments, dialysed FBS was used. Specific HeLa-derived clones were used at the following MTX concentrations: PD29424h (H, DM+, 60 nM MTX) and PD29426a (HSR+, 40 nM MTX).

Adult neuronal stem cells (aNSCs) from Myc^ec/+; Trp53^fl/fl (ecMyc) or control Myc^+/+; Trp53^fl/fl (control) mice were isolated and cultured in adherent condition on laminin-coated (Sigma-Aldrich, L2020) dishes in NeuroCult Stem Cell Basal Media with NeuroCult Proliferation Supplement (Mouse & Rat) (Stem Cells Technologies, 05702), 20 ng ml⁻¹ EGF (Stem Cells Technologies, 78006), 10 ng ml⁻¹ bFGF (Stem Cells Technologies, 78003), and 2 μg ml⁻¹ heparin (Stem Cell Technologies), ecMyc and control aNSCs were treated by Adeno-Cre (from ViraQuest (VQ-Ad-CMV-Cre; 1 × 1012 particles per ml; 091317)) to induce engineered ecDNA formation in ecMyc cells. Cells were maintained in culture for 3 weeks to accumulate Myc-containing ecDNAs. All cells were negative for mycoplasma contamination. Cells were maintained in a humidified 5% CO₂ atmosphere at 37 °C.

For proliferation assays, cells were seeded at clonal density in 12-well plates. The following day, cells were treated with a drug (Polθ inhibitors, MTX, CHK1 inhibitor or FGFR inhibitor). For HeLa experiments, MTX concentrations were applied as described above. In both the HeLa and SNU-16 experiments, cells were pre-treated with Polθi or vehicle for 1 week prior to replating for growth assessment. Medium with fresh drug was replaced every 72 h. Cell proliferation was monitored using the IncuCyte Live-Cell Imaging System (Essen BioScience/Sartorius), with confluency measurements collected every 24 h.

For proliferation assays involving FANCM loss and Polθ inhibition, FANCM-knockout cells were generated using the triple-guide CRISPR–Cas9 strategy described below, with AAVS1-targeted cells serving as controls. Two days after knockout, cells were plated at 2,000 cells per well in 48-well plates in three technical replicates for each condition, and Polθi-Pol treatment was initiated. Cell proliferation was monitored using the IncuCyte Live-Cell Imaging System with confluence measurements acquired every 12 h.

Drug treatments

The following compounds were used in this study: Polθ inhibitors: RP-2119 (Polθi-Hel; Repare Therapeutics; 100 nM), RP-6685 (Polθi-Pol; Repare Therapeutics; 10 µM); DNA-PKcs inhibitors: NU7441 (Selleck Chemicals, S2638; 1 µM), NU7026 (Selleck Chemicals, S2893; 10 μM); CHK1 inhibitor (CCT245737; Selleck Chemicals, S8253; 750 nM); methotrexate (MTX; Sigma-Aldrich, 454126; concentrations varied by cell line and clone as noted above); hydroxyurea (Sigma-Aldrich, H8627; 100 μM); PARP inhibitor (Olaparib; Selleck Chemicals, S1060; 1 μM); thymidine (Sigma-Aldrich, T1895; 2 mM); nocodazole (Sigma-Aldrich, M1404; 100 ng ml⁻¹); FGFR inhibitor (Infigratinib/BJ398; Selleck Chemicals, S2183; 6.25 nM); triptolide (Millipore, 645900; 1 µM); and CDK9 inhibitor (AZD4573; Selleck Chemicals, S8719; 200 nM).

For FISH experiments and proliferation assays, the media were replaced every 72 h with fresh medium containing the indicated drugs. Drug concentrations were maintained consistently throughout the assay.

CRISPR–Cas9 ribonucleoprotein complex assembly and nucleofection

CRISPR RNA (crRNA) and trans-activating CRISPR RNA (tracrRNA) (Integrated DNA Technologies, IDT) were each resuspended in IDTE buffer to a final concentration of 100 µM. Equimolar amounts were mixed to yield a 50 µM crRNA:tracrRNA (sgRNA) duplex, which was annealed by heating to 95 °C for 5 min, followed by cooling to room temperature. For ribonuclear protein (RNP) complex formation, purified Cas9 protein (Berkeley Facility; https://macrolab.qb3.berkeley.edu/) and sgRNA were mixed at a 1:1.2 molar ratio and incubated at room temperature for 20 min before use. In experiments using three sgRNAs simultaneously to knock out a gene, sgRNAs were designed to be between 35 and 300 bp apart. Each sgRNA was resuspended in IDTE buffer and pooled to a final concentration of 40 µM per guide. RNP complexes were assembled by mixing 104 pmol of Cas9 with 120 pmol of the combined sgRNA mix (1:1.15 molar ratio) and incubating at room temperature for 20 min. Target cells were prepared using the SF Cell Line 4D-Nucleofector X Kit (Lonza) according to the manufacturer’s instructions. Nucleofection was performed using the Lonza 4D-Nucleofector System with program CM158. Knockouts were validated by PCR amplifying the region around the sgRNA target site, followed by Sanger sequencing and TIDE⁴⁹ or ICE CRISPR Analysis (EditCo Bio., v3.0). For simultaneous RHINO and FANCM knockouts, a triple-guide CRISPR–Cas9 strategy was used for each gene, and RNP complexes targeting both genes were combined at a 1:1 ratio before nucleofection. Cells were resuspended in SF solution, mixed with the combined RNP complexes, and nucleofected as described above. For the corresponding single-knockout controls, IL25-targeting RNPs were included to match the total guide load in the double-knockout condition. Sequences for sgRNA in Supplementary Table 1. PCR primers for TIDE/EditCo ICE and CRISPR analysis in Supplementary Table 2.

Lentiviral production and transduction

Lentivirus was produced in HEK293T cells using a standard four-plasmid transfection system. For each transfection, 5 µg of RRE, 3 µg of VSV-G, 2.5 µg of REV, and 20.5 µg of a BFP-expressing BRCA2 shRNA transfer plasmid were combined with 62 µg ml⁻¹ polyethylenimine (PEI) and 150 mM NaCl. The DNA-PEI complexes were incubated at room temperature for 15 min before being added to 10-cm dishes of HEK293T cells at ~70–80% confluence. Cells were incubated overnight at 37 °C in a humidified 5% CO₂ incubator. The following day, the culture medium was replaced, and after a 6–8 h recovery period, the first viral supernatant was collected. This process was repeated twice at 24-h intervals to collect the second and third viral supernatants. Fresh medium was added after each collection.

The BRCA2 knockdown construct targeted exon 16 of the BRCA2 coding sequence and was designed using the Broad Institute’s RNAi design tool. The hairpin sequence (5′-GCGTTTCTAAACATTGCATAA-3′) was cloned into the pLKO.1-puro vector (Addgene #8453). Target cells were transduced with the lentiviral supernatant and selected with puromycin to generate stable knockdown lines. Functional validation of BRCA2 depletion was assessed by measuring cell proliferation in response to 1 µM olaparib. Cell growth was monitored using the IncuCyte Live-Cell Imaging System (Essen BioScience/Sartorius), with confluency measurements recorded every 24 h.

Generation of RHNO1 ^−/− cells

COLO320DM cells were stably infected with pLentiCRISPRV2 Cas9 – blast (Addgene #98293) and selected with blasticidin (5 μg ml⁻¹). Blasticidin-resistant cells were then stably infected with pLenti-GFP-2A-Puro-gRHINO1_A (Addgene #211613) and selected with puromycin at 3 μg ml⁻¹. RHNO1 knockout was confirmed with TIDE analysis (forward: TTTTGCTTGGTGGTTGTAGG, reverse: TATCTGGCATTCTCACCCAG).

Generation of Fucci4 cells

Lentivirus was prepared containing one of each of the following plasmids, gifts from M. Lin, to generate Fucci4 cells⁵⁰: pLL3.7m-Clover-Geminin(1-110)-IRES-mKO2-Cdt(30-120) (Addgene #83841), pLL3.7m-mTurquoise2-SLBP(18-126)-IRES-H1-mMaroon1 (Addgene #83842), mKO2-SLBP(18-126) (Addgene #83914) and Clover-Geminin(1-110) (Addgene #83915). COLO320DM and COLO320HSR cells were transduced with lentiviral supernatant containing either mTurquoise2-SLBP/H1-mMaroon1, mKO2-SLBP, or Clover-Geminin. Transgene-positive cells were isolated by FACS and used as colour controls in experiments. A split of mTurquoise2-SLBP/H1-mMaroon1-positive cells was transduced with lentiviral supernatant containing Clover-Geminin/mKO2-Cdt, and transgene-positive, fully labelled Fucci4 cells were isolated by FACS.

