Three-dimensional regulatory hubs support oncogenic programs in glioblastoma

SUMMARY

Dysregulation of enhancer-promoter communication in the 3D nucleus is increasingly recognized as a potential driver of oncogenic programs. Here, we profiled the 3D enhancer-promoter networks of patient-derived glioblastoma stem cells to identify central regulatory nodes. We focused on hyperconnected 3D hubs and demonstrated that hub-interacting genes exhibit high and coordinated expression at the single-cell level and associate with oncogenic programs that distinguish glioblastoma from low-grade glioma. Epigenetic silencing of a recurrent hub—with an uncharacterized role in glioblastoma—was sufficient to cause downregulation of hub-connected genes, shifts in transcriptional states, and reduced clonogenicity. Integration of datasets across 16 cancers identified “universal” and cancer type-specific 3D hubs which enrich for oncogenic programs and factors associated with worse prognosis. Genetic alterations could explain only a small fraction of hub hyperconnectivity and increased activity. Overall, our study provides strong support for the potential central role of 3D regulatory hubs in controlling oncogenic programs and properties.

Graphical Abstract

eTOC

Here, Breves et al. profile the 3D enhancer-promoter connectomes of patient-derived human glioblastoma stem cells (GSCs), identify hyperconnected 3D regulatory “hubs” relevant for GBM biology, examine the impact of 3D hub perturbation on the transcriptional program and oncogenic properties, and identify 3D hubs across cancer types associated with oncogenic programs.

INTRODUCTION

Despite extensive efforts to develop more efficacious therapies—including those targeted to mutational status—glioblastoma (GBM) remains a devastating disease with a five-year survival rate of less than ten percent.^1–4One of the main challenges in treating GBM is the high degree of inter-patient and intra-tumoral heterogeneity, due to both genetic alterations and epigenetic plasticity.^1,5–8Bulk RNA-seq analyses across hundreds of patients have identified three main GBM molecular subtypes (MESenchymal, CLAssical and PROneural) that are associated with—but not determined by—specific genetic alterations.⁹ Although widely used, the clinical value of this categorization remains unclear. More recently single-cell RNA-seq of primary IDH-wild type (IDH-wt) GBM has revealed striking intra-tumoral heterogeneity¹⁰ with multiple transcriptional states resembling neurodevelopmental cell types such as neural progenitor cell-like (NPC), oligodendrocyte progenitor-like (OPC), astrocyte-like (AC) and mesenchymal-like (MES).⁵ Importantly, these states are largely plastic and interconvertible—rather than hierarchic^5,10–12—suggesting the presence of a core regulatory logic that is preserved among states which enables transcriptional and phenotypic flexibility and increased fitness.

Patient-derived glioma stem cells (GSCs) constitute a powerful model for studying GBM biology since they can be manipulated in vitro while maintaining phenotypic, transcriptional and genotypic characteristics of the original tumor.^13–17 Moreover, ex vivo models have been developed to better recapitulate the complex host cellular environment of the human brain such as the GLICO (GLIoma Cerebral Organoid) model in which GSCs are grown within human cerebral organoids derived from human embryonic stem cells (ESCs).¹⁸ Using various single-cell technologies, GLICO more closely resembles the parental tumor biological behavior (e.g. cell invasion, tumorigenicity) and transcriptomic and epigenomic landscape compared to in vitro cultures or xenografts.^12,19 Therefore, both GSC and ex vivo organoid models have enabled important discoveries regarding the transcriptional and epigenetic states of GBM. However, important gaps remain in our understanding of the complex regulatory networks that govern these deadly tumors.

Since the invention of chromatin conformation capture methods, the three-dimensional (3D) structure of the genome has become increasingly appreciated to constitute an important regulatory layer for gene expression and cell identity.^20–24 The precise spatiotemporal regulation of gene expression is largely dependent on the activity of enhancers, which often reside at large linear distances from promoters.^25–29 How enhancers specifically modulate genes over tens or hundreds of thousands of base pairs of linear genomic distance remains an active area of investigation, but evidence suggests the requirement of physical proximity to target genes.^30,31 New technologies such as H3K27ac HiChIP³² that enable genome-wide mapping of putative active enhancer and promoter interactions in 3D space have revealed complex and dynamic 3D regulatory networks responsible for the spatiotemporal control of gene expression.³³ Although most genes are regulated through pairwise enhancer-promoter interactions, evidence from various cellular contexts supports the existence of highly interacting promoters and enhancers—referred to here as 3D regulatory hubs—which are associated with high levels of transcriptional activity and enrichment of genes critical for cell identity.^34–42 Here, we tested the hypothesis that the construction of 3D enhancer-promoter networks will enable the identification and targeting of core regulatory modules that dictate oncogenic programs and properties in GBM.

RESULTS

Heterogeneous patterns of enhancer activity and 3D interactivity support patient-specific and subtype-specific programs

Although previous studies have profiled the transcriptional heterogeneity of GSCs as well as their genetic and epigenetic landscapes, the 3D enhancer-promoter networks that support their tumorigenic programs has only started to be explored.^43–46 In this study, we profiled four patient-derived IDH-wt GSC samples using H3K27ac HiChIP, H3K27ac ChIP-seq, ATAC-seq, and RNA-seq to construct 3D enhancer-promoter maps for each patient and identify potential similarities or differences across patients and molecular subtypes (Figure 1A and Table S1). In parallel, we profiled human ESC-derived stable long-term neuroepithelial stem cells (lt-NSCs)⁴⁷ to distinguish among shared features that likely reflect the neural identity of both cell types.

Figure 1. Genes within hyperconnected 3D hubs associate with oncogenic programs, GBM biology and worse patient survival.

(A) Schematic of our experimental strategy and datasets created for this study (MES = mesenchymal subtype, CLA = classical subtype). (B) Principal Component Analysis (PCA) of all replicates based on RNA-seq profiles of the top 10% most variable genes across samples. (C) Plot showing the median normalized RNA-seq levels (transcripts per million, TPM) of all genes interacting with 10kb anchors ranked based on their connectivity in deciles from lowest to highest. (D) Top: Schematic of 3D regulatory hub definition, Left: Venn diagram displaying numbers of hyperconnected 3D hubs (top 10% by number of connections) that are either unique for each molecular subtype (common between GSC samples of the same subtype) or shared across all GSC samples or NSCs (see also Table S2) Right: Gene Ontology of interacting genes within MES-specific (teal) or CLA-specific (peach) hubs (see Table S2). (E) IGV example of common GSC 3D regulatory hub (highlighted in gray) SOX9 displaying H3K27ac ChIP-seq signals and the H3K27ac HiChIP arcs for each sample. (F) Heatmap depicting hierarchical clustering of a TCGA brain tumor patient cohort (n=673) based on the expression of common GSC hub-connected genes (from panel (D)). The different colored bars at the top indicate (top) different clusters (purple: high hub expression vs green: low-hub expression), (second) original TCGA classification-- GBM (red) or Low-Grade Glioma (LGG, blue), (third) patient IDH status: IDH-wt (black) or IDH-mutant (grey) and (bottom) the absence (light) or presence (dark) of at least one GBM mutant variant (TERT mutation, EGFR amplification/mutation, Trisomy 7 (partial or full)/deletion of chromosome 10, CDKN2A deletion) is also shown. (G) Bar plot depicting the percentages of GBM, LGG, IDH-wt, and GBM mutant patients within each 3D hub gene expression cluster from (F). (H) Boxplots showing the distribution and median expression of 3D hub genes per cluster split by tumor type (as originally assigned by TCGA). (I-K) Kaplan-Meier survival curves of (I) LGG patients (original TCGA classification) or (J) GBM patients (original TCGA classification) or (K) TCGA IDH-wt patients, each time clustered based on their expression of hyperconnected 3D GSC hub genes. Patients were split into quartiles based on their mean expression of hub-connected genes, and only patients with known survival outcomes were included. The numbers of patients in each cluster are shown on the top. P values from log-rank test are reported. (L) Heatmap depicting hierarchical clustering of expression of super-enhancer linear proximal genes (within 10kb) of TCGA cohort of GBM (red) and LGG (blue) patients. (M) Plot depicting the percentages of GBM, LGG, IDH-wt, and GBM mutant patients within each SE-cluster from (L). All statistics are provided in Table S7.

Consistent with the original characterization of these patient samples¹², our new RNA-seq analysis showed the expected enrichment for the subtype signatures (Figure S1A) and a clear separation of mesenchymal-like (MES-like, 320, 728) and classical-like (CLA-like, 810, 1206) GSC samples (Figure 1B). In agreement, ATAC-seq and H3K27ac ChIP-seq analysis revealed drastic remodeling of the chromatin and enhancer landscapes across samples and identified distinct groups of patient-specific, subtype-specific and GSC- or NSC-specific putative enhancers (Figure S1B–D). These results demonstrate extensive epigenetic and transcriptional rewiring among patients that partly reflects their molecular subtypes.

We next sought to investigate how these GSC-specific and subtype-specific enhancers and promoters communicate with each other in the 3D nucleus. We generated H3K27ac HiChIP data for all four patient samples and the NSCs in technical replicates (Figure S1E). We detected several tens of thousands significant interactions in each sample, which predominantly included connections between promoters (P, 10kb anchors with at least one TSS) and/or putative enhancers (E, anchors with at least one H3K27ac peak but no TSS) (Figures S1F and H, Table S1). The distributions of loop size (distance between interacting anchors) and connectivity (number of distinct interactions that each anchor forms) were similar across samples (Figure S1G). Overall, although most of our HiChIP anchors were involved in one loop per cell line, there were thousands of enhancer and promoter regions engaged in multiple distinct interactions (ranging from 2 to >100) (Figure S1G).

