Extrachromosomal DNA-Driven Oncogene Dosage Heterogeneity Promotes Rapid Adaptation to Therapy in MYCN-Amplified Cancers

Abstract

Extrachromosomal DNA (ecDNA) amplification enhances intercellular oncogene dosage variability and accelerates tumor evolution by violating foundational principles of genetic inheritance through its asymmetric mitotic segregation. Spotlighting high-risk neuroblastoma, we demonstrate how ecDNA amplification undermines the clinical efficacy of current therapies in cancers with extrachromosomal MYCN amplification. Integrating theoretical models of oncogene copy number–dependent fitness with single-cell ecDNA quantification and phenotype analyses, we reveal that ecDNA copy-number heterogeneity drives phenotypic diversity and determines treatment sensitivity through mechanisms unattainable by chromosomal oncogene amplification. We demonstrate that ecDNA copy number directly influences cell fate decisions in cancer cell lines, patient-derived xenografts, and primary neuroblastomas, illustrating how extrachromosomal oncogene dosage–driven phenotypic diversity offers a strong evolutionary advantage under therapeutic pressure. Furthermore, we identify senescent cells with reduced ecDNA copy numbers as a source of treatment resistance in neuroblastomas and outline a strategy for their targeted elimination to improve the treatment of MYCN-amplified cancers.

INTRODUCTION

Extrachromosomal DNA (ecDNA) poses a great challenge in oncology as a major driver of aberrant oncogene expression (1). It is a common form of oncogene amplification in most human cancer entities and affects women, men, children, and adults alike (2, 3). ecDNA does not only serve as a substrate for oncogene amplification but also contains more accessible chromatin and hijacks enhancer elements that facilitate oncogene expression (4–7). Due to their lack of centromeres, ecDNA molecules are relieved of hereditary constraints and conventional segregation laws (8). Furthermore, they can reintegrate into chromosomes as homogenously staining regions (HSR; refs. 5, 6). Therefore, dynamic changes in ecDNA abundance lead to rapid oncogene copy-number variation in dividing cancer cells and accelerate intratumoral heterogeneity (9–13). However, the impact of ecDNA-mediated intratumoral oncogene copy-number heterogeneity on cancer cells’ phenotypic plasticity is not well understood.

Clinical outcomes in patients with cancer whose tumors harbor ecDNA are significantly worse compared with other patients (2, 3, 14), but mechanistic connections between ecDNA-driven oncogene copy-number heterogeneity and principles of treatment resistance have not been investigated. This is particularly true for neuroblastomas, which develop in very young children as tumors of the sympathetic nervous system from neural crest cells (15). Although treatment intensification with combinations of multimodal chemotherapy, surgery, radiotherapy, differentiating agents, targeted therapies, and immunotherapy has been able to improve outcomes for patients with neuroblastoma in the last decade, high-risk neuroblastomas, which frequently harbor MYCN amplifications (MNA), are still very hard to cure (16). In neuroblastoma, MNA is predominantly maintained on ecDNA and not on HSR (17). A long-standing conundrum in pediatric oncology remains the discrepancy between the very good initial treatment responses of MYCN-amplified neuroblastomas followed by very rapid disease recurrence. This suggests the existence of yet undefined rapid adaptation processes in these tumors. We hypothesized that ecDNA-driven oncogene dosage heterogeneity may cause phenotypic variation among cancer cells harboring MYCN ecDNA at different copy numbers that enables a fast and plastic response to the changing selection pressures in the presence and absence of treatment.

RESULTS

ecDNA Links Genetic to Phenotypic Diversity in Neuroblastomas

Unequal mitotic segregation of ecDNA leads to substantial intercellular variability in copy number (18). To investigate copy-number heterogeneity in neuroblastomas, we first assessed MNA using FISH across primary and relapsed neuroblastomas, xenografts derived from patients (PDX) with neuroblastoma, and cell lines (n_total = 191 samples, n_MNA = 95). We then analyzed MYCN copy numbers on ecDNA and HSR within MYCN-amplified samples (Fig. 1A; Supplementary Fig. S1A). Quantification of MYCN copy numbers by either the Find Maxima function in Fiji (ImageJ) or the corrected total cell fluorescence (CTCF) intensity score indicated that ecDNA copy number varied greatly from cell to cell, whereas copy numbers from cells with linear HSR amplifications appeared more evenly distributed (Fig. 1B; Supplementary Fig. S1B and S1C). In line with the random segregation of oncogenes on ecDNA compared with linear chromosomal oncogene amplification, we observed a high variation in MYCN ecDNA copy numbers across cell lines, PDX, and primary neuroblastomas, indicated by a higher median absolute deviation for ecDNA samples than for HSR samples (Fig. 1C). In addition to the marked intertumor heterogeneity, ecDNA tumors also displayed a significantly higher intratumor heterogeneity, as demonstrated by the elevated mean coefficient of variation (CV) of the sample copy number over all samples for ecDNA than for HSR (65.19% ± 18.44% vs. 49.21% ± 12.64%, P < 0.001, Mann–Whitney U test; Fig. 1D and E). Thus, ecDNA can cause intratumoral oncogene dosage heterogeneity in neuroblastoma. This raises the question of whether such intercellular differences in oncogene copy number shape phenotypic variation.

Figure 1.

Extrachromosomal MNA impacts on oncogene copy-number and phenotypic heterogeneity in neuroblastoma. A, Schematic illustration of the experimental procedure to quantify ecDNA in MYCN-amplified neuroblastoma samples. B, Representative interphase FISH images of ecDNA and HSR neuroblastoma cell lines, PDX, and patient samples. For all pictures, FISH signals are presented in the following color scheme: MYCN (green) and CEP 2 as a reference probe for chromosome 2 (red) with Hoechst as a nuclear counterstain in blue. C, Box plot showing MYCN foci in MYCN-amplified neuroblastoma cell lines, PDX, and patient samples according to their amplification status (ecDNA vs. HSR). n = 10 cell lines, 8 PDX, and 52 patients (see Supplementary Fig. S1A for details; only untreated PDX and patient samples were considered). For all groups, mean absolute deviation values are presented in boxes. P values were calculated using a Fligner–Killeen test. *, P < 0.05 for cell lines and PDX; *, P < 0.0001 for patients. D, CV for all ecDNA samples as shown in A and C. For samples with the highest and lowest CV, MYCN copy number quantifications and distributions are highlighted in square boxes. $\overline{\text{x}}$ indicates the mean CV of the sample copy number over all samples. E, CV for all HSR samples shown in A and C. Square boxes demonstrate MYCN copy-number quantification and distribution for the samples with the highest and lowest coefficients of variation as in D. $\overline{\text{x}}$ indicates the mean CV of the sample copy number over all samples. F, Schematic illustration of the colony formation experiment. ecDNA and HSR cells were seeded as single cells and allowed to grow for 4 days. MYCN copy number was determined by FISH for several groups of colonies defined by cell number, as indicated. G, MYCN copy-number quantification and distribution in colonies with different cell numbers generated as illustrated in F using the near-isogenic cell line STA-NB-10 DM (left) and STA-NB-10 HSR (right). H, Schematic illustration of the computational model to assess the effect of ecDNA on cell fitness. The simulation starts from a random selection of n single cells harboring varying initial ecDNA copy numbers. These cells undergo clonal expansion for a short, defined period. At the end, the population size and ecDNA copy-number distribution within each clone are quantified. I, Representative simulation results of the ecDNA copy-number distribution across clonal populations of varying sizes under ecDNA copy number–dependent fitness (left) and ecDNA copy number–independent fitness (right) from the computational model described in H. J, Schematic illustration of ecDNA and HSR colony formation dynamics and the associated tumor heterogeneity. Cells with fewer ecDNA copies generate smaller colonies compared with cells with a higher initial number of ecDNA, whereas HSR cells with a stable MYCN copy number produce colonies that are more uniform in size.

We and others recently showed that intercellular ecDNA copy-number heterogeneity drives oncogene expression differences and affects intercellular gene expression diversity (19, 20). Neuroblastoma is a prototypical example of a tumor entity with transcriptionally heterogeneous cell populations (21, 22). Although superenhancer-driven transcription factor networks exist that influence lineage fidelity and cell identity, MYCN has been identified as the key factor that influences transcription in neuroblastoma both qualitatively and quantitatively in a dose-dependent manner. However, it is not known whether ecDNA-driven MYCN dosage differences directly contribute to phenotypic heterogeneity in neuroblastoma (23–25). As high MYCN or MYC expression results in increased cellular proliferation in neuroblastoma and other cancers, we hypothesized that ecDNA-driven intercellular differences in MYC or MYCN copy number could influence phenotypic diversity (26). To test this, we performed clonogenic assays with near-isogenic neuroblastoma and colon cancer cell lines that harbor amplifications of either MYC or MYCN on ecDNA or HSR and assessed their oncogene copy numbers using FISH (Fig. 1F). Indeed, ecDNA copy numbers were significantly higher in large (≥15 and 10–15) colony-forming cells than in smaller colony-forming cells (2–5, 6–9) or single cells (Fig. 1G; Supplementary Fig. S1D–S1H). We also investigated the proliferation dynamics of ecDNA and HSR cells using CellTrace Violet and the cytometry data modeling software ModFit LT. Although HSR cells from the near-isogenic neuroblastoma cell line STA-NB-10 proliferated more quickly than their ecDNA counterparts and progressed to daughter cell generations with little variation. In contrast, ecDNA cells displayed a marked heterogeneity in daughter cell generations only after a few cell divisions (Supplementary Fig. S1I–S1L). Therefore, high MYC or MYCN copy numbers seem to be linked directly to higher proliferation, particularly in HSR cells with little oncogene copy-number variation. Furthermore, the different proliferation potential of high versus low ecDNA copy-number cells suggests that ecDNA-driven genetic diversity directly affects phenotypic heterogeneity.

Modeling of ecDNA-Driven Gradual Selective Advantage

We previously demonstrated that ecDNA copy-number heterogeneity can be explained by the interplay of random ecDNA segregation and positive selection of ecDNA-containing cells, independent of their copy number (12). The observations above, however, suggest that cells with extrachromosomal amplifications of MYCN or MYC exhibit copy number–dependent fitness that links expression levels of oncogenes to ecDNA copy number. Therefore, we performed simulations mimicking the experimental clonogenic assays to test when and how copy number–dependent fitness affects ecDNA cell expansion.

We conducted individual-based stochastic simulations to evaluate two alternative hypotheses about how cell fitness depends on ecDNA copy number. The first hypothesis assumes copy number–independent fitness, in which all ecDNA-positive cells share the same fitness advantage, regardless of copy number. In contrast, the second hypothesis suggests that cell fitness is influenced by ecDNA copy number. To model this relationship, we used a heuristic approach, representing fitness dependence on copy number with an increasing sigmoid function:

$$s=\left\{\begin{array}{ll}1+\left(s_{max}-1\right)\left(1+\frac{k_e-k}{k_e-k_m}\right){\left(\frac{k}{k_e}\right)}^{\frac{k_e}{k_e-k_m}},\hfill & 0\le k<k_e\hfill \\ s_{max},\hfill & k\ge k_e\hfill \end{array}\right.$$

where $s$ is the fitness of a cell with $k$ ecDNA copies, $s_{max}$ the maximum fitness of a cell, and $k_e$ the number of ecDNA copies at which that maximum fitness is reached. This is a general parametrization of a sigmoid growth curve that can model both linear and S-shaped dependencies determined by the choice of the inflection point parameter $k_m$. For $k_e=1$ and $k_m=0$, we recover the special case of constant copy number–independent fitness.