Cell cycle analysis of Fucci4 cells by flow cytometry

COLO320DM or COLO320HSR Fucci4 cells were plated and treated for 72 h with either DMSO (0.1–0.3% as required per control), Polθi-Pol, Polθi-Hel, DNA-PKcs inhibitors, Chk1i, Chk1i plus Polθi-Pol, or Chk1i plus Polθi-Hel. Cell cycle controls included a double thymidine block using 2 mM thymidine or an overnight treatment with 100 ng ml⁻¹ nocodazole. For the experiment, CRISPR–Cas9-mediated sgRNA targeting of either AAVS1 or FANCM was performed in COLO320DM Fucci4 cells. Knockout cells were plated and treated for 72 h with either 0.1% DMSO, Polθi-Pol, or Polθi-Hel. Cells were trypsinized, collected in RPMI medium, washed with DPBS, pelleted, resuspended in DPBS with 3% FBS, and filtered using a 35-μm mesh into polystyrene tubes (Falcon; 38030) for flow cytometry. Cells were then analysed on an Aurora full-spectrum analyser (Cytek Biosciences), with 100,000 to 150,000 events collected per condition, and quantification was performed using FlowJo (v10.10.1). Single or dual-colour controls (mTurquoise2–SLBP/H1–mMaroon1, mKO2–SLBP or Clover–Geminin) and unstained cells were used to determine the gating strategy for G₁-, S–G₂- and M-phase cells (Supplementary Fig. 1).

Generation of H2B–GFP and H2B–RFP cells

Lentivirus was prepared containing one of each of the following plasmids, gifts from E. Fuchs: H2B–GFP (Addgene #25999) and H2B–RFP (Addgene #26001). COLO320DM and COLO320HSR cells were transduced with H2B–GFP and H2B–RFP separately, and transgene-positive cells were isolated by FACS. Labelling was confirmed by live-cell microscopy in 35 mm no. 1.5 glass-bottomed dishes (Mattek; P35G-1.5) using a Nikon Ti2 inverted microscope equipped with a TokaiHIT STX stage-top incubator and objective heating collar (37 °C, 5% CO₂), a Yokogawa CSU-W1 spinning disc confocal unit, and dual Hamamatsu C14440-20UP sCMOS cameras (Dual Fusion configuration). The system was controlled using NIS-Elements AR software version 6.10.02. Cells were imaged using an Apo TIRF 60× 1.49 NA oil immersion objective (Nikon). Four fluorescence channels were acquired sequentially using laser excitation from an LUN-F multi-laser unit and 300 ms exposures. Excitation wavelengths and laser powers were as follows: 405 nm (100 mW) at 30% power, 488 nm (100 mW) at 10% power, and 561 nm (100 mW) at 30% power. A quad-band dichroic mirror (Di01-T405/488/568/647) was used with the following emission filters: 455/50 nm (DAPI), 526/36 nm (GFP), and 605/52 nm (dsRed–mCherry), split across two detection paths using a 561 nm long-pass dichroic (DM A561LP). Images were saved in Nikon ND2 format and subsequently processed using Fiji (ImageJ version 2.16.0)⁵¹.

Fluorescence-based competition assays

CRISPR–Cas9-mediated targeting of IL25 or POLQ (see ‘CRISPR–Cas9 ribonucleoprotein complex assembly and nucleofection’) was performed in COLO320DM and COLO320HSR cells labelled with either H2B–GFP or H2B–RFP. Each cell type was mixed 1:1 in the following pairings and plated in triplicate at 2,000 cells per well in a 48-well tissue culture plate (Corning Costar; 3548):

1:1 mixtures were imaged on an IncuCyte Live-Cell Imaging System (Essen BioScience/Sartorius) for up to 2 weeks. Knockouts were confirmed during the experiments using DNA extraction with the DirectPCR Lysis Reagent (Viagen; 301-C), PCR amplification of the knockout regions, Sanger sequencing and EditCo’s ICE Analysis tool (v3). Cell confluency, the number of GFP-positive cells, and the number of RFP-positive cells were quantified using the IncuCyte Analysis software. Technical replicates were averaged, and the number of RFP- or GFP-positive cells per condition was normalized first to timepoint 0 (day 1), and then to the H2B–RFP (sgIL25) or H2B–GFP (sgIL25) control condition.

Live-cell imaging of mitotic timing

COLO320DM or COLO320HSR cells expressing H2B–GFP were plated on 35 mm no. 1.5 glass-bottomed dishes (Mattek; P35G-1.5) and treated for 48 h with either 0.1% DMSO or Polθi-Hel. At 48 h, cell medium was changed to RPMI without phenol red (supplemented with respective drugs or treatments). Time-lapse images were acquired on a Nikon Ti2 inverted microscope equipped with a TokaiHIT STX stage-top incubator and objective heating collar (37 °C, 5% CO₂), a Lumencor Sola II light engine, and a Hamamatsu C14440-20UP sCMOS camera (SN:000740). The system was controlled using NIS-Elements AR software version 5.21.03. Cells were imaged using a Plan Apo λ 100× 1.45 NA oil immersion objective (Nikon). Two fluorescence channels were acquired at each time point: GFP (ET-GFP; Nikon; 96366; ET470/40X; ET525/50 M; 200 ms exposure) and DIC. Time-lapse imaging was performed with a 3-min interval between frames, and the Perfect Focus System (PFS) was engaged. Images were saved in Nikon ND2 format and subsequently processed in Fiji (ImageJ version 2.16.0)⁵¹ to manually screen for mitotic events.

Polθ ATPase enzymatic assay

In a 384-well assay plate (Corning 3572), recombinant full-length human POLQ (3 nM) was combined with a 10-point concentration range of Polθi-Hel (highest concentration 0.1 mM, threefold serial dilutions) in a buffer containing 50 mM Tris-Cl pH 7.5, 10% glycerol, 5 mM DTT, 10 mM MgCl2, and 0.1 mg ml⁻¹ BSA. The plate was incubated at room temperature for 15 min, after which the reaction was initiated by adding a DNA substrate (20 nM) and ATP (100 µM). Plates were incubated for 90 min at room temperature, and ATPase activity was subsequently detected using an ADP-Glo assay kit (Promega) according to the manufacturer’s protocol. In brief, 15 μl of ADP-Glo reagent was added to each well, followed by a 1 h incubation at room temperature. 30 μl of Detection reagent was then added per well, the plates were incubated for 45 min at room temperature, and luminescence was read on an Envision plate reader (Revvity). The forked DNA substrate was prepared as follows: oligos 1–3 were mixed (1:1:1) in a buffer containing 10 mM Tris-HCl, pH 7.5, 50 mM NaCl, and 1 mM EDTA, heated to 95 °C for 5 min, and cooled to room temperature.

Oligo 1: 5′-GCACTGGCCGTCGTTTTACGGTCGTGACTGGGAAAACCCTGGCG-3′; oligo 2: 5′-TTTTTTTTTTTTTTTTTTTTTTCCAAGTAAAACGACGGCCAGTGC-3′; oligo 3: 5′-TTGGAAAAAAAAAAAAAAAAAAAAAA-3′.

Expression and purification of recombinant full-length Polθ

Four 1-l cultures of 293-6E cells were grown in F17 medium supplemented with 0.1% Pluronic F-68, 4 mM GlutaMAX, and 25 μg ml⁻¹ G418 in 2-l shake flasks. Cultures were transfected with 1 mg l⁻¹ of the phCMV1-2m6h-POLQ plasmid (lab of S. Doublié) using PolyPlus PEIPro at a 1:1 (w/v) DNA:PEI ratio. Transfected cultures were maintained at 37 °C with 5% CO₂ and shaken at 135 rpm. Cells were collected 4 days post-transfection by centrifugation at 1,000g for 5 min at 4 °C. For lysis, the resulting cell pellet was resuspended in 250 ml of lysis buffer (20 mM Tris-HCl, 250 mM NaCl, 250 mM KCl, 0.01% NP-40, 10% glycerol, pH 8.0) supplemented with 5 mM β-mercaptoethanol and EDTA-free protease inhibitors (Roche). Cells were lysed by freeze–thaw cycles followed by processing through a microfluidizer. The lysate was stirred at 4 °C for 30 min, then clarified by centrifugation at 20,000g for 60 min at 4 °C.

