Superenhancers shape the landscape and repair dynamics of transcription-associated DNA breaks in cancer

Abstract

Cancer is characterized by uncontrolled proliferation accompanied by oncogene hypertranscription, leading to transcription stress, a key source of DNA double-strand breaks (DSBs) that jeopardize genomic stability. Despite its importance, the landscape and consequences of transcription stress remain underexplored. Here, we used maps of DSBs identified through sBLISS (in-suspension break labeling in situ and sequencing) with transcription stress markers to delineate the transcription stress landscape in cancer. We found that transcription stress sites are shaped by the superenhancer regulatory landscape. Notably, γH2AX is enriched at transcription stress sites; however, not all DSB-enriched genes show similar γH2AX marking. Instead, genes with DSBs tied to transcription stress are distinctly marked. Genes with high DSBs marked by γH2AX exhibited substantially higher DSB turnover and repair than those with low γH2AX, and are associated with vulnerability to mutagenesis. These findings underscore superenhancer activity as a determinant of the transcription stress landscape in cancer, posing a threat to the genomic stability of oncogenes.

Superenhancers create hotspots of DNA fragility and mutations by intensifying transcription-associated breakage and repair.

INTRODUCTION

The most fundamental hallmark of cancer cells is their ability to sustain proliferation, which they acquire through genetic alterations during neoplastic transformation (1). Simultaneously, dysregulation of the transcriptional program emerges to accommodate the cell’s increased demand for cellular factors essential for proliferation, growth, and survival (2, 3). Consequently, genes encoding these factors, i.e., oncogenes, are hypertranscribed to fulfill the needs of cancer cells (4).

A well-established mechanism by which cancer cells up-regulate oncogenes involves the formation of oncogenic superenhancers (SEs) (5–12). In normal cells, SEs drive the expression of cell type–specific genes that encode master transcription factors, which shape cellular identity and the transcriptional program (2). In cancer cells, SEs drive the expression of cancer type–specific oncogenes and oncogenic master transcription factors that dictate their transcriptional programs and support their proliferation (2, 12). SEs can become dysregulated in cancer through various mechanisms: (i) Overexpression of enhancer transcription factors renders SEs much more effective at driving the expression of their associated genes, (ii) translocations can relocate SEs near an inactive oncogene in the corresponding normal cell, and (iii) focal amplifications of SEs can lead to the increased expression of target genes (12). These mechanisms contribute to the activation of proto-oncogenes by SEs, which drive the transcription of these genes to an exceptionally high level termed “hypertranscription.”

Transcription is increasingly viewed as a source of DNA damage (13–18). The most damaging type of DNA lesion produced during transcription is DNA double-strand breaks (DSBs) (19). These DSBs often result indirectly from the processing of transcription-associated intermediates, such as single-strand breaks that form during the repair of trapped topoisomerase 1 (TOP1) cleavage complexes (TOP1cc). DSBs resulting from transcription are primarily facilitated by the formation of pathogenic cotranscriptional R-loops and the aberrant activity of TOP1 (13, 20). R-loops tend to form more frequently at highly expressed genes, and their formation causes TOP1 to become trapped in its cleavage complex form (TOP1cc) (13).

Although transcription is a highly regulated process, it can occasionally become abortive and result in DSBs. These transcriptional aberrations occur more frequently as transcription levels increase in a phenomenon known as “transcription stress” (21). Characterizing the “hotspots” of transcriptional stress is crucial because transcription-related DSBs can lead to cell death, genomic instability, and mutations (22, 23). Moreover, given that cancer cells up-regulate genes vital for their proliferation and survival, a deeper understanding of transcriptional stress and transcription-associated DSBs associated with these oncogenes can aid in exploiting their vulnerabilities to target cancer cells more effectively. Nevertheless, while transcription stress has been increasingly studied (24–26), its landscape and consequences, particularly in cancer, remain incompletely characterized.

DSBs are impediments to transcription (27, 28) and must, therefore, be efficiently repaired to sustain continuous high transcriptional levels. However, it is not understood how regions of transcription stress are equipped to handle these deleterious DSBs. In this study, we characterize the landscape of transcription stress in cancer cells and reveal how these regions are structured to promote the rapid and efficient repair of transcriptional DSBs that sustain hypertranscription.

RESULTS

SE-regulated oncogenes are major transcription stress sites

Pathogenic R-loops preferentially form at highly expressed genes (29), facilitating aberrant TOP1 activity and TOP1cc trapping (13). The high TOP1 occupancy, TOP1cc trapping, and cotranscriptional R-loops all contribute to the formation of transcription-associated DSBs (20, 30). Therefore, we reasoned that these features, along with genome-wide maps of endogenous DSBs (hereafter referred to as “breakome”), could serve as markers of transcriptional stress. To determine transcription stress sites in the genome, we analyzed in-suspension break labeling in situ and sequencing (sBLISS) and chromatin immunoprecipitation sequencing (ChIP-seq) data for the breast cancer cell line MCF7 and applied a hidden Markov model (HMM)–based approach to identify regions in the genome with co-occurrence across the four datasets [sBLISS, DNA-RNA immunoprecipitation sequencing (DRIP-seq), TOP1cc, and TOP1 ChIP-seq]. As expected, the majority of these regions were genic (fig. S1A), while the rest were enriched mainly at intergenic SEs (fig. S1, B and C). To investigate which genes serve as transcription stress sites, we scored each gene on the basis of TOP1, TOP1cc, R-loop, and DSB scores (see Materials and Methods). This scoring yielded a shortlist of genes, many of which have well-established oncogenic functions (Fig. 1A). Upon closer analysis of this list, we noted that many of the genes were regulated by SEs in MCF7 cells, including cyclin D1 (CCND1), nuclear paraspeckle assembly transcript 1 (NEAT1), microRNA 21 (miR21), and MYC [SE-regulated genes were downloaded from SEdb 2.0 (31); see Materials and Methods] (Fig. 1A). A similar pattern was observed in the chronic myelogenous leukemia cell line K562 using publicly available sBLISS, TOP1cc CAD-seq (covalent adduct detection sequencing), and DRIP-seq datasets, supporting the broader relevance of these findings (fig. S1, D and E).

Fig. 1. SE-regulated oncogenes are major transcription stress sites.

(A) Transcription stress sites ranked by their scores on the basis of four transcription stress parameters: TOP1, R-loops, TOP1cc, and DSBs in MCF7 cells (see Materials and Methods). Genes with known oncogenic function in OncoKB are in bold and those from other resources in bold with an asterisk (see table S1). (B) Ratio of observed to expected numbers of genes in transcription stress sites. P values were calculated using hypergeometric testing (upper-tailed for enrichment in the top score and lower-tailed for the bottom score). (C) Genome browser snapshots depict DSBs, TOP1cc, and R-loops at selected transcription stress sites. (D) A box plot illustrates DSBs enrichment at the SE-regulated genes over other genes in each cell line [downloaded from SEdb^2.0 (31)]. Enrichment was calculated by dividing the break density of each SE-regulated gene by the median break density of non–SE-regulated genes, and the average values for replicates were plotted. In the box plot, the center line represents the median, the box limits indicate the upper and lower quartiles, and the whiskers denote 1.5× the interquartile range, while the dots represent outliers. P values were computed using the Mann-Whitney U test; ****P < 0.0001.

To statistically validate this observation, we systematically assessed the representation of multiple gene sets among transcription stress sites by performing gene set enrichment analysis (Fig. 1B; see Materials and Methods). SE-regulated genes were the most significantly enriched category, with SE-regulated oncogenes exhibiting the highest enrichment (~122-fold over expectation under a random model, P = 1.85 × 10⁻¹⁵; hypergeometric test). Similar results were obtained when quantifying the overlap between each gene set and the comarked regions identified by the HMM (fig. S1, D and E). Notably, the enrichment of SE-regulated genes was substantially higher than that of typical enhancer (TE)–regulated genes or highly expressed genes lacking enhancer regulation (Fig. 1B) despite having similar expression levels (extended data; Fig. 1F). To control for factors that might confound this analysis, we performed multivariate logistic regression using SE status, expression level, GC content, and gene length as predictors for transcription stress at genes. The results of this model demonstrate that SE status was the most statistically significant independent predictor of transcription stress (fig. S1G), indicating specificity to genes regulated by SEs.