We next tested the association between chromatin looping and gene expression. Genes whose promoters were engaged in HiChIP contacts showed significantly higher transcriptional levels compared to non-connected genes across all cell lines (Figure S1I). Moreover, a higher degree of connectivity or hubness (number of distinct contacts per anchor) resulted in higher transcriptional levels in the respective sample (Figure 1C), supporting the active regulatory nature of these interactions. Intriguingly, subtype-specific signature genes showed a significantly higher connectivity in the respective lines (Figure S1J), further suggesting that cell-type specific 3D interactivity and gene activity are tightly linked.

Genes within hyperconnected 3D hubs associate with oncogenic programs, GBM biology and worse patient survival

We next focused on hyperconnected 3D regulatory hubs, defined as 10kb genomic regions with the highest degree of connectivity/hubness in each cell line (top 10% by number of connections, ranging from >8 to >11 connections) (Figure S2A and Table S2). Although many hubs were patient-specific, there was a large overlap of hubs between samples of the same molecular subtype and fewer hubs common to all samples (Figure 1D). Hyperconnected 3D hubs involve previously established oncogenic drivers such as EGFR, MYC, MYCN, PI3KCA, PTEN and AKT2 (Figure 1D, Figure S2B and Table S2) either directly on the hub anchor or connected anchors. In accordance, gene ontology analysis for MES-like, CLA-like or common hubs revealed a strong enrichment for subtype-specific or universal oncogenic programs and signaling pathways (Figure 1D, Table S2). Specifically, hubs shared among the CLA-like GSC samples showed significant enrichment for PI3K/AKT/mTOR signaling and TNF-alpha signaling via NF-KB. MES-like hubs involved genes that enriched for subtype-characteristic processes such as epithelial-to-mesenchymal transition and hypoxia, in addition to universal oncogenic programs including KRAS signaling and TNF-alpha signaling via NF-KB. Finally, genes essential for GSCs —as determined by previously published CRISPR screens⁵¹— were also characterized by higher degree of hubness across all GSC samples but not in NSCs (Figure S2C) compared to non-essential genes, further supporting the biological relevance of hyperconnected hubs.

Finally, to further validate the significance of our findings for GBM biology and the transcriptional programs of primary tumors, we took advantage of the available RNA-seq and survival data from the TCGA brain cancer patient cohort.⁵² Clustering based on the expression of hub-connected genes—as detected in our four GSC samples—generated two main groups (“high hub” (purple) vs. “low hub” expression (green)) that largely separated GBM from low-grade gliomas (LGG) with ~90% sensitivity and specificity (Figures1F andG). Overall, although LGG patients had lower mean expression of hub-connected genes compared to the GBM patients across clusters, the “misclustered” LGG patients (47 out of 520) showed significantly higher levels—similar to GBM patients (Figure 1H). Moreover, the majority of these LGG patients within the “high hub” cluster had either IDH-wt status or carried GBM-defining mutations (TERT mutation, EGFR amplification/mutation, Trisomy 7 (partial or full)/deletion of chromosome 10 or CDKN2A deletion), suggesting that they were initially misclassified as LGG. The opposite was true for the few GBM patients within the “low hub” cluster. In agreement, GBM and LGG patients from the “high hub” expression cluster (purple) showed significantly worse survival compared to the LGG and GBM patients from the “low hub” expression cluster (green), respectively (Figures1I–J). When focusing on IDH-wt patients, we also noticed that patients from the “high hub” expressing cluster showed slightly but significantly worse survival (pval=0.033) compared to the patients from the “low hub” expression cluster (Figure1K). Of note, clustering of the patients based on the expression of SE-associated genes (by linear proximity), which have been previously linked to GSC identity and tumorigenesis^43,53 generated more mixed LGG/GBM clusters (~67% specificity for GBM patients) and showed no significant association with prognosis (Figures 1L and M). These analyses demonstrate that the common hyperconnected 3D hubs in GSCs harbor genes highly relevant for the biology of IDH-wt GBM and the aggressiveness of the disease. Together, these data support that highly interacting 3D hubs might operate as regulatory centers of GBM oncogenic programs.

3D hubs in GBM expand known oncogenic transcriptional networks

Given the widely reported inter-patient tumor heterogeneity in GBM, we next tested the degree of conservation of our hyperconnected hubs using recently published 3D genomics data (n=27 Hi-C⁵⁴, n=15 H3K4me3 HiChIP⁴⁴, and n=5 H3K27c HiChIP⁵⁵) from independent patient-derived GSCs or primary GBM samples. We consistently found that the common hyperconnected hubs detected in our four GSC samples showed significantly higher connectivity compared to low-connected anchors (bottom 10%) across all independent samples and datasets (Figure 2A). Importantly, the most highly recurrent hubs across all H3K27ac HiChIP GBM/GSC samples showed strong enrichment for oncogenic programs and included genes known in glioblastoma pathogenesis such as SOX9 and HEY1 (Figure S3A), further supporting that 3D hubness is tightly linked to the core regulatory logic of these tumors.

Figure 2. 3D hubs in GBM coordinate and expand known oncogenic transcriptional programs.

(A) Boxplots showing the distribution and mean connectivity of the most (top decile) or least (bottom decile) connected 3D hubs, as detected in our 4 GSC samples across three independent 3D genomics datasets in GSC/GBM samples (Yost et al, TCGA, 2024; Xie et al., 2024; Chakraborty et al., 2023). The numbers of samples are shown at the top. (B) Heatmap displaying the relative connectivity (z-score) of (proto)oncogene-interacting hubs across samples from this study and additional GSC/GBM 3D datasets (shown in (A)). (C) IGV track depicting H3K27ac HiChIP arcs of a 3D enhancer hub (red) interacting with JUN and other genes across multiple GSC (this study) and GBM (TCGA) samples. (D) Top: Schematic showing all gene promoters connected to the JUN hub (shown in (C)) and the location of the guide RNAs (lightning bolt) used for CRIPSRi targeting. Bottom: Relative mRNA levels of all JUN-connected genes upon CRISPRi silencing of JUN hub (48h) expressed relative (percentage) to the negative control values. Dots indicate independent replicates and experiments (n=3). Error bars indicate mean+/−standard deviation.

3D hubs can be classified into promoter or enhancer hubs depending on the nature of the hyperconnected anchor⁵⁶. Promoter-centric hubs potentially confer phenotypic robustness through redundant—or synergistic—enhancer regulatory input, as shown for developmental genes⁵⁷ or important oncogenes. For example, the high expression of MYC in various cancer types (where there is not a structural variant) seems to depend on interactions with multiple, functionally redundant enhancers over large distances^58–61. As expected, we observed that many of our hyperconnected promoter hubs—across samples and 3D genomics methods—are known glioblastoma-related oncogenes and proto-oncogenes (overlap of COSMIC⁶² with GBM DisGeNET gene lists⁶³) —including MYC (Figure S3B). On the other hand, we also observed that many oncogenes are not only hyperconnected themselves, but they are also consistently found interacting with hyperconnected regulatory regions (e.g. 3D enhancer hubs), which contact multiple other genes (Figure 2B). In this context, 3D regulatory hubs could not only promote the higher expression of well-known oncogenes but also coordinate the broader activation of larger previously unappreciated transcriptional networks^31,64–66. To test this hypothesis, we targeted one of our top hyperconnected and highly conserved 3D enhancer hubs (hyperconnected in 8/9 H3K27ac HiChIP samples), which has a direct interaction with the JUN proto-oncogene in addition to other genes of unknown significance to GBM (Figure 2C). Although JUN is not commonly mutated in GBM, higher JUN expression levels are associated with a worse prognosis⁶⁷, and CRISPR knock-out screens identified JUN as an essential gene for GSC survival in vitro⁶⁸. To experimentally perturb this enhancer, we generated a stable GSC line (GSC#320) that harbors a doxycycline (dox)-inducible dCas9-KRAB-P2A-GFP cassette (Figure S3C) and introduced guide RNAs (gRNAs) that target either the enhancer hub (red outline) or an intergenic region as a negative control (Figure 2D). RT-qPCR analysis 48 hours after induction detected statistically significant downregulation of hub-connected genes JUN, FGGY-DT, and FGGY in cells carrying the enhancer hub gRNA compared to cells targeted with the negative control gRNA (Figure 2D), providing a proof-of-concept for the role of 3D hubs in coordinating activation of broader oncogenic programs. However, not all oncogene-containing hubs demonstrated downregulation of all connected genes upon CRISPRi perturbations. Targeting of two additional hyperconnected 3D promoter hubs—one containing an oncogene (PIK3CA hub) and one contacting genes involved in glioblastoma and synaptic processes (BBS10 hub contacting GBM DisGeNET⁶³ gene GLIPR1 and gene involved in synaptic function, CAPS2⁶⁹)—resulted in only downregulation of the directly targeted gene with no impact on the connected genes (Figures S3D and E). As both hubs are in contact with multiple other enhancer and promoter regions (18 additional regulatory contacts for the PIK3CA hub and twelve for the BBS10 hub), this lack of observed downregulation of additional connected genes could be due to compensation from additional/redundant regulatory input.