The soluble fraction was batch-bound to 5 ml of amylose resin (New England Biolabs) in buffer A (20 mM Tris-HCl, 200 mM NaCl, 200 mM KCl, 0.01% NP-40, 10% glycerol, 5 mM β-mercaptoethanol, pH 8.0) for 1.5 h at 4 °C with gentle nutation. The resin was first collected in a 5-cm Econo-column, transferred to a 2.5-cm column, and washed with 30 column volumes of buffer A. Elution was carried out using buffer B (buffer A supplemented with 50 mM maltose). The eluate was diluted threefold with pre-heparin dilution buffer (10 mM HEPES, pH 7.5), then loaded at 2 ml min⁻¹ onto a 5 ml HiTrap Heparin Sepharose HP column (Cytiva, 17040703) pre-equilibrated in buffer D (20 mM Tris-HCl, 500 mM NaCl, 500 mM KCl, 0.01% NP-40, 5% glycerol, 5 mM β-mercaptoethanol, pH 8.0). After loading, the column was washed with 20 column volumes of buffer C (20 mM Tris-HCl, 100 mM NaCl, 0.01% NP-40, 5% glycerol, 5 mM β-mercaptoethanol, pH 8.0), and 5 ml fractions were collected. The protein was eluted using a linear 0–100% gradient of buffer D over 20 column volumes at 2 ml min⁻¹. The peak fractions (heparin pool 1) were concentrated to ~0.8 ml using a Vivaspin6 centrifugal device (10 kDa MWCO, PES membrane), then subjected to size-exclusion chromatography using a Superose 6 Increase column (Cytiva) equilibrated in destination buffer (20 mM Tris-HCl, 200 mM NaCl, 100 mM KCl, 2% glycerol, 3 μM sucrose monolaurate, 0.5 mM TCEP, pH 8.0). Fractions (0.5 ml) were collected over 1.1 column volumes and analysed by SDS–PAGE. Protein-containing fractions were pooled and concentrated using a Vivaspin6 centrifugal device (10 kDa MWCO, PES membrane) for downstream applications.

Metaphase FISH for detection of MYC, FGFR2 and DHFR ecDNA

Cells were seeded to reach approximately 50% confluence 36–48 h before collection. To enrich for mitotic cells, Colcemid (0.1 µg ml⁻¹) was added for 6 h, followed by trypsinization. Cells were pelleted by centrifugation (5 min at 1,000 rpm), gently resuspended in 5 ml of pre-warmed 0.075 M KCl, and incubated at 37 °C for 15–30 min with occasional mixing to induce swelling. After centrifugation, residual KCl was retained, and fixative (3:1 methanol:acetic acid) was added dropwise with gentle mixing. The total volume was adjusted to 10 ml with fixative, and samples were stored at 4 °C overnight or longer. For slide preparation, fixed cells were pelleted and resuspended in a minimal volume of fresh fixative. Cold, wet microscope slides were prepared by briefly soaking in water, and cells were dropped from a height onto tilted slides. Slides were immediately placed on a humidified 80 °C heating block for 1 min to promote chromosome spreading, then air-dried overnight. Slide quality was assessed by light microscopy.

To detect extrachromosomal MYC, FISH was performed using a locus-specific probe targeting the MYC genomic region (Empire Genomics: MYC FISH probe). Slides were rehydrated in PBS for 5 min, fixed in 4% formaldehyde for 2 min, and washed 3 times in PBS (5 min each), followed by three washes in 2× SSC (5 min each). Slides were then denatured in 70% formamide/2× SSC at 72 °C for 2 min, immediately dehydrated through an ethanol series (70%, 90% and 100%), and air-dried. The MYC probe, resuspended in hybridization buffer, was applied to the slide, denatured at 80 °C for 5 min, and hybridized overnight at 37 °C in a humidified chamber. The following day, excess probe was removed by sequential washes: 0.4× SSC at 72 °C for 5 min, followed by 2× SSC with 0.05% NP-40 at room temperature. Additional washes were performed in 50% formamide/2× SSC at 42 °C, 2× SSC, and 0.1× SSC. Slides were counterstained with DAPI and mounted using ProLong Gold antifade reagent. Fluorescence microscopy was used to visualize MYC-positive ecDNA, which appeared as distinct fluorescent foci located outside the main chromosomal arms in metaphase spreads.

Detection of additional ecDNA species was performed as follows. For FGFR2, a probe targeting the FGFR2 genomic region (Empire Genomics: FGFR2 FISH probe) was used; FISH was carried out as described above. For DHFR, a probe targeting the DHFR genomic region (Empire Genomics: DHFR FISH probe) was used; FISH was performed using the same protocol.

Detection of MYC-containing micronuclei by interphase FISH

To detect MYC-containing micronuclei, cells were seeded on poly-l-lysine-coated glass coverslips 96 h prior to fixation. Polθ inhibitor or vehicle was added 72 h before collection, followed by hydroxyurea or PBS 24 h before fixation. Cells were fixed in 4% paraformaldehyde (PFA) for 5 min, washed three times with PBS for 5 min each, and permeabilized with 0.5% Triton X-100 in PBS for 10 min. After two additional PBS washes (5 min each), coverslips were rinsed three times in 2× SSC for 5 min and then dehydrated through a graded ethanol series (70%, 90% and 100%). For FISH, the MYC probe (Empire Genomics) was mixed with hybridization buffer, applied to the coverslip, and denatured at 80 °C for 5 min. Hybridization was carried out overnight in a humidified chamber. The following day, coverslips were fixed in freshly prepared methanol: acetic acid (3:1) for 10 min at room temperature, air-dried, and incubated in 2× SSC. Denaturation was performed in 70% formamide/2× SSC at 72 °C for 2 min, followed by rapid dehydration in ethanol (70%, 90%, 100%). The fluorescently labelled MYC probe was then applied and hybridized overnight at 37 °C. The excess probe was removed by post-hybridization washes in 50% formamide/2× SSC at 42 °C, followed by rinses in 2× SSC and 0.1× SSC. A final wash was performed in 0.4× SSC at 72 °C for 5 min, followed by a rinse in 2× SSC containing 0.05% NP-40 at room temperature. Nuclei were counterstained with DAPI and mounted using ProLong Gold antifade reagent. Imaging was performed by fluorescence microscopy. MYC-positive micronuclei were identified as small, DAPI-stained structures that were spatially separated from the primary nucleus and displayed distinct MYC FISH signals.

DNA telomere FISH

Cells (2 × 10⁵) were plated onto poly-l-lysine-treated coverslips and treated with Polθi-Pol, Polθi-Hel, or 0.1% DMSO for 72 h or 0.5 μM Reversine for 24 h, prior to fixation. Cells on coverslips were washed 2 × 5 min with 1× DPBS, then fixed with 4% PFA for 10 min, followed by 3 × 5 min washes in 1× DPBS. Cells were then permeabilized with 0.5% Triton X-100 for 10 min and washed with DPBS, 2 × 5 min each. Coverslips were dehydrated in 70%, 95%, and 100% ethanol for 2 min each, then air-dried. Cells were incubated in 30 μl hybridization buffer (70% formamide, 1 mM Tris-Cl, and Roche blocking reagent at 1 mg ml⁻¹) and 100 nM TelC-Cy3 (PNAbio) on slides marked with a wax pen before denaturation at 80 °C for 5 min. Cells were incubated in the dark at room temperature for 2 h, followed by 4× 15 min formamide washes (70% formamide, 10 mM Tris-Cl in H₂O). Cells were washed 3× 10 min in PBST (0.1% Tween-20), with 5 μg ml⁻¹ DAPI added to the second wash. Coverslips were air-dried before being mounted with ProLong Glass antifade.