Stochastic simulations were implemented using a Gillespie algorithm. We ran 100 repeats of each simulation for different parameter settings, with each repeat initiated by a single cell carrying a random number of ecDNA copies drawn from a uniform distribution in the range of 1 to 30 (Fig. 1H). Under the copy number–dependent fitness hypothesis, the ecDNA copy number of the first assay-initiating cell strongly correlated with the average ecDNA copy number in the derived cell population after clonal expansion, and cells carrying a higher initial ecDNA copy number grew to larger clonal populations (Supplementary Fig. S2A and S2B). We also observed that the median ecDNA copy number was higher in larger cell populations, suggesting that higher ecDNA copy numbers confer observable clonogenic and proliferative selective advantages (Fig. 1I; Supplementary Fig. S2C). Interestingly, unlike the median, the variance of the ecDNA copy-number distribution does not increase monotonically with population size (Supplementary Fig. S2D). These trends were consistently observed across multiple simulation replications. Under ecDNA copy number–independent fitness, the ecDNA copy number of the first assay-initiating cell also strongly correlated with the average ecDNA copy number in the derived cell population after clonal expansion (Supplementary Fig. S2E). Conversely, we did not observe a linear relationship between ecDNA copy number and the final population size in this model (Fig. 1I; Supplementary Fig. S2F), nor did the median and variance of ecDNA copy-number distributions change in the same way as in the copy number–dependent fitness scenario (Supplementary Fig. S2G and S2H). Thus, a model of positive selection superimposed on the distribution of ecDNA at cell division was able to explain the frequency distribution of ecDNA. Overall, ecDNA samples, unlike HSR tumors, show copy number–dependent fitness, which drives genetic heterogeneity and instructs clonogenic growth (Fig. 1J).

ecDNA Drives Intercellular Molecular Heterogeneity

MYC family oncoproteins broadly influence the phenotype of cancer cells (13, 27). Furthermore, MYCN amplification on ecDNA shapes different transcriptional states of neuroblastoma cells (20). Therefore, we reasoned that neuroblastoma cells with different MYCN copy numbers on ecDNA should show variations in their proteomic profiles that could account for the observed differences in colony formation potential and proliferation. Indeed, MYCN copy numbers positively correlated with MYCN protein levels, as indicated by combined FISH and immunofluorescence staining in the neuroblastoma cell line CHP212 with extrachromosomal MNA (Supplementary Fig. S3A). To analyze the impact of MYCN ecDNA heterogeneity on phenotypic diversity in neuroblastoma, we directly compared proteome profiles of MYCN-high and MYCN-low cells from CHP212 using a FISH-guided, MYCN phenotype–resolved proteomic approach (Fig. 2A). For this purpose, individual cells were classified by their MYCN copy numbers and isolated by single-cell laser microdissection. Pooled fractions of cells with high (>40 MYCN copies) and low (<5 MYCN copies) MYCN copy numbers were subsequently analyzed by label-free quantitative proteomics (Fig. 2B). Although we could not quantify MYCN levels with this approach, we detected the DEAD-box helicase DDX1, which is coamplified together with MYCN on ecDNA in this cell line (13), as significantly enriched in cells with high MYCN copy numbers (Supplementary Fig. S3B). Therefore, our FISH-guided proteomics analysis enabled us to uniquely profile ecDNA high and ecDNA low neuroblastoma cells. As highlighted for individual proteins and by pathway analyses, cells with high MYCN copy numbers were characterized by protein modules associated with replication stress and DNA damage, whereas cells with low MYCN copy numbers showed an enhanced expression of pathways related to protein secretion, inflammation, autophagy, senescence, and mesenchymal identity (Fig. 2B and C; Supplementary Table S1). To directly link MYCN copy numbers to phenotypic diversity, we performed single-cell genome and transcriptome sequencing (scG&T-seq) on a neuroblastoma PDX with extrachromosomal MNA and interrogated published G&T-seq data of the neuroblastoma cell line TR14. As previously demonstrated for TR14, MYCN expression was also strongly correlated with ecDNA copy number in the neuroblastoma PDX (Supplementary Fig. S3C). Similar to our FISH-guided proteomics results, transcriptional programs representing MYCN target genes, proliferation, replication stress, and DNA damage signaling were more strongly expressed in neuroblastoma cells with high MYCN copy number, whereas gene expression modules of senescence were enriched in cells with low MYCN copy number (Fig. 2D; Supplementary Fig. S3D). In addition to neuroblastoma samples, we also analyzed the influence of MYCN copy-number variation on phenotypic diversity in a sample from a patient with MYCN-amplified medulloblastoma by single-cell multiome sequencing (Fig. 2E). Comparing the top and bottom 20% of MYCN ecDNA-containing medulloblastoma cells, we fully recapitulated the ecDNA copy number–driven phenotypic plasticity observed in neuroblastomas (Fig. 2F; Supplementary Fig. S3E and S3F). Next, we tested whether the different proteome profiles of high and low ecDNA cells were also reflected by a broader molecular and phenotypic heterogeneity between single ecDNA cells compared with HSR cells. Therefore, we quantified MYCN protein levels, the abundance of histone H2AX phosphorylated on serine 139 (γH2AX) as a measure of DNA damage, and the trimethylation of the amino acid lysine 9 on histone 3 (H3K9me3) as a marker of senescence by immunofluorescence in several ecDNA versus HSR neuroblastoma cell lines and PDX samples. Although γH2AX and H3K9me3 values were significantly different in MYCN-high versus MYCN-low ecDNA cells, as indicated by ImmunoFISH analysis, we uniformly observed a much higher variation in the abundance of MYCN, γH2AX, and H3K9me3 in the ecDNA samples (Fig. 2G and H; Supplementary Fig. S3G–S3J). Therefore, ecDNA-containing tumors do not only show a broader oncogene dosage difference arising from the MYCN copy-number differences on ecDNA but also display a greater molecular and phenotypic variance than HSR neuroblastomas.

Figure 2.

MYCN copy-number differences diversify cellular phenotypes in neuroblastoma. A, Schematic illustration of the FISH-guided proteomics workflow using image-guided laser microdissection followed by label-free quantitation of proteins with LC/MS in MYCN-high vs. MYCN-low copy-number cells in the neuroblastoma ecDNA cell line CHP-212. B, Volcano plot of differentially expressed proteins (P value ≤ 0.05) in pooled MYCN-high (>40 spots) vs. MYCN-low (<5 spots) cells by FISH-guided proteomics. Representative proteins of significantly enriched gene sets and pathways are indicated as follows: INTS13 and CTDSPL2 (DNA replication stress), STAG1 and PRDX2 (DNA damage), HLA-C (inflammation), CD44 and ANXA2 (mesenchymal phenotype), IGF2R and PIK3C3 (autophagy), and MMP2, MMP14, IGF2BP1, and MAP4K3 (senescence). C, Gene set and pathway enrichment analysis in MYCN-high vs. MYCN-low cells as in B. Tested gene sets and pathways were derived from the HALLMARK gene set of the human Molecular Signatures Database, WikiPathways, or reference 21 for the MES.signature.genes (van Groningen) and reference 85 for the senescence (Casella_up) signatures. D, Box plots comparing gene module enrichment scores for selected gene sets in MYCN-high vs. MYCN-low cells from scG&T-seq of a neuroblastoma PDX with extrachromosomal MNA. Enrichment scores were calculated for each cell, and box plots representing the IQR (25th–75th percentiles), with the median indicated by a central line, were plotted for cells with the highest (top 25%; red) vs. lowest (bottom 25%; blue) MYCN copy numbers. FDR-adjusted P values were calculated using a Wilcoxon rank-sum test. P < 0.05 for WP_DNA_DAMAGE_RESPONSE and senescence (Casella_up); P < 0.0001 for HALLMARK_MYC_TARGETS_V1, WP_CELL_CYCLE, REACTOME_Activation of ATR in response to replication stress, and the gene expression signature of cellular senescence (CellAge; ref. 28). E, Uniform Manifold Approximation and Projection for Dimension Reduction (UMAP) visualization of multiome (RNA + ATAC) single-cell data obtained from an SHH MB tumor (7316-178) containing an extrachromosomal MNA. Individual cells with high (n = 1225) and low (n = 1225) MYCN ecDNA amplifications are highlighted in blue and red, respectively, oligodendrocyte precursor cells (OPC) in purple and hematopoietic cells in green. F, Violin plots showing gene module enrichment scores for selected gene sets in single cells with high (red) and low (blue) extrachromosomal MNA. P values were calculated using the Student t test. Inset box plots represent the IQR (25th–75th percentiles), with the median indicated by a central line; outliers are displayed as individual points. P < 0.05 for senescence (Casella_up) comparison; P < 0.0001 for all the other comparisons. G, Relative frequency distribution of MYCN protein levels, γH2AX as a marker of DNA damage, and trimethylation of H3K9me3 as a marker of senescence in the near-isogenic neuroblastoma cell line pair STA-NB-10 HSR (red) and STA-NB-10 DM (blue). Mean (dashed lines) and median absolute deviation values are indicated in corresponding colors. P values were calculated using a Fligner–Killeen test. *, P < 0.05. H, Relative frequency distribution of MYCN, γH2AX, and H3K9me3 in 3 ecDNA PDX and 3 HSR neuroblastoma PDX. Mean (dashed lines) and MAD values are indicated in corresponding colors. P values were calculated using a Fligner–Killeen test. *, P < 0.05. BH, Benjamini–Hochberg.

ecDNA Copy-Number Heterogeneity Is Accompanied by Treatment Response Heterogeneity

Growing evidence suggests that genetically and phenotypically heterogeneous tumors are very difficult to treat (29). Based on our observations of ecDNA-driven oncogene dosage heterogeneity and phenotypic diversity in neuroblastomas, we reasoned that this may directly affect treatment responses. Genotoxic treatments currently represent the standard of care for most cancers, including MYCN-amplified neuroblastomas. They preferentially affect tumor cells with high MYC(N) levels, which show increased DNA replication stress and are sensitized to apoptosis (30–32). Thus, genotoxic treatment pressure is predicted to result in the positive selection of cells with low MYC(N) activity, as has been described for chemotherapy-resistant tumor cells, and the negative selection of tumor cells with high endogenous DNA damage levels (33, 34). We hypothesized that due to their oncogene copy-number variation, ecDNA cells may react more heterogeneously to genotoxic therapies than HSR cells, thereby driving adaptation to treatment more efficiently.

We first tested this hypothesis using our computational model by incorporating different possible effects of drug treatments. Again, two alternative hypotheses about the relationship between ecDNA copy numbers and cancer cell response to chemotherapy were tested. First, we considered that the effect of the drug was independent of ecDNA copy number (model A), meaning that the cytotoxicity of the drug or the sensitivity of the cancer cells is uniform across all cells, regardless of their ecDNA content (Supplementary Fig. S4A and S4B). Second, we tested the scenario in which cells with higher ecDNA copy numbers, due to their rapid proliferation and increased replication stress, were more vulnerable to chemotherapy (model B, Fig. 3A). In this scenario, cytotoxicity and ecDNA copy number are positively correlated. More specifically, we assumed that in addition to the drug’s baseline toxicity $\left(d_0\right)$, a cell would undergo apoptosis if the drug targets more than a certain threshold number of ecDNA copies $\left(k_d\right)$. This effect was modeled by a probability $\text{Pr}\left(\text{X}\ge \text{k\_d}\right)$, where $X$ is a binomial random variable representing the number of ecDNA copies hit by the drug (Fig. 3B).

Figure 3.

Fitness under chemotherapeutic selection reveals ecDNA-driven treatment response heterogeneity. A, Schematic representation of model B (ecDNA-dependent cytotoxicity). B, Growth rate, apoptosis rate, and net growth rate of model B, in which cell apoptosis is dependent on ecDNA copy numbers. C, Box plot showing mean ecDNA copy numbers before and after treatment in model B across all colony sizes (150 repetitions). The P value was calculated using the Mann–Whitney U test. *, P ≤ 0.05. D, Paired cell numbers before and after treatment in model B for colonies of different sizes (2–5 cells, 6–9 cells, and ≥15 cells) as defined in Fig. 1 (50 repetitions). E, Schematic overview of the experimental setup to test the theoretical models. ecDNA or HSR cell lines were seeded at low density and allowed to proliferate for 4 days before treatment with doxorubicin for 3 days. Colony size was defined by cell number as described in Fig. 1G and measured for 50 colonies prior to and after exposure to doxorubicin. F, Colony sizes of STA-NB-10 DM cells corresponding to cell numbers from 2 to 5 (left), from 6 to 9 (middle), and of ≥15 (right) in untreated and doxorubicin-treated colonies as described in E. G, Colony size of STA-NB-10 HSR cells corresponding to cell numbers of ≥15 in untreated and doxorubicin-treated colonies as described in E.

