Persistent chromatin alterations and gene expression reprogramming follow widespread DNA damage in glioblastoma

Abstract

DNA damage from routine cellular processes or exogenous insults can have a lasting impact on gene regulation beyond genetic mutations^1–5. The prevailing paradigm for the consequences of DNA damage repair revolves around restoration of the original genetic sequence, but long-term changes in chromatin configuration, gene expression and DNA modifications have not been analyzed. We introduce numerous, simultaneous Cas9-mediated DNA double strand breaks (DSBs) at defined locations in human glioblastoma cells and track both non-genetic and genetic alterations over time. Megabase-scale genomic alterations that endured two weeks after the initial damage were detected, involving a shift from transiently increased intra-TAD interactions to persistent long range cis and trans contacts, alterations in gene-expression and associated large structural variations. These findings reveal that widespread DNA damage, such as chemotherapy or radiotherapy, can trigger long-term genetic and non-genetic modifications which alter cellular function and may impact tumor outcome and the emergence of resistant cells.

How DNA damage reshapes cancer cells beyond genetic mutations is poorly understood. Here, the authors show that DNA double-strand breaks in glioblastoma cells trigger lasting non-genetic changes in 3D genome folding, gene expression and DNA modifications that persist for weeks.

Introduction

Following treatment with DNA damaging therapy, tumors recur with new phenotypic traits that are not fully explained by genetic alterations or clonal selection alone^6–8. This discrepancy underscores the hypothesis that other mechanisms, such as persistent changes in DNA modifications, chromatin configuration, and gene expression, may play a role in fueling therapy resistance. Indeed, the DNA damage response mobilizes chromatin modifiers^9,10, but whether these mediate changes that culminate in long-term gene expression shifts remains poorly understood.

DNA damage arises during physiological cellular processes and from exogenous sources such as chemotherapy, therapeutic or environmental radiation. Glioblastoma presents a particularly compelling model to investigate DNA damage–induced changes, as standard therapy includes 60 Gray of radiation, concurrent with the use of the chemotherapeutic agent temozolomide (TMZ) and DNA damage-generating Tumor Treating Fields^11,12. The radiation portion alone is estimated to make a total of ~1500 double-strand breaks per cell¹³. Despite multi-modal DNA-damaging therapy, tumors exhibit few new genetic lesions in driver oncogenes or tumor-suppressors, but recur with widespread, seemingly arbitrary, non-genetic alterations, as highlighted in recent clinical and preclinical work^7,14–16. Therefore, elucidating mechanisms underlying damage-induced phenotypes could illuminate new strategies to prevent or counteract the emergence of resistant glioblastoma subclones following therapy.

Previous work has shown that chromatin and DNA modifications can change acutely upon radiation or endonuclease-induced breaks, but these studies were performed over short time frames (hours to days), and thus it is not known whether there is a long-term impact (over weeks) on chromatin and its 3D organization that leads to sustained changes in gene regulation^1,2,9,17,18. Another limitation of these studies is the approach used for introducing DNA damage. For example, radiotherapy or chemotherapy can induce a variety of randomly localized DNA damage insults at different timepoints in individual cells, which prevents a clean analysis of longitudinal changes. Furthermore, the use of fixed target homing endonucleases systems, which are unable to target methylated DNA, likely produces a target selection bias, limiting the generalization and comprehensiveness of the mechanisms underlying the long-term effects of DSBs.

In order to dissect the dynamics of genetic and non-genetic events that follow DNA damage, it is important to study tumor evolution over controlled time periods without the confounding effects of the varied nature of chemoradiotherapy-induced breaks. The latter are randomly introduced into the genome in individual cells and cannot be tracked over time as part of bulk cell analysis. To address these limitations, we used a “multi-target” CRISPR–Cas9 approach that enables widespread pre-mapped DNA breaks at hundreds to thousands of loci in human glioblastoma stem cells¹⁹. This approach, in a similar fashion to prior efforts^20,21, introduces DSBs at pre-defined sites (while mimicking the widespread nature of chemoradiotherapy-induced DNA damage), allowing us to systematically track the downstream effects on the genome, with the goal of studying the mechanisms underlying short and long-term changes in gene regulation. Following the introduction of DNA breaks, we examined the interplay between changes in gene expression in the context of DNA and chromatin modifications, transient and persistent changes in 3D chromatin folding and the emergence of large-scale structural variations.

Results

Genome-wide DNA damage leads to dynamic alterations in genome-organization and gene-expression over time

To gain insight into how cancer genomes could be altered following DNA-damaging therapies such as chemotherapy or radiation, we performed time course experiments following the simultaneous induction of multi-targeted DSBs. Glioma-Stem Cell lines (GSCs)²² were co-infected with wt-spCas9 and two distinct doxycycline-inducible multi-target guide-RNAs (mtgRNA), termed mtg142 and mtg483, to simultaneously introduce numerous DSBs across the genome (Supplementary Fig. 1A)²³. These guides were selected because they have 142 and 483 predicted DSB target sites, respectively, which allows us to assess similar levels of damage to that incurred with cancer-directed therapy. Details of mtgRNA design and generation are previously described¹⁹. Induction of the mtgRNAs led to the formation of γH2AX foci as assessed by immunofluorescence at three hours, with each guide exhibiting similar toxicity to 2–4 Gy of radiotherapy as measured by colony formation assays in human glioma U87 cell lines (Supplementary Fig. 1B, C)²³. Because GSCs are floating neurospheres and incompatible with standard colony formation assays, we performed an extreme limiting dilution assay. Mtg142, mtg483, and 2 Gy did not significantly alter neurosphere stem cell frequency relative to vector-only, whereas 4 Gy produced a significant decrease in sphere formation (overall χ²(4) = 27.17, P = 1.84 × 10⁻⁵; Supplementary Fig. 1D). To create a double-strand break (DSB) map for back-tracing of genomic changes with damaged sites, we performed in-suspension break labeling in situ and sequencing (isBLISS) (Fig. 1A and Supplementary Fig. 1E), which enabled us to generate a map of DSB sites for mtg142 and mtg483, which had 339 and 1,386 total verified cut sites, respectively, shown alongside predicted DSB target sites (Fig. 1B). The increase of actual over predicted DSB sites occurs because of the presence of lower affinity sites with nucleotide mismatches. Genomic enrichment analysis of verified cut sites revealed enrichment at repetitive elements, including L1 and ERV1 (Supplementary Fig. 1F), consistent with the degenerate sequence design of these guide RNAs to generate widespread simultaneous DSBs. Importantly, this does not indicate that all L1 or ERV1 elements in the genome are cleaved; rather, just a subset of cut sites identified by isBLISS Fig. 1A and Supplementary Fig. 1E.

Fig. 1. Genome-wide DNA damage leads to distal and dynamic alterations in genome-organization and gene-expression over time.

A Double-strand break (DSB) sequencing (isBLISS) of control (Cas9 only), CRISPR-Cas9 guides mtg142 and mtg483. Bottom rows (sites) represent regions that have either 0, 1 or 2 mismatches (MM) in nucleotides (theoretical or possible cut sites due to lower affinity off-targeting). B Genome-wide DSB maps with isBLISS verified sites marked in red. Predicted sites (blue) represent a perfect guide to genome matches and do not include the lower-affinity or off-target sites that BLISS identifies. C Hi-C (genome-contact) maps of chromosome 17 showing contacts (top panel) and contact changes (bottom panel) over time (3 h and 2 weeks). Cut sites for mtg142 for this region are shown on the left y-axis in red. D RNA-seq of differentially expressed genes marked as a function of distance to the nearest cut site for mtg142 (LogFC > 0.05, P value < 0.05). E Association between genomic distance from the nearest cut site and differential gene expression. Points indicate odds ratios from logistic regression models in which differential-expression status was the binary outcome and distance from the nearest cut site was the predictor. Horizontal bars indicate 95% confidence intervals. P values were calculated using two-sided Wald tests for the distance coefficient, and n indicates the number of genes included in each model. F Pearson correlation matrix comparing TAD-aggregated log-fold changes between mtg142 and mtg483-treated cells across four genomic features (gene expression, H3K27ac, DNA methylation, Hi-C cis contacts) and two timepoints (3 h, 2 weeks).

We performed genome-wide Chromosome Conformation Capture with Hi-C at each of the three time points (control, 3 h, and 2 weeks) to examine the longitudinal impact of DSBs generated by the two mtgRNAs on chromatin organization. Changes in cis genomic contacts were observed Megabases (Mb) from verified cut sites that evolved over a period of two weeks, suggesting a persistent re-programming event (Fig. 1C and Supplementary Fig. 2A). Although the two guides target different sequences, the 3-h and 2-week samples exhibited similar overall variance in gene expression levels as shown by principal component analysis (Supplementary Fig. 2B). Significant gene expression alterations were found at promoters of genes proximal to cut sites and up to 20 megabases away at the 3 h time point and changes in mRNA levels persisted 2 weeks after the introduction of the DSBs (Fig. 1D and Supplementary Fig. 2C). Radiation produced a transcriptional response that was temporally distinct from the mtgRNA-induced DSB conditions. Whereas mtgRNA-induced DSBs generated early gene-expression changes, a subset of which persisted to 2 weeks, radiation produced comparatively few differentially expressed genes at 3 h but a larger delayed response at 2 weeks, with many genes becoming upregulated at this later time point (Supplementary Fig. 2D–F and Supplementary Data 1–4). Thus, radiation did not simply reproduce the transcriptional dynamics observed after recurrent, pre-mapped Cas9-induced DSBs. This distinction is expected because radiation introduces stochastic DNA damage across cells and generates multiple forms of cellular injury beyond DSBs, including single-strand breaks, base damage, oxidative damage, and broader effects on RNA, proteins, lipids, organelles, and stress-response pathways.

To quantify the relationship between distance to mtgRNA cut sites and differential expression, we performed logistic regression using differential expression status (P value < 0.05) as the binary outcome and distance to the nearest cut site as the predictor. Odds ratios (ORs) with 95% confidence intervals were derived from the regression coefficients (Fig. 1E). In nearly all conditions and time points, with the exception of mtg142 at 3 h, odds ratios were significantly less than 1, indicating that genes closer to cut sites were more likely to be differentially expressed. Gene-set enrichment analysis of mtgRNA RNA-seq data highlighted DNA damage repair pathways and, at both time points, processes related to chromatin and DNA modification (Supplementary Fig. 3A-D and Supplementary Data 5-12).

To determine the impact of DNA and chromatin-modifying processes, we performed genome-wide DNA methylation profiling and ChIP-seq for H3K27ac (Fig. 1F and Supplementary Fig. 4). Notably, the DNA methylation profiles at 3 h and 2 weeks were broadly similar, suggesting that at least a subset of these changes is established rapidly and persists over time. Genome-wide DNA damage biased methylation changes toward hypermethylation, with enrichment at distal regulatory elements and depletion at promoter-proximal elements, including promoters and proximal enhancers. Pearson correlation analysis was used to compare TAD-aggregated log-fold changes between mtg142- and mtg483-treated cells across gene expression, H3K27ac, DNA methylation, and Hi-C cis contacts at 3 h and 2 weeks, revealing both positive and negative relationships among these measurements (Fig. 1F). Because these comparisons are based on genomic regions aggregated at the TAD level, the correlations between H3K27ac, DNA methylation, and gene-expression changes should be interpreted as broad regional relationships rather than direct mechanistic links between individual regulatory elements and specific genes. Together, these findings indicate that DSB induction leads to context-specific rewiring of the genome at the level of 3D interactions, chromatin and DNA modifications, and gene expression, and that aspects of this reprogramming persist for at least 2 weeks. They also show that some changes detected at 3 h are distinct from the longer-term alterations observed at 2 weeks. Collectively, these data suggest that DNA damage engages chromatin-modifying processes both acutely and persistently.

Transient increases in cis chromatin contacts are followed by a delayed increase in trans chromatin interactions and compartment A-B intermingling

To investigate the connections between changes in 3D chromatin folding and chromatin activity, we used the H3K27ac and Hi-C data to determine whether shifting between transcriptionally active (A) and inactive (B) compartments occurred after the introduction of DSBs (Supplementary Fig. 5A–C). Despite differences in the number and location of targeted DSBs, similar broad mega-base-sized changes were detected in cells expressing either of the two different guides (Fig. 2A). A/B compartment traces were largely unchanged in placement between control and DSB-induced samples. However, it is noteworthy that the strength of the compartment score was attenuated upon the induction of DSBs, suggesting the dissolution of certain B compartments at some genomic locations. As shown in the example for chromosome 17, these were mostly positive compartment shifts, associated with mega-base scale alterations in chromatin contacts (Fig. 2B and Supplementary Fig. 5D). Compartment shifts can be associated with changes in regulatory marks, as illustrated in Fig. 2C, which shows an overall increase in H3K27ac, DNA methylation, and sustained gene-expression changes. Although major switches between compartments were not observed, a significant portion of compartments had score shifts (absolute change >0.2) that were widespread across the genome (Fig. 2D). Compartments with score shifts were also more prone to have alterations in chromatin contacts, DNA methylation and gene-expression, in contrast to unchanged compartments (compartments without a score shift at any timepoint) (Supplementary Fig. 5F).

Fig. 2. Transient increases in cis chromatin contacts are followed by a delayed increase in trans chromatin interactions and compartment A-B intermingling.

A Compartment scores over chromosome 17 for mtg142 and mtg483. Arrows point to areas of the genome that have a positive shift in compartment score. B Contact maps (HiC) (2 weeks vs. control) show similar patterns of alterations. Red circles highlight areas of the genome that have significant changes in long-range cis contacts near areas of cut sites. C Double-strand breaks induced by two different mtgRNAs are associated with increased surrounding H3K27ac, DNA hypermethylation, and longitudinal changes in gene expression across a 2 Mb region on chromosome 17. Values shown represent treated-minus-control differences in normalized signal, not log-fold changes, and may therefore be either positive or negative. D Proportion of 1 Mb genomic bins showing compartment score shifts relative to the untreated parental/control sample at 3 h and 2 weeks after mtg142 or mtg483 induction. Bins were classified by the change in compartment score relative to control (Δ > 0.2, more A-like; Δ < −0.2, more B-like). The proportion of bins in each category was calculated, and bars show mean ± SEM. Statistics were assessed by ANOVA with two-sided Dunnett’s multiple-comparisons test; **P < 0.0001; ns, not significant, n = 2 replicates per mtgRNA). E Saddle plots representing interaction score differences between A and B compartments across the genome. F Quantification of A/B interaction scores calculated along chromosome arms (dots). The data is represented as a ratio of treated/untreated (two-tailed one-sample t-test, * P value < 0.01, ns = not significant, n = 2 replicates per mtgRNA). G Contact proportions binned into cis or trans found in A compartments (ANOVA and two-sided Dunnett’s test, ** P value < 0.01, **** P value < 0.0001, bars show mean ± SEM, ns = not significant, n = 2 replicates per mtgRNA).

These results imply that the introduction of double-strand breaks induces sustained dissolution of A and B compartments, where regions originally maintained in separate domains begin to intermingle. Importantly, in cells expressing guide RNAs mtg142 or mtg483, saddle plots show that changes in A-B compartment interactions show subtle differences at 3 h, but by 2 weeks reveal significant increases, indicating an increase in A-B compartment intermingling that reflects changes in DNA and chromatin modifications and alterations in gene expression (Fig. 2E, F and Supplementary Fig. 6A, B). We further analyzed changes in cis and trans contacts in mtgRNA-treated samples. Cis contacts were transiently increased at 3 h and then decreased by 2 weeks, whereas trans contacts showed the opposite trend, with a higher final proportion of interactions in both A and B compartments (Fig. 2G and Supplementary Fig. 6C). A similar pattern was observed after radiotherapy, including diminished B-B interactions over time (Supplementary Fig. 6D, E). In radiation-treated samples, we found that B-compartment shifts were significantly elevated by 2 weeks (Supplementary Fig. 6F), but did not see differences in contact proportions between A/B compartments in cis or trans, which we hypothesize is likely due to the random nature of radiation damage that varies from cell-to-cell in this bulk population experiment.

In summary, induction of multiple double-strand breaks leads to transient increases in cis chromatin contacts associated with minimal compartment mixing, early sustained mega-base scale alterations in DNA methylation, H3K27 acetylation and gene expression (Supplementary Fig. 6G)²⁴. Persistent changes are accompanied by a delayed increase in trans chromatin interactions and a more prominent shift in compartment A-B intermingling.

Reprogramming of long-range 3D genome contacts and regulatory elements following DNA damage over time are associated with persistent alterations in gene-expression

To investigate the connection between changes in chromatin folding and gene expression at granular resolution, we first analyzed intra- and inter-TAD contacts over time. As expected, control cells exhibited more frequent intra-TAD than inter-TAD interactions, consistent with the insulating properties of TADs, a configuration that persisted shortly after DNA damage. However, by 2 weeks post-induction, we observed a marked and reproducible increase in long-range inter-TAD contacts spanning 10–20 Mb in both mtgRNA conditions (Fig. 3A and Supplementary Fig. 7A, B). Notably, this increase was only seen in cells that had DNA breaks, indicating that DSBs themselves drive these long-range contacts. Aggregate TAD analysis further revealed a transient increase in intra-TAD contacts at 3 h (consistent with prior observations over hours and days following DNA damage by other groups^1,17,25,26). This was followed by a consistent decrease at 2 weeks (Supplementary Fig. 7Ci), suggesting a temporal shift from local to more extensive chromatin reorganization. We also performed a parallel aggregate TAD analysis after radiation treatment. Similar to the mtg142 or mtg483 conditions, radiation-treated cells showed increased aggregate intra-TAD contact signal at 3 h. By 2 weeks, contacts had largely gone back to normal levels, in contrast to the mtgRNA-treated samples, which had reduced contacts at this time point (Supplementary Fig. 7Cii).

Fig. 3. Reprogramming of long-range 3D genome contacts and regulatory elements following DNA damage over time are associated with persistent alterations in gene-expression.

A Quantification of cis-contacts from TADs with DSBs that cross TAD borders over time. The inset shows 500 random TADs that do not have DSBs for comparison. B Differential cis-contacts ordered by distance to cut site show long-range (mega-base) alterations from the nearest cut site (LogFC > 1.2 and P value < 0.05). C Normalized cut-site–to–cut-site cis and trans contacts within 5 kb of verified cut sites. The axs represent the ordered list of cut sites by chromosome and genomic position, rather than genomic distance. Circles on the diagonal represent cis contacts for matched cut-site pairs, with circle size indicating the number of contacts. Off-diagonal circles represent trans contacts between cut sites on different chromosomes. D Trans-contact heatmaps using 500 kb trans-contact windows around cut sites. i Trans-contact heatmap-of-heatmaps (10 Mb regions) between cut sites for selected locations (mtg142 condition). ii Trans-contact heatmap between a single region over time (10 Mb regions), which is highlighted with a dash-lined box in (i). E Genomic enrichment analysis for differentially expressed genes with differential genomic contacts. F Genomic enrichment analysis for differential genomic contacts with regulatory regions. G Significantly up- and down-regulated genes over time. Alluvial plot shows the number of genes that are unchanged and remain differentially expressed as compared to control over time (LogFC > 1.2, P value < 0.05).

Consistent with these findings, differential analysis of cis contacts showed significant alterations, including long-range contacts that persisted over time in cells expressing guide RNAs mtg142 or mtg483 (Fig. 3B and Supplementary Fig. 7D). However, these analyses did not distinguish between new contacts that form between cut sites, alterations that could be important during the DNA damage response²⁷. Using a 5 kb window and restricting the analysis to contacts between pairs of cut sites, we quantified normalized cut-site–to–cut-site interactions over time (Fig. 3C and Supplementary Fig. 8A). In this analysis, cut-site–to–cut-site cis and trans contacts increased over time. As controls, we examined 500 random genomic regions and the non-induced cut-site set corresponding to the alternate mtgRNA in each condition. Neither control showed the marked accumulation of cis and trans contacts observed at induced cut sites over time (Supplementary Fig. 8A), supporting the conclusion that these changes are preferentially associated with induced damage sites. In this analysis, trans contact changes were not as pronounced as cis contact changes, likely because these occur less frequently.

To visualize trans contacts more clearly, we expanded our window size to 500 kb around cut sites and were able to identify the formation of significant trans contacts around damaged regions, as shown in Fig. 3D and Supplementary Fig. 9A–D. These were less prevalent at 3 h compared to the 2-week time point, emphasizing how chromatin organization changes over time after the introduction of DSBs. Our data suggest that architectural disruptions can be actively induced by targeted DSBs alone, without an underlying genetic disorder. These findings echo recent discoveries of pathological chromatin rewiring in disease contexts such as fragile X syndrome and pediatric ependymoma. Specifically, BREACHes (Beacons of Repeat Expansion Anchored by Contacting Heterochromatin)⁵ and TULIPs (ultra-long-range B compartment contacts)⁴ were recently identified as large-scale 3D genome remodeling phenomena that emerge in disorders with intrinsic genome instability. These findings imply that DSBs may be sufficient to trigger a breach of chromatin domain boundaries and induce genome-wide reorganization.

To determine if changes in gene expression could be linked to changes in chromatin organization, we analyzed changes in interactions involving differentially expressed genes and found a significant enrichment at both 3 h and 2 weeks in comparison to all contact pairs (Fig. 3E). A similar analysis for regulatory elements revealed that promoters and enhancer sequences were enriched for differential genomic contacts (Fig. 3F). These enrichment analyses indicate that differential contacts preferentially involve genes and regulatory elements after DNA damage, although they do not by themselves assign specific contact changes to individual gene-expression outcomes. We found that a substantial fraction of early gene-expression changes persisted from the 3 h time-point to the 2-week time-point: 30.1% (n = 377) of mtg142-responsive genes and 70.8% (n = 1011) of mtg483-responsive genes that were altered at 3 h remained up- or downregulated at 2 weeks (Fig. 3G). These findings indicate that the introduction of DSBs leads to transcriptional reprogramming in which a subset of gene expression changes persists until the 2-week timepoint. Transcriptional reprogramming is initially associated with a transient increase in intra-TAD contacts at 3 h, but by 2 weeks, there is a shift towards the formation of new, longer-range, inter-TAD and trans contacts that correlate with increased A-B compartment mixing.

Large structural variations accumulate over time and correlate with contact frequency and gene density and activity

DSB repair can lead to alterations in the genetic sequence and, less frequently, large structural variations (SVs), depending on the fidelity of the repair pathway involved. It is known that the emergence of SVs can be influenced by the non-genetic or chromatin context of the break^27–29, but this is typically challenging to study in a systematic fashion after the random introduction of radiation induced DSBs in individual cells. We capitalized on the uniformity of this Cas9-mediated DSB approach to examine the chromatin contexts of newly generated SVs after DNA damage induction.

We designed biotinylated DNA probes using complementary sequences that flanked mtg142 or mtg483 cut sites and performed high-coverage targeted sequencing to profile mutations (Supplementary Fig. 10A)²⁴. An inherent challenge in this process is that mtgRNAs, by definition, target the same sequence in the genome, and there is some sequence degeneracy around the cut sites that limits the number of unique cross-compatible DNA probes that can be designed. Working within these constraints, we created two panels of 402 and 979 probes targeting 36% of mtg142 sites and 32% of mtg483 sites, respectively, with one to four probes flanking each cut site (Supplementary Fig. 10B, Supplementary Data 13 and 14). As shown in Supplementary Fig. 10C, we were able to detect small deletions and insertions at known cut sites, with sequencing depths with a median of ~1100X after pooling replicates. As expected, nearly all cut sites had some level of detectible small indels, as highlighted in Supplementary Fig. 10D.

Large structural variations (SVs) were detected at nearly every sequenced cut site, with translocations, inversions, duplications and deletions seldom exceeding a frequency of 5% at any given cut site (Fig. 4A). The total frequency of any large SV at a given DSB site is shown in Fig. 4B and Supplementary Fig. 11A, B. We did not detect any large insertions. Interestingly, we did not see any direct correlation between cut site efficiency (as assessed by isBLISS sequencing) and the frequency of SVs (Supplementary Fig. 11C), suggesting that SV formation was influenced by other factors. Since these cut sites all shared very similar genetic sequences, we examined the local chromatin environment of the SV anchor sequences.

Fig. 4. Large structural variations accumulate over time and correlate with contact frequency and gene density.

A Circular genomic plots representing large structural variations (SVs) over time between two cut sites (compared to the control condition). Lines are shaded by cumulative SV frequency at each locus, with the color scale spanning 0–5% because individual large SV frequencies in this dataset only rarely exceeded 5%. B Cumulative large SV frequency (y-axis) ordered by total frequency along the x-axis. Each stacked bar represents a cut site. The two-week time points are shown. C Pearson correlations between 2-week SV frequency (mtg142) and anchor-specific genomic features measured at control, 3 h, and 2 weeks. Rows denote features at each end of the SV breakpoint (Anchor 1 or 2); colored dots represent SV classes and significance. Features are ordered by mean correlation across SV types (two-sided, P < 0.05 using Student’s t-distribution transformation of correlation coefficients). D Scheme highlighting that large structural variations increase over time and correlate with contact frequency and gene density or activity following DNA damage²³.

For each SV category, we compared the genomic characteristics of surrounding regions of sequenced cut sites over time, and calculated correlation coefficients at either end of the SV (anchor 1 or 2) (Fig. 4C and Supplementary Fig. 11D). This analysis included a complete cross correlation of the number of genes in the regions, RNA-seq, Hi-C, H3K27ac, DNA methylation and SV frequency (Supplementary Fig. 12–S15). Because alterations or shifts in RNA-seq, Hi-C, H3K27ac and DNA methylation are distinct from high or low basal levels, we included both log fold-change and average levels at all timepoints as separate variables. Although our results show statistically significant correlations, no single genomic characteristic alone could strongly predict SV formation. However, a high frequency of cis or trans contacts and high gene-density or activity within one or both anchors emerge as characteristics associated with their emergence (Fig. 4D)²³.

Discussion

In this study, we used pre-mapped DNA double-strand breaks (DSBs) in human glioblastoma stem cell lines to systematically track genetic and non-genetic changes over time. Our findings demonstrate that DNA damage can precipitate mega-base scale alterations in DNA and chromatin modifications, cis and trans genome interactions and change gene expression. Notably, these shifts remain detectable weeks after the initial DSBs, indicating that the cellular response to DNA damage can go far beyond the generation of transient repair phenotypes, with significant implications for cancer biology and normal cellular function.

While it is known that chromatin structure and gene expression change in the context of DNA damage stress^3,7,17,25,26, prior studies have focused on acute responses^1,2,9,26, typically within hours of damage, or have examined long-term consequences in clinical samples without a means to associate observed chromatic and gene-expression changes to specific, traceable DSB events^6,7,16,17. Our data builds on these studies, demonstrating the importance of progressing states that follow damage over time in cancer cells. Here we have shown that genomic reprogramming involves a transient gain of intra-TAD and cis interactions and the delayed formation of new long-range inter-TAD and trans contacts, that correlate with durable changes in DNA methylation, histone acetylation, gene expression and the emergence of large structural variations, all of which have important functional consequences.

The mechanisms by which widespread DSBs drive persistent 3D genome reprogramming remain unclear. DNA damage may trigger transient chromatin remodeling through the coordinated action of DNA damage response factors, chromatin remodelers, and transcriptional changes, which could initially alter contact probabilities and later become stabilized. Alternatively, damage may bias genome folding through shifts in compartment identity or broader architectural organization. Our current data do not distinguish between these possibilities. H3K27ac was profiled here as a broad indicator of active chromatin state and to support compartment annotation, rather than as a specific long-term mark of prior DNA damage; defining whether other activating or repressive chromatin pathways more directly stabilize persistent post-damage states will require further study. Although both mtgRNAs produced a shared temporal pattern of genome reprogramming, they diverged in their longer-term consequences. mtg142 was associated with stronger compartment-linked and long-range cis reorganization, whereas mtg483 retained a larger fraction of early gene-expression changes at 2 weeks. This divergence suggests that break density and genomic distribution can influence the type of persistent state that follows DNA damage.

More broadly, our findings suggest that the capacity of DSBs to induce 3D architectural reorganization may represent a unifying mechanism across both experimental and pathological contexts. While prior studies of fragile X syndrome and pediatric ependymoma have linked chromatin misfolding to disease phenotypes via structures like BREACHes⁵ and TULIPs⁴, our results indicate that reorganization can be context independent. This implies that chromatin architecture is not merely a passive reflection of sequence alterations or disease states, but it can be actively reshaped by damage itself. The ability of targeted DSBs to provoke inter-compartment mixing and long-range contact remodeling suggests a generalizable vulnerability of genome architecture, with potential relevance for understanding how genome topology contributes to plasticity, repair, and cellular identity.

Although our study focuses on widespread DSBs in a cancer and therapy context, DSBs also arise during normal cellular processes, including transcription, development, and signaling. For example, during lymphocyte development, programmed DSBs are introduced at antigen receptor loci, and the number of breaks per cell is tightly controlled to limit widespread damage and preserve genome integrity^30,31. In the nervous system, DSBs have also been reported during neural development and in post-mitotic neurons, where activity-induced DNA damage contributes to the expression of early response genes important for plasticity and memory formation^32–34. In post-mitotic neurons, activity-induced DNA damage contributes to the expression of early response genes important for plasticity and memory formation. However, whether physiological DSBs can drive persistent chromatin, architectural, or transcriptional remodeling remains incompletely understood.

Despite these growing insights, several limitations and questions remain. First, consideration should be given to the type and variety of damage. Inducing DSBs via multi-target CRISPR–Cas9 represents a specific type of controlled, damage, whereas clinical DNA insults (e.g., from irradiation or chemotherapy), or random damage from genomic instability, are often more stochastic and can vary widely in, number, density, genomic context and type of insult. Because the lesions studied here are Cas9-induced DSBs, they are likely repaired predominantly through end-joining pathways, although the pathway used at any given site may also depend on cell-cycle state, chromatin context, and locus-specific features. Importantly, our data define persistent chromatin and transcriptional outcomes following an initial mapped burst of DSB induction, but do not directly distinguish whether the later phenotypes reflect stable downstream consequences of those breaks or any contribution from ongoing DNA damage. Such variability may accentuate or dampen genomic reprogramming, and future work comparing different damage models could clarify a threshold at which persistent changes become locked in. Because matched normal DNA was not available for these patient-derived samples, we could not reliably determine whether variants in DNA repair genes were somatic and therefore did not interpret the observed phenotypes in terms of defined repair-pathway mutations. Additionally, while our focus centered on glioblastoma, it is unclear how broadly reprogramming events apply to other cancer types, disease states, or to nonmalignant cells, whose chromatin landscapes may respond differently to the introduction of widespread DNA damage.

Our data has broad implications for cancer biology as they suggest that cells sustaining mass DNA damage from genotoxic stress may adopt new chromatin configurations and gene expression programs that promote survival or drive therapy resistance. Understanding the precise pathways that stabilize reprogramming events could enable the development of therapeutic strategies aimed at preventing or reversing their occurrence to curb the emergence of resistant cell states. Overall, the findings presented in this study highlight DNA damage–induced chromatin rewiring as a potentially powerful mechanism underlying cancer cell plasticity, and they underscore the need to integrate studies of DNA repair with a deeper exploration of the longitudinal effects of damage.

Methods

Cell culture and multi-target guide RNA expression

Patient-derived glioma stem-like cultures (GSCs) were cultured in serum-free media as floating neurosphere cultures without doxycycline to enrich for stem-like states, as detailed in a publication from²². Cells were passaged and fed at the same time, and are not synchronized by any chemical or sorting means. was used for all mtgRNA experiments, unless stated otherwise, to compare timepoints and different multi-target guide RNAs. U87 human glioma cells were utilized for clonogenic assays only. U87 glioma cells were grown in DMEM containing 10% FBS (doxycycline-free). Lentivirus generation was performed using HEK 293 T cells grown in DMEM containing 10% FBS.

Multi-target guide RNA sequences (including the PAM sequence) are as follows: mtg483: 5’-AGTGGGTGCAGCCCATGG, mtg142: 5’-GGCTTCCAGGTCACAGGT. Guide RNAs were cloned into 3^rd-generation lentiviral vectors using standard techniques. The guide RNA was housed in a doxycycline-inducible construct, FgH1tUTCherry (addgene #85552). To generate stable cell lines, they were initially infected with lentivirus containing human codon-optimized S. pyogenes Cas9 protein with blasticidin resistance (lentiCas9-Blast, addgene #52962) and selected with Blasticidin for 7 days. Following lentiviral transfection with mtg142-, mtg483- or empty-FgH1tUTCherry, cells were sorted via FACS and cultured further for downstream experiments. For all experiments, to induce expression of an mtgRNA, cells were treated with 10 ng/uL of doxycycline for 3 h, followed by two centrifugation washes with media to remove doxycycline at the 3-h timepoint. They were then grown continuously in serum-free media under standard conditions.

Immunofluorescence

Glioma stem cells were dissociated into single-cell suspensions and seeded onto coverslips pre-coated with laminin and poly-L-ornithine. Following the specified treatments, cells were fixed in 4% paraformaldehyde (PFA), then blocked with 2% bovine serum albumin (BSA). Immunostaining was performed with a primary antibody against γH2AX (Millipore 05-636), and cell nuclei were counterstained with DAPI. Coverslips were subsequently mounted and imaged by fluorescence microscopy.

Clonogenic assays

Human U87 glioma cells were seeded for clonogenic assays and either exposed to radiation or subjected to mtgRNA induction. Radiation was performed using a CellRad per the manufacturer's instructions. After one week of incubation, surviving colonies were fixed and stained with crystal violet. Each condition was performed in triplicate (n = 3) to ensure reproducibility of the results.

Extreme Limiting Dilution Assay (ELDA)

GSC20 cells modified to stably express Cas9 and an empty gRNA, or mtgRNAs as detailed here in the methods section, were first treated with 10 µg/mL Doxycycline for 3 h. After treatment, the cells were washed with 1X PBS twice and dissociated with Accutase cell detachment solution for 5 min at 37 °C. Cells were then resuspended in Neural Basal Media (NBM) and filtered through a 100 μM cell strainer. Cell number and viability were then obtained via a Vi-Cell Blu Analyzer. 96-well plates were seeded at decreasing densities (100, 75, 50, 25, 20, 15, 10, and 5 cells/well), then supplemented with the appropriate volume of NBM to bring the total volume to 200 μL. Radiation (4 Gy in one dose) was performed using a CellRad per the manufacturer's instructions. Mock treatments were handled the same, but not put in the irradiator. Cells were then allowed to recover for 7 days before imaging at 4X magnification using brightfield imaging on the Evos M7000, and were analyzed in ImageJ. Further extreme limiting dilution analysis (ELDA) was performed in R using the statmod package to estimate stem cell frequency by fitting a single-hit Poisson model to limiting dilution data. Statistical significance between groups was determined using chi-square tests, including pairwise comparisons where indicated. ELDA curves were generated in R, and estimated stem cell frequencies with corresponding 95% confidence intervals were exported and plotted in GraphPad Prism.

In-suspension breaks labeling in situ and sequencing (sBLISS or isBLISS) and double-strand break definitions

Double-strand break sequencing in-suspension breaks labeling in situ and sequencing, or sBLISS or isBLISS) was performed (n = 2) as described in detail in Bouman et al.³⁵. Summarized here, the protocol begins with the crosslinking of cells with 2% PFA. Cells are lysed to expose the DSB ends, which are subsequently blunted in situ and ligated to the sBLISS adapter. Genomic DNA is then extracted and sonicated. Library preparation proceeds with in vitro transcription for linear amplification of the template. The library is then ligated to a new adapter and annealed to a complementary oligonucleotide to enable reverse transcription of each template. The entire library is amplified with sequencing primers that include indexes and unique barcodes. Following quantification and quality control checks via tapestation analysis, the libraries are sequenced via an Illumina NovoSeq6000 with 10% PhiX spiked in.

Raw sequencing data were first processed according to the author's instructions. Briefly, the blissNP pipeline³⁶ was used to generate a BED file that averaged both replicates with counts of every unique double-strand break that had been sequenced. Signals from the isBLISS data were normalized to achieve a mean signal of 1 across all bases. We calculated the total signal by multiplying each region’s length by its signal value and dividing by the total number of bases to determine the current average signal. We then applied a normalization factor to each region’s signal, bringing the average signal across each sample per base to 1.

To define the final set of mtgRNA-CRISPR-cas9 cut sites we defined a potential ‘universe’ of theoretical guide-RNA landing sides that includes a perfect match (0 mis-matches to the genome with the guide RNA sequence), 1 mis-match (1 nucleotide is not matched) or 2 mis-matches (2 nucleotides do not match), up to two RNA bulges on the genome and up to two DNA bulges on the guideRNA. This universe of potential target sites was intersected with the normalized isBLISS signals to only include sites that had a double-strand break, yielding 339 cut sites for mtg142 and 1386 sites for mtg483. Cut sites were visualized either directly by IGV (version 2.18) or by circlize (version 0.4.16) in R (version 4.1.2).

Transcriptomic profiling

RNA was extracted using the AllPrep genomic DNA/RNA extraction kit (Qiagen). Experiments were performed in duplicates (n = 2). RNA was subject to TapeStation for RIN quality assessment, with a minimum quality score of 9 required to proceed with library preparation. RNA libraries were prepared with TruSeq Stranded Total RNA Gold for whole-transcriptome sequencing according to manufacturer protocols. In brief, unwanted rRNA was depleted, mRNA/cRNA was converted to double-stranded cDNA, end-repaired, adapter-ligated, and PCR-amplified. The final libraries were purified, quantified, and assessed for integrity prior to sequencing with a TapeStation. Sequencing multiplexed and performed on a NovoSeq 6000 with a goal of achieving at least 75 million reads per sample using paired-end 50 bp sequencing.

For data processing, adapters along with low-quality bases were trimmed using Trimmomatic. The resulting reads were aligned to the human reference genome (hg38) with STAR, and additional alignments were performed against other species and common contaminants for quality control. Normalized genome browser tracks were generated to visualize coverage, and the 5′/3′ bias was assessed with Picard. Gene-level counts were then summarized to create a genes-by-samples matrix, and differential gene expression analysis was conducted using standard DESeq2 (version 1.46.0) procedures.

Gene set enrichment analysis was performed using the fgsea package. Genes were ranked by the −log₁₀(P value) signed by the direction of the log₂ fold change from the DESeq2 differential expression analysis, and the resulting pre-ranked list was tested against the Reactome canonical pathway gene sets from the Molecular Signatures Database (MSigDB collection C2:CP:REACTOME). Enrichment was assessed with the fgseaMultilevel procedure. For each pathway, an enrichment score (ES) and a normalized enrichment score (NES, accounting for gene set size) were computed. P values were corrected for multiple testing across all tested pathways using the Benjamini–Hochberg false discovery rate.

DNA-methylation profiling

DNA was extracted using the Qiagen AllPrep genomic DNA/RNA extraction kit (Qiagen). Experiments were performed in duplicates (n = 2). Genomic DNA was processed using the TruSeq Methyl Capture EPIC Library Prep Kit (Illumina) according to the manufacturer’s protocol. Briefly, high-quality DNA (as determined by tapestation analysis) was fragmented via a Covaris S220, followed by end-repair, 3′ adenylation, and ligation of methylation-compatible adapters. The resulting fragments underwent bisulfite conversion to convert unmethylated cytosines to uracils, while preserving methylated cytosines. The libraries were then hybridized to the TruSeq Methyl Capture EPIC probes for targeted enrichment of methylated regions, followed by bead-based capture. Finally, the enriched libraries were PCR-amplified, purified, and assessed for quality with a tapestation prior to sequencing. Sequencing multiplexed and performed on an Illumina NovoSeq 6000.

Partial genome bisulfite sequencing data from TruSeq Methyl Capture were trimmed for adapters and low-quality bases using Trimmomatic, then aligned to the bisulfite-converted hg38 reference genome via Bismark and Bowtie2. Duplicate reads were removed in Bismark, and methylation calls were subsequently extracted. Further processing was conducted in methylKit, beginning with coverage-based filtering to remove bases below 10× coverage and above the 99.9th percentile. Samples were merged and used to calculate percent methylation. Differential methylation analyses were performed on the merged data, applying thresholds for q-values (e.g., <0.01) and methylation differences (≥25%), and the resulting differentially methylated bases or regions were stored for further downstream interpretation.

Chromatin Immunoprecipitation

ChIP–seq (n = 2 per condition) was performed essentially as described in³⁷ using a dual crosslinking method and in situ tagmentation. In brief, ~1 million cells per replicate were first crosslinked in 1.5 mM ethylene glycol bis(succinimidyl succinate) (EGS) for 30 min, followed by 1% formaldehyde for 10 min, then quenched with 125 mM glycine. After washing, cell pellets were snap-frozen in liquid nitrogen and stored at −80 °C. Pellets were lysed in the presence of protease and deacetylase inhibitors, and the chromatin was sonicated to the desired size range. Drosophila spike-in chromatin (Active Motif) was added at 2% by mass for normalization, and immunoprecipitation was carried out with protein A magnetic beads (ThermoScientific #10001D) conjugated to H3K27ac (2.5 uL per 1 × 10⁶ cells) (Abcam #ab4729), or spike-in antibody (2% of ug of target Ab, Active Motif #104597). After extensive washing, in situ tagmentation was performed on bead-bound chromatin to incorporate sequencing adapters, and crosslinks were reversed by overnight proteinase K treatment. DNA was purified by column-based methods, then PCR-amplified to produce the final libraries. Libraries were size-selected with AMPure beads, checked for concentration and fragment distribution (Qubit and Tapestation), and quantified by qPCR prior to multiplexed sequencing on an Illumina NovaSeq 6000. Samples were sequenced to 40–60 million reads each. Adapter trimming was performed using fastp, followed by alignment to the hg38 reference genome using Bowtie2 with default parameters. PCR duplicates were removed using sambamba. Normalized genome browser tracks (BigWig format) were generated using deeptools for visualization. Peak calling was performed in a treatment-versus-control framework using MACS2, with peak type specified according to histone mark characteristics. Normalized signal tracks were generated using deepTools (bamCoverage) with reads per genomic content (RPGC) normalization.

Genomic enrichment analysis

To perform enrichment testing, we performed genomic enrichment analysis using a Locus Overlap Analysis (LOLA)³⁸. This method uses a Fisher’s exact test with false discovery rate correction to see if segments of the genome marked by differential Hi-C contacts, or gene-expression, are statistically enriched over other parts of the genome in a pairwise manner. Cis-regulatory element definitions are defined by ENCODE³⁹. Processing of all other genomic measures (RNA-seq and Hi-C contacts) is described elsewhere in the methods section.

Genomic contact (genome-wide chromosome conformation capture) sequencing and data processing

For Hi-C (genome-wide chromatin contact conformation), the ARIMA Standard Hi-C Kit was used following the manufacturer's instructions. In brief, cells were crosslinked with 2% formaldehyde for 10 min at room temperature, EGS (final concentration: 1.5 mM) and the reaction was quenched with glycine. Chromatin was digested with the restriction enzyme mixture provided in the ARIMA Standard Hi-C Kit, followed by a biotin fill-in step and proximity ligation according to the manufacturer’s protocol. Experiments were performed in duplicates (n = 2). After reversing cross-links and purifying the DNA, adapters were ligated to the fragment ends, and library amplification was performed. Sequencing multiplexed and performed on an Illumina NovoSeq 6000.

Standard pipelines were used to process cis-genome Hi-C contact data unless noted otherwise. Hi-C Bench was used in conjunction with standard bioinformatics tools to ensure reproducibility. Alignment was performed with bwa (version 0.7.17) using settings -A1 -B4 -E50 -L0. After alignment, reads were filtered with samtools (version 1.3) and gtools (version 3.0) using a mapping quality threshold of 20, no minimum distance, and a maximum offset of 500. Duplicate reads were also removed during this step. Filtered contact pairs were converted to format using juicer (version 1.5) and normalized using Knight-Ruiz (KR) normalization. These were then converted to cool format using hicexplorer (version 3.7.2) in various resolutions for further visualization and processing in GENOVA (version 1.0.0).

Hi-C data visualization with GENOVA using cool file outputs included the following steps. Centromeres were masked from analysis. Insulation scores for Topologically Associated Domain (TAD) inference were called using hic resolutions of 20 kb with replicates combined. Compartments were called using hic resolutions of 500 kb using combined replicates, and the signing of compartments was determined using H3K27ac ChIP-seq data. Interaction scores between arms of chromosomes were visualized in saddle plots, and data were exported for quantification.

Cross TAD-boundary cis-contact analysis: Cross-region contacts were quantified from the downsampled, deduplicated, sorted, and indexed BAM files (normalized to a common read depth) using the GxG R package together with bamUtils, rtracklayer, and GenomicRanges. Three region sets were analyzed: 5-kb windows centered on the 142 and 483 multi-target sgRNA cut sites, and 500 random 5-kb genomic windows (hg38), each reduced (GenomicRanges::reduce) to a non-redundant set of tiles. For each library, reads within tiles were imported (read.bam) and inter-tile contacts were enumerated by co-counting read pairs sharing a read name across distinct tiles (cocount, by = “qname”), yielding tile-pair contact counts. Self-contacts were removed, and for each remaining intrachromosomal (cis) contact, the anchor-to-anchor distance was computed as the absolute difference in anchor start positions; interchromosomal (trans) contacts were excluded. Each cis contact was then represented as a GRanges interval spanning its two anchors, and the number of TAD boundaries falling within each span was tallied by overlap (findOverlaps) against the TAD boundaries called from the GENOVA insulation-score analysis (20-kb resolution). After combining replicates, per-contact cross-boundary counts were aggregated by region set and condition and plotted as a function of the number of cis-TAD boundaries crossed (dplyr, tidyr, ggplot2).

Data processing of trans chromatin interactions from Hi-C data

Trans-chromatin contact maps were constructed at 1 Mb resolution separately for each replicate using Hi-C valid pairs from HiC-Pro (version 3.0.0) output. HiC-Pro processing options were as follows: ligation/digestion sequences (GATCGATC, GANTGATC, GANTANTC, GATCANTC), insert size range 100–800 bp, min and max fragment sizes set to 10 and 1,000,000, chromosomes XY were excluded and hg38 was used as the reference genome. Time point replicates were merged into a single chromatin map to reduce sparsity and enhance detection of trans interaction frequencies. Regions containing outliers were removed, and the contact maps were balanced using the Knight-Ruiz algorithm with default parameters⁴⁰. Subsequently, the contact maps were scaled using Hi-C contact distance-dependent median-of-ratios size factors to bring all samples to the same scale and enable direct comparisons⁴¹. For each pair of cut sites, trans contact maps were visualized with a 5 Mb window on either side. For downstream analysis, the matrix corresponding to the trans contact map with a 1 Mb window on either side was extracted. Summation and averaging of the matrices were performed, followed by transformation into a one-dimensional vector. Finally, P values were calculated using the Mann–Whitney U test.

DNA probe design and pull-down for deep sequencing of mtgRNA target sites

Two custom biotin-tagged capture panels (xGen DNA panels, Integrated DNA Technologies) were designed to enrich specific DNA regions, each featuring 120 bp probes with one to four probes flanking each targeted cut site, see Supplemental Data 13 and 14. One panel contained 402 probes (mtg142), and another contained 979 probes (mtg483). The degeneracy of genomic sequences around cut sites prevented the design of probes with high enough specificity to target all sites, therefore only probe/site combinations which could be targeted without significant overlap were included in the panel. Following library construction, the prepared libraries were hybridized with the panels according to the manufacturer’s guidelines, subjected to stringent washes to remove nonspecific fragments, and subsequently PCR-amplified. The enriched libraries were purified and assessed for quality with a TapeStation prior to sequencing. Sequencing was multiplexed and performed on a MiSeq 2000 with 10% PhiX spike-in.

Small mutation and large structural variant calling

To assess the frequency and distribution of small insertions and deletions (indels) and large structural variants (SVs) arising from CRISPR–Cas9-induced DNA double-strand breaks, we used separate, tailored analysis pipelines based on the sequencing assay. In both analysis pipelines, we constricted the analysis to verified cut sites and the pulled-down regions that achieved coverage levels exceeding 100X. ENCODE-defined blacklisted regions were filtered out. For small indel analysis, we applied the CRISPR-DART (Downstream Analysis and Reporting Tool) pipeline⁴². Paired-end Illumina reads were subjected to quality control using FastQC (version 0.11.7) and trimmed with Trim Galore, followed by alignment to the human reference genome (hg38) using BBMap. To enhance detection sensitivity, reads overlapping putative indels were realigned using the Genome Analysis Toolkit GATK (version 3.8.0) for local realignment around indels and base quality score recalibration. Indel frequency profiles were visualized using IGV (version 2.18) and aggregated into heatmaps with deeptools (v3.5.1) using BigWig files output by crispr-DART.

For large SV analysis, we followed a parallel preprocessing workflow. Raw paired-end reads underwent quality control using FastQC (version 0.11.7) and adapter trimming via Trimmomatic (version 0.36). Reads were aligned to hg38 with BWA-MEM (version 0.7.17), and duplicates were removed using sambamba (version 0.6.8). BAM files were processed with GATK (v3.8.0) for local realignment around indels and base quality score recalibration. These realigned and recalibrated BAMs were used for structural variant calling with DELLY (version 0.8.1). DELLY identified deletions, duplications, inversions, and translocations using split-read and paired-end evidence from tumor-versus-control BAM comparisons. Output.bcf files were converted to.vcf using bcftools, and high-confidence SVs were annotated and visualized using Circos and other genome plotting tools.

Correlation of structural variant (SV) relationships with genomic features

Pearson correlation coefficients were calculated between well-covered cut-sites from DNA-pull-down sequencing. Large Structural Variation (SV) frequencies and 58 genomic features across TADs spanning timepoints with cis-chromosomal contacts (Hi-C), RNA-seq attributes, histone modifications (H3K27ac), and DNA methylation attributes (processed as described in a separate section). For each trans-contact genomic region with valid data (a sequenced cut-site), we calculated the mean Log₂ fold-change across all overlapping observations at each time point. One-sample t-tests (comparing the sample mean to zero) were then performed using the stats (version 3.6.2) package in R, yielding a P value for each region/timepoint combination. We applied the Benjamini–Hochberg (BH) correction to control the false discovery rate (FDR).

Data were analyzed separately for two independent conditions (mtg142 and mtg483) at three timepoints (baseline, 3 h and 2 weeks post-treatment). For each SV type (deletions, duplications, inversions, insertions, and translocations), correlations were computed using pairwise complete observations to handle missing values. Statistical analysis was performed using R (version 4.1.2). Statistical significance was assessed using correlation tests with P-values adjusted for multiple comparisons (significance threshold P < 0.05), and insignificant correlations were masked in the visualization as noted in the figures. To identify patterns of relationships between genomic features and structural variant frequencies, we employed hierarchical clustering with complete linkage to organize correlation matrices. The resulting correlation matrices were clustered using a distance metric defined as d = sqrt(2(1–r)), where r is the correlation coefficient, producing a dendrogram-based organization of features. This clustering approach grouped features with similar correlation patterns together.

Correlation of genomic features between genomic regions

To assess the relationship between genomic distance to the cut site and differential expression, logistic regression was performed with differential expression status (P value < 0.05) as the binary outcome and distance as the predictor. Odds ratios (OR) with 95% confidence intervals were derived by exponentiating the regression coefficients. P values were obtained from the Wald test for the distance coefficient. An OR < 1 indicates that genes closer to the cut site have higher odds of differential expression.

To compare and correlate changes in multiple types of genomic data (RNA-seq, ChIP-seq, DNA methylation, Hi-C genomic contacts), the genome was first segmented into regions defined by Topologically Associated Domains (as defined elsewhere in this methods section). Genomic features were averaged within that domain, and average-log-fold change was calculated per region, with statistics performed to assess significance. This was done as follows:

RNA-seq differential expression results (genes with log₂ fold changes and P values) were reduced to their midpoints to facilitate TAD overlap identification. For each TAD, all genes fully or partially residing within it were identified, and a mean log₂ fold change was calculated across those genes per region. We used Fisher’s method to combine individual gene-level P values into a single P value for each TAD. Briefly, all gene-level P values belonging to the TAD in question are converted to their natural logarithms and summed. The test statistic is then defined as −2 times this summed log(p). Under the null hypothesis, this statistic follows a chi-square (χ2) distribution with degrees of freedom equal to 2 × (number of P values). The final combined P value is calculated as the upper-tail probability of observing a chi-square statistic at least as large as the one computed.

H3K27ac (ChIP-seq) normalized bigwig files were converted to bedgraph format, and an averaged region score was extracted per TAD region. Similarly, DNA methylation was averaged across TAD regions. Per-condition comparisons for ChIP-seq and DNA methylation data were then tested using a linear model framework in the limma package (version 3.62.2) using the fitting of the model lmFit; relevant contrasts (e.g., “treatment vs. vector”) were defined for each experimental group, and empirical Bayes shrinkage was applied (eBayes). This approach yielded estimates of log fold-changes and corresponding P values for regional comparisons.

To identify differential changes in cis-contact frequency by TAD regions, a DESeq2 (version 1.46.0) analysis was performed, fitting a negative binomial model with timepoint as the main factor. Empirical analysis of variance was done with DESeq, and results were extracted for each contrast of interest (3 h vs. control, 2 weeks vs. untreated). Log₂ fold changes, P values, and adjusted P values (via Benjamini–Hochberg correction) were extracted by genomic coordinates.

Reporting summary

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

Author contributions

These authors contributed equally: K.C. and C.D. Conceptualization: A.S.M., J.P.C., J.S. and E.P.S. Methodology: A.S.M., C.D., J.P.C., J.S. and E.P.S. Investigation: A.S.M., M.J.G.Q., K.C., C.D., R.E., T.S., J.K., Y.D., Z.Z., M.G. and G.C. Visualization: A.S.M., K.C. and C.D. Funding acquisition: A.S.M., J.S. and E.P.S. Writing – original draft: A.S.M and J.S. Writing – review & editing: all authors.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

National Institutes of Health – NCI grant K08CA263302 (ASM). V Foundation Scholar Grant V2024-034 (ASM). Donald E. and Delia B. Baxter Foundation Fellowship Award (ASM). American Brian Tumor Association grant (ASM). RSNA Research Resident/Fellow Grant (ASM). NYU NIH CTSI Pilot Project Grant (NIH/NCATS UL1TR001445) (ASM). National Institutes of Health – NIGMS grant R35GM122515 (JS). National Institutes of Health – NCI grant P01CA229086 (JS). U01 DK127405 (JPC). U01 DA052715 (JPC). R37 MH120269 (JPC). R01 MH120269 (JPC).

Data availability

The genomic data generated in this study have been deposited in the NCBI’s Gene Expression Omnibus and are accessible through GEO Series accession number: GSE300218 (RNA-seq datasets of glioblastoma stem-like cells treated multi-target gRNAs), GSE300769 (Genome-wide Hi-C datasets of glioblastoma stem-like cells treated multi-target gRNAs), GSE300833 (isBLISS datasets of glioblastoma stem-like cells treated multi-target gRNAs), GSE300834 (bisulfite sequencing datasets of glioblastoma stem-like cells treated multi-target gRNAs), GSE300868 (H3K27ac ChIP-seq datasets of glioblastoma stem-like cells treated multi-target gRNAs), and GSE301130 (Targetted DNA sequencing of cut sites of glioblastoma stem-like cells treated multi-target gRNAs).

Code availability

Source code is available at the following public Bitbucket repository. (https://bitbucket.org/creminslab/modrek_et_al_nat_communications_2026/src/main/).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77331-4.