DNA pan-centromere FISH

Cells (2 × 10⁵) cells were plated onto poly-l-lysine-treated coverslips and treated with Polθi-Pol, Polθi-Hel, or 0.1% DMSO for 72 h or 0.5 μM Reversine for 24 h, prior to fixation. Cells on coverslips were washed 2 × 5 min with 1× DPBS, then fixed with 4% PFA for 10 min, followed by 3 × 5 min washes in 1× DPBS. Cells were then permeabilized with 0.5% Triton-X for 10 min and washed with DPBS, 2 × 5 min each. Coverslips were washed 3 × 5 min in 2× SSC, then dehydrated in 70%, 85% and 100% ethanol for 2 min each, then dried completely. Coverslips were incubated with XCE pan-cen (MetaSystems), diluted 1:5 in hybridization buffer (Empire Genomics), and then sealed with CytoBond (SciGene) before denaturation at 80 °C for 10–15 min. After overnight incubation in a humidified chamber at 37 °C, the CytoBond sealant was removed from the coverslips, which were washed in 0.4× SSC (heated to 72 °C) for 5 min, then washed for 30 s in 2× SSC + 0.05% Tween-20. Coverslips were rinsed briefly with ddH₂O, then air-dried, and washed 3× 10 min in DPBS (with 5 μg ml⁻¹ DAPI added to the second wash). Coverslips were air-dried before being mounted with ProLong Glass antifade.

Immunofluorescence for the detection of 53BP1 and gH2AX

Cells were seeded on poly-l-lysine-coated glass coverslips and fixed 72 h later following the indicated treatments. Cells were fixed in 4% PFA for 15 min at room temperature, then permeabilized with 0.5% Triton X-100 for 10 min. Cells were then blocked for 1 h at room temperature using a blocking buffer consisting of 1 mg ml⁻¹ BSA, 3% goat serum, 0.1% Triton X-100, and 1 mM EDTA in PBS (pH 8.0). After blocking, cells were incubated for 90 min at room temperature with primary antibodies diluted in the same blocking buffer: rabbit anti-53BP1 (Novus NB100-304, 1:1,000) and mouse anti-γH2AX (Millipore JBW301, 1:1,000). Following three PBS washes, cells were incubated with Alexa Fluor–conjugated secondary antibodies (donkey anti-rabbit 568 and donkey anti-mouse 488, 1:500, invitrogen) for 30 min at room temperature in the dark. Coverslips were then washed three times with PCB and mounted using ProLong Gold Antifade Mountant with DAPI (Thermo Fisher). DNA damage foci were visualized using fluorescence microscopy (Nikon Si) and quantified based on the number of 53BP1 and γH2AX foci per nucleus.

For immunofluorescence experiments involving transcription inhibition, transcription was acutely inhibited for 3 h with triptolide or CDK9 inhibitor either 3 days after Polθ inhibitor or vehicle pretreatment, or 5 days after CRISPR–Cas9-mediated knockout of FANCM or AAVS1. For immunofluorescence experiments assessing DNA damage following CRISPR–Cas9-mediated targeting of the indicated nucleases and helicases, cells were treated with Polθ inhibitor or vehicle beginning 2 days after nucleofection and collected on day 5. For experiments involving CRISPR–Cas9-mediated co-targeting of RHNO1 and FANCM, cells were nucleofected and collected on day 5 for immunofluorescence.

Combined 53BP1 immunofluorescence and MYC-FISH

Cells were seeded onto glass coverslips and allowed to adhere overnight. The following day, cells were fixed in 4% PFA for 15 min at room temperature, then permeabilized with 0.5% Triton X-100 for 10 min. Cells were then blocked for 1 h at room temperature using a blocking buffer consisting of 1 mg ml⁻¹ BSA, 3% goat serum, 0.1% Triton X-100, and 1 mM EDTA in PBS (pH 8.0). After blocking, cells were incubated for 90 min at room temperature with primary antibody diluted in the same blocking buffer: rabbit anti-53BP1 (Novus NB100-304, 1:1,000). Following three PBS washes, cells were incubated with Alexa Fluor–conjugated secondary antibodies (donkey anti-rabbit 488 1:500, Invitrogen) for 30 min at room temperature in the dark. Coverslips were then washed three times with PBS. Cells were then fixed with 4% PFA at room temperature for 20 min, followed by re-permeabilization using 1× PBS containing 0.7% Triton X-100 and 0.1 M HCl on ice for 10 min. Coverslips were then subjected to acid denaturation using 1.9 M HCl for 30 min at room temperature, followed by washes in PBS and 2× SSC. Coverslips were dehydrated through an ethanol series (70%, 90% and 100%) and air-dried. The MYC probe, resuspended in hybridization buffer, was applied to the coverslip, denatured at 80 °C for 5 min, and hybridized overnight at 37 °C in a humidified chamber. The following day, excess probe was removed by sequential washes: 0.4× SSC at 72 °C for 5 min, followed by 2× SSC with 0.05% NP-40 at room temperature. Additional washes were performed in 50% formamide/2× SSC at 42 °C, 2× SSC, and 0.1× SSC. Slides were counterstained with DAPI and mounted using ProLong Gold antifade reagent.

Colony formation assay

COLO320DM cells were plated at clonal density in triplicate in 6-well plates with 1 μM NU7441 or DMSO. 48 h after plating, cells were then treated with 2 Gy of ionizing radiation. Cells were then allowed to grow for an additional 10 days, after which they were fixed in 4% PFA and stained with crystal violet. Colonies were counted with the Fiji ImageJ Colony Area plugin⁵².

Immunoblot to detect Cre recombinase expression

Cells were lysed in Laemmli buffer, supplemented with protease and phosphatase inhibitors (cOmplete and EDTA-free Protease Inhibitor cocktail; Roche, COEDTAF-RO and PHOSS-RO). Proteins were separated in NuPage Bis-Tris gels (Invitrogen, NP0322BOX), transferred on a nitrocellulose membrane (0.2μm; Bio-Rad), and blocked by incubation with Intercept (TBS) Blocking buffer (LI-COR: 927-60001). The following primary antibodies were used: anti-Cre recombinase (1:1,000; Cell Signaling, 15036) and anti-α-tubulin (1:5,000; Cell Signaling, 2144). The following secondary antibodies were used: IRDye 800 anti-rabbit (LI-COR, 926-32213) and IRDye 680 anti-mouse (LI-COR, 926-68072). Images were acquired using an Odyssey Imaging System (LI-COR). Full uncropped blots with molecular weight markers are available in the supplementary data (Supplementary Fig. 2).

AAVS1 and MYC Repair-seq

Cells were nucleofected using the Lonza 4D-Nucleofector System with the SF Cell Line Kit- following the manufacturer’s protocol. Cas9 RNP complexes were assembled by incubating purified Cas9 protein (Berkeley Facility; https://macrolab.qb3.berkeley.edu/) with synthetic sgRNAs (IDT) at a 1:2 molar ratio for 10–15 min at room temperature. sgRNAs were designed to target the AAVS1 locus and exon 2 of the MYC gene. Approximately 1 × 10⁶ cells were nucleofected per condition using program CM158 and collected 72 h post-nucleofection to allow sufficient time for editing events and genomic DNA recovery. Genomic DNA was extracted using the ZymoResearch Quick-DNA Genomic Kit according to the manufacturer’s instructions. DNA concentration and quality were assessed using a Qubit fluorometer and agarose gel electrophoresis. A ~450 bp region surrounding each CRISPR cut site was amplified by PCR using locus-specific primers (primer sequences provided in Supplementary Table 2). PCRs were performed using Taq DNA Polymerase (NEB) with the following cycling conditions: initial denaturation at 95 °C for 3 min; 35 cycles of 95 °C for 15 s, 53 °C for 15 s, 68 °C for 30 s; final extension at 68 °C for 5 min. Amplicons were quantified using Qubit and pooled in equimolar ratios. The IGO facility core of MSKCC prepared Pooled PCR products for sequencing. Libraries were sequenced on the Illumina MiSeq platform using 250 bp paired-end reads (MiSeq v2 PE250 kit). All primer pairs were pre-validated to produce amplicons of similar length and efficiency, ensuring uniform sequencing coverage. Sequence quality was monitored across all samples, and depth uniformity was confirmed during post-run analysis. For experiments evaluating MMEJ, a minimum of 700,000 paired-end reads per sample was obtained. For homology-directed repair quantification, a minimum of 30,000 paired-end reads per sample was achieved. Sequencing reads were demultiplexed and aligned with a reference sequence. The CRISPResso2 toolkit was used to quantify editing outcomes⁵³, characterize insertions and deletions, and reconstruct repair events. For each sample, allele fractions for these events were calculated by counting the number of reads with respective mutational signatures identified by CRISPResso2 and dividing the count by the total reads.