In the ecDNA-independent cytotoxicity model, the mean ecDNA copy number remained unchanged after drug treatment (Supplementary Fig. S4C). In contrast, the average number of ecDNA copies significantly decreased after treatment in the ecDNA-dependent cytotoxicity model (Fig. 3C). In paired simulated data from cell populations of small, medium, and large colonies, as defined in Fig. 1, the average ecDNA number remained mostly unchanged upon treatment in the ecDNA-independent cytotoxicity model (Supplementary Fig. S4D). Due to the proliferative advantage conferred by high ecDNA copy numbers, larger colonies tended to grow despite treatment (Supplementary Fig. S4E). In contrast, in the ecDNA-dependent cytotoxicity model, a different trend emerged. For small colonies, both the number of cells and the average ecDNA count remained stable or slightly decreased. For medium and large colonies, there was a significant reduction in both the average ecDNA copies and the overall cell count, particularly in larger colonies (Fig. 3D; Supplementary Fig. S4F). Thus, our computational model suggests that short proliferation and cytotoxicity assays can, in principle, distinguish treatment sensitivity and treatment outcome for different models of ecDNA-carrying cell populations. To test these theoretical predictions experimentally, we examined the effect of genotoxic treatment with doxorubicin, a commonly used anthracycline in neuroblastoma therapy known to primarily affect dividing cells, on colonies that had grown from single cells with MNAs on ecDNA or HSR, as outlined above for the clonogenic growth assay (Fig. 3E). We observed striking differences in treatment sensitivity between colonies derived from HSR and ecDNA cells in near-isogenic neuroblastoma cell line pairs: Although all HSR colonies were exquisitely sensitive to doxorubicin and large ecDNA colonies with high MYCN copy numbers showed a similar reduction in colony size, small ecDNA colonies seemed to be unaffected by treatment and retained or even increased their size (Fig. 3F and G; Supplementary Fig. S4G and S4H). Furthermore, doxorubicin treatment of additional ecDNA and HSR neuroblastoma cell lines demonstrated a significantly lower treatment sensitivity for ecDNA cells, as evidenced by less pronounced caspase-3/7 activation as a measure of apoptosis (Supplementary Fig. S4I). In contrast to HSR cells, which were almost fully eradicated by doxorubicin, ecDNA cells also entered senescence, as indicated by enhanced senescence-associated β-galactosidase activity (Supplementary Fig. S4J). Consequently, only cells with extrachromosomal MNA resumed proliferation upon drug withdrawal after 3 days of treatment (Supplementary Fig. S4K). To test whether the observed differences in treatment sensitivity between ecDNA and HSR cells extend beyond our simulations and experimental data, we interrogated drug sensitivity data from the GDSC repository (Supplementary Table S2). Using the recently released annotation of cancer cell line amplification states (35), we compared the sensitivity of ecDNA cell lines (n = 13) with that of HSR cell lines (n = 11) to a wide spectrum of cancer therapies, demonstrating that ecDNA cell lines, both with and without MYC/N/L-carrying amplicons, were significantly more resistant (Supplementary Fig. S4L; Supplementary Table S3). This difference was even more pronounced in cell lines with MYC/N/L amplifications without TP53 mutations, which are broadly recognized as crucial mediators of treatment resistance (36). In summary, theoretical modeling, published datasets, and experimental data indicate that ecDNA-driven oncogene copy-number heterogeneity contributes to heterogeneous treatment responses among neuroblastoma cells with different MYCN ecDNA levels and thereby facilitates tumor regrowth.

ecDNA Copy-Number Heterogeneity Dynamically Changes under Therapy

To test whether ecDNA copy number in cancer cells, PDX models, and patient samples responds dynamically to cytotoxic drug treatment, we assessed MYCN-containing ecDNA changes in three neuroblastoma cell lines (CHP212, SIMA, and UKF-NB-6) and MYC-containing ecDNA in other cancer lines, including the colon carcinoma cell line COLO320 DM, the mesothelioma line MSTO, and the gastric carcinoma line SNU-16 (Fig. 4A and B). Notably, we observed a rapid decline in both mean levels and overall distribution of MYCN and MYC copy numbers across these cell lines, which was also accompanied by a significant reduction in MYCN protein levels as early as 1 day after treatment, reaching a nadir at day 5 (Supplementary Fig. S5A–S5C). Importantly, we observed reduced MYCN copy numbers in all cell-cycle phases upon doxorubicin treatment, as indicated by cell cycle–resolved MYCN FISH analyses of CHP212 cells expressing a fluorescent ubiquitination–based cell-cycle indicator (Supplementary Fig. S5D), suggesting that the decrease in ecDNA copy numbers was not caused by a cell-cycle arrest. Following this decrease, MYCN and MYC copy numbers increased again as cells resumed proliferation (Supplementary Fig. S5A and S5C). We also measured MYCN copy-number dynamics in response to γ-irradiation and other cytotoxic therapies, such as mafosfamide (an in vitro active analog of ifosfamide that is part of the standard induction chemotherapy in the German high-risk neuroblastoma treatment protocol). Across these treatments, ecDNA-bearing neuroblastoma cell lines consistently exhibited marked reductions in ecDNA copy numbers (Supplementary Fig. S5E and S5F), indicating that ecDNA dynamics is a generalized adaptive response to various genotoxic stresses. In vivo analysis of neuroblastoma PDX models with extrachromosomal MNAs undergoing doxorubicin or ifosfamide treatment also revealed a significant decline in ecDNA copy numbers and a shift in ecDNA distribution (Fig. 4C; Supplementary Fig. S5G). Similar to cell lines, these PDX samples showed both an induction of apoptosis (evidenced by enhanced cleaved caspase-3 staining) and senescence (indicated by induced senescence-associated β-galactosidase activity and increased H3K9me3 levels) following treatment (Supplementary Fig. S5H). Conversely, HSR neuroblastoma cell lines and PDX samples showed no reduction in MYCN copy number after treatment (Supplementary Fig. S5I and S5J).

Figure 4.

Chemotherapy induces dynamic copy-number changes of MYCN and MYC in ecDNA-containing cancers. A, Dot plot analysis and representative pictures of MYCN copy numbers in three neuroblastoma ecDNA cell lines (CHP-212, SIMA, and UKF-NB-6) exposed to doxorubicin for 5 days or left untreated. Mean copy numbers are presented as horizontal lines. P values were calculated using a Mann–Whitney U test. *, P ≤ 0.05. For all pictures, FISH signals are presented in the following color scheme: MYCN (green) and CEP 2 as a reference probe for chromosome 2 (red) with Hoechst as a nuclear counterstain in blue. B, Dot plot analysis and representative pictures of MYC copy numbers in the colorectal carcinoma cell line COLO320 DM, the mesothelioma cell line MSTO, and the gastric adenocarcinoma cell line SNU-16 exposed to doxorubicin for 5 days or left untreated. Mean copy numbers are presented as horizontal lines. P values were calculated using a Mann–Whitney U test. *, P ≤ 0.05. For all pictures, FISH signals are presented in the following color scheme: MYC (green) and CEN 8 as a reference probe for chromosome 8 (red) with Hoechst as a nuclear counterstain in blue. C, Dot plot analysis of MYCN copy numbers in two ecDNA PDX before and at different time points after doxorubicin treatment. P values were calculated using a Mann–Whitney U test. *, P ≤ 0.05. D, Dot plot analysis of MYCN copy numbers from untreated (n = 36) or treated (n = 14) ecDNA neuroblastomas from the German High Risk Neuroblastoma Trial Protocol NB-2004-HR. P values were calculated using a Mann–Whitney U test. *, P ≤ 0.05. E, (Left) Dot plot analysis of MYCN copy numbers from four patient-matched untreated and treated ecDNA neuroblastomas as in D. P values were calculated using a Mann–Whitney U test. *, P ≤ 0.05. (Right) Representative FISH images of matched untreated vs. treated tumor samples from one patient with MYCN signal in green, CEP 2 as a reference probe for chromosome 2 in red, and Hoechst as a nuclear counterstain in blue.

To assess the effect of clinical treatment regimens on ecDNA dynamics, we analyzed tumor samples from a cohort of 56 high-risk patients with MYCN-amplified neuroblastomas with documented treatment history. We compared biopsies from treatment-naïve tumors to tumors after multimodal induction therapy with several treatment cycles, including doxorubicin, cisplatin, ifosfamide, vincristine, and etoposide, as used in the German High-Risk Neuroblastoma Trial Protocol NB-2004-HR (n = 48; ref. 37), and biopsies from neuroblastoma relapses that consecutively underwent combination treatment with irinotecan–temozolomide and dasatinib–rapamycin according to the RIST-rNB-2011 protocol (n = 8) (38). In four paired ecDNA cases with pre- and post-therapy biopsies and a larger cohort comparison of 36 treatment-naïve and 14 posttreatment ecDNA tumors, a significant reduction in MYCN copy number was observed, whereas MYCN copy numbers remained stable in two paired HSR neuroblastomas with pre- and post-therapy biopsies (Fig. 4D and E; Supplementary Fig. S5K).

To investigate whether dynamic changes in copy numbers could also result in resistance to other treatment principles, we incubated ecDNA neuroblastoma cells with a targeted inhibitor against PARP1, which has been shown to be particularly effective in MYCN-amplified neuroblastomas by sensitizing cells to DNA damage response (DDR)–mediated apoptosis (39). Indeed, exposure to the PARP1 inhibitor olaparib also reduced MYCN copy numbers in ecDNA neuroblastoma cells (Supplementary Fig. S5L). In contrast, incubation of ecDNA cell lines with the CDK4 inhibitor palbociclib, which acts by inducing a senescent cell-cycle arrest without provoking DNA damage–driven apoptosis, did not reduce MYCN copy numbers (Supplementary Fig. S5M and S5N). Thus, ecDNA copy-number dynamics facilitates rapid adaptation to genotoxic drugs under both standard-of-care chemotherapy and DDR-targeting therapies.

Alongside MYC(N) oncogenes, the epidermal growth factor receptor (EGFR) geneis frequently amplified on ecDNA in different cancers, particularly glioblastomas and lung adenocarcinomas, and drives tumor cell proliferation (3). Similar to MYCN-amplified neuroblastoma, FISH analysis of 23 glioblastoma samples (21 patients, two cell lines) revealed a higher CV for ecDNA (n = 18) than for HSR (n = 3) samples (Supplementary Fig. S6A and S6B). FISH analysis of two primary glioblastoma cell lines with extrachromosomal EGFR amplifications showed a significant reduction in EGFR copy numbers upon irradiation-induced cell death, highlighting the relevance of ecDNA copy-number dynamics in shaping treatment decisions also in non-MYC(N)–amplified cancers with extrachromosomal oncogene amplifications (Supplementary Fig. S6C and S6D). Although high ecDNA copy numbers impose a vulnerability under genotoxic pressure (40), the capacity for copy-number variation confers a selective advantage to ecDNA-harboring tumor cell populations that allows them to endure and adapt to different treatment conditions.