DSBs at these genes coincide with TOP1cc and R-loops (Fig. 1C). Analysis of sBLISS data for MCF7 cells depleted for TOP1 and R-loops shows that most of these genes have notably fewer breaks upon TOP1 and R-loop depletion (fig. S3A). In addition, DSBs are enriched at SE-regulated genes in both cycling and G₁-synchronized cells (fig. S3, C to F), validating their transcriptional origin and independence from replication.

The relationship between transcription and DSBs was previously established (13–15). However, our results suggest that this association may be more assertive in SE-regulated genes. This prompted us to further investigate the discrepancies in the relationship between transcription and DSBs in SE-regulated versus non–SE-regulated genes. The strong correlation between gene expression and DSBs is observed only in SE-regulated genes, not in others (fig. S3, G to I). Moreover, highly expressed genes enriched in DSBs showed significantly higher enrichment in TOP1, TOP1cc, and R-loops than similarly expressed genes lacking DSBs (fig. S4, A to E). In addition, SE-regulated genes were significantly more prevalent in highly expressed genes enriched in DSBs (fig. S4F). These findings indicate that the association between transcription and DSBs is primarily attributed to SE-regulated genes.

Given that SE activity is more potent in cancer cells than in normal cells, we next aimed to test whether the enrichment of DSBs at SE-regulated genes represents a universally conserved feature of cancer cells. To assess this, we mapped the breakome in 10 biologically diverse cell systems, including normal epithelial and neural stem cells, a premalignant line (MCF10A), and multiple cancer cell lines originating from distinct tissues and oncogenic programs (breast, kidney, and neural). This panel also includes hormone-responsive (ER⁺) and non–hormone-responsive (ER⁻) breast cancer models and both adult and pediatric tumor cells, enabling us to test whether SE-associated DSB accumulation depends on specific signaling contexts or is a general feature of malignant SE biology. The data showed that SE-regulated genes are more enriched in DSBs in all tested cancer cell types than in their normal counterparts (Fig. 1D), demonstrating the generality of this phenomenon and the specificity of transcriptional stress to active SE-regulated oncogenes in cancer cells.

The SE landscape shapes transcription stress sites in cancer

The landscape of SEs is cell type–specific, enhancing the transcription of different genes on the basis of cell identity (12). If SEs cause transcription stress in their target genes, one would predict that various cancer cell types exhibit distinct transcription stress sites reflective of their SE landscapes. To test this, we mapped the breakome of multiple cancer cell lines. For each cell line, SE-regulated genes were identified, common SE-regulated genes from other cell lines were excluded, and low to moderately expressed genes, as well as genes with low break counts, were filtered out. SE-regulated genes were enriched in DSBs, predominantly in the cell type where they are regulated by an SE (Fig. 2A). This pattern was also apparent when comparing cell types with very similar characteristics (MCF7 and T47D, both of which are ER⁺ luminal A breast cancer cell types) yet exhibiting differing SE landscapes, suggesting that transcription stress sites correspond to the SE landscape in cancer.

Fig. 2. The SE landscape shapes transcription stress sites in cancer.

(A) Heatmap showing the normalized break density row z-score for each cell line’s unique SE-regulated genes. For each cell line, SE-regulated genes were identified, common SE-regulated genes from other cell lines were excluded, and low to moderately expressed genes as well as genes with low break counts were filtered out. Each column corresponds to a biological replicate. Columns were normalized by median scaling. (B) Box plot illustrating the normalized break density of cells treated with DMSO (control), dBET6, or CDK9i (100 nM for 4 hours each) for SE-regulated genes in each cell line or for randomly selected genes. Values are averaged for three biological replicates. (C) Heatmap depicting the normalized break density row z-score for each cell line’s unique SE-regulated genes following treatment with DMSO (control), dBET6, or CDK9i. Each column corresponds to a biological replicate. Each sample was normalized by downsizing to 270,000 random breaks. (D) Box plot demonstrating the normalized break density of cells treated with vehicle (control) or estradiol (E2; 100 nM for 1 hour) for significantly affected (P < 0.05) ERSE target genes or for randomly selected genes. (E) Heatmap representing the normalized break density row z-score of significantly affected ERSE target genes. Box plot features include the center line representing the median, box limits denoting upper and lower quartiles, whiskers extending to 1.5× the interquartile range, and dots indicating outliers. P values were computed using the Mann-Whitney U test, with n.s. indicating P > 0.05 and with **P < 0.01 and ****P < 0.0001.

If SE activity causes the accumulation of DSBs at transcription stress sites, then inhibiting their activity would abolish SE-mediated DSB accumulation, while conversely, activating it would enhance DSB accumulation at target genes. To test this hypothesis, we used two strategies. First, we inhibited SE activity in MCF7 and HeLa cells through two approaches using the BRD4 degron dBET6 and the CDK9 inhibitor (CDK9i) NVP-2, both of which have been shown to lead to SE collapse and repress SE-mediated transcription (32, 33). Examining the breakome after a 4-hour incubation with these inhibitors, which effectively suppressed SE-regulated transcriptional activity as measured by spike-in real-time quantitative polymerase chain reaction (RT-qPCR) (fig. S4, G and H), revealed that SE inhibition significantly reduced DSBs at SE-regulated genes in both cell lines (Fig. 2B). Furthermore, SE inhibition using either dBET6 or CDK9i eliminated the cell type–specific effect of the SE landscape on the breakome (Fig. 2C). Second, we leveraged the fact that MCF7 cells are estrogen receptor–positive and that many of their SEs are estrogen–responsive and analyzed the breakome of these cells before and after estrogen treatment. We specifically focused on genes regulated by estrogen receptor–occupied SE-regulated genes (ERSE genes), which were previously characterized (34). Consistently, ERSE induction significantly increased DSBs in target genes but not random genes (Fig. 2, D and E), indicating that SE activity promotes the accumulation of DSBs at SE-regulated genes. Overall, these data demonstrate that the landscape of SE activity in cancer cells shapes the landscape of transcription stress in a cell type– and cancer type–specific manner.

Specificity of transcription stress at SE-regulated genes

While multiple studies have demonstrated the formation of DSBs as a consequence of transcription in general (35–37), our results indicate that this phenomenon is markedly amplified when transcription is driven by SEs (Fig. 1B and fig. S1, E and F). To further support this observation, we used two complementary approaches.

First, we examined predictors for the genomic location of transcription stress sites. Using SE-regulated genes, TE-regulated genes, and general highly expressed genes as predictors, multivariate least absolute shrinkage and selection operator (LASSO) regression revealed that SE regulation was the strongest predictor of transcription stress sites (odds ratio, 12.61; Fig. 3A), despite the highly expressed gene group exhibiting the highest transcriptional levels overall (Fig. 3B). Notably, SE regulation was also the best predictor for genomic regions enriched in DSBs (Fig. 3C).

Fig. 3. Specificity of transcription stress at SE-regulated genes.

(A) Multivariate LASSO regression identifying predictors of transcription stress site locations in 100-kb genomic bins. Odds ratios are shown. (B) Box plot showing normalized expression [transcripts per million (TPM)] of SE-regulated, highly expressed, and TE-regulated genes despite lower expression. (C) LASSO regression for prediction of DSB-rich genomic sites. The genome was binned in 10-kb bins, and the top 5% bins with the highest break density were assigned as DSB-rich. (D) Box plot showing the break density in estrogen-responsive genes upon E2 treatment, grouped by the degree of transcriptional up-regulation [fold change (FC)]. Non-SE genes were randomly downsampled to match SE gene numbers for comparable P values. (E) Box plot showing the break density in estrogen-responsive genes following CDK9 inhibition (CDK9i), grouped by the degree of transcriptional down-regulation (log₂ fold change). Non-SE genes were randomly downsampled to match SE gene numbers for comparable P values. Box plot features include the center line representing the median, box limits denoting upper and lower quartiles, whiskers extending to 1.5× the interquartile range, and dots indicating outliers. P values in box plots were calculated using the Mann-Whitney U test in (B) and Wilcoxon paired test in (D) and (E). n.s., P > 0.05; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.