Live-cell imaging of ecDNA tethering

Live-cell imaging to assess ecDNA tethering to mitotic chromosomes was performed using COLO320DM TetO-EGFP cells stably expressing H2B–mCherry and TetR–eGFP. Cells were seeded into four-well chamber slides (ibidi) 48 h prior to imaging and treated with DMSO, Polθi-Pol, or Polθi-Hel for 24 or 72 h. Imaging was conducted in a humidified, 37 °C chamber with 5% CO₂ using a Nikon SoRa Spinning Disk Confocal system equipped with a Borealis microadapter, Perfect Focus 4, motorized turret and encoded stage, and a 5-line laser launch (405 nm (100 mW), 445 nm (45 mW), 488 nm (100 mW), 561 nm (80 mW) and 640 nm (75 mW)). A PRIME 95B Monochrome Digital Camera and CFI Apo TIRF 60× 1.49 NA objective lens (W.D. 0.12 mm) were used in super-resolution mode. Laser power was set to 5% with a 600 ms exposure time. Images were acquired using NIS-Elements Advanced Research Software on a dual-Xeon imaging workstation and denoised using the default settings. Time-lapse imaging of mitotic cells was initiated at metaphase or anaphase and continued at 5-min intervals until reformation of daughter nuclei. z-stacks were processed as maximum intensity projections, and image adjustments (brightness and contrast) were performed using Fiji. To quantify cytosolic mis-segregation events, TetR⁺ ecDNA foci were tracked throughout mitosis; foci excluded from daughter nuclei post-mitosis were classified as mis-segregated ecDNA.

END-seq

To induce a control DSB at the MYC locus, COLO320DM and COLO320HSR cells were transfected with a Cas9 RNP complex targeting the MYC gene. The RNP complex was assembled by mixing purified Cas9 protein (Berkeley MacroLab) with an sgRNA targeting exon 2 of MYC (sequence: GCCGTATTTCTACTGCGACG; IDT), to a final concentration of 20 pmol Cas9 and 25 pmol sgRNA. The mixture was incubated at room temperature for 10–15 min. Cells were nucleofected using the Lonza 4D-Nucleofector System with the SF Cell Line Kit and program CM158. After overnight recovery, cells were treated with either the Polθi-Pol or vehicle control for 72 h, then collected for END-seq. For sgRNA-mediated depletion of AAVS1 or FANCM, cells were nucleofected with the triple-guide knockout strategy described above and treated 72 h later with Polθ-Pol inhibitor or vehicle (DMSO). Cells were collected 72 h after treatment for END-seq. END-seq was performed as previously described⁵⁴. In brief, COLO320 cells were resuspended in cell suspension buffer (10 mM Tris-HCl, pH 7.2, 50 mM EDTA, 2 mM NaCl) and embedded in 0.75% low-melting-point agarose plugs using the Bio-Rad CHEF Mammalian Genomic DNA Plug Kit. Approximately 5 × 10⁶ cells were used per plug, and two plugs were prepared per condition. Plugs were incubated in freshly prepared lysis buffer (10 mM Tris-HCl, pH 8.0, 50 mM EDTA, 150 mM NaCl, 1% SDS) containing Proteinase K (Qiagen) at 50 °C for 1 h, followed by overnight incubation at 37 °C. After lysis, plugs were rinsed and washed with plug wash buffer (10 mM Tris-HCl, pH 8.0, 50 mM EDTA) and TE buffer (10 mM Tris-HCl, pH 8.0, 1 mM EDTA), followed by treatment with RNase A (Qiagen) at 37 °C for 1 h. Additional washes in plug wash buffer were performed to remove residual enzymes. DNA DSB ends were blunted in-plug using Exonuclease VII and Exonuclease T, A-tailed with Klenow fragment (3′→5′ exo−), and ligated to a biotinylated hairpin adaptor. Plugs were then melted, digested with β-agarase I, and DNA was fragmented by sonication to an average size of ~175 bp. Biotinylated fragments were captured with streptavidin-coated beads, end-repaired, A-tailed, ligated to a second adaptor, treated with USER enzyme, and amplified by PCR with barcoded Illumina TruSeq primers. Libraries were purified with AMPure XP beads, size-selected by gel extraction, quantified, and sequenced on an Illumina platform⁵⁵.

END-seq analysis

Raw sequencing reads were trimmed using Trimmomatic (v0.39)⁵⁶. Filtered reads were aligned to the human reference genome (hg19) using Bowtie2 (v2.2.5.1)⁵⁷. Reads mapping to the mitochondrial genome or ENCODE blacklisted regions were excluded from downstream analysis. BAM files were converted to BED format using the bamtobedfunction in bedtools (v2.31.1)⁵⁸. The ecDNA region was defined as chromosome 8:127433703–129010006 based on the strong enrichment of END-seq signal across this interval. Peaks were called using MACS2 (v2.2.7.1)⁵⁹ with the parameters–keep-dup all–nomodel–shift −50–extsize 100 -q 0.001–fe-cutoff 5. Annotations for simple repeats were downloaded from the UCSC Genome Browser and overlaps between repeat elements and END-seq peaks were identified using the intersect function in bedtools⁶⁰. For visualization and downstream analysis, scaled coverage tracks normalized to reads per million were generated using BAMscale⁶¹.

Direct library preparation

COLO320DM cells were used for DLP+ experiments. Four populations were prepared: the parental starting population (T0), a DMSO-treated control population cultured for 2 months, a population cultured in parallel for 2 months in the presence of a Polθi-Hel, and a FANCM-knockout population cultured for 1 month. The FANCM knockout was generated using CRISPR–Cas9-mediated targeting, and knockout efficiency was monitored weekly by genomic DNA extraction, PCR amplification of the targeted cut site, Sanger sequencing, and EditCo’s ICE Analysis tool (v3). Single-cell whole-genome libraries were generated using DLP+, as described previously⁶². In this approach, individual cells are dispensed into nanowell chips containing preloaded primers and processed directly for single-cell library construction. Cells were subjected to lysis and protease treatment, followed by heat lysis, transposition-based fragmentation and adapter insertion, reaction neutralization and indexed PCR amplification. The resulting single-cell libraries were recovered, pooled, size-selected with AMPure XP beads, and sequenced. Single-cell DNA sequencing data was analysed using the publicly available Mondrian pipeline (https://github.com/mondrian-scwgs/mondrian) within the Isabl platform⁶³. This included alignment using bwa-mem⁶⁴ and structural variant calling using DeStruct⁶⁵. DeStruct was run jointly on all four samples. Analysis was restricted to those structural variants that were identified exclusively in a single sample, as these were thought to be newly arising somatic events. For events outside of the amplicon-implicated chromosomes (6, 8 and 13), only those events with at least 3 supporting reads were included.

Sequence analysis of the breakpoint of rearrangement junctions in human tumours

Focal amplifications in human tumour samples from PCAWG (Pancancer Analysis of Whole Genomes), Hartwig (Hartwig Medical Foundation)⁴¹, PedPanCan (tumour samples from St Jude and PBTA), GLASS (Glioma Longitudinal AnalySiS) and CUGA (Chinese Urothelial Carcinoma Genome Atlas), detected and classified by Amplicon Architect and Amplicon Classifier, publicly available through the Amplicon Repository (https://ampliconrepository.org/), were analysed^41,42,66,67. Only amplifications from whitelisted primary tumour samples were retained for the analysis.

To identify amplicon junctions occurring within TA repeat regions, junction coordinates were intersected with TA/AT-rich repeats (for example, (TA)_n, (TATA)_n, etc., including reverse complements), extracted from the UCSC Table Browser RepeatMasker track. The intersection was performed using BEDTools intersect v2.31.1. Each dataset was processed individually with the TA-rich repeat annotation from their corresponding reference genome: hg19 (PCAWG, GLASS) and hg38 (CUGA, PedPanCan).