MYCN ecDNA Copy-Number Heterogeneity Determines Cell Fate Diversity under Treatment

Building on our findings that ecDNA dosage differences drive phenotypic variation and facilitate treatment resistance, we next investigated how these affect cell fate decisions in response to therapy. For this purpose, we examined apoptosis and senescence, two cell fates known to be influenced by the MYC family of transcription factors (41, 42), in cells harboring varying amounts of ecDNA. Cell lines were stained for γH2AX to detect DNA damage, for cleaved PARP1 or propidium iodide (PI) to detect apoptosis, or for H3K9me3 to detect senescence, together with MYCN FISH to determine ecDNA copy numbers before and at several time points after exposure to doxorubicin by ImmunoFISH. In untreated neuroblastoma cells, high MYCN copy numbers positively correlated with elevated levels of γH2AX. This positive correlation persisted after doxorubicin treatment, which further increased DNA damage levels, particularly in cells with high MYCN copy numbers (Fig. 5A). As early as 4 hours after treatment initiation, cells with high MYCN copy numbers also displayed significantly elevated levels of cleaved PARP1, and after 24 hours, they also displayed higher levels of PI as signs of cell death (Fig. 5B; Supplementary Fig. S7A). Cells with lower MYCN copy numbers that survived the treatment predominantly showed positive staining for H3K9me3 (~80% after doxorubicin treatment vs. 5% before doxorubicin treatment) as a marker of senescence 5 days after exposure to doxorubicin, whereas a small remaining fraction of H3K9me3-negative cells displayed markers of cell division, as measured by proliferating cell nuclear antigen (PCNA) and Ki67 staining (Fig. 5C; Supplementary Fig. S7B and S7C). These cells had significantly elevated MYCN copy numbers compared with PCNA- or Ki67-negative cells (Fig. 5C; Supplementary Fig. S7B and S7C). Although MYCN copy numbers did not change in cells that retained their H3K9me3 signal, indicative of a senescent proliferation arrest, MYCN copy numbers gradually increased in Ki67-positive, H3K9me3-negative cells (Fig. 5C; Supplementary Fig. S7B). This suggests that the increase in ecDNA copy number upon tumor regrowth was driven by random segregation of proliferating cells exhibiting copy number–dependent fitness, which was in line with results from a population-level simulation (Supplementary Fig. S7D).

Figure 5.

MYCN copy-number status determines treatment response to cytotoxic therapies. A, Correlation plot and representative images showing the association of MYCN copy numbers with γH2AX intensity as a marker of DNA damage in single cells of the ecDNA cell line CHP-212 by ImmunoFISH. Each dot represents a cell before (red) or 4 hours after (blue) doxorubicin treatment. Correlation for the untreated and treated conditions is assessed using Pearson’s correlation coefficient and its associated P value. B, Correlation plot and representative images showing the association of MYCN copy number and cleaved PARP1 as a marker of apoptosis for single cells as in A. Correlation for the untreated and treated conditions is assessed using Pearson’s correlation coefficient and its associated P value. C, Box plot showing MYCN copy numbers in cells that stain either negative or positive for the senescence marker H3K9me3, as determined by ImmunoFISH, before and 5, 10, or 20 days after doxorubicin treatment. P values were calculated using a Mann–Whitney U test. *, P ≤ 0.05. All comparisons between H3K9me3-positive cells are not significant. D, Gene set and pathway enrichment analysis comparing MYCN-high and MYCN-low cells 5 days after doxorubicin treatment using FISH-guided proteomics. Tested gene sets and pathways were derived from the Hallmark gene set of the human Molecular Signatures Database, WikiPathways, or reference 21 for the MES.signature.genes (van Groningen) and reference 85 for the senescence (Casella_up) signatures. E, Mean senescence score for individual neuroblastoma cells in untreated (n = 32) and chemotherapy-exposed (n = 38) MYCN-amplified tumors. Boxes from first to third quartile; whiskers indicate data at a maximum of 1.5-fold IQR from the box. P value were calculated using Welch’s two-sample t test (two-sided; mean score before chemotherapy −0.065, mean score after chemotherapy −0.035; t = −3.27, d_f = 65.35, P = 0.0017). F, Mean MYCN expression and senescence score for neuroblastoma cells in MYCN-amplified samples (n = 70). Correlation is assessed using Pearson’s correlation coefficient and its associated P value. G, Box plot showing MYCN copy number and the senescence marker H3K9me3, as determined by ImmunoFISH, in cells before treatment, treated with doxorubicin for 5 days or sequentially with doxorubicin for 5 days followed either by vehicle (DMSO) or tranylcypromine (2-PCPA) every 3 days until day 20. P values were calculated using a Mann–Whitney U test. *, P ≤ 0.05. H, Detection of MYCN copy numbers by FISH in untreated vs. doxorubicin-treated CHP-212 cells stably transduced with either BCL2 or an empty vector as control. Boxes from the first to third quartile; whiskers indicate data at a maximum of 1.5-fold IQR from the box. P values were calculated using a Mann–Whitney U test. *, P ≤ 0.05. I, Detection of MYCN copy numbers by FISH in untreated vs. doxorubicin-treated CHP-212 cells stably transduced either with a short hairpin against TP53 or against GFP as a control. Quantification as in H. For all pictures, FISH and immunofluorescence signals are presented in the following color scheme: MYCN (green); CEP 2 as a reference probe for chromosome 2 (red); immunofluorescence signal for γH2AX, cleaved PARP1, H3K9me3, or PCNA (purple); and Hoechst as a nuclear counterstain (blue). Scale bar, 10 μm. Simple linear regression was performed to analyze the relationship between MYCN copy number and γH2AX, cleaved PARP1, or H3K9me3. BH, Benjamini–Hochberg.

The influence of oncogene copy number on treatment response and tumor cell regrowth was also reflected in the different proteome profiles of MYCN-high versus MYCN-low ecDNA cells 5 days after doxorubicin treatment, as determined by FISH-guided proteomics. Although MYCN-low ecDNA cells were characterized by the enrichment of protein modules defining senescence and senescence-associated proinflammatory and metabolic pathways, doxorubicin-exposed MYCN-high ecDNA cells expressed proteins associated with proliferation (e.g. MCM proteins and KIF1C), replication stress, and DNA repair (e.g. ATM and MRE11), in line with the ImmunoFISH staining of MYCN with Ki67 or PCNA (Fig. 5D; Supplementary Table S1). Furthermore, they also upregulated the expression of proteins with senescence-overriding H3K9me3 demethylation activity [e.g. lysine-specific histone demethylase 1A (KDM1A); Supplementary Fig. S7E], which can promote senescence escape and treatment resistance, particularly in MYC-driven cancers (43). Similarly, scG&T-seq showed that KDM1A expression was stronger in cells with a high MYCN copy number, emphasizing the influence of oncogene dosage–driven tumor cell fate decisions under treatment (Supplementary Fig. S7F).

As single-cell transcriptome analyses have identified different tumor cell populations and treatment-induced changes in both tumor biology and the tumor environment in high-risk neuroblastomas (44, 45), we examined whether tumor cell senescence is also correlated with low ecDNA content in MYCN-amplified neuroblastomas in vivo. Therefore, we merged count data for 10 single-cell/single-nucleus RNA sequencing (RNA-seq) datasets (see “Methods” for references) and explored single-cell RNA-seq datasets from 10 MYCN-amplified neuroblastomas, for which matched tumor biopsies prior to and after genotoxic therapy were available. In line with our mechanistic exploration, we observed an enrichment of senescent tumor cells in posttreatment samples in both datasets (Fig. 5E; Supplementary Fig. S7G). Additionally, MYCN expression and senescence were inversely correlated in untreated and chemotherapy-exposed individual tumor samples, supporting the results from FISH-guided proteomics and ImmunoFISH analyses (Fig. 5F).

The results suggest that ecDNA-driven MYCN copy-number heterogeneity directly affects cell fate so that the execution of apoptosis or senescence in response to genotoxic stress and the subsequent regrowth of tumor cells are mechanistically linked to ecDNA copy number. In fact, combination therapies with doxorubicin and the KDM1A inhibitor tranylcypromine (2-PCPA) to prevent H3K9me3 demethylation delayed tumor cell proliferation with cells displaying significantly lower MYCN copy numbers and higher H3K9me3 levels as compared with cells only treated with doxorubicin (Fig. 5G; Supplementary Fig. S7H). Furthermore, we modified ecDNA neuroblastoma cells genetically by retroviral overexpression of either BCL2 or by introducing a short hairpin against TP53 to interfere with doxorubicin-induced cell death. In both cases, neuroblastoma cells with apoptotic defects showed significantly higher MYCN copy numbers after therapy and resumed proliferation more quickly compared with the apoptosis-sensitive cells infected with control vectors (Fig. 5H and I; Supplementary Fig. S7I and S7J). Similarly, a neuroblastoma cell line carrying both a MYCN and an MDM2 amplification on ecDNA, which suppresses a TP53-mediated DDR (46), did not undergo apoptosis in response to therapy and did not show MYCN copy-number alterations upon doxorubicin treatment (Supplementary Fig. S7K). Therefore, ecDNA cancer cells can maintain higher MYCN copy numbers once the pressure to adapt their ecDNA content to genotoxic stress has been alleviated. Additionally, we utilized cell lines without MNA and transduced them with doxycycline-inducible MYCN expression vectors to directly test the effect of MYCN on cell fates under therapy. In line with our observations and previous reports (47–49), MYCN expression was sufficient to increase the sensitivity of cancer cells to genotoxic treatment and led to increased apoptosis (Supplementary Fig. S7L). Furthermore, MYCN nonamplified cell lines with MYCN expression exhibited greater sensitivity to doxorubicin compared with MYCN nonamplified cells without MYCN expression (Supplementary Fig. S7M). In particular, two related neuroblastoma cell lines with different MYC(N) levels that were derived from the same non-MYCN–amplified parental cell line reacted strikingly differently to doxorubicin treatment: Although the clone with high MYC(N) levels was very sensitive to doxorubicin-induced cell death, the clone with low MYC(N) levels displayed a much lower sensitivity to doxorubicin and underwent senescence (Supplementary Fig. S7N and S7O). Within a heterogeneous ecDNA tumor cell population, ecDNA copy number, therefore, shapes cellular treatment responses: Although genotoxic therapies effectively eliminate high MYCN copy-number cells, low MYCN copy-number cells resist their elimination due to senescence reprogramming, from which they emerge with steadily increasing MYCN copy numbers to fuel tumor regrowth.

Targeting Therapy-Resistant Tumor Cells with Low ecDNA Copy Numbers

Although current clinical treatment regimens have been geared toward maximizing cytotoxicity for the treatment of high-risk neuroblastomas, our results identify ecDNA copy-number dynamics as a major contributor to treatment resistance. As cytotoxic therapies fail to eliminate ecDNA cells with low MYCN copy numbers, we sought to identify selective vulnerabilities of this tumor cell population by targeting their senescent phenotype. For this purpose, we performed a small combinatorial inhibitor screen using doxorubicin to induce senescence and five commonly employed senolytic drugs with different pharmacologic properties: the BCL2 and BCL-xL inhibitor navitoclax (ABT263; ref. 50), the tyrosine kinase inhibitor nintedanib (51), the lysosomal V-type ATPase inhibitor bafilomycin A1 (52), the cardiac glycoside digoxin (53), and the flavonoid fisetin (Fig. 6A; Supplementary Fig. S8A; ref. 54). Although all five inhibitors weakly induced apoptosis as single agents when used at low concentrations (≤IC₁₀), they showed a strong senolytic effect on doxorubicin-exposed ecDNA cells. Although the combination therapy of doxorubicin with all senolytic inhibitors showed synergistic effects in ecDNA cells, the treatment did not potentiate cell death in HSR cells (Fig. 6B; Supplementary Fig. S8B and S8C). As MYC directly suppresses BCL-xL expression (55), we wondered whether senescent ecDNA cells with low MYCN copy numbers presented with higher BCL-xL levels, which would make them an attractive target for navitoclax treatment. Indeed, BCL-xL levels inversely correlated with MYCN copy numbers in untreated cells (Fig. 6C) so that doxorubicin-treated ecDNA cells showed a stronger expression of the antiapoptotic protein Bcl2-like protein 1 BCL-xL (Supplementary Fig. S8D). Navitoclax treatment of neuroblastoma cell lines induced cell death only in neuroblastoma cells with low MYCN copy numbers, as evidenced by an increased ecDNA copy number and a negative correlation of cleaved PARP1 levels and MYCN copy numbers by ImmunoFISH analysis upon treatment (Fig. 6D and E). Similarly, cancer cells with extrachromosomal MYC amplification succumbed to combination therapies of doxorubicin and navitoclax (Supplementary Fig. S8E). This indicated that cells with low MYC and MYCN ecDNA were sensitive to senolytic, “one-two punch” therapies (56). Consequently, a treatment strategy using sequential administrations of doxorubicin and navitoclax resulted in reduced tumor growth of a neuroblastoma PDX with extrachromosomal MNA compared with single-agent treatments, which induced the expected ecDNA copy-number changes but could not block tumor progression (Fig. 6F; Supplementary Fig. S8F). Therefore, combinational treatment strategies that target copy number–dependent vulnerabilities could represent a therapeutic principle to counteract the ecDNA-driven adaptation to therapy.