Second, to experimentally assess whether transcriptional activation through SEs promotes greater transcription stress and DSB accumulation, we performed a deeper analysis on the MCF7 breakome before and after estradiol treatment. Estrogen activates transcription via both promoter-proximal binding and estrogen-responsive SEs (38), enabling us to compare the impact of these two mechanisms. We grouped genes by their level of transcriptional up-regulation [on the basis of global run-on sequencing (GRO-seq) fold change (FC)] and further subdivided each group into SE-regulated and non–SE-regulated categories. Consistently, SE-regulated genes accumulated more DSBs than their non-SE counterparts. This effect was especially prominent in the highest induction group (fold change >10; Fig. 3D). A similar pattern was observed upon CDK9 inhibition, where SE-regulated genes exhibited a greater reduction in DSBs compared to non–SE-regulated genes with matched transcriptional down-regulation (measured by RNA-seq) (Fig. 3E).

Together, these results demonstrate that SE-mediated transcription is a major determinant of transcription stress site formation and leads to amplified DSB accumulation compared to non-SE transcriptional activity, even at equivalent expression levels.

γH2AX marks transcription stress sites

Given that DSBs are transcription-blocking lesions (28), transcription stress sites must be rapidly and efficiently repaired to maintain high expression levels. γH2AX is a marker for DNA damage response (DDR) activation and is among the first responders to DSBs, as it is phosphorylated within a few minutes after DSB formation (39). To further explore the role of DDR activation in the signaling of DSB repair at transcription stress sites, we conducted ChIP-seq for γH2AX on untreated MCF7 cells and analyzed high-quality available γH2AX ChIP-seq data for HeLa and the nontransformed IMR90 cell lines.

In both MCF7 and HeLa cells, the γH2AX signal is predominantly localized to gene bodies, whereas in IMR90 cells, the distribution is more balanced between genic and intergenic regions (fig. S5A). This pattern is consistent with a greater burden of transcription-associated stress in cancer cells compared to nontransformed cells. Moreover, γH2AX is distributed across the gene bodies of top targets (fig. S5, B to D). Ranking genes on the basis of their γH2AX enrichment revealed that most exhibit relatively low γH2AX levels, while a subset displays high to very high γH2AX density (Fig. 4A), indicating a nonuniform distribution of γH2AX, with preferential enrichment at specific genes. All transcription stress sites are among the top 8% of γH2AX targets (Fig. 4, A and B). Notably, this enrichment in the top γH2AX genes was much less for the general high-DSB genes (Fig. 4C). In addition, top γH2AX targets are enriched in SE-regulated genes across each respective cell line (P = 2.3 × 10⁻⁴¹, 2.9 × 10⁻³⁹, and 8.5 × 10⁻⁸ for MCF7, HeLa, and IMR90, respectively; hypergeometric test) (Fig. 4, D to F).

Fig. 4. γH2AX marks transcription stress sites.

(A) Ranked plot of genes in MCF7 cells based on γH2AX density. The γH2AX density per gene was computed using bigWigAverageOverBed. (B) Box plot showing γH2AX density for transcription stress sites and randomly selected genes. (C) Ratio of the observed to expected number of transcription stress sites or top 1000 break-prone genes in MCF7 among the top γH2AX-enriched genes. (D to F) Ratio of observed to expected numbers of SE-regulated genes in the top and bottom 300 genes enriched in γH2AX for each cell line. P values were determined using the hypergeometric test: n.s., P > 0.05; *P < 0.05; ***P < 0.001; ****P < 0.0001. (G) Genome browser snapshots of representative common γH2AX targets across MCF7, HeLa, and IMR90 cell lines. (H) Genome browser snapshots of representative unique γH2AX targets in MCF7, HeLa, and IMR90 cell lines. (I) Heatmap displaying the γH2AX density column z-score for each cell line’s unique SE-regulated genes. For each cell line, SE-regulated genes were identified, common SE-regulated genes from other cell lines were excluded, and genes with low γH2AX counts were filtered out. (J) Scatter plot showing the correlation between break density and γH2AX density for each gene. (K) Coefficient of variation of γH2AX density as a function of mean DSB signal. Genes were binned into deciles on the basis of their DSB signal intensity (log-transformed), and for each bin, the mean γH2AX density, standard deviation, and coefficient of variation were calculated. Box plot features include the center line representing the median, box limits denoting upper and lower quartiles, whiskers extending to 1.5× the interquartile range, and dots indicating outliers. P values in box plots were calculated using the Mann-Whitney U test; ****P < 0.0001.

The observed γH2AX enrichment across the three cell lines was consistent for several oncogenes, including miR21, NEAT1, MALAT1 (metastasis-associated lung adenocarcinoma transcript 1), and ACTG1 (actin γ1) (Fig. 4G). Many prominent γH2AX targets were unique to specific cell lines, most with established activity in their respective cell types (Fig. 4H). For example, two top γH2AX targets in MCF7, TFF1 (trefoil factor 1) and GATA3, are well-known oncogenic transcription factors primarily active in breast cancer and were exclusively enriched in γH2AX in MCF7 cells. KRT7 (keratin 7) and CCAT1 (colon cancer–associated transcript 1), which are overexpressed and linked to poor prognosis in cervical cancer, are enriched in γH2AX only in HeLa cells. Consequently, we questioned whether these variations in γH2AX enrichment stem from cell type–specific SEs. To investigate this, we compared the γH2AX enrichment of top targets that are regulated by a cell type–specific SE in each respective cell line. SE-regulated γH2AX targets are predominantly enriched in the cell line where they are controlled by an SE (Fig. 4I), highlighting that γH2AX marks sites of transcriptional stress shaped by the unique SE landscape of each cancer type.

γH2AX levels vary widely across genes enriched in high endogenous DSBs

Next, we examined the correlation between DSBs and γH2AX in genes. This analysis revealed a high variability of γH2AX among genes with high DSBs (Fig. 4, J and K, and fig. S6, A to E). Some genes enriched in breaks are marked by γH2AX, while others are not, in both MCF7 and HeLa cell lines (fig. S6, F and G). This discrepancy in γH2AX enrichment at various DSB-enriched genes might suggest that regions enriched in endogenous DSBs in the genome are signaled with varying efficiency and potentially repaired unevenly depending on their context.

γH2AX preferentially marks endogenous DSBs of active chromatin

Although genes exhibit a moderate correlation between DSB density and γH2AX density (Spearman’s correlation, 0.57), nonlinear regression analysis using generalized additive models (GAMs) indicates that the transcription stress score explains a greater proportion of the variance in γH2AX levels (R² = 0.73) than any individual component (Fig. 5A and fig. S7). To further investigate the underlying factors contributing to variability in γH2AX at high-DSB genes, we defined two groups of genes with similar break density but different γH2AX levels (Fig. 4J): DSBs^high γH2AX^high genes and DSBs^high γH2AX^low genes. DSBs^high γH2AX^high genes are more expressed and enriched in transcription stress markers (TOP1, TOP1cc, and R-loops) (fig. S8, A to E). Furthermore, 64 of the 65 transcriptional stress sites were among DSBs^high γH2AX^high genes (Figs. 4J and 5B). This group also included all SE-regulated transcription stress sites and is significantly enriched for active SE-regulated genes. In contrast, DSBs^high γH2AX^low genes lack active SE-regulated genes and transcriptional stress hotspots (Fig. 5C and fig. S8F).

Fig. 5. γH2AX preferentially marks endogenous DSBs of active chromatin.

(A) Bar plot showing R² values for the different variables tested regarding their effect on γH2AX variance across all genes. (B and C) Ratio of observed to expected number of transcription stress sites (B) and active SE-regulated genes (C). (D) γH2AX signal at the two defined sets of DSBs. (E) Percentage of DSBs enriched in γH2AX (γH2AX^high DSBs) or lacking γH2AX (γH2AX^low DSBs) in each chromatin state. Chromatin states were ordered on the basis of what is typically known about their transcriptional activity (75). (F) Ratio of the observed to expected numbers of high-DSB genes overlapping heterochromatin and repressed regions in both gene groups in MCF7 cells. (G) Percentage of overlaps between each gene group and chromatin states in MCF7. For each group, the number of overlaps was counted for each chromatin state and then divided by the total number of overlaps. (H) Bar plot showing R² values for the effect of DSBs on γH2AX variance for genes overlapping different chromatin states. P values were determined using the hypergeometric test: ****P < 0.0001.

Chromatin context was previously shown to affect the phosphorylation of H2AX, with heterochromatin being refractory to this modification (40). Therefore, we wondered whether the observed DSBs^high γH2AX^low genes are close to, or within, heterochromatic regions. To gain further insights into whether γH2AX preferentially marks endogenous DSBs on the basis of the chromatin context, we defined two groups of individual DSB sites across the genome according to their γH2AX signal (Fig. 5D and fig. S8G) and mapped them to the different chromatin states of MCF7 and HeLa. γH2AX^high DSBs were biased toward active chromatin states, while γH2AX^low DSBs were biased toward silent regions (Fig. 5E and fig. S8H). These results made us question whether the differential γH2AX marking at DSB-enriched genes was due to their chromatin context. Indeed, the DSBs^high γH2AX^low gene group, but not the DSBs^high γH2AX^high group, is enriched for DSB-enriched genes that overlap with heterochromatin and repressed chromatin states (Fig. 5F and fig. S8I). In addition, an analysis of the chromatin state contexts of DSBs^high γH2AX^high and DSBs^high γH2AX^low in both MCF7 and HeLa showed that DSBs^high γH2AX^high genes primarily overlap with chromatin regions associated with strong transcriptional activity. Conversely, DSBs^high γH2AX^low have a higher overlap with heterochromatin, repressed, and weak transcription chromatin states (Fig. 5G and fig. S8J).

Our results illustrate that DSBs alone do not fully account for the variation in γH2AX across genes (Fig. 5A). Conducting GAM on genes categorized by their chromatin context revealed that DSBs primarily account for variation in γH2AX in genes that overlap with strong enhancer regions and least in genes that overlap with repressed or heterochromatin regions (Fig. 5H). This suggests that breaks and γH2AX are most closely associated at genes spanning strong enhancer elements.

Collectively, these results indicate that γH2AX preferentially marks endogenous DSBs in active chromatin compared to inactive chromatin. This differential marking may influence the repair efficiency of distinct endogenous DSBs depending on their chromatin context.

Genes marked with γH2AX exhibit high DSB turnover and repair efficiency

sBLISS, like other methods used to map DSBs, captures a snapshot of the DSB landscape at the moment of fixation (41–45). Therefore, a gene that shows high levels of DSBs could experience frequent DSB generation and repair cycles (high DSB turnover) or infrequent DSB generation and repair cycles (low DSB turnover). To estimate the rate of endogenous DSB repair across the genome, we mapped DSBs in HeLa cells after disrupting the two major DSB repair pathways, nonhomologous end joining (NHEJ) and homologous recombination (HR), by using small interfering RNAs (siRNAs) that target x-ray repair cross complementing 4 (XRCC4) and RAD51, respectively (fig. S9A). To identify genes with significant DSB accumulation following the knockdown (KD) of each factor, we applied DESeq2 (46) to the DSB count matrices. Hierarchical clustering of these genes revealed three distinct response groups: Cluster 1 genes were primarily affected by XRCC4 depletion, cluster 2 genes were affected by both XRCC4 and RAD51 (with a stronger effect from RAD51), and cluster 3 genes affected mainly by RAD51 depletion (fig. S9B). Interestingly, no cluster was exclusively dependent on RAD51, consistent with a previous report showing that XRCC4 localizes to both RAD51-bound and unbound breaks (47).

Further analysis revealed that all three clusters are associated with transcription stress, with cluster 1 (XRCC4-dependent genes) showing the strongest enrichment (fig. S9C). These genes also exhibit the highest baseline break density (fig. S9D), suggesting that NHEJ is particularly engaged at the most DSB-prone, transcriptionally stressed loci. Collectively, these findings support the involvement of both NHEJ and HR in repairing transcription-associated DSBs while indicating a prominent role for XRCC4-mediated repair at the most vulnerable genomic sites. On the basis of this, we proceeded to quantify genome-wide DSB turnover using XRCC4 KD–induced break accumulation as a proxy for repair efficiency.

Regions that are frequently repaired should accumulate excessive DSBs when repair is impaired. In contrast, regions with inherently low repair rates should show minimal DSB accumulation (Fig. 6A). A total of 429 genes showed significant DSB accumulation upon XRCC4 KD (Fig. 6B). Stratifying these genes by their baseline DSB density revealed a trend in which groups with a higher baseline DSB density showed a progressively greater DSB accumulation following XRCC4 KD (Fig. 6C), supporting the notion that DSB-prone genes are normally repaired more effectively. Enrichment analysis of genes that significantly accumulated DSBs upon XRCC4 KD indicated a strong enrichment of DSBs^high γH2AX^high genes (Fig. 6D), along with the top DSB-enriched and active SE-regulated genes. In contrast, DSBs^high γH2AX^low genes and the bottom DSB-enriched and non–SE-regulated genes were not enriched. Directly examining the effects of repair impairment demonstrated that NHEJ impairment leads to more pronounced accumulation of DSBs at DSBs^high γH2AX^high and SE-regulated genes than DSBs^high γH2AX^low and non-SE genes (Fig. 6E). To investigate further how γH2AX patterns influence the repair turnover of endogenous DSBs, we used the accumulation of DSBs following repair impairment as a proxy for repair efficiency (Fig. 6F). Within DSBs^high γH2AX^high genes, repair efficiency markedly increases with higher DSB levels; however, within DSBs^high γH2AX^low genes, this relationship of repair efficiency is notably weaker (Fig. 6G). These findings suggest two distinct types of endogenous DSBs: γH2AX-marked DSBs, which exhibit high turnover, and poorly γH2AX-marked DSBs, which exhibit low turnover.

Fig. 6. Genes marked with γH2AX show high DSB turnover and repair efficiency.

(A) Schematic representation of the rationale behind the experiment to estimate the rate of endogenous DSB repair and turnover. (B) Volcano plot depicting the results of the differential expression analysis of DSBs following XRCC4 KD, assessed using the DESeq2 algorithm. The experiment was performed on two biological replicates per condition. (C) Box plot illustrating the log₂ fold change of break density upon XRCC4 KD for significantly affected genes stratified according to their baseline break density into 10 groups. (D) Ratio of the observed to expected number of genes in our defined gene sets. P values were calculated using the Fisher exact test. (E) Box plot of break density fold change. P values were calculated using the Mann-Whitney U test. (F) Genome browser snapshots of DSBs^high γH2AX^high and DSBs^high γH2AX^low gene examples, showing DSBs, γH2AX, and repair efficiency. The repair efficiency was estimated by pooling replicates, tiling the genome into 100-bp tiles, subtracting the break score of each tile in the siXRCC4 sample from siSc (control), and smoothing the segments through sliding window averaging. (G) Line plot for repair efficiency (log₂ fold change after XRCC4 KD) for DSBs^high γH2AX^high versus DSBs^high γH2AX^low genes ordered by their break density. Lines were generated using LOESS smoothing. Box plot: The center line indicates the median, box limits represent upper and lower quartiles, whiskers extend to 1.5× the interquartile range, and dots indicate outliers. *P < 0.05; ****P < 0.0001.

Efficient repair at transcriptional stress sites is crucial for SE-regulated oncogene expression

High γH2AX marking and repair efficiency at transcription stress sites under physiological conditions suggest that these regions rely on DSB repair to sustain their elevated transcriptional levels. To test this, we inhibited DNA-PKcs (DNA-dependent protein kinase catalytic subunit), a core component of the canonical NHEJ pathway, using a selective small-molecule inhibitor (DNA-PKci) in MCF7 cells. We then performed sBLISS to map genome-wide DSBs and spike-in RT-qPCR to quantify transcript levels.

To assess genome-wide DSB accumulation, we divided the genome into 10-kb bins, quantified DSBs per bin, and used DESeq2 to identify bins with significant DSB enrichment in DNA-PKci–treated cells. We then overlapped significantly differential bins with annotated gene regions to determine whether specific gene classes were disproportionately affected. Indeed, significantly up-regulated bins were strongly enriched at SE-regulated genes and not at non–SE-regulated genes (fig. S10, A and B).

Measuring mRNA levels of the top γH2AX-marked SE-regulated genes upon DNA-PKci treatment indicated a significant decrease in their expression levels (fig. S10C). To determine whether the observed transcriptional reduction was due to repair failure rather than secondary signaling effects, we also knocked down XRCC4 under the same conditions as in Fig. 6. This similarly led to down-regulation of highly γH2AX-marked SE-regulated genes that accumulated DSBs (fig. S10D). Together, these results demonstrate that efficient DSB repair via the NHEJ pathway is required to sustain the transcription of SE-regulated oncogenes.

Efficiently repaired transcriptional DSBs cluster in the 3D genome

Induced DSBs cluster in repair foci (48, 49), especially when introduced within active genes (49). We hypothesized that endogenous transcription-associated DSBs might also cluster within the three-dimensional (3D) genome as a mechanism to enhance repair. To investigate this, we analyzed high-quality Hi-C data generated from MCF7 cells (50).

To quantify the clustering of genes associated with break density, we introduced two Hi-C metrics that measure how a group of genes with a common characteristic (e.g., break density) is clustered (Fig. 7A): intragroup interactions, which represent the sum of genomic interaction scores between genes within the same group, and intergroup interactions, the sum of genomic interaction scores between genes from different groups. A clustered group of genes is expected to exhibit higher intragroup interactions and lower intergroup interactions compared to a nonclustered group (Fig. 7A). To investigate whether DSB-enriched genes exhibit clustering, genomic bins were ranked according to the break densities of their constituent genes and categorized into nine groups, from low (group 1) to high (group 9) break density. Bins containing DSB-enriched genes demonstrate higher intragroup interactions and lower intergroup interactions than bins with low-DSB genes (Fig. 7B), indicating the clustering of high-DSB genes.

Fig. 7. Efficiently repaired transcriptional DSBs cluster in the 3D genome.

(A) Schematic illustrating inter- and intragroup interactions and their use to determine the clustering of genes with a common characteristic (high DSB density in this instance). (B) Line plot depicting intergroup (dashed line) and intragroup (solid line) genomic bins arranged according to the break density of their overlapping genes. Lines were created using LOESS smoothing, with shaded areas representing the 95% confidence intervals. (C) Percentage of strong interactions within genomic bins containing DSBs^high γH2AX^high compared to bins with DSBs^high γH2AX^low genes. Strong interactions are defined as a log₂ interaction score >2. The percentage is calculated by dividing the number of strong interactions by the total number of interactions. P values were calculated by the chi-square test. (D) Percentage of strong interactions calculated as in (C). (E) Box plot reflecting the break density of common SE-regulated genes between MCF7 and MCF10A cell lines. P values were calculated by the Mann-Whitney U test. (F) Heatmap showing log₂ interaction scores for genomic bins containing common SE-regulated genes from both MCF7 and MCF10A cell lines. Arrows indicate distinct sites of variation in interactions. (G) Count of interchromosomal strong interactions shown in the heatmap from (F). Random bins serve as a reference. Box plot: The center line indicates the median, box limits represent upper and lower quartiles, whiskers extend to 1.5× the interquartile range, and dots indicate outliers. P values were calculated using the chi-square test in (C), (D), and (G) and Wilcoxon paired test in (E). ***P < 0.001; ****P < 0.0001.

The percentage of strong interactions among DSBs^high γH2AX^high genes is significantly higher than that among DSBs^high γH2AX^low genes (Fig. 7C and fig. S10A). In addition, transcription stress sites exhibit a substantially higher percentage of strong interactions compared to random sites with similar chromosomal distribution (Fig. 7D and fig. S10B), suggesting that efficiently repaired transcription stress sites are clustered together. To avoid the confounding effects of SE regulation on 3D interactions, we analyzed Hi-C data from the MCF10A cell line, which was conducted parallel to MCF7 (50). MCF10A is a precancerous cell line that shares many common SE-regulated genes with MCF7 cells. However, these common genes are significantly more enriched in DSBs within MCF7 cells (Fig. 7E). This enabled a comparison of the same regions with the same SE-regulation status but differing DSB enrichment. Bins containing common SE-regulated genes, as opposed to random regions, showed significantly higher strong interchromosomal interactions in MCF7 cells than in MCF10A cells (Fig. 7, F and G), indicating that transcription stress sites undergo increased clustering.

The repair of transcription stress sites is mutagenic

Given that transcription stress sites are more likely to be in close proximity (Fig. 7), we suspected that recurrent DSBs at these sites could serve as mutational hotspots. To explore this, we analyzed mutational data from the COSMIC database (51), which contains extensive mutational data on human cancers. While mutations within coding regions are subject to selective pressures on the basis of their impact on protein function and cellular fitness (52, 53), noncoding mutations—particularly those in intronic regions—are less constrained by selection and can provide a more direct readout of a region’s intrinsic susceptibility to mutation. To minimize the confounding effects of selection, we focused on noncoding intronic mutations in breast cancer and intersected them with our breakome data from the breast cancer cell line MCF7. Grouping genes by their DSB or γH2AX density revealed that genes with a higher break or γH2AX density exhibit greater intronic mutation density (Fig. 8, A to C). This association suggests that transcription-associated DSBs, potentially through their repair, lead to the accumulation of mutations over time.

Fig. 8. Transcriptional DSBs with high turnover are especially susceptible to mutations.

(A to C) Intronic deletion density (A), insertion density (B), and SNV density (C) for genes categorized by their break or γH2AX density (excluding zero break genes). Random genes are included for reference. (D to F) Rate of intronic mutations across genes categorized by their break density and further divided into three groups on the basis of their γH2AX response to DSBs. (G to I). Box plot of intronic deletion density (G), insertion density (H), and SNV density (I) for DSBs^high γH2AX^high versus DSBs^high γH2AX^low versus randomly selected genes. P-values were computed using the Mann-Whitney U test, with n.s. indicating P > 0.05, *P < 0.05, **P < 0.01, and ****P < 0.0001.

To disentangle the effects of DSBs from their repair and turnover, we modeled the relationship between γH2AX density and break density using nonlinear least squares regression. The residuals from this model were then used to classify genes in each group into three subgroups: underresponsive genes (negative residuals), moderately responsive genes (residuals near zero), and overresponsive genes (positive residuals). Notably, overresponsive genes showed the most increase in mutation density with increased DSBs (Fig. 8, D to F). This suggests that excessive γH2AX accumulation at DSB sites, which reflects recurrent DSB formation and turnover and heightened repair activity, leads to increased mutation accumulation. Furthermore, DSBs^high γH2AX^high genes, but not DSBs^high γH2AX^low genes, demonstrated intronic mutation enrichment compared to randomly selected genes (Fig. 8, G to I). Collectively, these results suggest that transcription stress sites are prone to mutations, likely due to a combination of high DSB turnover, clustering, and faulty repair. These findings have important implications for understanding how transcription-associated genome instability contributes to cancer evolution, as recurrent DSB formation and error-prone repair may drive the accumulation of oncogenic mutations over time.

DISCUSSION

Cellular processes—particularly those up-regulated in cancer such as replication and transcription—play an important role in genomic instability (54). Replication-dependent instability, caused by replication stress (4, 55), has been extensively studied, revealing its landscape (56, 57), causes, consequences (58), and potential for therapeutic targeting (59, 60). Transcription-replication–dependent genomic instability (61), particularly transcription-replication conflicts (62), has also been well characterized, with studies examining its genomic distribution, influencing factors, and the impact of transcription-replication orientation (63). Recently, attention has shifted toward transcription-dependent, replication-independent genomic instability, namely, transcriptional DSBs. While research has begun to uncover their causes and implications in disease (17, 35), several key aspects remain underexplored, including genomic distribution of hotspots, repair mechanisms, and their impact on cancer evolution. By using multiomic approaches, we delineated the landscape of transcription stress and revealed that SEs shape the transcriptional program and contribute to the formation of genomic instability hotspots in cancer (Fig. 9).

Fig. 9. Schematic model summarizing the interplay between SE-driven hypertranscription, DSB accumulation, γH2AX marking, repair dynamics, and mutagenesis.

The unique SE regulatory landscape of each cancer type defines distinct transcription stress hotspots. These loci undergo frequent DNA DSBs, marked by γH2AX, which facilitates the recruitment of DNA repair factors. Because of the predominant use of the error-prone NHEJ pathway at these sites, repair is often imperfect, leading to the accumulation of mutations. This mechanism helps explain the enrichment of mutations at SE-regulated oncogenes in cancer genomes.

SEs determine cell identity in normal cells (31). In cancer cells, SEs enhance the transcription of oncogenes, promoting cell growth and proliferation (6, 8, 64, 65). Previous studies have established that highly transcribed genes are prone to transcription-associated genome instability (13, 15, 20, 30, 66). However, our data reveal that this relationship is significantly amplified in SE-regulated genes, suggesting that transcriptional stress is not merely a function of high gene expression but is modulated by SE activity. The enrichment of highly expressed SE-regulated genes among transcription stress sites was more than 30-fold higher than expected by chance, far exceeding that of highly expressed TE-regulated genes or highly expressed genes lacking enhancer regulation. This specificity suggests that SEs impose a unique transcriptional burden, predisposing their target genes to DNA damage. The functional implications are significant, as many SE-regulated genes encode critical oncogenes, including CCND1, NEAT1, miR21, and MYC, which play essential roles in tumorigenesis.

The pronounced enrichment of DSBs at SE-regulated oncogenes was considerably higher in cancer cell lines compared to their normal counterparts, indicating that SEs primarily drive genomic instability in cancer cells. SEs have previously been recognized as hypermutated regions that can alter the expression of target genes and contribute to carcinogenesis (6). Our findings advance this understanding by demonstrating that SEs also dictate mutational susceptibility at their target genes by driving hypertranscription.

Our findings reveal that SE-regulated oncogenes are not only transcriptionally hyperactive but also prone to transcription-associated DSBs, requiring efficient repair to sustain their expression. This dependency presents a therapeutic opportunity: Targeting DNA repair pathways in SE-addicted tumors could destabilize transcriptional programs essential to tumor growth.

Several therapeutic strategies may exploit this vulnerability. For instance, inhibiting core NHEJ factors such as DNA-PKcs (e.g., with nedisertib) can impair repair at SE loci, leading to DSB accumulation and suppression of oncogene expression. This concept has shown promise in preclinical models, including Ewing sarcoma and MYC-driven tumors (67), and may extend to other SE-reprogrammed cancers such as ER⁺ breast cancer and diffuse large B cell lymphoma. Another option is using epigenetic modulators like histone deacetylase inhibitors or proteolysis-targeting chimeras, which have been shown to boost transcription from SEs, which could further increase transcription-related DNA damage, overwhelming cancer cells under certain conditions (68, 69).

Although sBLISS captures a snapshot of the DSB landscape at the time of fixation, we were able to estimate the repair and turnover of DSBs across genes, revealing how transcription stress sites particularly undergo high DSB turnover and repair efficiency. To accommodate this, these regions experience constant activation of the DDR and H2AX phosphorylation. Not all DSB-enriched genes exhibit such efficient repair signaling, suggesting two distinct types of DSB-enriched genes in the genome: genes that undergo frequent breakage-repair cycles and are actively recognized by the repair machinery and genes that, despite being DSB-rich, experience infrequent breakage-repair cycles. These persistent DSBs, found in regions with low transcriptional activity, have a lower demand for repair. We speculate that the latter genes experience breakage in a more programmed manner compared to genes with high turnover, which break as a by-product of transcription. The nature of these DSBs requires further investigation. In addition, we demonstrate that genes with efficient DSB repair signaling are more vulnerable to mutations than genes with comparable DSB levels but less effective repair signaling, emphasizing the importance of the repair process itself in mutagenesis at transcription stress sites.

In summary, our findings highlight SEs as critical determinants of transcriptional stress and genomic instability in cancer. By integrating multiomic data, we demonstrate that SE-driven hypertranscription uniquely predisposes oncogenes to DNA breakage and mutation, revealing a previously underappreciated mechanism by which SEs shape the cancer genome. These insights not only deepen our understanding of transcription-dependent DNA damage but also point to SE-associated transcription stress as a potential vulnerability that could be therapeutically exploited. Further investigation into the distinct dynamics of DSB repair at these loci may uncover new strategies to modulate genome stability in cancer.

MATERIALS AND METHODS

Cell culture and experimental conditions

MCF7 (HTB-22) and T-47D (HTB-133) cells were grown in RPMI supplemented with 10% (v/v) fetal bovine serum (FBS; Gibco), glutamine, and penicillin/streptomycin. Human embryonic kidney (HEK) 293 (CRL-1573), MDA-MB-468 (HTB-132), MDA-MB-231 (HTB-26), and HeLa (CCL-2) cells were grown in Dulbecco’s Modified Eagle’s Medium (DMEM) supplemented with 10% (v/v) FBS, glutamine, and penicillin/streptomycin. A-4098 cells (HTB-44) were grown in Eagle’s minimal essential medium supplemented with 10% FBS, glutamine, and penicillin/streptomycin. SH-SY5Y (CRL-2266) were grown in DMEM/F12 supplemented with 10% FBS, glutamine, and penicillin/streptomycin. HMLE cells were grown in Promocell mammary epithelial cell basal media (C-21010) with added supplements (c-93110), whereas MCF10A cells were grown on DMEM/F12 supplemented with 5% horse serum, epidermal growth factor (20 ng/ml), hydrocortisone (0.5 mg/ml), cholera toxin (100 ng/ml), insulin (10 mg/ml), and penicillin/streptomycin. MCF-7 cells were synchronized in the G₁ cell cycle stage by culturing in serum-free media for 72 hours followed by replacing the media with 5% FBS media for 5 hours. Cells were grown at 37°C under a humidified atmosphere with 5% CO₂. Cells were routinely authenticated by short tandem repeat profiling and tested for mycoplasma, and cell aliquots from early passages were used.

Cells were treated with dimethyl sulfoxide (DMSO; Sigma-Aldrich), dBET6 (Sigma-Aldrich, SML2638; 100 nM), or CDK9i (NVP-2) (MCE, cat. no. Y-12214A; 100 nM) for 4 hours. Cells were treated with DMSO or DNA-PKci (NU-7441, Selleckchem; 2 μM) for 5 hours.

Transcription stress score

Our sBLISS, DRIP-seq, TOP1cc, and TOP1 ChIP-seq data used for transcription stress score calculation are already published (13) and are available at Gene Expression Omnibus (GEO): GSE241309. Publicly available K562 data were downloaded from GEO (GSM3444988 for sBLISS, GSM4478670 for DRIP-seq, and GSM5502572 for TOP1cc CAD-seq). For each dataset, the signal density per gene was first converted to an empirical P value on the basis of its rank within the distribution of all genes. Higher signal values correspond to lower P values. These P values were then transformed into z-scores using the standard normal inverse cumulative distribution function

$$Z_k={\mathrm{\Phi }}^{-1}(1-p_k)$$

where p_k is the empirical P value for dataset (K), and Φ⁻¹ denotes the inverse of the standard normal cumulative distribution function.

The individual z-scores from all four datasets were combined per gene using the Liptak weighted Z-test (unweighted version)

$$\text{TS}{\mathrm{S}}_{\text{raw}}=\frac{Z_1+Z_2+Z_3+Z_4}{\sqrt{4}}$$

For better representation, the combined score was further min-max normalized and then stretched by a power transformation to emphasize differences

$$\text{TS}{\mathrm{S}}_{\text{final}}=\text{TS}{\mathrm{S}}_{\text{scaled}}^4$$

Genes not among the top 1000 in one or more parameters were excluded.

Identification of transcription stress–associated domains using the multivariate HMM

To identify genomic regions simultaneously enriched for markers of transcription stress, we applied a multivariate HMM using the mhsmm R package. The hg38 genome was divided into 1-kb nonoverlapping bins, and mean signals for TOP1, TOP1cc, R-loops, and DSBs were calculated per bin. Signal intensities were capped at the 0.5th and 99.5th percentiles to minimize the influence of outliers. A two-state multivariate HMM was initialized with uniform state priors and a transition matrix favoring self-transition (0.95 probability). Emissions were modeled using multivariate Gaussian distributions with means offset by ±0.5 standard deviations and shared covariance. The model was trained by maximum likelihood estimation, and Viterbi decoding was used to assign each bin to a state. The state with the highest average signal across all datasets was designated as the “comarked” state. To refine the identified regions, we selected bins showing a high signal (above the 95th percentile) in at least three of the four datasets, merged adjacent bins separated by ≤1 kb, and filtered them to retain regions ≥3 kb.

Gene set compilation and enrichment analysis

Gene sets tested for enrichment here were defined as follows: SE-regulated genes for each cell line were downloaded from SEdb2.0 (31). Samples IDs used were as follows: Sample_01_0046 for MCF7, Sample_02_1273 for HeLa, Sample_00_0015 for IMR90, Sample_02_0434 for HEK-293, Sample_02_0097 for T-47D, Sample_02_0095 for MDA-MB-468, Sample_02_0219 for MDA-MB-231, Sample_02_0643 for MDA-MB-436, Sample_02_1812 for SHSY5Y, Sample_02_0056 for HMLE, and Sample_02_0268 for A4098. A list of annotated oncogenes was downloaded from OncoKB (70). Highly expressed genes are defined as the top 1000 expressed genes. TE-regulated genes are defined as genes that overlap enhancer regions and are not annotated by an SE. Nonenhancer (non-E) genes are genes that do not overlap any enhancer element. Enrichment was calculated by dividing the number of observed overlapping genes by the theoretical number of genes predicted under a random model. A one-tailed (upper-tailed for values >1 and lower-tailed for values <1) hypergeometric test was performed using the phyper function in R to determine the statistical significance of the observed enrichment.

In-suspension break labeling in situ and sequencing (sBLISS)

sBLISS was conducted as previously described (41, 71). In summary, 10⁶ cells were fixed in 2% paraformaldehyde in 10% FBS/phosphate-buffered saline (PBS) for 10 min at room temperature. The fixation was quenched with 125 mM glycine for 5 min at room temperature, followed by another 5 min on ice and two washes in ice-cold PBS. Cells were lysed for 60 min on ice, and their nuclei were permeabilized for 60 min at 37°C. Next, nuclei were rinsed twice with CutSmart Buffer containing 0.1% Triton X-100 (CS/TX100), and DSB ends were blunted in situ using NEB’s Quick Blunting Kit for 60 min at room temperature. The blunted nuclei were then washed twice with 1× CS/TX100 before in situ ligation of the sBLISS adapters to the DSB ends. Adaptor ligation was carried out with T4 DNA ligase for 20 to 24 hours at 16°C, with bovine serum albumin and adenosine 5′-triphosphate added. Following ligation, the nuclei underwent two washes with 1× CS/TX100, and genomic DNA was extracted using proteinase K at 55°C for 14 to 18 hours while shaking at 800 rpm. Proteinase K was then heat inactivated for 10 min at 95°C, followed by extraction using phenol:chloroform:isoamyl alcohol, chloroform, and ethanol precipitation. The purified DNA was sonicated in 100 μl of ultrapure water using Covaris M220 for 60 s. Sonicated samples were concentrated with AMPure XP beads (Beckman Coulter), and fragment sizes were evaluated using a BioAnalyzer 2100 (Agilent Technologies), targeting a range of 300 to 800 base pairs (bp) with a peak around 400 to 600 bp. The sonicated DNA was then in vitro transcribed using the MEGAscript T7 Kit for 14 hours at 37°C. After RNA purification and ligation of the 3′-Illumina adaptors, the RNA underwent reverse transcription. The final library indexing and amplification step was performed with NEBNext Ultra II Q5 Master Mix.

sBLISS fastq files were initially demultiplexed using sample barcodes. Quality control was performed with Trim Galore to eliminate residual adapters, trim reads to a base quality of at least 20, and filter out short reads smaller than 20 bp. Initial and final sample qualities were assessed with fastqc. Quality-processed fastq files were aligned to the GRCh38 assembly with hisat2, then sorted, and indexed using samtools. The resulting bam files were deduplicated with umi-tools using genomic coordinates and unique molecular identifiers. Custom Python and R scripts were used to identify read start positions and convert bam files to bigwig format for downstream analysis. An additional custom R script was used to discard a blacklist of positions, primarily within centromeres. sBLISS for MCF7 cells transfected with siTOP1, or overexpressing ribonuclease H, as well as treated with E2, are already published and available at GEO: GSE241309.

RNA-seq and GRO-seq data processing

Bulk RNA-seq processed data for untreated MCF7, HeLa, and T-47D cell lines were obtained from ENCODE (72, 73), while data for MDA-MB-436 and MDA-MB-468 were downloaded from GEO (GSE212143) (74). For MCF7, the following datasets were used: ENCFF967AOT, ENCFF328CKZ, ENCFF921PJP, ENCFF930TQB, ENCFF030IUO, ENCFF878CJF, ENCFF057ITQ, ENCFF317XCR, ENCFF985KJE, and ENCFF456OYZ. For HeLa, we used ENCFF750TQI, ENCFF622TLZ, ENCFF689BFU, ENCFF151QTZ, ENCFF262EJQ, ENCFF878CJF, ENCFF964OAC, ENCFF346JDS, and ENCFF638MLA. For T-47D, we used ENCFF234TQB, ENCFF412DJX, ENCFF475SHD, and ENCFF846PLI. For K562, we used ENCFF172GIN. Transcripts per million (TPM) values were averaged across replicates for each cell line to determine the expression rank of each gene.

RNA-seq data for MCF7 after CDK9i treatment were downloaded from GSE129012 (4-hour time point). In addition, GRO-seq data for MCF7 after E2 treatment were downloaded from GSE27463 (40-min time point).

RT-qPCR and spike-in RT-qPCR

Total RNA was isolated using TRI reagent (Biolab), following the manufacturer’s guidelines for the phenol/chloroform extraction technique. cDNA was synthesized from 1 μg of RNA using the QScript cDNA synthesis kit (Quantabio). The SYBR Green PCR Master Mix (Applied Biosystems) was used for RT-qPCR. All assays were conducted in triplicate. Experiments were performed to assess the levels of cDNA using primers targeting the following: chicken RPL4 (forward, GAGTGACTACAACCTGCCGA; reverse, TTGGCGTATGGGTTCAGCTT), SIK1 (forward, CTCCGGGTGGGTTTTTACGAC; reverse, CTGCGTTTTGGTGACTCGATG), ID3 (forward, GAGAGGCACTCAGCTTAGCC; reverse, TCCTTTTGTCGTTGGAGATGAC), XBP1 (forward, CCCTCCAGAACATCTCCCCAT; reverse, ACATGACTGGGTCCAAGTTGT), CCND1 (forward, GCTGCGAAGTGGAAACCATC; reverse, CCTCCTTCTGCACACATTTGAA), HES1 (forward, CCTGTCATCCCCGTCTACAC; reverse, CACATGGAGTCCGCCGTAA), FOS (forward, TCCCATCGGTCCACTAGGTTT; reverse, AGGGCTGCACTGAGTTCTTTG), JUNB (forward, ACGACTCATACACAGCTACGG; reverse, GCTCGGTTTCAGGAGTTTGTAGT), TFF1 (forward, CCCTCCCAGTGTGCAAATAAG; reverse, GAACGGTGTCGTCGAAACAG), UBC (forward, ATTTGGGTCGCGGTTCTTG; reverse, TGCCTTGACATTCTCGATGGT), HPRT (forward, TGACACTGGCAAAACAATGCA; reverse, GGTCCTTTTCACCAGCAAGCT), MYC (forward, ATGCCCCTCAACGTGAACTTC; reverse, CGCAACATAGGATGGAGAGCA), XRCC4 (forward, ATGTTGGTGAACTGAGAAAAGC; reverse, GCAATGGTGTCCAAGCAATAAC), GREB1 (forward, TGGTCCGTAATGCACAAGGG; reverse, CTGCGTTTAGTGAGGGGTGA), GATA3 (forward, GCCTCTGCTTCATGGATCCC; reverse, CACACTCCCTGCCTTCTGTG), ACTB (forward, TTTTGGCTATACCCTACTGGCA; reverse, CTGCACAGTCGTCAGCATATC).

In RT-qPCR experiments, cycle threshold values were normalized to HPRT levels. For spike-in RT-qPCR, chicken RNA was added in proportion to the number of cells in each sample, and cycle threshold values were normalized to chicken RPL4.

γH2AX ChIP-seq

MCF7 cells (~10⁶) were cross-linked with 1% formaldehyde (methanol-free; Thermo Fisher Scientific, 28906) for 10 min at room temperature and quenched with glycine at a 125 mM final concentration. Fixed cells were washed twice in PBS and incubated in sonication buffer [0.5 M NaCl and 0.5% SDS radioimmunoprecipitation assay (RIPA) buffer containing phenylmethylsulfonyl fluoride, protease, and phosphatase inhibitors] for 30 min on ice. Cells were sonicated using Covaris M220 for 10 min to produce chromatin fragments of ~200 to 300 bp. The sheared chromatin was centrifuged for 10 min at a maximum speed. From the supernatant, 50 μl was saved as input DNA, and the rest was diluted in RIPA buffer without NaCl and SDS. The chromatin was immunoprecipitated by incubation with 5 μg of γH2AX (phospho-S139; Abcam, AB2893) antibody premixed with 100 μl of Dynabeads Protein G (Thermo Fisher Scientific, 10004D) and incubated overnight at 4°C with rotation. Immunoprecipitates were washed twice with RIPA buffer and 150 mM NaCl, twice with RIPA buffer and 300 mM NaCl, and twice with tris-EDTA buffer. The chromatin was eluted from the beads with 200 μl of direct elution buffer (10 mM tris, pH 8, 0.3 M NaCl, 5 mM EDTA, and 0.5% SDS) and incubated overnight at 65°C to reverse the cross-linking. Samples were treated with ribonuclease for 1 hour at 37°C and with proteinase K at 55°C for 2 hours. DNA was cleaned up by a QIAquick PCR purification column (Qiagen), according to the manufacturer’s instructions, and eluted in 30 μl of elution buffer. The ChIPed (chromatin-immunoprecipitated) DNA and input DNA were used for real-time PCR or to prepare libraries by Truseq (Illumina) and sequenced in Nextseq (Illumina).

ChIPed and input fastq files were quality controlled with Trim Galore to remove residual adapters, trim reads to base quality of at least 20, and filter out short reads with a size smaller than 20 bp. Initial and final sample qualities were evaluated with fastqc. The quality-processed fastq files were aligned to the GRCh38 genome with hisat2 followed by sorting and indexing with samtools. The utility bamCompare of deepTools was applied on the bam files to create a peak bigwig file. This operation was set to ignore duplicates, discard a blacklist of regions, normalize by RPKM (reads per kilobase per million mapped reads), and compare by ratio in tiles of 50 bp. γH2AX ChIP-seq data for HeLa and IMR90 were downloaded from GEO: GSM4564685 and GEO: GSM2067927, respectively. The γH2AX density per gene was computed using bigWigAverageOverBed, which calculates the average signal from the γH2AX bigWig track over annotated gene intervals.

Statistical modeling and residual-based classification

For multivariate logistic regression, genes were categorized on the basis of whether they overlapped transcription stress sites identified by the HMM. Logistic regression was performed using the glm function in R (family = binomial).

Multivariate LASSO regression was conducted using the glmnet package in R, with 10-fold cross-validation (cv.glmnet) to select the optimal regularization parameter (λ). The model was trained to classify genes on the basis of transcription stress site overlap or DSB enrichment. Odds ratios were derived from the final model coefficients to assess the relative predictive power of each feature.

GAM was applied to model the relationship between each variable and γH2AX across all genes in Fig. 5A and for genes classified into their overlapping chromatin states in Fig. 5H. The model was constructed using the gam function in the mgcv package in R.

To dissect the effects of absolute DSBs from DSB turnover on mutation frequency, we modeled the relationship between break density and transcription stress by nonlinear least squares regression using the following equation

$$\mathrm{\gamma }\mathrm{H}2\text{AX}\text{density}=a\times \text{break densit}{\mathrm{y}}^b$$

The model was fitted using the nls function in R, with initial parameter estimates for “a” and “b” set to 0.1. After fitting the model, residuals were extracted and used to classify genes into three response categories on the basis of predefined thresholds: Genes with residuals >0 were classified as “overresponsive,” genes with residuals <0 as “underresponsive,” and genes between −0.3 and 0.3 as “moderately responsive.”

Classifying individual DSB sites on the basis of γH2AX

Each break site was assigned a γH2AX coverage level (γH2AX rank). To ensure a comparable distribution of break scores (i.e., the number of DSBs at the same site) between the two groups, break sites were classified into four categories: those with a break score of 1 (the large majority), those with a break score between 2 and 4, those with a break score between 5 and 10, and those with a break score between 11 and 50. Within each of the four categories, 1% of the top and bottom γH2AX-ranked breaks were selected. All categories were then recombined so that the top set of breaks and the bottom set of breaks exhibit a similar break score distribution. Mapping to chromatin states was then computed and plotted for these sets.

Chromatin states

Chromatin states bed files were downloaded from ENCODE (72, 73). For MCF7, the ENCFF631HCR file was used. For HeLa, the ENCFF555HXM file was used.

Transient transfection

Transient transfection of siXRCC4 (Dharmacon; SMARTpool siRNA; catalog ID: L-004494-00-0005) and siSc (Dharmacon; D-00181010) was carried out using Lipofectamine (Thermo Fisher Scientific). On the day of transfection, cells were seeded to achieve a confluency of 60 to 70%. The transfection solution was prepared by mixing 1.5 ml of serum and antibiotic-free RPMI with 30 μl of Lipofectamine and incubating for 5 min before adding a mixture of 1.5 ml of RPMI and 0.6 nmol of siRNA, followed by a 20-min incubation at room temperature. This mixture was then added to cells in a 10-cm plate cultured in antibiotic- and serum-free media. The media were replaced with fully supplemented media after 5 to 6 hours, and the cells were incubated in the incubator for 48 hours.

Hi-C data analysis

Normalized Hi-C contact matrices for MCF7 and MCF10A cells were obtained from GEO (GSE66733). Genomic bins were intersected with all genes, and each bin was assigned to a single gene. Bins overlapping multiple genes were excluded from the analysis.

In inter- and intragroup metrics calculation, genes were stratified into nine groups on the basis of their break density, excluding genes with zero breaks. Intra- and intergroup interaction scores were computed as follows: In intragroup interactions, for each gene, interaction scores with all other genes within the same group were summed and normalized by the total number of genes in that group. In intergroup interactions, for each gene, interaction scores with all genes outside its group were summed and normalized by the total number of genes in all other groups. Genes, along with their interaction scores, were ordered on the basis of break density. The data were visualized using locally estimated scatterplot smoothing (LOESS).

COSMIC mutation data analysis

The file Cosmic_NonCodingVariants_Tsv_v101_GRCh37.tar, containing all noncoding variants, was downloaded from COSMIC. This file was first filtered for breast cancer samples and separated on the basis of the type of mutation into single-nucleotide variant (SNV), deletion, and insertion mutation files. These files were then converted into genomic range objects and filtered to contain only mutations within introns. To calculate mutation density, mutations per gene were counted and normalized by the sum of intron length.

Supplementary Materials

This PDF file includes:

Figs. S1 to S11

Table S1

References