A two-sided Pearson’s chi-squared test (R base function prop.test v4.3.2) was used to determine whether there is a significant difference between the fraction of junctions overlapping TA repeats on ecDNA vs other focal amplification types.

MYC copy number assay using TaqMan

MYC gene amplification was evaluated using TaqMan copy number variation assays on a QuantStudio 6 Flex real-time PCR system. Myc copy number in mouse neuronal stem cells was assessed using the TaqMan Copy Number Assay (probe Mm00734221_cn) and TaqMan Copy Number Reference Assay (Tfrc, 4458367). MYC copy number in human COLO320(DM and HSR) cells was assessed using the TaqMan Copy Number Assay (probe Hs01764918_cn) and TaqMan Copy Number Reference Assay (RnaseP, 4403326). Genomic DNA was isolated using the Zymo Quick-DNA Miniprep kit and amplified using the TaqPath ProAmp Master Mix (Applied Biosystems, A30865), following the supplier’s instructions. Relative copy number variations were calculated using the ΔΔC_t method.

RT–qPCR to detect MYC expression

Total RNA was isolated using the NucleoSpin RNA Clean-up kit (Macherey-Nagel) following the manufacturer’s instructions. Columns were treated with DNase I to degrade genomic DNA. 1 µg of purified RNA was reverse transcribed using the iScript gDNA Clear cDNA Synthesis Kit (Bio-Rad) following the suppliers’ instructions. MYC expression was assessed using a 1:10 dilution of cDNA with PowerUp SYBR Green Master Mix (Applied Biosystems) and standard cycling conditions on a QuantStudio 6 Flex real-time PCR system. MYC expression was normalized to ACTB and quantified using the ΔΔC_t method.

RT–qPCR to detect cre recombinase expression

Total RNA was isolated using RNeasy Mini Kit (QIAGEN, 74106) following the manufacturer’s instructions. After treatment with DNAse I (Ambion, AM2222), 0.5 μg of purified RNA was retro-transcribed with random hexamers by using SuperScript IV First-strand System (Invitrogen), cre recombinase expression was assessed using a 1:20 dilution of cDNA with PowerUp SYBR Green Master Mix (Applied Biosystems) and standard cycling conditions on a QuantStudio 6 Flex real-time PCR system. Cre recombinase expression was normalized to GAPDH and quantified using the ΔΔC_t method. Quantitative PCR with reverse transcription (RT–qPCR) primers are provided in Supplementary Table 3.

Reporting summary

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

Custom code used for the DLP+ single-cell whole-genome sequencing analysis is available at Zenodo (https://doi.org/10.5281/zenodo.21725678 (ref. ⁶⁹)). No other custom code was generated for this study. All remaining analyses were performed using publicly available software, and all scripts used for analysing sequencing data are available on GitHub as part of previously published studies. Specifically, END-seq data were processed as described previously⁵⁴ (https://github.com/Gang-Zhen/ENDseq). Repair-seq data were analysed using CRISPResso2 (https://github.com/pinellolab/CRISPResso2). Breakpoint junctions of ecDNA and non-ecDNA amplicons were obtained using AmpliconArchitect (https://github.com/AmpliconSuite/AmpliconSuite-pipeline) and AmpliconClassifier (https://github.com/AmpliconSuite/AmpliconClassifier). Amplicon junction coordinates were intersected with TA/AT-rich repeat annotations using BEDTools v2.31.1 (https://github.com/arq5x/bedtools2).

1. Wu, S., Bafna, V., Chang, H. Y. & Mischel, P. S. Extrachromosomal DNA: an emerging hallmark in human cancer. Annu. Rev. Pathol. 17, 367–386 https://doi.org/10.1146/annurev-pathmechdis-051821-114223 (2022).

   Article  CAS  PubMed  Google Scholar 

2. Hung, K. L. et al. Coordinated inheritance of extrachromosomal DNAs in cancer cells. Nature 635, 201–209 https://doi.org/10.1038/s41586-024-07861-8 (2024).

   Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

3. Nichols, A. et al. Chromosomal tethering and mitotic transcription promote ecDNA nuclear inheritance. Mol. Cell 85, 2839–2853.e2838 https://doi.org/10.1016/j.molcel.2025.06.013 (2025).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

4. Luebeck, J. et al. Extrachromosomal DNA in the cancerous transformation of Barrett’s oesophagus. Nature 616, 798–805 https://doi.org/10.1038/s41586-023-05937-5 (2023).

   Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

5. van Wietmarschen, N. et al. Repeat expansions confer WRN dependence in microsatellite-unstable cancers. Nature 586, 292–298 https://doi.org/10.1038/s41586-020-2769-8 (2020).

   Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

6. Kowalski, D. & Eddy, M. J. The DNA unwinding element: a novel, cis-acting component that facilitates opening of the Escherichia coli replication origin. EMBO J. 8, 4335–4344 https://doi.org/10.1002/j.1460-2075.1989.tb08620.x (1989).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

7. Shoshani, O. et al. Chromothripsis drives the evolution of gene amplification in cancer. Nature 591, 137–141 https://doi.org/10.1038/s41586-020-03064-z (2021).

   Article  ADS  CAS  PubMed  Google Scholar 

8. Engel, J. L. et al. The Fanconi anemia pathway induces chromothripsis and ecDNA-driven cancer drug resistance. Cell 187, 6055–6070.e6022 https://doi.org/10.1016/j.cell.2024.08.001 (2024).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

9. Rose, J. C. et al. Disparate pathways for extrachromosomal DNA biogenesis and genomic DNA repair. Cancer Discov. 15, 69–82 https://doi.org/10.1158/2159-8290.Cd-23-1117 (2025).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

10. Chung, O. W. et al. BRCA1-A and LIG4 complexes mediate ecDNA biogenesis and cancer drug resistance. Proc. Natl Acad. Sci. USA 123 e2530443123 https://doi.org/10.1073/pnas.2530443123 (2026).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

11. Jaworski, J. J. et al. ecDNA replication is disorganized and vulnerable to replication stress. Nucleic Acids Res. 53, gkaf711 https://doi.org/10.1093/nar/gkaf711 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

12. Tang, J. et al. Enhancing transcription–replication conflict targets ecDNA-positive cancers. Nature 635, 210–218 https://doi.org/10.1101/2024.03.29.586681 (2024).

13. Oobatake, Y. & Shimizu, N. Double-strand breakage in the extrachromosomal double minutes triggers their aggregation in the nucleus, micronucleation, and morphological transformation. Genes Chromosomes Cancer 59, 133–143 https://doi.org/10.1002/gcc.22810 (2020).

    Article  CAS  PubMed  Google Scholar 

14. Schoenlein, P. V. et al. Radiation therapy depletes extrachromosomally amplified drug resistance genes and oncogenes from tumor cells via micronuclear capture of episomes and double minute chromosomes. Int. J. Radiat. Oncol. Biol. Phys. 55, 1051–1065 https://doi.org/10.1016/s0360-3016(02)04473-5 (2003).

    Article  CAS  PubMed  Google Scholar 

15. Sfeir, A., Tijsterman, M. & McVey, M. Microhomology-mediated end joining chronicles: tracing the evolutionary footprints of genome protection. Annu. Rev. Cell Dev. Biol. 40, 195–218 https://doi.org/10.1146/annurev-cellbio-111822-014426 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

16. Matos-Rodrigues, G. et al. S1-END-seq reveals DNA secondary structures in human cells. Mol. Cell 82, 3538–3552.e3535 https://doi.org/10.1016/j.molcel.2022.08.007 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

17. McClellan, J. A. & Lilley, D. M. A two-state conformational equilibrium for alternating (A-T)_n sequences in negatively supercoiled DNA. J. Mol. Biol. 197, 707–721 https://doi.org/10.1016/0022-2836(87)90477-3 (1987).

    Article  CAS  PubMed  Google Scholar 

18. Inagaki, H. et al. Chromosomal instability mediated by non-B DNA: cruciform conformation and not DNA sequence is responsible for recurrent translocation in humans. Genome Res. 19, 191–198 https://doi.org/10.1101/gr.079244.108 (2009).