Figure 6.

Targeted elimination of chemotherapy-resistant low MYCN copy-number cells improves treatment outcome. A, Bar plot showing cell viability as assessed by CellTiter-Glo in the ecDNA cell line CHP-212 without treatment, exposed to one of the indicated senolytic drugs (navitoclax, nintedanib, bafilomycin A1, digoxin, or fisetin) at their respective IC₅₀ or IC₁₀ individually for 3 days, to doxorubicin for 5 days, or to the sequential treatment with doxorubicin for 5 days followed by the indicated senolytic drugs at their respective IC₅₀ or IC₁₀. Values are presented as mean ± SD, n = 3 technical replicates. P values were calculated using the Student t test. *, P ≤ 0.05. B, Zero interaction potency (ZIP) synergy score from cell viability measurements with CellTiter-Glo in the near-isogenic neuroblastoma cells STA-NB-10 DM (left) and STA-NB-10 HSR (right) exposed sequentially to doxorubicin and navitoclax as in A, n = 3 technical replicates. C, Correlation plot and representative image showing the association of MYCN copy number and BCL-xL in single cells of the ecDNA cell line CHP-212 by ImmunoFISH. Each dot represents a cell before doxorubicin treatment. Correlation is assessed using Pearson’s correlation coefficient and its associated P value. For imaging, cells were stained with MYCN FISH (green), CEP 2 as a reference for chromosome 2 abundance (red), BCL-xL by IF (purple), and Hoechst as a nuclear counterstain (blue). D, Dot plot indicating MYCN copy numbers in CHP-212 cells left untreated or exposed to navitoclax for 3 days. Mean copy numbers are presented as horizontal lines [x = 23.5 (untreated) and x = 36 (navitoclax)]. *, P ≤ 0.05. E, Correlation plot showing the association of MYCN copy numbers with cleaved PARP1 as a marker of apoptosis for single cells of the ecDNA cell line CHP-212 by ImmunoFISH. Each dot represents a cell 16 hours after navitoclax treatment. Correlation is assessed using Pearson’s correlation coefficient and its associated P value. F, Tumor volumes in neuroblastoma PDX with extrachromosomal MNA treated with doxorubicin, navitoclax, or both drugs in sequential combination. Changes in tumor volume were monitored for 20 days and compared with mice receiving vehicle treatment (n = at least 3 independent mice per treatment group). *, P < 0.05 for all comparisons.

DISCUSSION

Our findings highlight ecDNA as a key driver of intercellular oncogene dosage variation and phenotypic diversity that significantly contributes to treatment resistance in neuroblastomas and other MYC(N)-driven cancers. Unlike the stable copy number and the uniform treatment response observed in HSR tumors, the dynamic copy-number shifts in ecDNA tumors in response to cytotoxic therapies provide a mechanistic explanation for the rapid treatment evasion and frequent relapses of MYCN-amplified neuroblastomas. The higher prevalence of ecDNA amplifications over HSR amplifications in neuroblastomas, as corroborated by our data, may reflect a selective advantage conferred by ecDNA copy number–driven phenotypic heterogeneity. Intratumor oncogene dosage heterogeneity has also been demonstrated for MYC protein levels in pancreatic ductal adenocarcinoma and breast cancer, in which cells with different MYC levels coexist in a functional codependency that significantly improves overall tumor fitness but also creates dosage-dependent therapeutic vulnerabilities (57, 58). Advancing our understanding of ecDNA-driven tumor heterogeneity could thus inform the development of ecDNA-targeted therapies aimed at halting cancer progression or achieving more effective tumor eradication.

We demonstrate that ecDNA enables tumors to adapt more effectively to treatment by increasing the likelihood that tumor cell subpopulations will express the driving oncogene at different levels, thereby maximizing cell survival in different environments. Using theoretical models and mechanistic experimental studies, we illustrate how DNA damage tolerance affects ecDNA dynamics in a tumor cell population, particularly under genotoxic treatment pressure, and leads to the elimination of cells with high copy numbers rather than the reduction of ecDNA molecules. Although ecDNA presents a major source of DNA damage and provides an opportunity for ecDNA-specific therapies, for example, by S-phase checkpoint kinase CHK1 inhibition, replication stress and DNA damage are also directly induced by MYC(N), which sensitizes cells to apoptosis or other forms of cell death independent of cell proliferation (39, 40, 48, 49, 59). As MYCN protein levels correlate with MYCN copy numbers in ecDNA cells, the increased vulnerability of MYCN-high cells to cytotoxic therapies could arise due to either elevated ecDNA copy numbers or increased MYCN protein levels. We demonstrate that MYCN overexpression in non-MYCN–amplified neuroblastoma cells, which do not contain ecDNA, also increases their sensitivity to cytotoxic therapies. Therefore, our results suggest that the observed vulnerability of high copy-number cells to genotoxic therapies and the drug resistance of low copy-number cells are predominantly orchestrated by ecDNA-driven MYC(N) dosage differences and the associated phenotypic plasticity. Similarly, in EGFR-amplified glioblastomas, in which EGFR copy number also correlates with EGFR expression, proliferation, and DNA repair, ecDNA copy numbers decline upon irradiation (60, 61). However, ecDNA copy numbers seem to directly affect DNA damage levels and proliferation in pancreatic ductal adenocarcinoma (58). Therefore, future studies are required to assess the contribution of ecDNA quantity to treatment sensitivity, particularly for ecDNA amplicons with oncogenes other than MYC(N) or EGFR that do not directly induce DNA damage or proliferation or for ecDNA amplicons without oncogenes.

In contrast, tumor cells with low MYC(N) copy numbers persisted after therapy. Although previous reports have described dynamic changes in ecDNA upon cytotoxic and targeted therapies (12, 62, 63), our results for the first time investigate mechanistically how tumor cell populations with heterogeneous ecDNA levels rapidly adapt to different treatment pressures by driving oncogene dosage–dependent cell fate decisions. Following different cell fates upon doxorubicin treatment, we demonstrate that MYC(N)-high cells undergo apoptosis, whereas MYC(N)-low cells, which display increased levels of the antiapoptotic protein BCL-xL, enter therapy-induced senescence, from which cells with enhanced expression of the senescence-overriding H3K9me3-active demethylase KDM1A can readily escape. Although our data suggest that the observed increase in MYCN copy numbers in cells that have resumed proliferation arises from random mitotic segregation of ecDNA, we cannot exclude that cells with defined copy numbers or unique phenotypic profiles could also contribute to the observed escape from senescence. Although the senescent phenotype is organized on many different cellular levels, our data suggest that ecDNA-driven oncogene dosage heterogeneity affects the induction and maintenance of therapy-induced senescence. Consequently, a combinatorial treatment approach using doxorubicin and senolytic drug action, for example, with the BCL2/BCL-xL inhibitor navitoclax, strongly reduced the viability of ecDNA cancer cells in vitro and blocked tumor relapse in a neuroblastoma PDX with extrachromosomal MNA in vivo. As our data suggest that the relapsed tumor cell population quickly restores ecDNA copy-number heterogeneity to pretreatment levels through random mitotic segregation, it will be important to further characterize the oncogene dosage–driven plasticity of relapsing tumors with a specific emphasis on elucidating ecDNA-driven dynamics.

In different cancers, such as breast cancer, prostate cancer, and neuroblastoma, tumor cell persistence upon cytotoxic treatment has been associated with the reversible suppression of MYC and altered tumor cell plasticity primarily through nongenetic mechanisms (26, 27). In contrast, our results support an ecDNA-driven mechanism of treatment resistance and document an oncogene copy number–dependent underpinning for the observed intratumor heterogeneity and phenotypic plasticity. In fact, treatment resistance driven by tumor cell plasticity, which is commonly explained by phenotype switching to drug-tolerant cell states (64), could be strongly influenced by dynamic changes in oncogene copy numbers, highlighting ecDNA dynamics as a previously overlooked genetic mechanism of treatment resistance. Furthermore, our results also suggest that ecDNA-driven tumor remodeling under therapy is not organized by the interaction of tumor cells with defined MYCN levels but rather along a dynamic MYCN dosage continuum.

Overall, we demonstrate that ecDNA-driven MYC(N) copy-number differences regulate phenotypic tumor heterogeneity and shape treatment resistance. Although neuroblastoma therapy has been geared toward maximizing cytotoxicity for the treatment of high-risk neuroblastomas, MYCN-amplified tumors that predominantly harbor MNAs on ecDNA frequently relapse despite their high initial treatment sensitivity. In our work, we identify ecDNA dynamics, which drives copy number–dependent cell fate decisions and subsequently restores oncogene dosage heterogeneity through random mitotic segregation, as a contributor to treatment resistance. Although the presented data demonstrate mechanistically why existing therapies efficiently eliminate a large fraction of the ecDNA-carrying tumor cell population, the results also emphasize that our attention should be directed toward targeting residual tumor cells with low ecDNA copy numbers to improve treatment outcomes for children with high-risk, MYC(N)-amplified cancers.

METHODS

Tumor Sample Collection

Tumor specimens were collected from patients who were treated in the Department of Pediatric Oncology and Hematology or the Department of Neurosurgery at the Charité-Universitätsmedizin Berlin, Germany, between 2014 and 2024 or provided by the National Neuroblastoma Biobank (University Children’s Hospital Cologne). Treatment was performed according to national trial protocols of the German Society of Pediatric Oncology and Hematology. The study was conducted in accordance with the Declaration of Helsinki and Good Clinical Practice, and written informed consent was obtained from all patients or their guardians. Collection and use of patient specimen were approved by the Institutional Review Boards of Charité-Universitätsmedizin Berlin and the Medical Faculty, University of Cologne. Specimen and clinical data were archived and made available by the Charité-Universitätsmedizin Berlin or the neuroblastoma trial registry (University Children’s Hospital Cologne) of the German Society of Pediatric Oncology and Hematology. Anonymized patient identifiers, MNA status, disease status (primary, relapse, or metastasis), and therapy details are provided in Supplementary Table S4. MNA was determined as a routine diagnostic procedure of the trial registry to stratify patients into different risk groups. Samples were taken by open surgical biopsy either at initial diagnosis or during tumor resection after neoadjuvant chemotherapy. Samples were immediately snap-frozen in liquid nitrogen and stored at −80°C or preserved as formalin-fixed, paraffin-embedded tissue in the hospitals’ pathology departments. From each single tumor sample, two to seven biopsies were taken from geographically separate areas of the tumor body. From these biopsies, tumor regions with a high tumor cell content (>70%) were macrodissected from sequential cryosections. A pathologist confirmed the diagnosis and the tumor purity of the macrodissected regions by evaluating parallel hematoxylin and eosin–stained sections.

Cell Culture

Human tumor cell lines were obtained from the DSMZ-German Collection of Microorganisms and Cell Cultures or from ATCC or were a gift from Carol J. Thiele (Center for Cancer Research, National Cancer Institute, Bethesda, USA). UKF-NB-6 cells were provided by the Interdisciplinary Laboratory for Paediatric Tumour and Virus Research (Dr. Petra Joh-Research House, Frankfurt, Germany). The STA-NB-10 DM and STA-NB-10 HSR primary cultures were established at the St. Anna Children’s Cancer Research Institute from a patient tumor biopsy with informed consent under the ethic vote of the Ethics Committee of the Medical University of Vienna (CCRI Biobank: EK#1853/2016; project-specific: EK#1216/2018). Written informed consent was obtained from all patients or their guardians. The study was conducted in accordance with the Declaration of Helsinki and Good Scientific Practice and was approved by an Institutional Review Board. The identity of all cell lines was verified by short tandem repeat genotyping (Genetica DNA Laboratories and/or IDEXX BioResearch), and the absence of Mycoplasma sp. contamination was determined with a Lonza MycoAlert system. The following neuroblastoma cell lines were used: CHP-212 (RRID: CVCL_1125), SIMA (RRID: CVCL_1695), TR-14 (RRID: CVCL_B474), UKF-NB-6 (RRID: CVCL_M474), STA-NB-10 DM (RRID: CVCL_DG49), LAN-5 (RRID: CVCL_0389), KCN-R (RRID: CVCL_7134; ecDNA), NGP (RRID: CVCL_2141), CHP-134 (RRID: CVCL_1124), IMR-5/75 (RRID: CVCL_M473), STA-NB-10 HSR (RRID: CVCL_DG50; HSR), SHEP (RRID: CVCL_0524), SH-SY5Y (RRID: CVCL_0019), NBL-S (RRID: CVCL_2136), SK-N-SH (RRID: CVCL_0531), CLB-GA (RRID: CVCL_9529), and GIMEN (RRID: CVCL_1232). The following MYC-amplified cell lines were used: COLO320 DM (RRID: CVCL_0219), MSTO (RRID: CVCL_1430), SNU-16 (RRID: CVCL_0076; ecDNA), and COLO320 HSR (RRID: CVCL_0220; HSR). All neuroblastoma and MYC-amplified cell lines, with the exception of UKF-NB-6, STA-NB-10 DM, and STA-NB-10 HSR, were cultured in RPMI 1640 medium (Thermo Fisher Scientific) supplemented with penicillin, streptomycin, and 10% FCS. LAN-5 medium was additionally supplemented with 20% FCS. STA-NB-10 DM and STA-NB-10 HSR medium were additionally supplemented with 100 mmol/L sodium pyruvate and 1 mol/L HEPES, as previously described (65). STA-NB-10 HSR was generated by serial passaging for >20 passages from STA-NB-10 DM, and the MYCN status was determined by interphase FISH and high-density SNP array (Affymetrix). UKF-NB-6 was cultured in Iscove’s Modified Dulbecco’s Medium supplemented with penicillin, streptomycin, and 10% FCS. The CHP-212 cell line, stably transduced with the fluorescent ubiquitination–based cell-cycle indicator reporter, was cultured in RPMI 1640 medium supplemented with penicillin, streptomycin, 10% FCS, and 100 μmol/L hygromycin. The primary glioblastoma cell lines E26 and E28 were cultured in DMEM:Ham’s F12 media (Sigma-Aldrich) with supplementary N2 and B27 (Life Technologies), laminin-1 (2 μg/mL; Cultrex), penicillin–streptomycin [1% (v/v); Gibco], and EGF and FGF-2 (10 ng/mL; PeproTech), with E26 plated on prelaminated plates (laminin-1 10 μg/mL in PBS). For all experiments involving isogenic cells, we used clonal cell lines containing either pure ecDNA or HSR populations. The purity of ecDNA and HSR amplification was routinely verified by metaphase FISH. To assess the number of viable cells, cells were trypsinized, resuspended in medium, and sedimented at 300 × g for 5 minutes. Cells were then resuspended in medium, mixed in a 1:1 ratio with 0.02% trypan blue (Thermo Fisher Scientific), and counted with a Bio-Rad TC20 cell counter.

Lentiviral and Retroviral Transduction

Lentiviral and retroviral production and transduction were performed as previously described (66, 67). For lentiviral production, HEK293T (RRID: CVCL_0531) cells were transfected using TransIT-LT1 (Mirus Bio) according to the manufacturer’s protocol, with a 2:1:1 ratio of lentiviral plasmid, psPAX2, and pMD2.G plasmids. Viral supernatant was collected 48 and 72 hours after transfection, pooled, filtered, and stored at −80°C. Cells were then transduced for 1 day in the presence of 8 μg/mL polybrene (Sigma-Aldrich). To produce retroviruses, Phoenix retroviral producer cells (RRID: CVCL_H716) were treated with chloroquine 5 minutes before transfection and subsequently transfected with a DNA/HEPES mixture using Ca₂PO₄. Retroviral supernatant was collected at 48 and 72 hours after transfection, pooled, filtered, and stored at −80°C. The cells were then transduced for 1 day in the presence of 8 μg/mL polybrene (Sigma-Aldrich). Retroviral transduction was performed using a system developed and provided by Dr. Garry Nolan (Stanford University, Stanford, CA). More detailed information about this system is available at www.stanford.edu/group/nolan/.

FISH of Cell Lines

Cells were cultured on coverslips and fixed in MeOH/acetic acid at the end of the experiment. Subsequently, coverslips were washed in PBS for 5 minutes at room temperature and in 0.5× saline sodium citrate (SSC) buffer. Coverslips were dehydrated by washing in 70%, 90%, and 100% cold ethanol for 3 minutes each. Dried coverslips were stained with Vysis LSI N-MYC SpectrumGreen/CEP 2 SpectrumOrange Probes (Abbott), mounted on a glass slide (Thermo Fisher Scientific), and sealed with rubber cement. Denaturing occurred in a ThermoBrite system (Abbott, RRID: SCR_010477) for 5 minutes at 75°C, followed by a 37°C overnight incubation. The coverslips were washed in 2× SSC/0.1% IGEPAL, followed by 3 minutes at 60°C in 0.4× SSC/0.3% IGEPAL (Sigma-Aldrich) and an additional wash in 2× SSC/0.1% IGEPAL for 3 minutes at room temperature. Dried coverslips were stained with Hoechst 33342 (Thermo Fisher Scientific) for 10 minutes and washed with double-distilled H₂O for 5 minutes. After drying, coverslips were mounted on a glass slide and sealed with ProLong Gold Antifade (Thermo Fisher Scientific). Images were taken using a Leica STELLARIS 8 Confocal microscope (Leica Microsystems).

FISH of PDX and Patient Tumor Tissue

FISH was performed on 4-μm sections of formalin-fixed, paraffin-embedded blocks using the ZytoLight FISH-Tissue Implementation Kit (ZytoVision). Slides were deparaffinized, dehydrated, and incubated in pretreatment solution for 15 minutes at 98°C. Samples were treated with pepsin solution for 15 minutes at 37°C. For hybridization, the ZytoLight SPEC MYCN/2q11 Dual Color Probe (ZytoVision) was used. Specimens were denatured at 75°C for 10 minutes, and hybridization took place overnight at 37°C, followed by counterstaining with DAPI. Image analysis was performed with Fiji (ImageJ2, RRID: SCR_002285) software. MNA was defined as an MYCN/2q11.2 ratio >4.0, as described in the International Neuroblastoma Risk Group (INRG) report (68). EGFR amplification was defined as an EGFR/7p11.2 ratio >6.0, in accordance with the 2018 American Society of Clinical Oncology/College of American Pathologists guideline for HER2 copy number (69). Images were acquired using a Leica Stellaris 8 Confocal microscope (Leica Microsystems, RRID: SCR_024662). Quantification of FISH foci was performed using the ImageJ Find Maxima plugin in a supervised fashion, as previously described (12). Briefly, the Find Maxima plugin in Fiji (ImageJ) detects local intensity peaks—pixels that are brighter than their immediate surroundings—allowing precise identification of ecDNA foci. To reduce false positives and improve specificity, we applied a prominence threshold. This threshold sets a minimum required difference in brightness between a peak and its surrounding pixels, ensuring that only strong, well-defined signals are detected, whereas background noise and weak fluctuations are filtered out. To quantify MYCN fluorescence intensity, the CTCF was evaluated using the following formula: CTCF = Integrated density − (Area of selected cell × Mean fluorescence of background readings), as previously described (70, 71).

Immunofluorescence and ImmunoFISH

Coverslips were fixed with 4% paraformaldehyde (PFA) for 10 minutes at room temperature, washed with PBS/glycine for 5 minutes, and then blocked with 3% BSA for 30 minutes at room temperature. The primary antibody was incubated overnight at 4°C, followed by three washes with PBS. The secondary antibody (Alexa Fluor 647 or Alexa Fluor 750) was incubated for 1 hour at room temperature. To preserve the immunofluorescence signal during the FISH protocol, the coverslips were fixed again in 4% PFA for 20 minutes. Subsequently, the coverslips were washed once with 0.7% Triton X-100/0.1 N HCl for 10 minutes at 4°C and once with 2 N HCl for 30 minutes at room temperature and rinsed with PBS/glycine for 5 minutes. This was followed by the FISH protocol as previously described. Protein detection was carried out with antibodies against MYCN (Santa Cruz Biotechnology, B8.4.B, 1:100, RRID: AB_831602), c-MYC (Cell Signaling Technology, D84C12, 1:400, RRID: AB_1903938), BCL2 (Cell Signaling Technology, #2876, 1:300, RRID: AB_10693462), phospho-Histone H2A.X Ser139 (Sigma-Aldrich, clone JBW301, 1:200, RRID: AB_309864), Histone H3-trimethyl K9 (Abcam, ab8898, 1:500, RRID: AB_306848), PARP1 (Santa Cruz Biotechnology, F-2, 1:100, RRID: AB_628105), PCNA (Cell Signaling Technology, D3H8P, 1:100, RRID: AB_2636979), and Ki67 (eBioscience, SolA15, 1:100, RRID: AB_10854564).

Microscopy

Confocal microscopy was performed using a Leica Stellaris 8 microscope (Advanced Light Technology, Max Delbrück Center for Molecular Medicine). All neuroblastoma cell lines and tissue specimens were imaged with a ×63 1.4 oil lens. Specimens used for comparison were imaged under identical microscope settings, and all analyses were conducted on the unprocessed original files.

Clonogenic Assay

Neuroblastoma cell lines were seeded at a density of 1,000 cells and allowed to grow for 4 days until colony formation. To evaluate ecDNA copy numbers in individual colonies, FISH staining was performed as previously described. For assessing the treatment response in colonies, the colony area was measured using Fiji software before and after a 72-hour treatment with 25 ng/mL doxorubicin.

Cell Irradiation

E26 and E28 cells were irradiated with 4 Gy [a clinically relevant dose of IR used in other in vitro glioblastoma studies (72, 73)] using the Faxitron X-ray machine and fixed 24 hours later in methanol/acetic acid (3:1). CHP-212 was irradiated with 5 Gy and fixed 24 hours later in methanol/acetic acid (3:1).

Cellular Proliferation, Cell Death, and Cellular Senescence Assays

To evaluate cellular proliferation during doxorubicin treatment, cell numbers were analyzed using trypan blue dye exclusion, and growth was calculated as the fold change of $d_t/d_0$. Cell death detection was performed using the Caspase-Glo 3/7 Assay System (Promega) according to the manufacturer’s instructions. The Caspase-Glo 3/7 Reagent was added to the culture medium at a 1:1 ratio, mixed for 30 seconds, and incubated for 30 minutes at room temperature. Luminescence of each sample was measured using the Synergy LX plate reader luminometer (BioTek, RRID: SCR_019763). Assessment of SA-β-gal activity at pH 6.0 in cryosections or cells was carried out as previously described (74, 75). Cell proliferation of untreated neuroblastoma cells was assessed using the CellTrace Cell Proliferation Kit (Thermo Fisher Scientific) according to the manufacturer’s instructions. Cells were labeled with CellTrace dye prior to treatment and cultured under experimental conditions. At the indicated time points, cells were harvested and analyzed by flow cytometry using a BD LSRFortessa (BD Biosciences, RRID: SCR_018655). Proliferation profiles and cell generations were quantified using ModFit LT software (Verity Software House, RRID: SCR_016106), which calculates division indices and proliferation parameters based on dye dilution.

IHC

IHC was carried out using the BenchMark ULTRA PLUS (Roche, RRID: SCR_026335). Slides were preheated to 70°C and incubated for 8 minutes, with the mixer activated before processing. The temperature was set to 72°C and then increased for deparaffinization, followed by a 4-minute incubation. EZ Prep solution was used for rinsing, volume adjustments, and coverslip applications throughout the process. Cell Conditioner No. 1 was applied at 95°C with multiple cycles of incubation and coverslip application, ensuring no barcode blow-off. Following this, the slides were rinsed with reaction buffer, and UV inhibitor was applied, with a subsequent 4-minute incubation at 36°C. Primary antibody incubation was performed for 32 minutes, followed by application of the ultra-View Universal DAB Detection Kit (Roche). The slides were counterstained with hematoxylin for 12 minutes, rinsed, and treated with Bluing Reagent for a final 12-minute incubation. The protocol concluded with a final rinse and heating off. Antigen detection was conducted using antibodies against caspase-3 (Zytomed, RBK009-05, 1:400, RRID: AB_2864707) and p16 (Ventana, 805-4713, 1:2, RRID: AB_3675558).

In Vivo Treatments

All animal experiments were approved by local authorities under license E0023/23 and conducted in compliance with national laws and animal welfare regulations. The establishment of neuroblastoma PDX models was performed as previously described in collaboration with Experimental Pharmacology and Oncology GmbH (13). Briefly, freshly isolated tumor fragments from neuroblastoma PDX were serially transplanted subcutaneously to the left flank of 6- to 8-week-old female NOD/SCID gamma-F (NOD.Cg-Prkdcscid Il2rgtm1Sug/JicTac, RRID: IMSR_TAC:HSCFTL-NOG) mice. Animals were housed in individually ventilated cages under sterile and standardized conditions (22°C ± 1°C, 50% relative humidity, 12-hour light/dark cycle) with autoclaved food, bedding material, and tap water provided ad libitum. Tumor growth was monitored by caliper measurements, and tumor volume was calculated using the following formula: (length × width²)/2. For PDX model selection, the MNA status, determined by FISH at the time of diagnosis, was used as a biomarker. For the single-agent study design, randomization of seven experimental groups (n = 1) was initiated when tumor size reached a volume of 0.2 to 0.25 cm³. Treatment with doxorubicin (2 mg/kg, intravenously) and ifosfamide (70 mg/kg, orally) was administered once or repeated weekly for the progressive disease groups. For the combination therapy with doxorubicin and navitoclax, mice were randomized into four groups (n = 3 per group) and treated with doxorubicin (2 mg/kg, intravenously) on day 1, with oral administration of navitoclax (75 mg/kg) on days 4, 5, and 6, with a sequential combination of both agents or with both vehicle formulations in the corresponding control group. Mice were individually euthanized by cervical dislocation upon reaching ethical endpoints for tumor volume or body weight loss.

Protein Blotting

Whole cell extracts for immunoblot analysis were generated by lysing cells in protein lysis buffer (150 mmol/L NaCl, 50 mmol/L Tris-HCl, 1% NP-40, pH 8.0), including proteinase inhibitors (cOmplete, Mini, EDTA-free Protease Inhibitor Cocktail, Sigma-Aldrich) and phosphatase inhibitors (PhosSTOP, Sigma-Aldrich). Equal aliquots corresponding to 30 μg of protein were resolved on the Novex Tris-Glycine gels and transferred to PVDF Transfer Membranes (Thermo Fisher Scientific). Antigen detection was carried out with antibodies against MYCN (Santa Cruz Biotechnology, B8.4.B, 1:500), p53 (Santa Cruz Biotechnology, DO-1, 1:500), BCL-xL (Cell Signaling Technology, (54H6) #2764, 1:500, RRID: AB_2228008), and vinculin (Proteintech, 26520-1-AP, 1:1,000, RRID: AB_2868558). Corresponding anti-mouse (RRID: AB_2534069) and anti-rabbit (RRID: AB_2556544) peroxidase-conjugated secondary antibodies (Thermo Fisher Scientific, 1:2,500 and 1:5,000, respectively) were used.

Single-Cell RNA-Seq Analysis

We merged count data for 10 single-cell/single-nucleus RNA-seq datasets (bioRxiv 2024.01.07.574538; refs. 18, 34, 76–82). We filtered for at least 400 unique molecular identifiers per barcode, at most 20% mitochondrial reads, and at least 200 barcodes per sample. We corrected for ambient RNA contamination using decontX (83) and discarded doublets using doubletFinder (84). Genes with no expression in any individual dataset were discarded. We used Seurat version 5 (85) for normalization, sketching, Harmony integration (using dataset, experimental platform, and single cell vs. single nucleus to define batches), and clustering. Neuroblastoma cell clusters were identified based on the cluster-level overexpression of marker genes PHOX2B and HAND2. For the comparison of matched neuroblastoma samples before and after genotoxic therapy, preprocessed count data were downloaded from the Human Tumor Atlas Network portal (https://humantumoratlas.org). Cells with less than 1,000 detected genes were excluded from this dataset, and we focused only on annotated tumor cells from MYCN-amplified samples. MYCN expression was calculated as the mean log1p-normalized count data across all neuroblastoma cells in each sample. We used Seurat’s AddModuleScore (RRID: SCR_016341) function to calculate gene signatures for neuroblastoma cells, specifically using the Casella_up (86) gene set to compute a senescence score. Cells with a positive senescence score were considered senescent.

Modeling

We developed an agent-based stochastic computational model using Python (version 3.11.8, RRID: SCR_008394) to simulate the expansion of small cancer cell populations and their response to drug treatment. The model starts by randomly selecting individual cancer cells, each containing varying numbers of ecDNA, distributed according to a uniform distribution. After a short period of clonal expansion, a population of clones derived from single cells is generated. We then tracked both the number of cells within each clone and the distribution of ecDNA they carry. The number of ecDNA copies $k$ within a cell provides a selective advantage $s$, which follows a sigmoidal function (for more details, see the main text). When the number of ecDNA copies exceeds a threshold, the selective advantage reaches its maximum $s_{max}$ and does not increase further. During each cell division, the parent cell’s ecDNA copy number doubles and is randomly segregated into the two daughter cells (12). As a control, we assume in an alternative scenario that the number of ecDNA copies does not confer any additional selective advantage and remains constant across all ecDNA-positive cells. Thus, the clonal growth reaction can be expressed as follows:

$$T_k\stackrel{s}{\to }{T}_{k_1}+T_{k_2}$$

For simulations under treatment, we assumed a baseline toxicity $d_0$, which affects all cells equally, regardless of their ecDNA content. Cells containing ecDNA, however, experience additional toxicity. During cell division, each ecDNA copy in a cell has a probability $p$ of being targeted by the drug. If the total number of targeted ecDNA copies exceeds $k_d$, the cell undergoes apoptosis. These reactions could be represented by the following equations:

$$T_k\stackrel{s\cdot \mathrm{Pr}\left(X<k_d\right)}{\to }{T}_{k_1}+T_{k_2}$$

$$T_k\stackrel{d_0+s\cdot \mathrm{Pr}\left(X\ge k_d\right)}{\to }\varnothing$$

As a control, we again considered an alternative scenario in which the number of ecDNA copies does not influence drug sensitivity, with cytotoxicity being constant across all cells, defined by $d_0$, the baseline toxicity. The time for clonal growth is defined as $t_1$, and the duration of drug treatment is set as $t_2=\frac{3}{4}{t}_1$, corresponding to the time ratio used in experimental conditions. All parameters used in the simulations are given in the corresponding figures. We implemented the stochastic simulation using the Gillespie algorithm (87, 88). The code for the model and plotting scripts is available on GitHub: https://github.com/AprilTu/ecDNA-model-for-neuroblastoma. For the tumor regrowth after treatment model, we performed simulations on a large population. Starting from a single ecDNA-positive cell carrying k_0 copies of ecDNA, the population was allowed to expand according to the previously defined fitness equation until it reached a size of 10⁵ cells. We then simulated a period of drug treatment using the same method described above. After the treatment phase, the growth dynamics were restored to the original ecDNA-dependent fitness model. We tracked the ecDNA distribution at three time points: prior to treatment, immediately after treatment, and following the subsequent regrowth phase.

FISH-Guided Proteomics

For FISH-guided proteomics, a modified Deep Visual Proteomics (89) approach was implemented. Briefly, CHP212 cells were seeded on polyphenylene sulfide (PPS) membrane slides (Leica) and fixed with 4% PFA for 10 minutes at room temperature. A confocal microscope (Operetta, PerkinElmer, RRID: SCR_018809) was used for whole-slide imaging. Images were acquired under 40× magnification with z-stacks and imported into the BIAS software (Single-Cell Technologies Ltd.) for downstream nucleus and cell segmentation, MYCN count recognition, and statistical analysis. The contours of cells with the highest and lowest MYCN counts were exported from BIAS for laser microdissection. An LMD7 system (Leica) was used to isolate single-cell contours from the membrane slides (approximately 200 contours per replicate). One microliter of lysis buffer (0.2% n-dodecyl-β-D-maltoside in 100 mmol/L triethylammonium bicarbonate) was added to the samples using a Mantis liquid dispenser (Formulatrix) and incubated at 75°C for 30 minutes. To digest the proteins into peptides, 1 μL of enzyme mix (trypsin/Lys-C, 2 ng) was added, and the samples were incubated at 37°C overnight. The digested peptides were loaded onto Evotips (Evosep) for mass spectrometric measurement. LC/MS analysis was performed on an Evosep One system using the 30 samples per day method coupled with a timsTOF SCP mass spectrometer (Bruker, RRID: SCR_025639). All mass spectrometry measurements were performed using the diaPASEF (90) method (factory default). We used DIA-NN (version 1.8.1, RRID: SCR_022865; ref. 91) and an in silico spectral library for raw data analysis. Statistical analysis was conducted with Perseus (version 1.6.15.0, RRID: SCR_015753; ref. 92) and R (version 4.4.1, RRID: SCR_001905).

Data were filtered to keep the proteins valid in more than 50% of samples. A column-based median subtraction was used for normalization. Missing values were imputed based on a normal distribution (width = 0.3, downshift = 1.8). A row-based z-score was implemented within each treatment group for comparison across groups. One-dimensional pathway enrichment analysis (93) was performed based on the Kyoto Encyclopedia of Genes and Genomes, Hallmark, Reactome Pathway Database, and customized pathway lists generated from the article (significance cutoff value was Benjamini–Hochberg FDR <0.05).

Synergy Analysis

The synergy between doxorubicin and senolytic drugs was evaluated using the SynergyFinder web application (https://synergyfinder.fimm.fi/, RRID: SCR_019318; ref. 94). Drug response data were analyzed using the zero interaction potency model, which quantifies drug interactions by comparing the observed combination response to the expected outcome under the assumption of no interaction (95).

Drug Sensitivity Analysis

Fitted dose–response values were downloaded from the Genomics of Drug Sensitivity in Cancer website, https://www.cancerrxgene.org/, release 8.4 (96–101). For our analysis, we used the reported natural logarithms of the inferred IC₅₀ values for both GDSC1 and GDSC2 and stratified cell lines into TP53 wild type/knockout and MYC family/non-MYC family amplified using data from the Dependency Map (100) and the Cancer Cell Line Encyclopedia (101). We performed a two-sided Welch’s t test to test for differences in mean response between ecDNA and HSR cell lines as implemented in R’s t. test function in R 4.2.3.

scG&T-seq

We performed G&T-seq to enable parallel sequencing of the transcriptome and genome from individual cells, following an optimized and miniaturized version of the protocol described by Chamorro and colleagues (19). All samples were processed using a Biomek FXP Laboratory Automation Workstation (Beckman Coulter).

Single-Cell Suspension from PDX and Sorting

PDXs were dissociated into single cells using the Papain Dissociation System (Worthington). Tissues were placed in the papain solution and incubated at 37°C with constant agitation. The resulting cell suspension was transferred to sterile screw-capped tubes and centrifuged at 300 × g for 5 minutes at room temperature. The cell pellet was resuspended in DNase-diluted albumin-inhibitor solution, layered on top of a discontinuous density gradient, and centrifuged at 70 × g for 6 minutes at room temperature. The pelleted cells were then resuspended in 1× PBS for single-cell sorting. Cells were stained with PI (Thermo Fisher Scientific), and viable cells were selected based on forward- and side-scattering properties and PI staining. Viable cells were sorted using a FACSAria Fusion Flow Cytometer (BD Biosciences) into 2.5 μL of RLT Plus buffer (QIAGEN) in low-binding 96-well plates (4titude), sealed with foil (4titude), and stored at −80°C until processing.

Bead Preparation and RNA Capture

Dynabeads MyOne Streptavidin C1 were washed sequentially with buffers A and B, followed by binding with 100 μmol/L biotinylated oligo-dT30VN (IDT) for 20 minutes at room temperature under rotation. Beads were washed and resuspended in resuspension buffer supplemented with SUPERase•In RNase Inhibitor (Thermo Fisher Scientific). A volume of 10 μL of bead suspension was added to each lysed cell sample to capture polyadenylated mRNA. Following a 20-minute hybridization, the supernatant containing genomic deoxyribonucleic acid (gDNA) was collected for downstream processing.

Genomic DNA Amplification

The collected genomic deoxyribonucleic acid (gDNA) was purified using AMPure XP beads (Beckman Coulter) and subsequently amplified using the REPLI-g Advanced Single Cell Kit (QIAGEN) via multiple displacement amplification. The reaction was carried out at 30°C for 2 hours, followed by a 3-minute heat inactivation step at 65°C. Purified gDNA was quantified using the Qubit dsDNA HS Assay (Invitrogen) and evaluated for fragment size distribution using the Agilent TapeStation Genomic DNA ScreenTape.

mRNA Reverse Transcription and cDNA Amplification

Bead-bound mRNA underwent reverse transcription using a modified Smart-seq2 protocol. Reverse transcription was performed with SuperScript II reverse transcriptase (Invitrogen) in the presence of a template-switching oligo and SUPERase·In. The reaction was cycled at 42°C and 50°C to enhance yield and cDNA integrity. Full-length cDNA was amplified using KAPA HiFi HotStart ReadyMix (Roche) with in situ (IS)-PCR primers. PCR conditions included an initial denaturation at 98°C for 3 minutes, followed by 20 to 22 amplification cycles (optimized per cell type), and a final elongation step at 72°C for 5 minutes.

Library Preparation and Sequencing

Amplified cDNA and gDNA were fragmented and processed using the NEBNext Ultra II DNA Library Prep Kit for Illumina (New England Biolabs) with quarter-reaction volumes. Fragmentation and end-repair reactions were followed by adaptor ligation with diluted NEBNext dual-index adaptors. After ligation, libraries underwent uracil-specific excision reagent (USER) enzyme treatment, cleanup using AMPure XP beads, and PCR enrichment using NEB Q5 High-Fidelity DNA Polymerase. Final libraries were assessed for concentration (Qubit) and size distribution (TapeStation HS D1000). Equimolar pooling was performed based on Qubit and fragment length data, and libraries were diluted to 10 nmol/L for sequencing on an Illumina platform.

Quality Control

For both gDNA and cDNA, quality and yield were assessed using the Qubit dsDNA HS assay. Fragment lengths were analyzed using the Agilent 4200 TapeStation system. Samples falling outside the expected range (0.01–1 ng/μL) were excluded. Negative controls were included at each step to ensure the absence of contamination or reagent-derived artifacts.

Single-Cell Sequencing Data Processing

scG&T-seq data were generated and processed as previously described (18). For the PDX dataset, FASTQ files were trimmed using TrimGalore version 0.6.10 (RRID: SCR_011847). The proportion of human and mouse reads was assessed with BBMap version 39.08 (RRID: SCR_016965) by mapping the reads to both the human (hg38) and mouse (mm39) reference genomes to identify and exclude mouse-contaminated cells. Transcriptome reads for both PDX and TR14 datasets were aligned to the human hg38 reference genome using STAR version 2.7.11b (RRID: SCR_004463) with default parameters. Gene expression counts were generated with HTSeq-count version 2.0.5 (–mode = union) and annotated using GENCODE version 44. For the PDX dataset, genome reads were aligned to hg38 using BWA-MEM version 0.7.18 (RRID: SCR_010910). Duplicate reads in both transcriptome and genome BAM files were marked using bammarkduplicates from biobambam2 version 2.0.185. Quality control and RNA normalization were performed using the Seurat package version 5.1.0. For PDX, cells were excluded based on the following criteria: more than 20% of mouse reads, more than 10% mitochondrial gene content, or fewer than 100 detected genes. Copy number profiles were generated for G&T-seq PDX DNA data using AneuFinder with a bin size of 200 kB. For both datasets, single cells were stratified by MYCN expression: the top 25% of cells were labeled “MYCN-high,” and the bottom 25% were labeled “MYCN-low.”

Signature Scoring Analysis

Gene signatures were defined based on external gene lists. Signature scores were calculated using the AddModuleScore function in Seurat version 5.1.0 (R version 4.4.2), generating a module score for each cell reflecting the expression level of the defined signatures. Only MYCN-high and MYCN-low cells were used in downstream comparative analyses.

Statistical Analysis and Visualization

Statistical comparisons of signature scores between MYCN-high and MYCN-low cells were conducted using nonparametric Wilcoxon rank-sum tests. Resulting P values were corrected for multiple hypothesis testing using the Benjamini–Hochberg FDR method. Data visualization was generated with the ggstatsplot R package. Pairwise comparisons were displayed directly on the plots, indicating significance levels after FDR adjustment.

Single-Cell Multiome ATAC + Gene Expression Sequencing

The sonic hedgehog (SHH) medulloblastoma tumor 7316-178 was received through the Childhood Brain Tumor Network. From the patient tumor, dissociated cryopreserved cells stored in 10% DMSO/FBS were used. At least 50 mg of tissue (1 mol/L cells) was used for both samples. Dissociated cells were prepared for Single Cell Multiome ATAC + Gene Expression sequencing (10x Genomics) according to the manufacturer’s instructions. Sequencing was performed on an Illumina NovaSeq S4 200 to a depth of at least 250 mol/L reads for single-nucleus ATAC sequencing and 200 mol/L reads for single-nucleus RNA-seq.

Single-Cell Data Processing and Clustering

The SHH medulloblastoma tumor 7316-178 sequencing data were processed using Cell Ranger ARC version 2.0.0 (RRID: SCR_023897) with default parameters, followed by Seurat version 4.0.47. Cell barcodes that passed the following quality thresholds were retained: RNA mitochondrial fraction less than 0.2; ATAC read count between 1,000 and 25,000; and RNA read count between 1,000 and 20,000. Doublets were identified and removed using Doublet-Finder version 2.0 (RRID: SCR_018771) using default parameters. Single-cell transcription data were normalized using regularized negative binomial regression, implemented in the sctransform package (102) included with Seurat. Clustering was performed using the weighted nearest neighbors algorithm (85) with a resolution of 0.5 and the other parameters set to default. To label cell clusters with cell type identities, the expression patterns of Seurat-generated clusters were cross-referenced against known cell type marker genes (103).

Identification of ecDNA-Containing Cells

ecDNA-containing cells were identified as previously described (11). Permutation tests were used to compare single-nucleus ATAC sequencing read coverage at the ecDNA regions with read coverage of random regions elsewhere in the genome for each cell. Empirical P values were estimated, and z-scores were calculated using the standard formula, comparing the average read coverage at the ecDNA-amplified region to the mean and variance of the permutations.

Classification of ecDNA Subpopulations and Gene Module Analysis

Tumor cells were first divided from detected normal cell types. Cells with nonsignificant empirical P values were labeled as ecDNA tumor cells. All the other ecDNA⁺ cells were ranked by z-scores, and the top 20% were designated as ecDNA high cells, whereas the bottom 20% were labeled as ecDNA low cells. The remaining cells are ecDNA middle cells.

Differential gene expression analysis between ecDNA high and low populations was performed using Seurat, with a log fold change threshold set to 0 and a minimum expression fraction of 0.05. Selected gene sets were scored using the AddModuleScore function in Seurat and visualized with violin plots. Statistical comparisons were performed using the R package ggpubr version 0.6.0 (RRID: SCR_021139), and inset box plots were generated using ggplot2 version 3.5.1 (RRID: SCR_014601; ref. 104).

Gene Set Enrichment Analysis

Gene set enrichment analysis (GSEA) was performed using the fast GSEA (FGSEA) algorithm implemented in the R package fgsea version 1.32.2 (Korotkevich, G. Sukhov V. Sergushichev, A. Fast gene set enrichment analysis, bioRxiv preprint, 2019, RRID: SCR_020938). Ranked gene lists were generated based on the log₂ fold change from differential expression analysis between ecDNA-high and ecDNA-low tumor cell populations. Enrichment was assessed using selected customized gene sets, with default parameters.

Statistics

If not stated otherwise, data are presented as arithmetic means ± SD or SEM, and statistical analyses of the generally normally distributed (as assessed by the Shapiro–Wilk test) and variance homogeneous (as assessed by the Levene test) data were based on paired or unpaired t tests. The CV for each sample was evaluated to assess the relative variability within the datasets. For each sample, the mean (μ) and SD (σ) of MYCN copy number were calculated, and the CV for each sample was computed using the following formula: $CV=\left(\frac{\sigma }{\mu }\right)\times 100$. All quantifications from staining reactions were carried out by an independent and blinded second examiner and reflect at least three samples with at least 200 events counted (typically in three different areas) each. Unless otherwise stated, a P value < 0.05 was considered statistically significant.

Data Sources

Jansky and colleagues (76), EGAS00001004388

Dong and colleagues (77), GSE137804

Verhoeven and colleagues (78), GSE147766

Kildisiute and colleagues (79), https://www.neuroblastomacellatlas.org/

Yuan and colleagues (80), GSE192906

Patel and colleagues, humantumoratlas.org

Wienke and colleagues (81), GSE218003

Nian and colleagues (82), GSE223374

Stöber and colleagues (18), EGAS50000000509

Grossmann and colleagues (34), humantumoratlas.org

Walentynowicz and colleagues (105), phs003100.v1.p1

Supplementary Material

Supplementary data for this article are available at Cancer Discovery Online (http://cancerdiscovery.aacrjournals.org/).

SIGNIFICANCE:

ecDNA-driven tumor genome evolution provides a major challenge to curative cancer therapies. We demonstrate that ecDNA copy-number dynamics drives treatment resistance by promoting oncogene dosage–dependent phenotypic heterogeneity in MYCN-amplified cancers. Exploiting phenotype-specific vulnerabilities of ecDNA cells, therefore, presents a powerful strategy to overcome treatment resistance.

Data Availability

Neuroblastoma single-cell data are available from Gene Expression Omnibus (GSE137804, GSE147766, GSE192906, and GSE218003), EGA (EGAS00001004388, EGAS50000000509, and EGAS50000001040), neuroblastomacellatlas.org, and humantumoratlas.org. Single-cell data from GSE223374 are available from the corresponding author of the study upon request. The mass spectrometry proteomics data have been deposited in the ProteomeXchange Consortium via the PRIDE (106) partner repository with the dataset identifier PXD064049.

Scripts used to analyze sequencing data have been uploaded to www.github.com/henssenlab.