3D hubs display coregulation of connected genes at the single cell level

In addition to the inter-patient heterogeneity, GBM is characterized by intra-tumoral heterogeneity as documented by scRNA-seq analyses^5,11. As our HiChIP analysis is performed on bulk populations, we wondered to what degree the detected hyperconnected hubs could also reflect co-regulation at a single cell level. To test this systematically, we used published scRNA-seq data for each of our GSC patient samples¹² and calculated the spearman correlation coefficient (rho value) of scRNA levels for each pair of genes that are connected in the same hub (in-hub pairs). We observed a significantly higher spearman correlation (rho value) of in-hub gene pairs relative to random control groups. The latter included either random gene pairs within the same Topologically Associating Domain (TAD), or within TADs with the highest H3K27ac signal (top 10%), or random pairs of genes that were not directly interacting with each other or in the same hub but of similar linear distance as the in-hub pairs (Figure 3A). These results support that gene pairs within hubs have higher probability of coregulation than expected by chance based on their linear distance or topological domain.

Figure 3. 3D hubs display coregulation of connected genes at the single cell level.

(A) Boxplots showing the distribution and median Spearman correlation rho for gene-gene pairs connected to the same hub (in hub) compared to (i) gene pairs within highly active TADs (top 10% by H3K27ac signal), (ii) gene pairs within the same TAD or (iii) random, non-hub pairs with matched linear distances per sample. P-values were calculated by Wilcoxon rank test. (B) Ranking of 3D hubs in GSC#1206 by mean Spearman’s rho of all hub-connected gene pairs per hub. The red dashed line indicates the inflection point of the curve. Oncogenes (COSMIC gene census) are annotated. Red circles highlight the TCF3 hub (highly coregulated) and the JUND hubs (low coregulation score) which are shown in (D). (C) Top: Scatterplots of scRNAseq counts and respective Spearman’s rho for each pair of hub-connected genes (black) and non-hub connected/skipped genes (gray) in a highly coregulated hub (mean Spearman’s rho) in GSC#320. Bottom: UMAP of GSC#320 with kernel density estimator projection of individual hub genes (black) and skipped genes (gray) from top scatterplots. (D) UMAP of GSC#1206 scRNA-seq data with kernel density estimator projection of individual hub-connected genes of the (left) highly coregulated TCF3 hub or (right) the low-coregulated JUND.

We next checked the degree of coregulation among all hub-connected genes by calculating the mean spearman correlation of all in-hub gene pairs. Stratification of hubs based on this score revealed a broad distribution of coregulation scores (Figures 3B and S4A). Genes within top coregulated hubs showed highly correlated scRNA-seq counts and similar distribution of expression across cell subpopulations in the UMAP, suggesting that this hub is either active or inactive as a unit in the different cellular states. Importantly, genes that were either skipped or outside of the hub showed low correlation with hub-connected genes and different distribution of expression on the UMAP, as shown for the example of the TMPO-APAF1 hub in the GSC#320 sample (Figure 3C). Among the most highly coregulated hubs in each sample, we found known oncogenes (COSMIC gene census), such as TCF3, EGFR, MYCN, JUN and CDK6, connected to many other genes of lesser or unknown importance in cancer (Figures 3B and S4A). For example, the TCF3 gene in GSC#1206 was part of a highly interacting and coregulated hub involving five additional genes (MEX3D, MBD3, UQCR11, TCF3, REXO1, and KLF16) which showed similar distribution of expression across cell subpopulations in the UMAP (Figure 3D). On the other hand, genes within hubs with low coregulation scores exhibited discordant expression patterns across subpopulations such as in the example of the JUND hub in GSC#1206 (Figure 3D). Of note, hubs with high mean correlation scores—compared to the ones with low coregulation—had a significantly higher number of connected genes and higher proportion of promoter interactions compared to enhancers (Figure S4B). This suggests that spatial clustering of multiple gene promoters might facilitate their coordinated expression, as recently shown in other contexts.^70–72

Together, these results support that 3D regulatory hubs can function as centers of transcriptional coregulation promoting robust and coordinated expression of gene pairs or—in some cases—of broader gene networks including well-recognized oncogenes/drivers.

Silencing of a highly recurrent 3D hub with an unknown role in glioblastoma alters the transcriptional programs and cellular properties of GSCs

The strong association of hyperconnected 3D hubs with multiple GBM-associated genes and pathways led to the hypothesis that targeting specific hubs could disrupt the regulatory logic of GSCs and their oncogenic properties. To functionally test this hypothesis, we focused on 3D enhancer hubs that (i) show high degree of 3D connectivity across our GSC samples and low/no connectivity in NSCs and (ii) highly recurrent H3K27ac signal in an independent cohort⁴³ of n=44 patient-derived GSCs compared to n=10 normal NSCs (iii) did not contain known oncogenic drivers or genes previously associated with GBM (Figures 4A and S5A). We also prioritized hubs that overlapped with GSC-specific superenhancers. Among our top candidates was an intronic enhancer (Chr 3:67,241,590–168,608,307) located ~10kb away from the GOLIM4 transcriptional start site, which forms complex interaction networks with up to six different genes (GOLIM4, SERPINI1, PDCD10, WDR49, ZBBX, LINC01997) in all four GSC lines but not in NSCs (Figure 4A). Consistently, it shows strong enhancer activity in the independent cohort of GSC samples, while is inactive in NSCs (Figure S5A). Although none of the genes within this hub have been linked to GBM pathogenesis, some of the target genes are involved in cell proliferation, epithelial-to-mesenchymal transition, and tumor growth in other cancer types.^73–77

Figure 4. Targeting of a recurrent 3D hub with unknown role in glioblastoma causes transcriptional shifts and reduced clonogenicity.

(A) HiGlass visualization of the GOLIM4 3D enhancer hub (yellow) showing the H3K27ac HiChIP contact matrices, HiChIP arcs and H3K27ac ChIP-seq peaks for all samples. Of note, this region is not active nor connected in NSCs. (B) Top: Schematic showing all connected gene promoters and the location of the guide RNAs (lightning bolt) used for CRISPRi targeting either the GOLIM4 enhancer hub (HUB CRISPRi) or a nearby, intergenic negative control region (neg ctrl). Bottom: Relative mRNA levels of all GOLIM4-connected genes (black) or non-hub control genes (grey) upon CRISPRi silencing of GOLIM4 hub (48h) expressed as percentage relative to the negative control values. Dots indicate independent replicates and experiments (n=5). Error bars indicate mean+/−standard deviation (s.d.). (C) Schematic of the cerebral organoid glioma model system (GLICO)¹⁷ where hESC-derived cerebral organoids are co-cultured with our CRISPRi targeted GSC cells in the presence of doxycycline (for dCas9-KRAB expression) for seven days prior to imaging and FACS of GFP+ cells for scRNA-seq analysis. (D) Left: UMAP and clustering of all HUB CRISPRi and negative control GFP+ sorted cells. Right: Bar graph showing the percentage of cells representing each cluster per condition. HUB CRISPRi cells(red), neg ctrl(blue). Dashed line represents expected proportion of neg ctrl samples per cluster if evenly distributed between clusters. (E) UMAPs with clustering displaying scRNA-seq data of either HUB CRISPRi cells (left) or neg ctrl cells (right). The two most differential clusters (5 and 6) are highlighted in circles. (F) Top: UMAP with projections of kernel density estimators of expression of GOLIM4 hub-connected genes. Bottom: Violin plots comparing the distribution of scRNA-seq levels of each hub-connected gene between the HUB CRIPSRi sample and the negative control. (G) Gene ontology for differentially expressed genes between clusters 5 and 6 (p adj. < 0.05) upon GOLIM4 hub silencing. (H) Extreme limiting dilution assays comparing HUB CRISPRi GSCs with negative control after 12 days in doxycycline (n= 24 replicates per dilution per condition, n=5 independent experiments). P value was calculated based on the difference between groups for the binomial generalized linear model fit per condition.

To experimentally perturb this enhancer, we used the stable, dox-inducible CRISPRi GSC#320 line (MES-subtype) that we generated to test the JUN hub (Figure S3C) and introduced gRNAs that target either the GOLIM4 enhancer hub or a nearby inactive, intergenic region as a negative control (Figure 4B). RT-qPCR analysis 48 hours after dox induction detected a significant downregulation of all hub-connected genes in cells carrying the GOLIM4 hub gRNA (HUB CRISPRi) compared to cells targeted with the negative control gRNA (neg ctrl) (Figure 4B). Importantly, the nearby gene PLD1, which is not connected to the hub or other housekeeping genes, such as RPL13, remained unaffected. To confirm these results in an independent patient sample, we generated a second stable dox-inducible dCas9-KRAB-P2A-GFP system in the GSC#810 line and observed significant downregulation of hub-connected genes (Figure S5B).

We next investigated the global transcriptional consequences of GOLIM4 hub perturbation on GSCs at the single cell level using the GLICO model¹⁸ in which our stable CRISPRi GSCs were co-cultured with normal hESC-derived brain cerebral organoids for seven days (Figure 4C). By imaging the distribution of GFP (dCas9-KRAB) signal as a qualitative assessment of the ability of GSCs to invade GLICO, we did not observe differences between the GOLIM4 HUB CRISPRi and the neg ctrl (Figure 4C). After GLICO dissociation, we sorted dCas9-KRAB GFP-expressing GSCs (Figure S5C) and performed scRNA-seq. Upon standard filtering, scaling, and log-normalization with Seurat,⁷⁸ we clustered ~4,330 high-quality cells through a shared nearest neighbor (SNN) modularity optimization-based algorithm and performed dimensionality reduction through Uniform Manifold and Projection (UMAP) (Figure 4D–F). As expected, we observed significant downregulation of GOLIM4 hub-connected genes in the HUB CRISPRi cells compared to the neg ctrl (Figure 4F). The UMAP of cells separated by condition showed a clear transcriptional shift in the HUB CRISPRi cells manifested as a preferential gain of cells in cluster 5 and loss of cells in cluster 6 (Figure 4D–F). These shifts could not be explained by differences in cell cycle stage (Figure S5D). Moreover, these changes did not reflect shifts in the Neftel et al. meta-module cell state assignment⁵ (Figure S5E, see Methods). In agreement with the mesenchymal nature of the parental GSC#320 line, cells from both samples (HUB CRISPRi and neg ctrl) were predominantly in the MES-like state (>75%), including the differential 5 and 6 clusters. Overall, all detected states showed nearly identical proportions in both conditions except for the OPC-state, which was slightly increased in HUB CRISPRi sample (~3% as opposed to ~1.5% in the control cells) (Figure S5E). To understand the nature of the global transcriptional changes between HUB CRISPRi and neg ctrl samples, we focused on the differentially expressed genes (p.adj. <0.05) between clusters 5 vs. 6 (n=916). Gene ontology analysis for these genes showed significant enrichment for oncogenic programs such as MYC targets, mTORC1 signaling, and oxidative phosphorylation (Figure 4G). These results indicate that persistent silencing (7 days) of a single hyperconnected hub in GSCs is sufficient to shift the global transcriptional program of GSCs leading to downregulation of oncogenic pathways involved in aggressive disease.

To interrogate the degree to which GOLIM4 hub perturbation in GSCs leads to altered functional properties, we measured the impact on GSC clonogenicity using extreme limiting dilution assays. These experiments consistently documented that HUB CRISPRi GSCs had a significantly lower clonogenic capacity compared to negative controls and required >5 times higher input cells per well to form new spheres (Figure 4H). The reduced clonogenicity upon GOLIM4 hub perturbation was also confirmed in the independent GSC#810 Hub CRISPRi sample (Figure S5B). Importantly, CRISPRi silencing of individual hub-connected genes (Figure S5F), also resulted in reduced GSC clonogenicity compared to the neg ctrl. However, the impact of individual gene perturbations was significantly weaker than the one induced by silencing the entire hub (Figure S5G).

Together, these findings demonstrate that epigenetic silencing of a single multiconnected 3D regulatory hub is sufficient to perturb not only the activity of hub-connected genes but also the transcriptional network and cellular properties of GSCs. This provides proof-of-concept for the role of 3D hubs as potential centers of the GSC regulatory logic as well as highlights its potential to nominate novel regulatory nodes central to transcriptional programs and cellular behavior.

3D regulatory hubs across cancer types enrich for cancer-specific and universal oncogenic programs

We next investigated the presence and nature of 3D regulatory hubs across different cancer types by analyzing published H3K27ac HiChIP cancer datasets with sufficient quality to call enhancer-promoter interactions and more than one patient sample or cell line per cancer type. This analysis covered 16 cancer types (adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), colon adenocarcinoma (COAD), Ewing sarcoma (ES), esophageal carcinoma (ESCA), glioblastoma (GBM, all IDH-wt), kidney renal clear cell carcinoma (KIRC), liver hepatocellular carcinoma (LIHC), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), prostate adenocarcinoma (PRAD), skin cutaneous melanoma (SKCM), stomach adenocarcinoma (STAD), thyroid carcinoma (THCA), uterine corpus endometrial carcinoma (UCEC)) for a total of 88 samples (Figure S6A and see Table S4 for sample details). Consistently across all H3K27ac HiChIP samples, we observed complex 3D interaction networks of putative active enhancers and promoters. We then focused on the top 10% hyperconnected hubs from each cancer type (common in >50% of samples, see Methods) and performed dimensionality reduction via UMAP and clustering by a SNN approach. This generated several clusters of hubs, which we annotated as cancer-specific or MULTI-cancer by projecting their mean connectivity per cancer type (Figures 5A and S6B–C and Table S4).

Figure 5. 3D hubs across cancer types enrich for cancer-specific and universal oncogenic programs.

(A) Left: UMAP and clustering (SNN) of all hyperconnected 3D regulatory hubs across all cancer types (top 10% in >50% of samples, n=88, per cancer) based on their connectivity. Clusters were annotated as MULTI or cancer-type specific based on the projection of mean connectivity across cancer type (see Figures S5B and C, single-sample clusters removed for clarity). Right: Projection of connectivity values in ACC/PRAD/SKCM samples as examples of cluster annotation method. Adrenocortical carcinoma (ACC), Bladder Urothelial Carcinoma (BLCA), Breast invasive carcinoma (BRCA), Colon adenocarcinoma (COAD), Ewing Sarcoma (ES), Esophageal carcinoma (ESCA), glioblastoma (GBM), Kidney renal clear cell carcinoma (KIRC), Liver hepatocellular carcinoma (LIHC), Lung adenocarcinoma (LUAD), Lung squamous cell carcinoma (LUSC), Prostate adenocarcinoma (PRAD), Skin Cutaneous Melanoma (SKCM), Stomach adenocarcinoma (STAD), Thyroid carcinoma (THCA), Uterine Corpus Endometrial Carcinoma (UCEC). Details about the samples are provided in Figure S6A and Table S4. See also Table S8 for all hyperconnected 3D hubs per cancer type. (B) Gene Ontology of all hub-connected genes per cancer type for each respective cluster. See Table S5. (C) Left: Example of the multi-cancer MYC promoter hub showing high connectivity and heterogeneity of interactions (shown as H3K27ac HiChIP arcs) across tumors. Right: Example multi-cancer 3D enhancer hub interacting with the HIF1A gene and others across tumors.

Gene ontology analysis of all hub-connected genes per cluster showed that MULTI-cancer cluster hubs were associated with universal oncogenic pathways and signatures such as the p53 pathway, MYC targets, and TNFα signaling via NFκB. (Figure 5B and Table S5). On the other hand, cancer type-specific hubs enriched for more specialized processes, such as UV response, EMT, and the inflammatory response for the melanoma-specific cluster, fatty acid metabolism, xenobiotic metabolism, and adipogenesis programs for the LIHC regulatory networks, and androgen response, peroxisome, and protein secretion for the PRAD cluster (Figure 5B). In agreement with our findings in GSC samples, these analyses suggest that hyperconnected hubs in each cancer type associate with critical genes for the regulation of the identity and oncogenic properties of the tumor.

Consistent with gene ontology analysis, we observed that hyperconnected hubs across samples involved drivers and genes known to support oncogenic processes such as MYC, JUN, VEGFA, AKT2, NOTCH1, CDK4 and HIF1A (Figure 5C). Although these regulatory regions appear multiconnected in all samples, the specific 3D interaction networks vary substantially among patients and cancers (Figure 5C). Based on the observed gene coregulation within GSC hubs, we could postulate that hubs with distinct interaction patterns foster different gene coregulation networks in each patient and cancer type, leading to changes beyond the higher expression of well-known oncogene/drivers. For example, the MYC promoter hub interacts with several enhancers but also with different combinations of gene promoters across cancer types (MYC, CASC11, LINC00824, LRATD2, PVT1, TMEM75, CASC19, CASC8, CCAT2, POU5F1B, PCAT1, PRNCR1, LINC00861, CCDC26) (Figure 5C). Aside from MYC itself, none of the interacting genes appear on Hallmark MYC targets gene lists (V1 and V2),⁷⁹ suggesting the potential value in using 3D enhancer-promoter interactivity data to gain a deeper understanding and uncover previously unappreciated genes involved in canonical cancer signaling pathways.

Both genetic and epigenetic mechanisms could contribute to 3D hyperconnectivity and nominate factors associated with worse prognosis

We next investigated potential genetic and epigenetic mechanisms that could be associated with highly interacting 3D regulatory hubs. Structural variants (SVs) have been reported to alter local 3D chromatin organization and nearby gene and enhancer activity either by amplifying the regulatory input and/or by enabling aberrant communication with new enhancers (enhancer hijacking).^80–85 To address the degree to which SVs can explain patient-specific or tumor-specific hub formation, we applied the EagleC pipeline⁸⁹ (see Methods) on all HiChIP across cancers to call SVs in each sample. EagleC was successful in detecting SVs across all samples (Figure 6A and Table S4), including the previously detected EGFR amplification in GSC#810 and 1206 (Figure S6D), which frequently forms extrachromosomal DNA in cancer^86–88. A large fraction (77%) of EagleC-predicted SVs overlapped with hyperconnected hubs in the respective sample. These overlaps included some high-confidence, candidate driver SVs, as previously nominated using Whole Genome Sequencing data from a larger cohort of ICGC/PCAWG cancer patients by the CSVDriver pipeline, which uses a generalized additive model to identify regions with SV breakpoints that occur more frequently than random expectation indicating positive selection (Figure S6E).⁹⁰ Importantly, the vast majority (80–99.6%) of hyperconnected hubs did not overlap with any predicted SV in the respective samples (Figure 6A). Examples of hubs with or without SVs as detected by low genomic coverage sequencing are shown in Figure S6F. These analyses indicate that genetic alterations associate with—and might drive formation of—some hyperconnected hubs involving important oncogenes but cannot explain the majority of detected 3D hubs.

Figure 6: Both genetic and epigenetic mechanisms could contribute to 3D hyperconnectivity and nominate factors associated with worse prognosis.

(A) Top: Bar plots of EagleC-detected SV per sample with breakdown by SV type. Overlapping SV predictions were merged and counted as one. See also Table S4. Bottom: Bar graph displaying numbers and percentages of hyperconnected 3D hubs that overlap (black) with EagleC predicted SVs in each sample. (B) ChIP Enrichment Analysis (ChEA) for select protein factors (among the top ten by p adj.) that are significantly enriched on hub-connected genes specific for each cancer type cluster as defined in Figure 5A. For full list of factors, see Table S5. (C) ChEA for top 5 protein factors (by p adj.) significantly enriched on hub-connected genes for each MULTI cluster as defined in Figure 6A. For full list of factors, see Table S5. (D) Kaplan-Meier curves showing example MULTI cluster ChEA-nominated factors (outlined in red in panel (C)) that associate with worse prognosis in various cancers (left) and IDH-wt GBM (right). Patients with very high/low expression were derived using 1^st and 4^th quartiles. The numbers of patients in each group are provided at the top right.

The poor overlap of hubs with genetic alterations prompted us to consider epigenetic mechanisms of hub organization. For example, aberrant expression and/or binding of specific TFs or cofactors might nucleate the assembly of 3D hubs possibly through biomolecular condensation as previously described.^{31,56,91–93} Association analysis using published ChIP-seq datasets through the EnrichR (ChIP-X Enrichment Analysis, ChEA)⁹⁴ revealed a large number of lineage-specific TFs, architectural factors (CTCF) and co-activators (SMARCA4) as significantly enriched either across all hubs or in specific clusters (Figures 6B and C and Table S5). Some of them were previously proposed to mediate enhancer-promoter interactions (e.g. CTCF, FOXP and RUNX proteins)^95–98 in GBM or other cancer lines. In some cases, the enriched factors have known links to the biology of the specific tumor, such as FOXA1 and the androgen receptor (AR) in the PRAD hub cluster^99–102, MITF in melanoma¹⁰³, and SMARCA4¹⁰⁴ and OLIG2¹⁰⁵ in GBM (Figure 6B). When focusing on proteins enriched in the MULTI-cancer hub clusters through ChEA, we found TFs whose high expression levels were significantly associated with worse prognosis across different cancer types (Figures 6C and S6G). Specifically, among the top ChEA-enriched factors across the MULTI-cancer hub clusters (top 10 per cluster, p adj.), we found 16 factors in which high expression is associated with worse prognosis in at least one cancer type with 13 associating with worse prognosis in at least two (see Figure S6G). This analysis also revealed three TFs associated with worse prognosis in IDH-wt GBM (RUNX2, PGR, and FOXA1) (Figure 6D).

In conclusion, these analyses nominate potential mechanisms of hub formation both by genetic alterations and epigenetic factors. Overall, in extending our study to include 16 different cancer types across 88 samples, we found highly conserved and cancer-specific hyperconnected 3D regulatory hubs and established important links with oncogenic programs and cell identity as well as nominated novel factors associated with more aggressive disease.

DISCUSSION

Cell-type specific transcriptional programs are dictated by the ability of enhancer elements to regulate their target genes in the 3D nucleus. Dysregulation of enhancer-promoter communication through genetic and epigenetic mechanisms is emerging as an important process in oncogenic transformation and progression.^106–111 In this study, we generated a genome-wide atlas of enhancer landscapes and 3D interactomes in patient-derived-GSC samples to gain insights into the regulatory logic of this deadly cancer and identify potential central nodes of gene regulation. Based on the integration of bulk and scRNA-seq data from the same patients, the association analyses using published datasets of independent patient cohorts, and the proof-of-concept perturbation experiments, our findings strongly support that 3D regulatory hubs represent regulatory centers of tumorigenic programs, fostering robust expression and coregulation of cancer-associated genes and pathways. Importantly, genes within 3D hubs seem to share four key properties: (i) high transcriptional levels compared to low connected genes (ii) high degree of gene-pair coregulation compared to non-hub genes (iii) strong enrichment for oncogenic programs and (iv) —often—significant association with worse patient outcomes. Based on these properties, we speculated that targeting 3D hubs could have profound effects on the cancer regulatory logic and oncogenic properties. As hypothesized, proof-of-concept perturbations of a previously uncharacterized hub in GSCs showed concordant downregulation of hub-connected genes, significant transcriptional reprogramming and changes in cellular behavior. However, our functional experiments also demonstrated that perturbations of individual promoters/enhancers within complex, hyperconnected hubs will often be compensated by multiple other interacting regulatory elements and/or genes. Future high-throughput perturbations of multiple nodes within hubs followed by single-cell analysis (e.g. PERTURB-seq)¹¹⁵ combined with continuously improving machine learning models^24,116 will enable a deeper understanding and a more accurate prediction of central regulatory nodes and their downstream effects on oncogenic programs and properties.

Although the higher transcriptional activity and the coregulation potential within hubs have been reported before in other contexts, here we provide genome-wide evidence of coregulation at the single-cell level. By integrating genome-wide 3D enhancer-promoter interactivity with matched scRNA-seq data, we demonstrated that gene pairs within hubs have significantly higher correlation of expression at a single cell level compared to non-hub gene pairs of similar linear distances or located within the same highly active TADs. Importantly, although individual gene pairs within hubs display a high degree of correlation, the mean correlation scores across all hub-connected genes were highly variable across hubs. This variability could partly reflect our inability to distinguish whether the 3D hubs detected by bulk HiChIP analysis have the same interaction network in all cells or represent the sum of simpler “structures” and interactions across different subpopulations. This is supported by the observation that genes in highly coregulated hubs show a similar distribution of expression across cell subpopulations on the UMAP, while genes in hubs with low coregulation scores have more random expression patterns. Future studies using single-cell 3D approaches or predictive models¹¹⁹ might help deconvolute the 3D complexity of hubs and better assess the degree of gene coregulation within subpopulation-specific subnetworks. Despite these limitations, our analysis revealed a set of hyperconnected hubs with high single-cell correlation scores among all connected genes, suggesting that these assemblies are regulated as a unit across subpopulations. These highly coregulated hubs had a higher proportion of interacting gene promoters (as opposed to enhancers), bringing—among others—important oncogenes in proximity with other genes of lesser or unknown roles in cancer. The concordant (hyper)activation of different, seemingly unrelated genes reveals complex oncogenic networks which could lead to the discovery of unappreciated interconnections and interdependencies, potentially enabling better predictions of treatment response or resistance and the development of combinatorial therapies.

Disruption of 3D genomic architecture through structural variants has long been theorized to be a potential mechanism of oncogenic transformation and progression,^80,82,120 and was recently shown to support the formation of hyperconnected 3D hubs or neo-hubs. In agreement, our study showed that a large fraction of predicted structural variants (77%) overlapped with hyperconnected hubs. However, SVs could only explain a minority of our detected 3D hubs suggesting that most hubs are assembled through epigenetic mechanisms, including TFs whose overexpression associates with worse prognosis. More work will be required to determine mechanisms of de novo hub formation and to dissect the complex interplay between somatic mutations, structural variants, and epigenetic mechanisms regulating oncogenic programs.

In conclusion, our study provides support that hyperconnected 3D hubs might operate as regulatory centers of oncogenic programs in glioblastoma and other tumors. Therefore, identifying and targeting complex 3D hubs or the factors that support their organization and function could enable a deeper understanding of genes and pathways that are central to tumorigenic programs and nominate novel actionable therapeutic targets.

LIMITATIONS OF THE STUDY

Since the cell of origin of GBM remains uncertain^45,121,122 there is no perfect control for determining 3D enhancer-promoter rewiring during gliomagenesis and for identifying cancer-specific hubs. Therefore, the comparison between GSCs and NSCs reveals shared features associated with the “original” primitive neuroectodermal cell program but does not guarantee that GSC-unique features represent true rewiring during oncogenesis. Given the high inter-patient heterogeneity in GBM, the small number of samples we characterized—especially per subtype—is unlikely to capture all relevant 3D regulatory networks and hubs. To partially overcome this limitation, we utilized published datasets from other GSC and primary GBM samples to check the degree of conservation and prioritize recurrent hubs as well as TCGA patient data to validate the relevance of our detected hubs on GBM biology. Another consideration when interpreting bulk HiChIP data from GSC samples is the high intra-tumoral heterogeneity, which makes it hard to interpret whether hyperconnected hubs have the same interaction network in all cells or represent the sum of simpler “structures” and interactions across different subpopulations. Finally, although our hub analysis across cancer types reveals only a small overlap with predicted SVs by EagleC, more complex or subclonal SVs might be missed with this approach.

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Effie Apostolou (efa2001@med.cornell.edu)

Materials availability

CRISPRi GSC lines and constructs can be available upon request.

Data and code availability

1. All genomics data generated in this study have been deposited at GSE262089 and are publicly available as of the date of publication.

2. This paper does not report original code.

3. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

STAR Methods

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

There were no new samples collected for this study and all GSC samples are already de-identified. Patient characteristics of all study participants is provided in Tables S1 and S4.

METHOD DETAILS

Cell Culture

Patient-Derived GSCs.

Patient-derived GSCs were obtained from the Fine lab and previously characterized into molecular subtypes as described in Pine et al., 2020.¹² Briefly, following informed consent, tumors classified as GBM based on WHO criteria were obtained from patients undergoing treatment at Weill Cornell Medicine/New York-Presbyterian Hospital in accordance with an Institutional Review Board-approved tissue-acquisition clinical protocol. Following surgical removal, tumors were dissociated and cultured in Neurobasal medium A (NBE medium, Thermo Fisher Scientific) supplemented with N2 (Thermo Fisher Scientific), B27 (Thermo Fisher Scientific), human recombinant bFGF and EGF (25 ng/mL each, R&D Systems), heparin (Sigma), Penicilin-Streptomycin solution (LifeTech), and L-glutamine (200 mM/100X. LifeTech). Mycoplasma screening was performed using the ABM Mycoplasma PCR detection kit (Applied Biological Materials Inc).

Human long-term Neuroepithelial Stem Cells (ltNSCs)

Human long-term Neuroepithelial Stem Cells (ltNSCs) were generated as previously described.⁴⁷ Briefly, embryonic stem cells (H1 or H9) were dissociated using accutase and seeded on Matrigel (1:25 in DMEM/F12 HEPES) coated dishes at a density of 10k/cm² in E8 medium supplemented with 5 μM XAV 939 and 10 μM Y27632. The next day, the medium was switched to Neurobasal:DMEM/F12 HEPES 1:1 supplemented with 1xGlutaMAX, 1:200 N2 supplement, 1:100 B27 supplement (-RA), 0.1 mM ascorbic acid, 2 μM DMH-1, 1 μM dorsomorphin, 250 nM LDN193189, 12.5 μM SB431542, 4 μM CHIR99021 and 0.5 μM purmorphamine. Prior to cells approaching 90% confluency, cells were split 1:3 onto Matrigel coated dishes (1:50 in DMEM/F12 HEPES) in the presence of 10 μM Y27632. On day 7 after start of differentiation, the media was switched to Neurobasal:DMEM/F12 1:1 supplemented with 1xGlutaMAX, 1:200 N2 supplement, 1:100 B27 supplement (-RA), 0.1 mM ascorbic acid, 4 μM CHIR99021 and 0.5 μM purmorphamine. Cells were split 1:3 upon 90% confluency onto Matrigel coated dishes (1:50 in DMEM/F12 HEPES) using 0.05% trypsin-EDTA and trypsin inhibitor (0.5 mg/ml in PBS). On day 13, medium was switched to N2 medium (DMEM/F12 w/o HEPES, 1.6 glucose g/L, 100 μg/ml transferrin, 25 μg/ml insulin, 20 nM progesterone, 100 μM putrescine, 30 nM sodium selenite) supplemented with 20 ng/ml FGF2, 10 ng/ml EGF and 1 μM CHIR99021. On day 16, FGF2 concentration was reduced to 10 ng/ml and cells were split at 90% confluency onto poly-L-ornithine(PO)/laminin(Ln) coated dishes at a density of 295k/cm². To remove any possible neural crest cell contaminants, cells were subjected to MACS using the Neural Crest Stem Cell Microbeads together with LS-columns following the manufacturers protocol when passaging for first three passages. For maintenance, ltNSC were cultured in N2 medium supplemented with 10 ng/ml FGF2, 10 ng/ml EGF and 1 μM CHIR99021 on PO/Ln coated dishes and re-seeded using trypsin and trypsin inhibitor upon confluency.

Lentiviral production and infection

293T cells were transfected with overexpression constructs along with packaging (psPAX2, Addgene, 12260) and envelope vectors (VSV-G, Addgene, 1488) using PEI reagent (PEI MAX; Polyscience, 24765–2). The supernatant was collected at 48h and 72h and concentrated using polyehtylglycol (Sigma, P4338). For infection, GSC samples were dissociated and infected in medium containing 5 μg ml⁻¹ polybrene (Millipore, TR-1003-G) for 6 hours.

CRISPRi

GSCs were first infected with lentiviruses harboring pLenti CMV rtTA3 Blast (Addgene, 26429) and underwent blasticidin selection for 5 days. GSCs were next infected with the TRE-KRAB-dCas9-IRES-GFP vector (Addgene, 85556).¹²³ GFP+ cells were selected by three consecutive rounds of FACS sorting. The resulting GSCs were then infected with a lentivirus harboring the pLKO5.GRNA.EFS.PAC vector (Addgene, 57825)¹²⁴ containing a guide RNA targeting the region of interest. For each region of interest, guide RNAs were designed to target the center of prominent ATAC-seq peaks of the region interest using CRISPOR.¹²⁵ Cells were selected with puromycin (LifeTech, K210015) for 5 days, expanded, and underwent an additional FACS sorting for GFP+ prior to collection of RNA for RT-qPCR analysis, clonogenicity assays and cerebral organoid co-culture experiments. Guide RNA sequences and RT–qPCR primers are listed in Table S6.

Clonogenicity Assay

Clonogenicity was measured by in vitro extreme limiting dilution assay, as previously reported.¹²⁶ Briefly, GSC samples were dissociated into a single cell suspension and decreasing numbers of cells per well (50, 25, 16, 11, 7, 5, 3, 2, and 1) were plated (24 replicates per condition) into U-bottom 96-well plates with the addition of doxycycline. Doxycycline and media were replenished every other day. The presence and number of colonies in each well were recorded 12 days after plating and doxycycline induction. Extreme limiting dilution analysis was performed using software available at http://bioinf.wehi.edu.au/software/elda.

Co-Culture of GSCs and Cerebral Organoids (GLICO model)

12 GLICOs from 4.5-month-old cerebral organoids were made as previously described¹⁸ for the 320 dCAS9-KRAB negative guide sample and the 320 dCAS9-KRAB GOLIM4 enhancer sample using 100K GSCs each. Briefly, organoids were plated one per well in a 96-well U-bottom plate, excess medium was removed and 100,000 GSCs in 150 uL of NBE were added to each well. After GLICO formation through stationary co-culture incubation at 37°C for 24h, 1ug/ml of doxycycline was added to the GLICO media and replenished with fresh media after 96h. Pictures were taken with fluorescent microscope at 72h. After 1 week, GLICOs were dissociated and single-cell and GFP+/DAPI- cells were sorted. Cells from both samples were submitted for 10x scRNAseq.

RNA sequencing and library preparation

Total RNA was prepared with TRIzol (Life Technologies, cat. no. 15596018) following the manufacturer’s instructions. Libraries were generated by the Weill Cornell Genomics core facility using an Illumina TruSeq stranded total RNA kit (cat. no. 20020596) and sequenced on an Illumina Novaseq6000 platform on PE50 mode.

ATAC-seq

ATAC-seq was performed as previously described.¹²⁷ In brief, a total of 50,000 cells were used as input for the protocol. In order to minimize PCR bias an aliquot of each ATAC-seq library was first subjected to five cycles of amplification to determine by quantitative PCR the suitable number of cycles required for optimal library amplification. Samples were then subjected to a dual size selection (0.55x–1.5x) using SPRIselect beads (Beckman Coulter, B23317). Fragment distribution of libraries was assessed with an Agilent Bioanalyzer and, finally, the ATAC libraries were sequenced on an Illumina Nextseq2000 platform on PE100 mode.

H3K27ac ChIP-seq

ChIP-seq was performed as previously described.³¹ In brief, 5 million cells per replicate in each cell line was crosslinked with 1% formaldehyde and quenched with 125mM glycine for 5 min at room temperature. The cell pellets were washed twice in PBS and resuspended in 400 μl lysis buffer (10 mM Tris pH 8, 1 mM EDTA and 0.5% SDS). The cells were sonicated in a Bioruptor device (30 cycles of 30 s on/off; high setting) and spun down at the maximum speed for 10 min at 4 °C. The supernatants were diluted five times with dilution buffer (0.01% SDS, 1.1% Triton X-100, 1.2 mM EDTA, 16.7 mM Tris pH 8 and 167 mM NaCl) and incubated overnight with an H3K27ac antibody with rotation at 4 °C (Abcam, ab4729). Protein G Dynabeads (Thermo Scientific) pre-blocked with BSA protein (100 ng per 10 μl Dynabeads) were added (10 μl blocked Dynabeads per 10 × 10⁶ cells) the following day and incubated for 2–3 h at 4 °C. Samples were washed twice in low-salt buffer (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 150 mM NaCl and 20 mM Tris pH 8), twice in high-salt buffer (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 500 mM NaCl and 20 mM Tris pH 8), twice in LiCl buffer (0.25 M LiCl, 1% NP-40, 1% deoxycholic acid (sodium salt), 1 mM EDTA and 10 mM Tris pH 8) and once in TE buffer. The DNA was then eluted from the beads by incubating with 150 μl elution buffer (1% SDS and 100 mM NaHCO₃) for 20 min at 65 °C (in a thermomixer at high speed). The supernatants were collected and reverse crosslinked by incubation overnight at 65 °C in the presence of proteinase K. After RNase A treatment for 1 h at 37 °C, the DNA was purified using a Zymo kit. 10–30ng of the immunoprecipitated material was used for ChIP-seq library preparation using the KAPA Hyper prep kit (KAPA Biosystems) and applying 7–8 cycles of amplification. Libraries were then subjected to a dual size selection (0.6x–0.8x) using SPRIselect beads (Beckman Coulter, B23317) and sequenced on an Illumina Nextseq2000 platform on PE50 mode.

H3K27ac HiChIP

HiChIP experiments were performed in duplicates for samples GSC320, GSC728, GSC810, GSC1206 and NSC using the Arima-HiC+ kit (Arima, A101020). The protocol was performed with 5M cells (per sample) and using an H3K27ac antibody (Abcam, ab4729 was used for GSC320 and GSC728; active motif, 91193 was used for samples GSC810, GSC1206 and NSC) according to the manufacturer’s instructions with few modifications. The efficiencies of the two H3K27ac antibodies were tested by ChIP-seq in GSC1206 and NSC, and both antibodies resulted in similar distribution and number of peaks. To improve the sonication efficiency, a modified lysis buffer was used containing 10 mM Tris pH 8, 1 mM EDTA and 0.5% SDS. Before overnight incubation with the antibody, the sample was diluted in a buffer to bring it back to the original composition of the Arima R1 buffer (10 mM Tris pH 8, 140 mM NaCl, 1 mM EDTA, 1% triton, 0.1% SDS, 0.1% sodium deoxycholate). Libraries were generated using the Swift Accel-NGS 2S Plus DNA Library Kit (Swift Biosciences, 21024) according to the manufacturer’s instructions and performing between 8 and 14 cycles of PCR amplification. The samples were quantified by Qubit and sent for Bioanalyzer to check the quality and final size of the library. All HiChIP libraries were sequenced using the Illumina Nextseq 2000 on PE50 mode.

In-situ Hi-C

The protocol was performed as previously described¹²⁸ with minor modifications. Hi-C was performed starting with 4 million cells per condition and using the Arima-HiC kit (Arima, A510008) according to the manufacturer’s instructions (A160134 v01). Approximately 2 ug of DNA was used for each Hi-C sample to prepare libraries using the KAPA Hyper Prep Kit (KAPA, KK8502) and performing eight cycles of amplification. Libraries were sequenced using the AVITI sequencer in PE150 mode.

Computational Methods

RNA-seq analysis

Paired-end read alignment to the human genome (hg38 version) was performed with TopHat2¹²⁹ (version 2.11) with default setting and ‘-r 200 –mate-std-dev 100’ option. Samtools¹³⁰ was used for filtering and sorting aligned reads before annotation to “Homo_sapiens.GRCh38.104” gene version with htseq-count¹³¹ and ‘-m intersection-nonempty’ option. Only protein-coding and long-non-coding RNA transcripts were used for annotation and downstream differential expression analysis was performed with R package DESeq¹³² where we set a log2 fold change of 1 and P-adjusted of <10.01 as a cut off for calling deferentially expressed genes. Average TPM values were calculated for all replicates and genes with TPM above 1 were considered as expressed.

ChIP-seq analysis

Paired-end reads were aligned to human genome (hg38 version) with bowtie2¹³³ (version 2.3.4.1) and ‘–local –very-sensitive-local’ option active. Picard tools (http://broadinstitute.github.io/picard/), Samtools¹³⁰ and Bedtools¹³⁴ were used for filtering duplicate reads (‘MarkDuplicate’ option), low quality reads (MAPQ<20), chrM and blacklisted regions and converting files into sam, bam, bed bigWig format. All filtered reads were used to call both narrow and broad peaks with MACS2¹³⁵ and default options. All peaks within 147 bp (one nucleosome) were merged into one peak for each experiment and only common peak between replicates were considered as peaks in each experiment.

ATAC-seq analysis

We used the same pipeline and steps as in ChIP-seq for ATAC-seq datasets with the addition of ‘-I 10 ×2000’ in bowtie2¹³³ when aligning reads to human genome and correction of Tn5 insertions at each read end of the filtered paired end reads by shifting +4 bp or −5 bp from the positive and negative strands, respectively.

HiChIP analysis

HiC-Pro pipeline (version 3.0.0)¹³⁶ was used for processing paired-end files with default setting. Aligned filtered reads were assigned to MboI restriction fragments and valid read pairs (interactions) were used for generating binned interaction matrices with Juicer-tools¹³⁷ and loop calling at 10kb resolution with FitHiChIP (release 9.0)⁵⁰ and coverage bias regression and “Peak2All” option active where peaks form ChIP-seq for each experiment were used as input alongside valid-pairs. Interactions with a distance of more than 10-kb between the start of both anchors and p-adjusted <0.05 were considered significant and called loops. For each loop we classified each anchor into ‘promoter’, if a TSS was present within the 10kb, ‘enhancer’, if there was a H3k27ac peak and no TSS and ‘X’, if none TSS and H3k27ac peak was present. Multi-connected anchors (n>=2) were considered as ‘hubs’ and hubs were annotated as ‘promoter’, ‘enhancer; and ‘X’ depending on the type of the multi-connected anchor.

Hi-C analysis

Fastq reads were processed using the HiC-Pro pipeline using default settings and aligned to the GRCh38 genome. The generated valid pairs were converted to .hic files for visualization using juicertools. TAD and compartment analysis was performed using HOMER. The HOMER command makeTagDirectory was used to generate HOMER tag directories from valid pair files. TADs were called using the command findTADsAndLoops with the parameters -res and -window set to 5kb and 10kb respectively. For selecting TADs with high H3K27ac, TAD overlapping H3K27ac ChIP-seq peaks were identified using the bedtools intersect command. Size normalized number of peaks per TAD was calculated by dividing total number of peaks overlapping a TAD by the TAD size. TADs were then ranked in descending order of size normalized peaks per TAD, and top 10% TADs were chosen for downstream analysis. Compartments were called using the command runHiCpca.pl with both the parameters -res and -window set to 100kb. H3K27ac ChIP-seq peaks were supplied to the parameter -active, for assigning the appropriate sign to PC1 values.

scRNAseq coregulation analysis

Previously published scRNAseq data of 320, 728, 810 and 1206 GSCs were re-analyzed (accessible from PRJNA595375).¹⁷ The Cell Ranger 2.0.1 pipeline was used to align reads to the GRCh38 human reference genome and produce count matrices for downstream preprocessing and analysis using the Seurat v4.0 R package.⁷⁸ For quality control, cells with fewer than 500 or more than 6000 genes detected, or greater than 15% mitochondrial gene expression were removed. Expression values were library size corrected to 10,000 reads per cell and log1p transformed. Next, for each sample mean expression in single cells of set of genes in the top (CON10) or bottom (CON1) deciles of hubs ranked by connectivity. For gene-gene correlation analysis, zero-preserving imputation of the data using ALRA was performed.¹³⁸ We then performed spearman correlation between imputed expression values of within hub and control gene-gene pairs using the correlatePairs() function of the scran v1.28.2 package. As control gene-gene pair sets we considered within-TAD gene-pairs, the top 10% k27 acetylated within-TAD gene-pairs, and a random set of gene pairs drawn from the same distribution of genomic distances as the query gene-gene pairs sets. Within-hub gene pairs also found in control sets were removed. Mean hub spearman correlation rho values per GSC sample are summarized in Table S2.

scRNA-seq Processing and analysis

scRNA-seq data resulting from the GFP+ FACS sorted cells from the GLICO cerebral organoid co-culture CRISPRi experiment were processed with the 10x Genomics Chromium Single Cell Platform, and count matrices were generated using their Cell Ranger pipeline version 3.0 with the GRCh38 reference genome used to align and quantify the reads (10x Genomics). scRNA-seq data were preprocessed and largely analyzed using Seurat version 4.3.0.⁶⁴ For quality control, genes detected in less than 3 cells and cells with fewer than 200 genes were excluded. Cells with percentage of mitochondrial genes outside of 5 M.A.D. were removed. Doublet detection was conducted via DoubletFinder version 2.0.3, and doublets were removed.¹³⁹ Expression values were further library size corrected to 10,000 reads per cell and log1p transformed. After quality control, we used the standard analysis pipeline of the R package Seurat including using the FindVariableFeatures (nfeatures = 2500), RunPCA, FindNeighbors (dims= 1:50), FindClusters(resolution = 0.5), and the RunUMAP (dims = 1:50) functions. Projections of cell cycle scores on the UMAP were calculated via the CellCycleScoring function in Seurat. UMAP projections of kernel density estimators of individual hub-connected genes were conducted via Nebulosa version 1.8.0.¹⁴⁰ Differentially expressed genes were determined per cluster and for clusters 5 vs. 6 using the Seurat FindAllMarkers function with default parameters (Full results inTable S3). Gene ontology of differentially expressed genes was conducted via EnrichR (MSigDB Hallmark 2020) with a background of all expressed genes (full results in Table S3).¹⁴¹ For Neftel et al. meta-module cell state assignment, individual cell meta-module scores were calculated for NPC-like, OPC-like, MES-like, and AC-like signatures given in Neftel et al., 2019 (NPC1 and NPC2 and MES1 and MES2 signatures were combined into respective NPC and MES signatures) using the AddModuleScore Seurat function. Each cell was assigned to its maximum score.

Enrichment Analysis

Gene ontology, pathway and TF and or motif analysis were performed with the use of EnrichR¹⁴¹ and LOLA¹⁴² R packages. For each enrichment analysis, we defined a ‘background’ peak or gene group tailored for the group of genes or peaks tested. For motif analysis with LOLA all accessible regions were merged to form a common ‘background’ atlas, while for EnrichR for each gene set tested we generated the appropriate background as described in each ‘figure legend’.

Multi Cancer H3K27ac HiChIP dataset collection

Both GEO¹⁴³ and SRA¹⁴⁴ databases were used to screen for HiChIP H3k27ac cancer datasets. We selected at least two datasets of high quality from the above databases for BRCA (n=3), SKCM (n=3), ewing sarcoma (n=2), UCEC (n=2), LUSC (n=2) and LIHC (n=3) which we analyzed together with our GSC HiChIP datasets using the same pipeline. Sample accession codes are provided in the Star methods. For datasets with no matched H3k27ac ChIP-seq experiments we utilized “PeakInferHiChIP.sh” which infers peaks from the HiChIP dataset with default parameters and the output of this algorithm was used in FitHIChIP as a peak file to call loops. All TCGA H3K27ac HiChIP datasets (n=69) were analyzed in the same manner as the HiChIP datasets in this manuscript, starting from filtered paired-end reads (.allvalidPairs files) and H3K27ac peak files (.narrowPeak files) provided by Howard Chang.

Multi Cancer H3K27ac HiChIP dataset dimensionality reduction and clustering

To perform dimensionality reduction and clustering for our H3K27ac HiChIP multicancer data set, the top 10% of hubs (10 kb genomic regions) by connectivity per sample were selected and filtered to retain hubs that were present in >50% of samples in at least one cancer type. This resulted in a list of 9,958 hubs which was used to construct a matrix of connectivity values across all 88 samples. UMAP was performed on this matrix using the umap command from the R package ‘umap’ with default parameters. To identify clusters, a shared nearest neighbor graph was constructed from this matrix using the command makeSNNGraph from the R package ‘bluster’ followed by clustering using the command cluster_louvain from the R package ‘igraph’.

Structural Variant Analysis

HiChIP data was processed using the HiCPro pipeline¹³⁶ to generate valid pairs as previously described. Valid pairs were converted to cool format using the hicpro2higlass.sh script. Structural variants were predicted using cool files at 5k,10k and 50k resolution with the EagleC⁸⁹ function ‘predictSV’ using default settings (—prob-cutoff-5k 0.8 —prob-cutoff-10k 0.8 —prob-cutoff-50k 0.99999). To call neoloops, cool files were processed using the NeoLoopFinder commands ‘calculate-cnv’, ‘segment-cnv’ and ‘correct-cnv’. CNV normalized cool files were then processed using the command ‘assemble-complexSVs’ to generate SV assemblies at 10k resolution. Neoloops were called with the command ‘neoloop-caller’ using SV assemblies at 10k resolution and CNV normalized cool files at 5k,10k and 50k resolution as input. Plots were generated using visualization scripts included in the EagleC and NeoLoopFinder pipelines. All called structural variants and characteristics are included in Table S4. In addition, using CSVDriver⁷³ we evaluated whether our analyzed regions of interest are located within significantly rearranged genomic regions potentially evidence of tumoral positive selection observed in their respective cancer types. This method investigates the tissue-specific covariates of the somatic breakpoint empirical proximity curve to understand the pattern of significantly breakpoint clustering.

Survival Analysis

Survival analysis was performed in R with ‘survminer’ and ‘survival’ R packages and all clinical and expression data were collected from TCGA (https://www.cancer.gov/tcga) and GDC portal.¹⁴⁵ Patients with matched RNA and clinical data that contained information regarding their survival status were used for survival analysis. We matched our GBM HiChIP with TCGA’s GBM-LGG dataset and for all cancer types apart from Ewing sarcoma due to the lack of available expression data, we extracted information from following studies: ACC, BLCA, BRCA, COAD, ESCA, KIRC, LIHC, LUAD, LUSC, PRAD, SKCM, STAD, THCA, and UCEC. For each ‘gene signature’ or factor that we tested we stratified patients’ expression profile based on mean expression levels of our gene signature into 4 equally sized groups and performed log-rank test to compare survival outcome between the top and bottom 25% patient groups for each individual cancer type.

QUANTIFICATION AND STATISTICAL ANALYSIS

R language was used for all statistical analysis and comparisons among groups in this manuscript. Two-tailed Wilcoxon rank sum test was used to compare medians between two groups and two-tailed Student’s t-test for comparison of the means. K-means was used to find group of peaks or hubs with similar patterns across our datasets. All statistics are provided in Table S7.

Supplementary Material

1

Document S1. Figures S1–S6 and legends

2

Table S1. All Genomics Data Sets for GSCs and NSCs from this study, related to Figure 1.

3

Table S2. GSC Hyperconnected 3D Hub Information: coordinates, interacting genes and coregulation values, related to Figures 1–3.

4

Table S3. CRISPRi GLICO scRNAseq information: cell characteristics, differentially expressed genes, gene ontology, related to Figure 4.

5

Table S4. Multi cancer H3K27ac HiChIP 3D hub information: sample characteristics, 3D hubs by cluster, EagleC SV predictions per sample, related to Figures 5 and 6.

6

Table S5. Multi cancer H3K27ac HiChIP 3D hub cluster gene ontology and ChEA, related to Figures 5 and 6.

7

Table S6. RT-qPCR primers, guide RNAs for CRISPRi experiments, antibody information, related to STAR Methods.

8

Table S7. Statistics Source Data for all figures, related to Figures 1–6.

9

Table S8. Coordinates of 3D Hubs per cancer type, related to Figure 5.

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
H3K27ac	Abcam	ab4729
H3K27ac	active motif	91193
Critical commercial assays
Arima-HiC+ kit	Arima	A101020
TruSeq stranded total RNA kit	Illumina	20020596
KAPA Hyper prep kit	KAPA Biosystems	KK8500
Swift Accel-NGS 2S Plus DNA Library Kit	Swift Biosciences	21024
Deposited data
H3K27ac HiChIP, H3K27ac ChIP-seq, ATAC-seq, RNA-seq, and scRNA-seq from this study	This study	SuperSeries GSE262089
Published wt GSC scRNA-seq: 320, 728, 810, 1206	Pine et al.12	PRJNA595375
Published GSC and NSC H3K27ac ChIP-seq	Mack et al.43	GSE119776
H3K27ac HiChIP: T-47D an Ishikawa cell lines	Ginley-Hidinger et al.146	GSE227242
H3K27ac HiChIP: MDA-MB-453 and MCF cell lines	Watt et al.147	GSE157381
H3K27ac HiChIP: ARK1 cell line	O’Mara et al.148	GSE137936
H3K27ac HiChIP: A673 cell line	Surdez et al.149	GSE133227
H3K27ac HiChIP: TC71 cell line	Surdez et al.149	GSE133228
H3K27ac HiChIP: WM-266–4 and COLO829 cell lines	Donohue et al.150	GSE188401
H3K27ac HiChIP: M14 cell line	Chu et al.151	GSE156772
H3K27ac HiChIP: NCI-H889 cell line	Pongor et al.152	GSE206351
H3K27ac HiChIP: NCI-H524 cell line	Pongor et al.152	GSE206352
H3K27ac HiChIP: 3 patient-derived HCC tumors	Jeon et al.153	GSE212055
H3K27ac HiChIP TCGA dataset	Yost et al.55	https://www.cancer.gov/tcga
Published GDC dataset	Grossman et al.145	https://portal.gdc.cancer.gov/
Experimental models: Cell lines
Patient-Derived Glioma Stem Cells: 320, 728, 810, 1206	National Institutes of Health and Weill Cornell Medicine/NYU Presbyterian Hospital	N/A
NIH-registered human H1 (WA01) embryonic stem cells	WiCell Research Institute	Cat# WA01; RRID: CVCL_9771
NIH-registered human H9 (WA09) embryonic stem cells	WiCell Research Institute	Cat# WA09; RRID: CVCL_9773
Oligonucleotides
For CRISPRi guide RNA sequences and RT-qPCR primer sequences, see Table S8
Recombinant DNA
psPAX2	N/A	Addgene #12260
VSV-G	N/A	Addgene # 1488
pLenti CMV rtTA3 Blast	N/A	Addgene# 26429
TRE-KRAB-dCas9-IRES-GFP	Fulco et al.123	Addgene# 85556
pLKO5.GRNA.EFS.PAC vector	Heckl et al.124	Addgene #57825
Software and algorithms
TopHat2 (version 2.11)	Trapnell et al.129	https://github.com/DaehwanKimLab/tophat2
Samtools	Li et al.130	https://github.com/samtools/
htseq-count	Anders et al.131	https://github.com/htseq
DESeq2	Anders et al.132	https://github.com/thelovelab/DESeq2
bowtie2 (version 2.3.4.1)	Langmead et al.133	https://github.com/BenLangmead/bowtie2
Bedtools	Quinlan et al.134	https://github.com/arq5x/bedtools2
MACS2	Zhang et al.135	https://github.com/macs3-project/MACS
HiC-Pro pipeline (version 3.0.0)	Servant et al.136	https://github.com/nservant/HiC-Pro
Juicer-tools	Durand et al.137	https://github.com/aidenlab/JuicerTools
FitHiChIP (release 9.0)	Bhattacharyya et al.50	https://github.com/ay-lab/FitHiChIP
Cell Ranger 2.0.1 pipeline	10X Genomics	https://github.com/10XGenomics/cellranger
Seurat v4.0	Stuart et al.78	https://github.com/satijalab/seurat
ALRA	Linderman et al.138	https://github.com/KlugerLab/ALRA
DoubletFinder (version 2.0.3)	McGinnis et al.139	https://github.com/chris-mcginnisucsf/DoubletFinder
Nebulosa (version 1.8.0)	Leinonen et al.144	https://github.com/powellgenomicslab/Nebulosa
EnrichR	Chen et al.141	https://github.com/wjawaid/enrichR
LOLA	Sheffield et al.142	https://github.com/nsheff/LOLA
scran	Lun et al.154	https://github.com/elswob/SCRAN
umap	McInnes et al.155	https://github.com/tkonopka/umap
bluster	Lun et al.156	https://www.bioconductor.org/packages/release/bioc/html/bluster.html
igraph	Csardi et al.157	https://igraph.org
HOMER	Heinz et al.158	http://homer.ucsd.edu/homer/motif/
Extreme limiting dilution analysis software	Hu et al.126	http://bioinf.wehi.edu.au/software/elda
Other

HIGHLIGHTS.

• GBM 3D regulatory hubs enrich for coregulated genes and expand oncogenic networks.

• CRISPRi silencing of 3D hubs alter transcriptional states and cellular properties.

• Hubs across cancers associate with tumor-specific and universal oncogenic programs.

• Most hyperconnected 3D hubs do not overlap with structural variants.