    Article  CAS  PubMed  Google Scholar 

19. Alitalo, K., Schwab, M., Lin, C. C., Varmus, H. E. & Bishop, J. M. Homogeneously staining chromosomal regions contain amplified copies of an abundantly expressed cellular oncogene (c-myc) in malignant neuroendocrine cells from a human colon carcinoma. Proc. Natl Acad. Sci. USA 80, 1707–1711 https://doi.org/10.1073/pnas.80.6.1707 (1983).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

20. Mateos-Gomez, P. A. et al. Mammalian polymerase theta promotes alternative NHEJ and suppresses recombination. Nature 518, 254–257 https://doi.org/10.1038/nature14157 (2015).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

21. Ceccaldi, R. et al. Homologous-recombination-deficient tumours are dependent on Poltheta-mediated repair. Nature 518, 258–262 https://doi.org/10.1038/nature14184 (2015).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

22. Bubenik, M. et al. Identification of RP-6685, an orally bioavailable compound that inhibits the DNA polymerase activity of Polθ. J. Med. Chem. 65, 13198–13215 https://doi.org/10.1021/acs.jmedchem.2c00998 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

23. Mochirian, P. et al. The discovery of RP-2119: a potent, selective, and orally bioavailable Polθ ATPase inhibitor. J. Med. Chem. 68, 19726–19745 https://doi.org/10.1021/acs.jmedchem.5c02103 (2025).

    Article  CAS  PubMed  Google Scholar 

24. Kang, X. et al. Extrachromosomal DNA replication and maintenance couple with DNA damage pathway in tumors. Cell 188, 3405–3421.e27 https://doi.org/10.1016/j.cell.2025.04.012 (2025).

    Article  CAS  PubMed  Google Scholar 

25. Pradella, D. et al. Engineered extrachromosomal oncogene amplifications promote tumorigenesis. Nature 637, 955–964 https://doi.org/10.1038/s41586-024-08318-8 (2025).

    Article  ADS  CAS  PubMed  Google Scholar 

26. Wyatt, D. W. et al. Essential roles for polymerase essential roles for polymerase θ-mediated end joining in the repair of chromosome breaks. Mol. Cell 63, 662–673 https://doi.org/10.1016/j.molcel.2016.06.020 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

27. Brambati, A. et al. RHINO directs MMEJ to repair DNA breaks in mitosis. Science 381, 653–660 https://doi.org/10.1126/science.adh3694 (2023).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

28. Sankar, V. et al. Genetic elements promote retention of extrachromosomal DNA in cancer cells. Nature 649, 152–160 https://doi.org/10.1038/s41586-025-09764-8 (2026).

    Article  ADS  CAS  PubMed  Google Scholar 

29. Lange, J. T. et al. The evolutionary dynamics of extrachromosomal DNA in human cancers. Nat. Genet. 54, 1527–1533 https://doi.org/10.1038/s41588-022-01177-x (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

30. Wu, S. et al. Circular ecDNA promotes accessible chromatin and high oncogene expression. Nature 575, 699–703 https://doi.org/10.1038/s41586-019-1763-5 (2019).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

31. Canela, A. et al. DNA breaks and end resection measured genome-wide by end sequencing. Mol. Cell 63, 898–911 https://doi.org/10.1016/j.molcel.2016.06.034 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

32. Kim, H. et al. Extrachromosomal DNA is associated with oncogene amplification and poor outcome across multiple cancers. Nat. Genet. 52, 891–897 https://doi.org/10.1038/s41588-020-0678-2 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

33. Haniford, D. B. & Pulleyblank, D. E. Transition of a cloned d(AT)n-d(AT)n tract to a cruciform in vivo. Nucleic Acids Res. 13, 4343–4363 https://doi.org/10.1093/nar/13.12.4343 (1985).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

34. Dayn, A., Malkhosyan, S. & Mirkin, S. M. Transcriptionally driven cruciform formation in vivo. Nucleic Acids Res. 20, 5991–5997 https://doi.org/10.1093/nar/20.22.5991 (1992).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

35. Liu, L. F. & Wang, J. C. Supercoiling of the DNA template during transcription. Proc. Natl Acad. Sci. USA 84, 7024–7027 https://doi.org/10.1073/pnas.84.20.7024 (1987).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

36. Feng, S. et al. Profound synthetic lethality between SMARCAL1 and FANCM. Mol. Cell 84, 4522–4537.e4527 https://doi.org/10.1016/j.molcel.2024.10.016 (2024).

    Article  CAS  PubMed  Google Scholar 

37. Lu, S. et al. Short inverted repeats are hotspots for genetic instability: relevance to cancer genomes. Cell Rep. 10, 1674–1680 https://doi.org/10.1016/j.celrep.2015.02.039 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

38. Mengoli, V. et al. WRN helicase and mismatch repair complexes independently and synergistically disrupt cruciform DNA structures. EMBO J. 42, e111998 https://doi.org/10.15252/embj.2022111998 (2023).

    Article  CAS  PubMed  Google Scholar 

39. Zhu, K. et al. CoRAL accurately resolves extrachromosomal DNA genome structures with long-read sequencing. Genome Res. 34, 1344–1354 https://doi.org/10.1101/gr.279131.124 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

40. Lee, J. J.-K. et al. Evolution of oncogene amplification across 86,000 cancer cell genomes. Preprint at bioRxiv https://doi.org/10.64898/2026.02.12.705658 (2026).

41. Kim, H. et al. Mapping extrachromosomal DNA amplifications during cancer progression. Nat. Genet. 56, 2447–2454, https://doi.org/10.1038/s41588-024-01949-7 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

42. Luebeck, J. et al. AmpliconSuite: an end-to-end workflow for analyzing focal amplifications in cancer genomes. Preprint at bioRxiv https://doi.org/10.1101/2024.05.06.592768 (2024).

43. Turner, K. M. et al. Extrachromosomal oncogene amplification drives tumour evolution and genetic heterogeneity. Nature 543, 122–125 https://doi.org/10.1038/nature21356 (2017).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

44. Chedin, F. & Benham, C. J. Emerging roles for R-loop structures in the management of topological stress. J. Biol. Chem. 295, 4684–4695 https://doi.org/10.1074/jbc.REV119.006364 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

45. Inagawa, T. et al. C-terminal extensions of Ku70 and Ku80 differentially influence DNA end binding properties. Int. J. Mol. Sci. 21, 6275 https://doi.org/10.3390/ijms21186725 (2020).

    Article  CAS  Google Scholar 

46. Deng, L. et al. Mitotic CDK promotes replisome disassembly, fork breakage, and complex DNA rearrangements. Mol. Cell 73, 915–929.e916 https://doi.org/10.1016/j.molcel.2018.12.021 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

47. Fielden, J. et al. Comprehensive interrogation of synthetic lethality in the DNA damage response. Nature 640, 1093–1102 https://doi.org/10.1038/s41586-025-08815-4 (2025).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

48. Bailey, C. et al. Origins and impact of extrachromosomal DNA. Nature 635, 193–200 https://doi.org/10.1038/s41586-024-08107-3 (2024).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

49. Brinkman, E. K., Chen, T., Amendola, M. & van Steensel, B. Easy quantitative assessment of genome editing by sequence trace decomposition. Nucleic Acids Res. 42, e168 https://doi.org/10.1093/nar/gku936 (2014).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

50. Bajar, B. T. et al. Fluorescent indicators for simultaneous reporting of all four cell cycle phases. Nat. Methods 13, 993–996 https://doi.org/10.1038/nmeth.4045 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

51. Schindelin, J. et al. Fiji: an open-source platform for biological-image analysis. Nat. Methods 9, 676–682 https://doi.org/10.1038/nmeth.2019 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

52. Guzman, C., Bagga, M., Kaur, A., Westermarck, J. & Abankwa, D. ColonyArea: an ImageJ plugin to automatically quantify colony formation in clonogenic assays. PLoS ONE 9, e92444 https://doi.org/10.1371/journal.pone.0092444 (2014).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

53. Clement, K. et al. CRISPResso2 provides accurate and rapid genome editing sequence analysis. Nat. Biotechnol. 37, 224–226 https://doi.org/10.1038/s41587-019-0032-3 (2019).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

54. Wong, N., John, S., Nussenzweig, A. & Canela, A. in Homologous Recombination (eds Aguilera, A. & Carreira, A.) 9–31 https://doi.org/10.1007/978-1-0716-0644-5_2 (Springer 2021).

55. Tubbs, A. et al. Dual roles of poly(dA:dT) tracts in replication initiation and fork collapse. Cell 174, 1127–1142.e1119 https://doi.org/10.1016/j.cell.2018.07.011 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

56. Bolger, A. M., Lohse, M. & Usadel, B. Trimmomatic: a flexible trimmer for Illumina sequence data. Bioinformatics 30, 2114–2120 https://doi.org/10.1093/bioinformatics/btu170 (2014).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

57. Langmead, B. & Salzberg, S. L. Fast gapped-read alignment with Bowtie 2. Nat. Methods 9, 357–359 https://doi.org/10.1038/nmeth.1923 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

58. Quinlan, A. R. & Hall, I. M. BEDTools: a flexible suite of utilities for comparing genomic features. Bioinformatics 26, 841–842 https://doi.org/10.1093/bioinformatics/btq033 (2010).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

59. Zhang, Y. et al. Model-based analysis of ChIP-Seq (MACS). Genome Biol. 9, R137 https://doi.org/10.1186/gb-2008-9-9-r137 (2008).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

60. Casper, J. et al. The UCSC Genome Browser database: 2026 update. Nucleic Acids Res. 54, D1331–D1335 https://doi.org/10.1093/nar/gkaf1250 (2026).

    Article  PubMed  PubMed Central  Google Scholar 

61. Pongor, L. S. et al. BAMscale: quantification of next-generation sequencing peaks and generation of scaled coverage tracks. Epigenet. Chromatin 13, 21 https://doi.org/10.1186/s13072-020-00343-x (2020).

    Article  CAS  Google Scholar 

62. Laks, E. et al. Clonal decomposition and DNA replication states defined by scaled single-cell genome sequencing. Cell 179, 1207–1221.e1222 https://doi.org/10.1016/j.cell.2019.10.026 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

63. Medina-Martínez, J. S. et al. Isabl Platform, a digital biobank for processing multimodal patient data. BMC Bioinform. 21, 549 https://doi.org/10.1186/s12859-020-03879-7 (2020).

    Article  Google Scholar 

64. Li, H. & Durbin, R. Fast and accurate short read alignment with Burrows-Wheeler transform. Bioinformatics 25, 1754–1760 https://doi.org/10.1093/bioinformatics/btp324 (2009).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

65. McPherson, A., Shah, S. & Sahinalp, S. C. deStruct: accurate rearrangement detection using breakpoint specific realignment. Preprint at bioRxiv https://doi.org/10.1101/117523 (2017).

66. Grobner, S. N. et al. The landscape of genomic alterations across childhood cancers. Nature 555, 321–327 https://doi.org/10.1038/nature25480 (2018).

    Article  ADS  CAS  PubMed  Google Scholar 

67. Lv, W. et al. Spatial-temporal diversity of extrachromosomal DNA shapes urothelial carcinoma evolution and tumor-immune microenvironment. Cancer Discov. 15, 1225–1246 https://doi.org/10.1158/2159-8290.CD-24-1532 (2025).

    Article  ADS  CAS  PubMed Central  Google Scholar 

68. Sfeir, A. MMEJ repair of breaks at TA repeats maintains ecDNA and cancer fitness. Mendeley Data https://doi.org/10.17632/2vh4pg9y6n.1 (2026).

69. Myers, M., McPherson, A., Sfeir, A. & Selvaraj, M. Code to generate figures from DLP+ data, supporting “MMEJ Repair of Breaks at TA Repeats Maintains ecDNA and Cancer Fitness”. Zenodo https://doi.org/10.5281/zenodo.21725678 (2026).

70. Ciszewski, W. M., Tavecchio, M., Dastych, J. & Curtin, N. J. DNA-PK inhibition by NU7441 sensitizes breast cancer cells to ionizing radiation and doxorubicin. Breast Cancer Res. Treat. 143, 47–55 https://doi.org/10.1007/s10549-013-2785-6 (2014).

    Article  CAS  PubMed  Google Scholar 

Download references

Author notes

1. Gabriel Matos-Rodrigues

   Present address: CNRS-UMR9019, Gustave Roussy, Université Paris-Saclay, Villejuif, France

2. Alessandra Brambati

   Present address: Department of Pharmacology, Anschutz Medical Campus, University of Colorado School of Medicine, Aurora, CO, USA

3. Davide Pradella

   Present address: Italian Institute for Genomic Medicine, Candiolo, Italy

4. These authors contributed equally: David Billing, Monica Selvaraj

Authors and Affiliations

1. Molecular Biology Program, Sloan Kettering Institute, Memorial Sloan Kettering Cancer Center, New York, NY, USA

   David Billing, Monica Selvaraj, Megan E. Kelley, Benton Bailey, Alessandra Brambati, Oluchi Ezekwenna, Ashley Nichols, Yi-Zhen Jiang, Lucia Wang, Shih-Chun Wang, John Maciejowski & Agnel Sfeir

2. Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, USA

   David Billing

3. Laboratory of Genome Integrity, National Cancer Institute, NIH, Bethesda, MD, USA

   Gabriel Matos-Rodrigues, Gang Zhen, Elsa Callen & André Nussenzweig

4. Department of Pediatric Oncology and Hematology, Charité–Universitätsmedizin Berlin, Berlin, Germany

   Finnja Becker, Elias Rodriguez-Fos & Anton George Henssen

5. Experimental and Clinical Research Center (ECRC), Max Delbrück Center and Charité–Universitätsmedizin Berlin, Berlin, Germany

   Finnja Becker, Elias Rodriguez-Fos & Anton George Henssen

6. Halvorsen Center for Computational Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, USA

   Matthew A. Myers, Andrew McPherson & Sohrab P. Shah

7. Computational Oncology, Department of Epidemiology and Biostatistics, Memorial Sloan Kettering Cancer Center, New York, NY, USA

   Matthew A. Myers, Andrew McPherson & Sohrab P. Shah

8. Repare Therapeutics, Montréal, Québec, Canada

   Marie-Claude Mathieu, Hugo Poirier, Stephen J. Morris & Michal Zimmermann

9. Cancer Biology and Genetics Program, Sloan Kettering Institute, Memorial Sloan Kettering Cancer Center, New York, NY, USA

   Davide Pradella & Andrea Ventura

10. Repare Therapeutics, Cambridge, MA, USA

    Michael Zinda

Contributions

D.B. and A.S. conceived the study and designed the experiments. D.B. and M.S. performed the majority of the experimental work, with help from M.E.K., B.B., O.E., Y.-Z.J., S.-C.W. and L.W. M.S. performed END-seq experiments together with G.M.-R., G.Z. and E.C. A. Nussenzweig supervised and interpreted the END-seq analyses. A.B. performed and analysed the Repair-seq experiments. D.P. analysed MYC ecDNA levels in genetically engineered cells, supervised by A.V. A. Nichols performed live-cell ecDNA imaging experiments, supervised by J.M. F.B. and E.R.-F. analysed ecDNA breakpoints from whole-genome sequencing of human tumours, supervised by A.G.H. M.-C.M., H.P., S.J.M., M. Zimmermann and M. Zinda designed, performed and interpreted the experiments to develop RP-2119. M.A.M. performed the single-cell whole-genome sequencing (DLP+) analysis, with A.M. and S.P.S. contributing to analysis design and interpretation. A.S. supervised the study and acquired funding. A.S. wrote the manuscript, with input and feedback from all authors.

Corresponding author

Correspondence to Agnel Sfeir.

Competing interests

A.S. is a co-founder, consultant, and shareholder for Repare Therapeutics. M.-C.M., H.P., S.J.M., M. Zimmermann and M. Zinda are current or former employees of Repare Therapeutics and receive salary and equity compensation. S.P.S. receives research funding from AstraZeneca and Bristol Myers Squibb, unrelated to this work. A.G.H. is a founder, consultant and shareholder for Econic Biosciences. The other authors declare no competing interests.

Peer review information

Nature thanks Jan Korbel who co-reviewed with Nikolaus Watson; Sergei Mirkin and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Reporting Summary (download PDF )

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions