Chromatin rewiring in endothelial cells upon cohesin reduction redistributes genome-wide FLI1 binding

ABSTRACT

Three-dimensional genome organization is crucial for gene regulation and cell identity. The architectural proteins CTCF and the cohesin complex mediate chromatin looping, shaping cell-type-specific transcription. While their functions have been extensively characterized in embryonic stem cells and other proliferative systems, the regulation of transcription by these two ubiquitously expressed factors in terminally differentiated cells, such as endothelial cells (ECs), remains poorly understood. Moreover, interpretation of prior studies is further complicated by the essential requirement of CTCF and cohesin for cell viability, as most rely on near-complete depletion. Consequently, the extent to which CTCF and cohesin function redundantly, cooperatively, or independently to regulate cell-type-specific programs in differentiated cells remains a major knowledge gap. To address this, we reduced CTCF or the core cohesin subunit RAD21 to approximately 50% protein levels in an isogenic system of human induced pluripotent stem cells (iPSCs) and iPSC-derived endothelial cells (iEcs), enabling direct comparison between undifferentiated and lineage-committed cell states. CUT&Tag and RNA sequencing revealed a striking, EC-specific loss of CTCF and RAD21 occupancy at gene promoters following partial depletion; RAD21 reduction had a more pronounced impact on endothelial gene expression than CTCF depletion. RAD21 loss also led to a marked redistribution of FLI1 binding, a transcription factor essential for endothelial identity and function. Collectively, our findings highlight critical contributions of RAD21 in EC-specific gene expression through coordinated control of chromatin architecture and modulating transcription factor occupancy, providing mechanistic insight into cohesin-dependent regulation of vascular transcriptional programs.

Introduction

The three-dimensional (3D) organization of the genome is a fundamental regulator of gene expression and cell identity [1,2]. This hierarchical structure partitions regulatory interactions to support constitutive, cell-type-specific, and inducible transcriptional programs. Chromatin is hierarchically organized into topologically associating domains (TADs), which are largely conserved across cell types [1]. These TADs are further divided into sub-TADs, chromatin loops that not only bring enhancers into proximity with their target promoters to activate transcription but also insulate promoters from non-cognate distal enhancers to prevent inappropriate gene activation, and are critical for cell-type-specific gene expression programs. These higher-order structures are primarily governed by the DNA-binding protein CTCF and the cohesin complex, both of which are essential for cell viability. In addition to their well-characterized roles in essential cellular processes such as DNA repair and cell-cycle regulation, CTCF is critical in maintaining chromatin organization. CTCF defines TAD boundaries, anchors some chromatin loops, and regulates cohesin-mediated loop extrusion in a DNA tension-dependent manner [3]. This regulation shapes sub-TAD structure, chromatin loops, and enables activation or repression of gene expression in a cell-type-specific context. The cohesin complex, composed of RAD21, SMC3, SMC1A, and STAG2, is essential for the formation and maintenance of chromatin loops that organize the 3D genome structure [4]. RAD21 interaction with the cohesin loader complex plays a critical role in cohesin loading onto chromatin, and its mutation or deficiency reduces chromatin-bound cohesin levels by ~50% [4]. In addition to their cooperative roles in shaping chromatin architecture, CTCF and cohesin also exert independent regulatory effects on chromatin dynamics [5]. These independent functions also regulate lineage-specific transcriptional programs during development, extending their role beyond genome organization. For example, CTCF is essential for cardiac development, where it promotes enhancer–promoter interaction to coordinate cardiomyocyte differentiation and maturation [6]. Deletion of CTCF in cardiac progenitor cells disrupts chromatin interactions and gene expression, leading to severe cardiac defects and embryonic lethality [6]. Likewise, it has been reported that RAD21 plays a critical role in the development of thymocytes by organizing the chromatin architecture at the T-cell receptor α (Tcra) locus [7]. Loss of RAD21 disrupts long-range promoter–enhancer interactions required for Tcra rearrangement and impairs thymocyte differentiation without affecting cell survival, highlighting another cell-type-specific regulatory function of cohesin [7]. Although the architectural roles of CTCF and cohesin have been extensively characterized in embryonic stem cells (ESCs) and proliferative systems [5,8–10], their cell-type-specific functions in terminally differentiated cells, such as endothelial cells (ECs), remain poorly understood. Uncovering these mechanisms is crucial, as they may reveal novel, context-dependent functions of cohesin beyond its canonical architectural roles.

However, the extent to which CTCF and cohesin function redundantly, cooperatively, or independently in shaping cell-type-specific versus more general transcriptional outputs remains unclear. This represents a critical knowledge gap, since it is difficult to interpret studies where there is a complete loss of either complex, because both are required for cell viability. These findings underscore the need to investigate the distinct and overlapping functions of CTCF and cohesin in lineage-specific contexts. To address this knowledge gap, we systematically depleted CTCF and the core cohesion subunit RAD21 to 50% protein levels to better mimic what occurs in disease states in both isogenic induced pluripotent stem cells (iPSCs) and iPSC-derived endothelial cells (iEcs), representing a generalized versus a lineage-committed cell type, respectively, in an isogenic system. This allowed us to compare two isogenic cell-types, iPSCs, and iEcs, and to compare the effects of depletion on ‘undifferentiated’ and ‘differentiated’ cells simultaneously. Using Cleavage Under Targets and Tagmentation (CUT&Tag) [11] and RNA sequencing, we compared chromatin architecture and gene expression across these cell types to delineate shared versus unique contributions of CTCF and cohesin in a cell-type-specific context. We observed a striking loss of promoter occupancy of CTCF and RAD21 following their depletion, unique to ECs. Notably, RAD21 depletion led to a redistribution in the binding of a critical transcription factor, FLI1, in ECs. Moreover, our data uncover that CTCF and RAD21 are critical for maintaining long-range promoter–enhancer interactions that reinforce EC-specific transcriptional programs. Together, these findings suggest the essential, cell-type-specific roles of CTCF and RAD21 in organizing chromatin and orchestrating vascular transcriptional networks.

Materials and methods

iEcs differentiation from iPSCs

We followed the same protocol used by Pandey et al. [12] to differentiate the human iPSC line 039B into iEcs. Briefly, iPSCs were cultured in mTeSR™ Plus medium (STEMCELL Technologies) on Matrigel (Corning)-coated dishes. For differentiation, cells were dissociated using Accutase (STEMCELL Technologies) and seeded onto Matrigel-coated plates at 45,000–50,000 cells/cm² in mTeSR™ Plus medium with 10 µM ROCK inhibitor Y-27632 (STEMCELL Technologies). Mesoderm induction was performed in an N2B27 medium containing 8 µM CHIR99021 (Selleck Chemicals) and 25 ng/ml BMP4 (PeproTech) for 3 days. Subsequently, cells were cultured in StemPro-34 SFM medium (STEMCELL Technologies) with 200 ng/ml VEGF (PeproTech) and 2 µM forskolin (Abcam) for 2 days to induce endothelial lineage. On day 6, CD144-positive cells were purified using CD144 (VE-Cadherin) magnetic beads (Miltenyi Biotec) and cultured in StemPro-34 SFM medium with 50 ng/ml VEGF for an additional 6 days before harvest.

Lentiviral-mediated knockdown of CTCF and RAD21

We followed the HEK293T transient transfection protocol to deplete CTCF/RAD21 in both iPSCs and iEcs described in Liu et al [13]. To prepare the lentiviral particles targeting CTCF/RAD21, HEK293T cells were transiently transfected with the packaging vectors VSV-G and pΔ8.9, alongside lentiviral vectors targeting CTCF, RAD21, or the control vector pLKO.1 puro (#8453, Addgene). We used two different shRNA constructs targeting CTCF (TRCN0000218498 and TRCN0000014548) and RAD21 (TRCN0000148110 and TRCN0000281202) (Millipore Sigma) separately. Viral supernatants were collected at 48 and 72-h post-transfection and concentrated by centrifugation at 6000g overnight at 4°C. Fifty microliters of viral concentrate was taken from the final 500 µl of the concentrated solution to infect the iPSCs and iEcs growing in the Matrigel-coated six-well plates. Media were replaced with fresh medium the next day. After 48 h of infection, cells were selected using 1 µg/ml puromycin for an additional 48 h.

RNA-sequencing and preliminary data analysis

Total RNA was extracted from the puromycin‑selected CTCF/RAD21-targeting-shRNA or control-viral vector‑infected iPSCs and iEcs. One microgram of total RNA from three replicates of control/shRNA targeting CTCF/RAD21‑infected iPSCs and iEcs was sent to Novogene for library preparation and subsequent sequencing on a NovaSeq 6000 sequencer with paired-end reads of 2 × 150 bp, generating approximately 86 million reads for each library.

The preliminary data analysis was performed as described by Liu et al [13]. Briefly, after the removal of adapters and reads with low quality (base quality < 20) by Trim Galore (v0.6.2) [14], raw RNA-seq reads were aligned to the human genome (hg38) using HISAT2 (v2.2.1) [15]. Transcript assembly and quantification were performed using StringTie2 (v2.2.1) [16], based on the GENCODE reference annotation (v38) [17]. To minimize the variability among different replicates, we used Combat-seq to perform batch correction.

Differential gene expression analysis

We detected differentially expressed genes (DEGs) between CTCF/RAD21-depleted iPSCs and iEcs compared to the respective pLKO (control virus, also referred to as Empty Vector-EV) infected cells by using DESeq2 [18]. We considered significant genes (padj value is ≤0.05) ‘UP’ regulated while the Log2FoldChange > 0.5 and ‘DOWN’ regulated while the Log2FoldChange < 0.5. All the volcano plots showing the distribution of UP and DOWN-regulated genes are made using ggplot2. We extracted a list of Endothelial-specific genes from PanglaoDB [19] to detect the EC-specific genes in our differentially expressed genes list.

CUT&Tag in control and depleted iPSCs and iEcs

We performed CUT&Tag to map the genome-wide binding of CTCF, RAD21, H3K27me3, H3K4me3, and IgG (negative control) in the pLKO-infected and CTCF/RAD21-depleted iPSCs (three replicates) and iEcs (two replicates). FLI1 CUT&Tag was performed in pLKO and RAD21-depleted iEcs (three replicates), while H3K27Ac and H3K4me1 CUT&Tag (three replicates) were performed only in uninfected iPSCs and iEcs. The protocol was adapted from Henikoff et al. [11]. We harvested cells by spinning down at 600g for 3 min and then isolated nuclei by incubating the cell pellet with nuclei isolation buffer (20 mM HEPES pH 7.9, 10 mM KCl, 0.1% Triton-X 100, 20% glycerol, 0.5 mM Spermidine, 1X EDTA-free protease inhibitor) for 10 min on ice, followed by centrifugation at 1300g for 4 min at 4°C. Nuclei were resuspended in wash-150 buffer (20 mM HEPES, pH 7.5, 150 mM NaCl, 0.5 mM spermidine and 1X protease inhibitor cocktail). We used 50,000 nuclei for each antibody reaction. Nuclei were mixed with activated Concanavalin A beads for 10 min at room temperature (RT), magnetized to remove the liquid, and resuspended in the appropriate primary‑antibody solution. We used every primary antibody at 1:50 dilution, including CTCF (EpiCypher, #13–2014), RAD21 (GTX, #106012), H3K27me3 (CST, #9733), H3K4me3 (Active Motif, #39060), H3K27Ac (Active Motif, #39133), H3K4me1 (Abcam, #ab8895), FLI1 (Abcam, #ab133485), IgG (CST, #2729S), and incubated for overnight at 4°C on an end-over-end rotator. After washing away the primary antibody solution, nuclei attached beads were incubated with 1:100 diluted secondary antibody (Guinea Pig anti-Rabbit IgG antibody Abcam #ab46540) in wash-150 buffer and incubated on end-over-end rotator for 1 h at RT. After washing the secondary antibody solution, beads were resuspended with 1:20 diluted pAG-Tn5 (Epicypher #15–1017) in Wash-300 buffer (300 mM NaCl in wash-150 buffer) and incubated for 1 h at RT. Next, the beads were resuspended with 50 μl of tagmentation buffer (10 mM MgCl₂ in wash-300 buffer) and incubated at 37°C for 1 h. To stop tagmentation, 10 mM TAPS buffer with 0.2 mM EDTA was added. After removing this TAPS solution, we added 10 mM TAPS buffer with 1% SDS, and 20 mg/ml Thermo labile Proteinase K to the beads and incubated them at 37°C for 1 h, followed by an additional hour at 58°C to deactivate the Proteinase K. To amplify the tagmented DNA on beads, we proceeded directly to PCR by adding 15 μl of 6% Triton-X-100, 25 μl of 2X NEB Q5 HiFi polymerase, and 2 μl of 10 μM of unique i5 and i7 sequencing primers for each antibody/sample. Each sample was placed in a thermocycler with a heated lid using the following cycling conditions: 58°C for 5 min, 72°C for 5 min (gap filling); 98°C for 5 min; 14 cycles of 98°C for 10 sec, and 60°C for 10 sec; final extension at 72°C for 1 min and hold at 8°C. Post-PCR clean-up was performed by adding 1.3 volume of Ampure XP beads (Beckman Counter), and libraries were incubated with the beads for 15 min at RT, washed twice gently with 80% ethanol, and eluted in 25 µL 10 mM Tris, pH 8.0. The concentration of each library was measured at Tapestation by using the High‑Sensitivity D1000 tape. An equal amount of each library was taken to make a pool, and finally, the concentration of this library pool was quantified by the KAPA Quant kit by following the manufacturer’s protocol. The iPSC libraries were sequenced on the NextSeq 500 with paired-end 2 × 75 bp reads, while the iEC libraries were sequenced on the NovaSeq 6000 with paired-end 2 × 150 bp reads at Novogene. Additionally, three replicates of H3K4me1 and H3K27ac libraries for both iPSCs and iEcs were sent to Novogene and sequenced under the same conditions as iEC samples.

CUT&Tag data analysis

For each of the iPSCs CUT&Tag libraries, we first concatenated four FASTQ files generated from each lane of the NextSeq 500. For each of the iEC CUT&Tag libraries, the adapters and low-quality sequences (>20 bp) were trimmed by Cutadapt [20] using the criterion ‘-a CTGTCTCTTATACACATCT ‑m 20 ‑q 20.’ Then all the raw reads were aligned to the human reference genome (GRCh38, NCBI) using Bowtie2 –end-to-end – very-sensitive – no-mixed – no-discordant [21]. The PCR duplicate reads were removed using Picard [22]. After removing PCR duplicates, replicate alignment files for each sample were merged and sorted by using $PICARD MergeSamFiles. These merged alignment files were used for peak calling by MACS2 (2.2.9.1) [23]. We called the ‘broad’ peaks for H3K27me3 and H3K4me1, and for the rest of the marks, ‘narrow’ peaks were called.

Overlap peak analysis

The global reduction of the numbers of CTCF/RAD21 binding sites after their depletion in iPSCs and iEcs was detected by the ChIPpeakAnno (3.34.1) package (findOverlapsOfPeaks function) [24]. The corresponding Venn diagrams were prepared using the makeVennDiagram function from the same package. The corresponding genomic features for all peaks were annotated by using the ChIPseeker package (1.36.0) [25] with TxDb.Hsapiens.UCSC.hg38.knownGene (3.17.0) [26] and org.Hs.eg.db (3.17.0) [27].

CTCF/RAD21-dependent chromatin-loop analysis

To analyze chromatin‑loop dependency on CTCF/RAD21, we implemented a computational pipeline using R. We examined MicroC data from the iEC control condition generated by Liu et al. 2024 [13], focusing on loop anchors under both control and CTCF/RAD21 knockdown conditions. Our approach extended anchor regions by 5 kb in both directions and quantified CTCF‑binding signals within these windows using the GenomicRanges package [28] for genomic‑interval manipulation and findOverlaps for identifying peak intersections. For each anchor, we calculated the log2 fold change between knockdown and control conditions (with a pseudocount of 1 to handle zero values). Statistical significance was assessed using z-score transformation with the scale function in R, followed by two-tailed p-value calculation. We classified chromatin loops into four distinct categories based on explicit criteria: (i) CTCF/RAD21-independent loops: loops where fold change at both anchors is greater than −1 (ii) Not-significant loops: loops where p-values at both anchors > 0.05 (iii) Strong CTCF/RAD21‑dependent anchors (both anchors lost): loops where fold change at both anchors is less than −1, with p-value < 0.05 (iv) Partial CTCF/RAD21 dependent (one anchor lost): loops where fold change at anchor 1 or fold change at anchor 2 is less than −1, with p-value < 0.05. This classification framework was implemented using dplyr’s case_when function [29], allowing for systematic categorization of loops based on their dependency patterns. Visualization was performed with ggplot2 [30], creating scatter plots of log2 fold changes at each anchor. For iPSC, we downloaded 4DNFI31XBNMI.mcool (iPSC WTC-11 AAVS1-GFP C6 MicroC) from 4D Nucleosome [31,32]. We used Mustache 5 kb resolution to call the loops [33]. Then, the same analysis was performed to analyze chromatin‑loop dependency on CTCF/RAD21 in iPSC.

Genomic feature analysis of CTCF/RAD21-sensitive loops

We converted the above-mentioned categories of chromatin loops (both anchors lost , one anchor lost , no anchor lost) into genomic ranges using the GenomicRanges [28] package, separating each loop into its constituent anchors (anchor1 and anchor2). Loop anchors were then annotated using the ChIPseeker package (1.36.0) [25] with a TSS region defined as ±2kb, referencing the UCSC hg38 known gene database (TxDb.Hsapiens.UCSC.hg38.knownGene (3.17.0)) [26] and Homo sapiens gene annotation database (org.Hs.eg.db (3.17.0)) [27]. For simplified analysis, we categorized the genomic features into five main groups: ‘Promoter,’ ‘Exonic,’ ‘Intronic,’ ‘Distal Intergenic,’ and ‘Other,’ using a custom function implementing dplyr’s case_when logic [29]. We then analyzed the co-occurrence of these features at loop anchor pairs, creating interaction pairs (e.g., ‘Promoter-Enhancer’) for each loop. To ensure unbiased analysis of interaction types, we standardized anchor‑pair nomenclature by combining equivalent orientations (e.g., ‘Promoter-Enhancer’ and ‘Enhancer-Promoter’ were merged). For statistical comparison between loop categories, we constructed contingency tables and performed pairwise Fisher’s exact tests with Benjamini–Hochberg correction for multiple comparisons. This allowed us to identify significant differences in the distribution of genomic element interactions among both anchors lost , one anchor lost , and no anchor lost chromatin loops. Results were visualized using ggplot2, with bars representing the percentage of each interaction type within each loop category and statistical significance indicated by asterisks (*p < 0.05, **p < 0.01, ***p < 0.001).

CTCF and RAD21 co‑dependency at loop anchors

Using the CTCF and RAD21 dependent and independent chromatin loop information, the common loops were detected, which are present in both CTCF and RAD21 datasets, to ensure direct comparability. A contingency table of the co-occurrence of these dependency categories was constructed to examine potential associations between CTCF and RAD21 binding dynamics at the same genomic locations. The relationship between CTCF and RAD21 dependency classes was statistically evaluated using a chi-square test [34] for independence to determine whether the pattern of CTCF dependency was significantly associated with the pattern of RAD21 dependency. A color-coded heatmap using ggplot2 was created, where each cell represented the count of loops in the corresponding CTCF-RAD21 dependency category pair. Cells were colored according to standardized residuals from the chi-square test [34], with red indicating positive associations (more loops than expected by chance) and blue indicating negative associations (fewer loops than expected). Statistical significance levels were indicated with asterisks (*p < 0.05, **p < 0.01, ***p < 0.001).

Loop prediction of CTCF/RAD21-depleted iEcs

A computational framework was developed to analyze chromatin loops and predict their maintenance following CTCF/RAD21 depletion in iEcs. The model integrated chromatin interaction data from Micro-C with CUT&Tag profiles of chromatin architecture proteins (CTCF, RAD21) and histone modifications (H3K27me3, H3K4me3) from control conditions. For each loop anchor, features were extracted within 5kb windows using the GenomicRanges package [28], capturing both local and flanking binding patterns. To train the model, a matched set of negative examples was generated, preserving the genomic distances and chromosomal distribution characteristics of real loops. A Random Forest classifier (2000 trees) was implemented using the randomForest package [35], with stratified 5-fold cross-validation to ensure robust prediction of loop maintenance. Model performance was assessed through ROC analysis (pROC package) [36] and visualized using ggplot2. The trained model was then applied to CUT&Tag profiles from CTCF/RAD21-depleted iEC as well as the control, and CTCF/RAD21-depleted iPSCs to predict loop maintenance probabilities. Statistical validation included Wilcoxon signed-rank tests for comparing prediction scores and chi-square tests [34] for loop count differences between conditions. Reproducibility was maintained through consistent random seed settings (seed = 42) across all analyses. We validated the model performance by predicting chromatin-loop maintenance in the control and CTCF-depleted mouse embryonic stem cells (mESCs), using previously published data by Nora et al. 2017 [9]. To get the actual number of chromatin loops in mESCs, two replicates of Hi-C contact matrices files (.cool files) at the control (GSM2644945 and GSM2644946) and the CTCF-depleted (GSM2644947 and GSM2644948) conditions were merged, and loops were called using Mustache [33] at 20 kb resolution.

Analysis of loop-associated DEGs

A computational approach was developed to identify and characterize DEGs associated with chromatin loops predicted to be either maintained or lost following CTCF/RAD21 depletion in iEcs. For each chromatin loop, two distinct regions were analyzed: the loop body (the region between loop anchors) and flanking regions (extending 50 kb upstream and downstream of loop boundaries). Using the hg38 RefGene annotation (hg38.refGene.gtf), genes were categorized based on their spatial relationship to loops as either ‘Inside’ (overlapping the loop body) or ‘Flanking’ (within the 50 kb flanking regions). Differential gene expression analysis was performed by comparing CTCF/RAD21-depleted conditions to iEC controls, with significant DEGs defined by an adjusted p-value < 0.05 and |log2FoldChange| >0.5. The analysis pipeline, implemented in R using the GenomicRanges [28] package, systematically classified genes into four distinct categories: (i) genes inside lost loops, (ii) genes in flanking regions of lost loops, (iii) genes inside maintained loops, and (iv) genes in flanking regions of maintained loops.

RAD21 and FLI1 peak distribution relative to TSS

Genomic coordinates and distance-to-TSS information for Rad21 lost promoter-specific sites and FLI1 binding sites were obtained from their corresponding annotated peak files. Genes present in both datasets were identified, and only these common genes were retained for paired analysis. For each common gene, the Rad21 binding site closest to the TSS was selected (if multiple sites were present). All corresponding FLI1 binding sites for each gene were paired with the selected Rad21 site. To visualize the spatial distribution of binding sites, distances to the TSS were binned using variable-width bins: 10 kb bins from −50 kb to −11 kb and from 10 kb to 50 kb, and finer 1 kb bins from −10 kb to +9 kb, providing higher resolution near the TSS. Each binding site was assigned to a bin, and the number of sites per bin was counted for Rad21, FLI1 in control, and FLI1 in Rad21-depleted conditions. For visualization, the binned counts were normalized within each binding site type to the maximum count (expressed as a percentage). This normalization allowed direct comparison of distribution shapes across conditions, independent of absolute binding site numbers. The resulting data were plotted as line graphs using ggplot2, with the x-axis representing distance to the TSS (in kb) and the y-axis showing the percentage of maximum binding. Distinct colors were used for each binding site type (Rad21: grey, FLI1-Control: blue, FLI1-shRad21: red).

Micro-C loops between RAD21-lost promoters and FLI1

GRanges objects were created using Rad21 binding sites (lost upon shRAD21) and FLI1 binding sites (in both control and shRAD21 conditions). Overlaps between Rad21 binding sites and loop anchors (derived from iEC control Micro-C data) were identified using the findOverlaps function from GenomicRanges. Loops were classified based on whether Rad21 binding was present at either anchor. For each loop with Rad21 at one anchor, the presence of FLI1 binding at the opposite anchor was assessed separately for control and shRAD21 conditions. This was accomplished by iterating over the relevant loops and using findOverlaps to detect FLI1 binding at the corresponding anchor. Loops with Rad21 and FLI1 at opposite anchors were cataloged for both conditions. The total number of loops with Rad21 and FLI1 at opposite anchors was determined for each condition. Overlap between conditions was assessed by identifying loops present in both control and shRAD21 datasets, as well as those unique to each condition. For loops with Rad21 and FLI1 at opposite anchors, the genomic distance between Rad21 and FLI1 binding sites was calculated for each pair within the same loop and chromosome. These distances were compiled for both control and shRAD21 conditions. Bar plots were generated using ggplot2 to compare the number of loops with Rad21-FLI1 at opposite anchors between conditions, and to visualize the overlap loops between the two conditions. Violin and box plots were used to display the distribution of distances between Rad21 and FLI1 binding sites within loops for each condition.

Enrichment plot for FLI1 binding categories

RAD21-responsive genes following depletion were ranked based on their differential expression following RAD21 depletion. For each gene, a significance score was calculated as -log₁₀(padj) * ± (log₂FC). This metric incorporates both the magnitude and statistical significance of gene expression changes, with positive values indicating significant upregulation and negative values indicating significant downregulation. Genes were then ordered from most significantly upregulated to most significantly downregulated. Gene sets were defined according to FLI1 binding change categories (Increased, Decreased, Unchanged), following RAD21 depletion in iEcs. For each category, the set of genes with the corresponding FLI1 binding change was extracted.

A running enrichment score was calculated for each FLI1 binding category using a custom R function. For a given gene set, the algorithm iterates through the ranked gene list, incrementing a running sum when a gene is present in the set (a ‘hit’) and decrementing it otherwise (a ‘miss’). Hits are weighted by the absolute value of the significance score, so that more significant genes contribute more to the enrichment score. The running sum is normalized by the total weight of the gene set and the number of genes outside the set, following the approach of the original GSEA algorithm [37]. The resulting enrichment score profile reflects the degree to which a gene set is concentrated at the top or bottom of the ranked list. To quantify the overall enrichment, the area under the enrichment score curve (AUC) was calculated using the trapezoidal rule, with the x-axis normalized to the range [0, 1]. In this analysis, positive AUC values indicate enrichment of the gene set among significantly upregulated genes (left side of the ranked list), while negative AUC values indicate enrichment among significantly downregulated genes (right side of the ranked list). An AUC value near zero indicates no strong enrichment in either direction. Enrichment score profiles for each FLI1 binding category were visualized using ggplot2.

Enhancer analysis

The binding sites of H3K4me1 and H3K27Ac were used to identify the iPSCs and iEcs-specific enhancers. The common binding sites that do not fall at the promoter between H3K4me1 and H3K27Ac were considered active enhancers, whereas unique H3K4me1 sites (not H3K27Ac) were treated as poised enhancers. GenomicRanges (1.52.1) [28] was used to convert the peak files to genomic ranges to perform the downstream analyses. ChIPpeakAnno (3.34.1) package [24] (findOverlapsOfPeaks function) was used to identify the common and unique peaks between these two marks. The genomic features for all peaks were annotated by using the ChIPseeker package (1.36.0) [25] with TxDb.Hsapiens.UCSC.hg38.knownGene (3.17.0) [26] and org.Hs.eg.db (3.17.0) [27]. The Dplyr package [29] was used to select only non-promoter peaks to mark the poised vs active enhancer. The non-promoter-specific H3K27Ac binding sites were used to detect the iEC-specific super-enhancer by using the Rank Ordering of Super-Enhancers (ROSE) package [38,39]

Statistical information

Statistical analyses were performed using GraphPad Prism and R. For qPCR, two-tailed unpaired t-tests were used to compare FLI1 and target gene expression between control and CTCF-/RAD21-depleted iEcs (n = 3 biological replicates per condition), with Holm-Šidák correction for multiple comparisons. Error bars represent mean ± SEM. The p-value is indicated as follows: *p < 0.05, **p < 0.01, ***p < 0.001. Fisher’s exact test was used to assess the enrichment of lost RAD21/CTCF promoter peaks versus non-promoter peaks in iEcs and iPSCs, upon CTCF and RAD21 depletion (* p < 10⁻⁴, ** p < 10⁻⁶, *** p < 10⁻⁸). Fisher’s test with Benjamini–Hochberg correction was used to compare RAD21-sensitive loop categories (Distal Intergenic, Exon, Intron, Promoter). The p-value is indicated as follows: *p < 0.05, **p < 0.01, ***p < 0.001. Chi-square tests were used to evaluate the co-occupancy of RAD21- and CTCF-sensitive loop classes and to test differences between observed and predicted loop numbers after depletion. Predictions were based on a Random Forest model trained on control iEC Micro-C data (n = 1). The p-value is indicated as follows: *p < 0.05, **p < 0.01, ***p < 0.001.

Results

CTCF/RAD21 differentially regulated cell-type-specific gene expression

We hypothesized that CTCF and cohesin differentially regulate gene expression in pluripotent (undifferentiated, generalized) versus endothelial states (lineage-committed). To dissect context-dependent functions, we performed independent ~50% knockdown of either CTCF or RAD21 in both iEcs and iPSCs using lentiviral delivery of two independent shRNAs (short hairpin RNA) for each gene (Fig. S1, supplementary section).

After confirming the cell identity of iEcs and iPSCs based on cell type-specific marker gene expression (example: POU5F1, NANOG, etc., for iPSCs, and EDN1, ANGPT2 for iEcs, Fig. S2A) and pathway enrichment (Fig. S2B, supplementary section), we analyzed transcriptional changes in both cell types following independent depletion of CTCF and RAD21. We detected that RAD21 depletion had a greater impact on gene expression than CTCF depletion in both iEcs and iPSCs (compare Figure 1(A,B) for iEcs; Fig. S2C, D for iPSCs). We also validated RNA-seq results by performing quantitative RT-PCR on selected genes following independent depletion of CTCF and RAD21 in iPSCs and iEcs, relative to their respective controls (Fig. S2E-H). Comparison of DEGs between iPSCs and iEcs following CTCF (14) or RAD21 (240) depletion revealed a minimal overlap between the two cell types (Figure 1(C,D)), suggesting that both factors exert highly cell-type-specific regulatory functions.

Figure 1.

CTCF/RAD21 regulates distinct transcriptional programs in iEcs and iPSCs:.

A, and B. Volcano plots showing DEGs in CTCF-depleted (A) and RAD21-depleted iEcs (B) relative to control iEcs. C, and D. Venn diagrams showing shared and cell-type-specific DEGs following CTCF depletion (C) and RAD21 depletion (D) in iPSCs and iEcs. E, and F. GSEA heatmaps showing enriched pathways in iEcs and iPSCs following CTCF depletion (E), and RAD21 depletion (F) compared to their cell-type-matched controls. The color scale represents Normalized Enrichment Score (NES), with values indicated in each cell.

To further dissect the functional consequences of independent loss of CTCF or RAD21, we performed Gene Set Enrichment Analysis (GSEA) based on their respective empty vector (EV)-infected controls. Both shared and distinct pathway alternations were observed in CTCF/RAD21-depleted iEcs and iPSCs (Figure 1(E,F)). CTCF depletion activated inflammatory and disrupted cell-cycle pathways in both cell types (Figure 1(E)), but uniquely suppressed metabolic activity and stress tolerance in iEcs (Figure 1(E)). By contrast, RAD21 depletion broadly suppressed inflammatory pathways in both cell types (Figure 1(F)), implying its role as an activator, while CTCF may act as a repressor under basal conditions. Interestingly, RAD21 depletion had opposing effects on the TGF-β and epithelial-to-mesenchymal transition (EMT) pathways in iEcs versus iPSCs (Figure 1(F)). RAD21 depletion in iEcs was also associated with reduced enrichment of gene sets related to angiogenesis, glycolysis, and Notch signaling in iEcs, indicating that RAD21 may support EC-associated transcriptional program in a cell type restricted manner (Figure 1(F)). Direct comparison of CTCF- and RAD21-regulated DEGs showed minimal overlap within each cell type (Fig. S3A, B), indicating independent roles. Collectively, these observations demonstrate that both proteins function in a cell-type-dependent manner, and can modulate independent gene expression programs.

CTCF/RAD21 binding dynamics orchestrate iEC-specific chromatin architecture

We hypothesized that CTCF and cohesin control cell-type-specific transcriptional programs by shaping unique chromatin-binding landscapes essential for maintaining cellular identity. To this end, we performed CUT&Tag for CTCF and RAD21 in iEcs and iPSCs in control (EV-infected) and depleted conditions. In iEcs, depletion caused a substantial loss of CTCF and RAD21-bound sites, with a small subset of novel ‘gained’ sites (~10% for CTCF and ~7% for RAD21) (Figure 2(A,B), top Venn diagrams). One possible explanation for these gained sites is that depletion of CTCF or RAD21 may alter chromatin accessibility, thereby exposing previously inaccessible regions to residual protein binding. A similar binding pattern was observed in iPSCs following depletion (Fig. S4A,D). Metaprofiles centered on called peaks showed that relative to control, CTCF and RAD21 signals were lower under corresponding depletion not only at lost binding sites but also at preserved sites. This indicates a reduced average signal of CTCF/RAD21 at preserved sites without complete loss of enrichment (Figure 2(A,B), bottom profile plot). Genomic annotation revealed that promoter regions were the most affected in iEcs, unlike iPSCs (Figure 2(C,D); Figure S4B, E). To quantify selective loss of promoter occupancy, we performed Fisher’s exact tests to compare CTCF/RAD21 binding loss at promoter versus non-promoter regions in iEcs and iPSCs after depletion (Figure 2(E)). In iEcs, both RAD21 and CTCF showed significantly greater loss at promoters, whereas in iPSCs the loss is more evenly distributed across promoter and non-promoter regions (Figure 2(E)), suggesting that CTCF/RAD21 may play a more pivotal role in maintaining promoter architecture in differentiated iEcs.

Figure 2.

CTCF/RAD21 depletion reshapes cell-type-specific chromatin architecture:.

A, and B. Top. Venn diagrams showing preserved and unique CTCF (A) and RAD21 (B) binding sites in control (lost) and depleted (shCTCF/shRAD21) (gained) iEcs. Bottom. Profile plots of average CTCF (A) and RAD21 (B) signal at the lost (blue), gained (sky-blue), and preserved (green) binding sites at control and depleted conditions. C, and D. Genomic distribution of lost, gained, and preserved binding sites of CTCF (C), and RAD21 (D). E. Forest plot showing log₂ odds ratios (points) and their 95% confidence intervals (bars) for enrichment of lost promoter peaks versus non-promoter peaks in iEcs (orange) and iPSCs (green), upon independent depletion of CTCF and RAD21; Significance by Fisher’s exact test (* p < 10^–4, ** p < 10^–6, *** p < 10^–8). F, and G. Enriched TF motifs identified by de novo HOMER analysis in lost, gained, and preserved CTCF (F) and RAD21 binding sites (G); Number indicates motifs p-values.

To gain further insight into the functional implications of altered CTCF/RAD21 binding, we performed de novo motif analysis using HOMER on lost, gained, and preserved sites in iEcs, and iPSCs (Figure 2(F,G). This analysis revealed enrichment of the SRSF1 motif at lost sites in both cell types, implicating a potential role of CTCF in alternative splicing regulation (Figure 2(F) and Figure S4C) [40,41]. Unlike iPSCs, RAD21 lost sites in iEcs were enriched with TFs reported to play critical roles in EC-specific functions, angiogenesis, and EC-differentiation and proliferation, including FLI1 and NFYA (Figure 2(G), and Figure S4F) [42,43]. This finding suggests a possible regulatory link between RAD21 and endothelial identity. While RAD21-FLI1 interactions are known in oncogenic contexts [44,45], their relationship in ECs remains unexplored.

RAD21 depletion diminishes promoter occupancy beyond CTCF-RAD21-shared sites

Given the largely non-overlapping gene targets of CTCF and RAD21 in iEcs, we overlapped their genome-wide occupancy. This analysis revealed that ~65% of CTCF sites were co-occupied by RAD21 across all conditions (Figure 3(A)), consistent with CTCF anchoring cohesin loops [46]. However, many unique RAD21 binding sites suggested a CTCF-independent contribution in regulating EC-specific gene expression (Figure 3(A)). Our genomic‑distribution study detected promoter enrichment of shared CTCF-RAD21 binding sites (Figure 3(B)). While depletion of either factor led to reduced occupancy at their respective unique promoter sites in iEcs, RAD21 loss had a more pronounced effect, reducing promoter binding at unique RAD21 sites to 25.1% (vs. 43.4% in control, 44.3% CTCF-KD), compared to a modest change observed at unique CTCF sites (30.3% vs 33.8% in control) (Figure 3(C,D)). Together, these findings indicated that independent depletion of CTCF and RAD21 preferentially disrupts promoter binding in iEcs, with RAD21 exhibiting a more pronounced promoter loss. These disruptions were specific to iEcs and absent in iPSCs, a more generalized and less lineage-committed cell type (Fig. S5). This contrast again suggested a cell-type-specific role of CTCF and RAD21 in maintaining endothelial chromatin architecture and gene expression.

Figure 3.

Genome-wide co‑occupancy of CTCF and RAD21:.

A. Venn diagrams showing shared and unique CTCF and RAD21 binding sites under control conditions (left), following CTCF depletion (middle), and RAD21 depletion (right) in iEcs. B, C, and D. Genomic distribution of shared CTCF-RAD21 sites (B), CTCF-only sites (C), and RAD21-only sites (D) in control, CTCF-depleted, and RAD21-depleted iEcs. E and F. Average RAD21 peak density within ±2000 bp of transcription‑start sites in iEcs (E) and iPSCs (F) for upregulated (light blue), and downregulated (dark blue) DEGs under control conditions, and after RAD21 depletion (iEC (E), and iPSC (F) (upregulated: orange, downregulated: red)).

Based on their context-specific role, we hypothesized that CTCF and cohesin may exert bidirectional control over gene expression. We analyzed gene expression changes (red tracks) and corresponding binding patterns (blue tracks) at representative loci under control and depletion conditions (Fig. S6A-D). Depletion of CTCF or RAD21 caused gene-specific up- or downregulation depending on their binding context. Specifically, CTCF loss increased GATA6 and decreased SLC8A2 expression (Fig. S6A, B), while RAD21 loss increased CYP1B1 and decreased COL1A2, each associated with reduced promoter occupancy of the respective factor (Fig. S6C, D).

To assess genome-wide effects of independent depletion of CTCF and RAD21 on promoter occupancy, we analyzed average binding around transcription start sites (TSS; ±2 kb) for all significant DEGs. Profile plots showed a substantial loss of promoter binding in RAD21- (Figure 3(E)) and CTCF-depleted iEcs (Fig. S6E), irrespective of gene up- or downregulation. In iPSCs, RAD21 depletion did not result in a comparable reduction in promoter binding (Figure 3(F)), and CTCF depletion produced only a modest effect (Fig. S6F), suggesting that these factors regulate gene expression differently in iPSCs, likely through distal rather than promoter-centric mechanisms.

These results indicate context-dependent roles for CTCF and RAD21. In iEcs, they stabilize promoter architecture to modulate gene expression, whereas in iPSCs, they may function through alternative mechanisms that are not primarily promoter-centric. Overall, both factors orchestrate chromatin structure to balance gene activation and repression in a context-specific manner.

CTCF/RAD21 chromatin loops selectively organize promoter contacts with distinct genomic elements

Cell-type-specific chromatin-loop organization regulates gene expression. Thus, we hypothesized that CTCF- and RAD21-mediated bidirectional gene regulation in iEcs occurs through their functions at chromatin-loop anchors. To define loops in control iEcs, we analyzed Micro-C data from Liu et al. 2024 [13], and assessed independent occupancy of CTCF and RAD21 at loop anchors under control and depleted conditions. Binding changes at loop anchors were quantified as the log2 fold changes in depleted versus control cells, and loops were classified as losing binding at both anchors, one anchor, or neither. This analysis revealed distinct requirements for RAD21 and CTCF in maintaining chromatin-loop structures in iEcs (Figure 4(A) and Figure S7A). RAD21 depletion caused largely symmetric disruption, with a similar number of loops losing binding at one (498) or both anchors (455), consisting a broad role in loop stability (Figure 4(A)). By contrast, CTCF loss predominantly affected one anchor (820) more than both (377), suggesting partial compensation by adjacent CTCF sites (Fig. S7A). Despite ~50% global loss, many loops retained RAD21 (2,273) or CTCF (3,447) binding at both anchors, indicating a stable subset of protected loop anchors (Figure 4(A) and Figure S7A). To assess coordinated loss, we compared RAD21 and CTCF signal loss across loop anchors using a contingency table (Figure 4(B)). Each cell shows the number of loops in that category (both, one, no anchors) and the standardized residuals from the Chi-square test; red indicates more loops than expected under independence, and blue indicates fewer than expected. Loops where both factors were retained (5,461 loops) or where both were lost (64 loops) at both anchors were strongly enriched (Figure 4(B), red cells), indicating coordinated retention or loss. In contrast, categories in which only one factor was lost were underrepresented (Figure 4(B), blue cells; negative standardized residuals). That is, asymmetric loss of a single factor at loop anchors occurred less often than expected by chance, consistent with a cooperative role for CTCF/RAD21 in maintaining loop architecture. Together, these findings suggested that while CTCF and cohesin regulate gene expression in a cell-type-specific manner, their structural role at loop anchors is largely cooperative, with partial depletion affecting transcription more than loop stability. However, as this conclusion is derived from our predictive model, a direct experimental validation will be necessary to substantiate these findings, which is a limitation of our study.

Figure 4.

Mechanistic and predictive analysis of CTCF/RAD21-driven chromatin loops and gene regulation in iEcs.

A. Scatter plot of Micro-C-detected loops in control iEcs showing RAD21 binding loss after depletion, categorized by loss at both anchors (red), one anchor (yellow), or neither (blue); grey indicates non-significant changes (log₂FC < −1, p-values < 0.05). B. Heatmap showing overlap between CTCF- and RAD21-sensitive loop categories; numbers indicate shared loops, with chi-square significance (*p < 0.05, **p < 0.01, ***p < 0.001) and standardized residuals (red: positive, blue: negative). C. Bar plots showing the distribution of genomic interaction types, Distal Intergenic (DI), Exon (E), Intron (I), and Promoter (P) across RAD21-sensitive loop categories; significance by Fisher’s exact test with Benjamini–Hochberg correction (*p < 0.05, **p < 0.01, ***p < 0.001). D. Actual (grey) versus Random Forest-predicted (red) loop numbers in control and CTCF-depleted mESCs (Nora et al. [9]), using a model trained on control iEC Micro-C loops. E. Bar graph showing predicted loop numbers after CTCF (green) or RAD21 (blue) depletion compared to control (red); significance by chi-square test (*p < 0.05, **p < 0.01, ***p < 0.001). F. Bar plots showing DEGs located within ±50 kb of loop boundaries (flanking) that are lost or maintained following CTCF (left panel) and RAD21 (right panel) depletion compared to control.

We assessed the distribution of genomic features (promoters (P), introns (I), distal intergenic regions (DI) and exons (E)) at loop anchors, and examined how these distributions change upon loss of RAD21 or CTCF binding. Loops predicted to lose RAD21 at both anchors were enriched for promoter-associated interactions (P–P, P–I, P-DI) compared to other loop categories (Figure 4(C)), suggesting a potential role for RAD21 in stabilizing promoter–promoter, as well as promoter-enhancer (intronic and distal intergenic) contacts in iEcs. In contrast, predicted CTCF loss was more frequently associated with promoter–promoter loops (Fig. S7B). Loops without predicted loss of binding were relatively enriched for non-promoter interactions. Together, these results suggest that RAD21- and CTCF-associated loops may differentially contribute to promoter-focused chromatin organization in iEcs.

Comparison with iPSCs Micro-C data [31,32] showed that while CTCF loss similarly impacted promoter loops, RAD21 loss disrupted promoter contacts only in iEcs (Fig. S7C-F, supplementary information), indicating a more specialized role for RAD21 in iEcs.

We developed a random forest-based model trained on control iEC Micro-C data and CTCF/RAD21 binding data to predict loop loss. The model was validated using a published dataset from Nora et al. 2017 [9], which includes genome-wide CTCF ChIP-seq from control and CTCF-depleted mouse embryonic stem cells (mESCs). The model’s predicted loop numbers closely matched the loop numbers detected experimentally by Hi-C (Figure 4(D)), supporting its accuracy. Therefore, we applied the model to our depleted iEC datasets. This analysis predicted that chromatin loops would decrease by ~35.1% and ~41.8% following independent depletion of CTCF and RAD21, respectively (Figure 4(E)), supporting their essential roles in loop formation in iEcs. We applied the same model to predict chromatin-loop numbers in control and CTCF- or RAD21-depleted iPSCs. We observed estimated reductions of ~22% and ~18% in loops following independent depletion of CTCF and RAD21, respectively (Fig. S7G).

To assess functional impact, we analyzed DEGs within predicted lost loops and flanking regions (±3 kb). CTCF depletion was associated with more upregulated DEGs (Figure 4(F)), consistent with gene derepression, whereas RAD21 depletion led to more downregulated DEGs (Figure 4(G)), indicating a role in maintaining active transcription, likely through enhancer-promoter interaction (Figure 4(C)).

EC-specific association of FLI1-RAD21

To explore how RAD21 may sustain EC-specific transcription, we examined its potential interaction with FLI1. Previous motif analysis revealed enrichment of FLI1 motifs at RAD21 binding sites lost upon depletion (Figure 2(G)), suggesting a RAD21-FLI1 regulatory axis in iEcs. Although FLI1 expression (at both mRNA and protein levels) remained unchanged after RAD21 loss (Figure 5(A)), several direct and indirect target genes (e.g., TEK, VWF, ENG) were downregulated, indicating functional disruption without transcriptional loss (Figure 5(A)).

Figure 5.

Distinct RAD21-FLI1 binding profile in iEcs.

A. Left: qPCR analysis of FLI1, and target gene expression in control and RAD21-depleted iEcs; significance by multiple unpaired t-tests with Holm-Šidák correction (*p < 0.05, **p < 0.01, ***p < 0.001). Right: Western blot of FLI1 in control and RAD21-depleted iEcs. B. Venn diagram showing preserved and unique FLI1 binding sites in control versus RAD21-depleted iEcs (top), with genomic distribution in each condition (bottom). C. Average FLI1 signal (CPM) across ±2 kb around RAD21 peak summits at lost (left), and preserved (right) RAD21 sites, in control (blue) and RAD21-depleted (red) iEcs (top). Shaded bands denote 95% confidence intervals of the mean estimated by bootstrap resampling of regions (B = 1000). padj values are Bonferroni-corrected Wilcoxon rank-sum tests comparing FLI1 signal between conditions at the peak center bin. D. Quantification of chromatin loops with one anchor overlapping RAD21 lost-promoter sites and the other overlapping FLI1 binding sites in control (blue), RAD21-depleted (orange), or both (green) (top). Violin and box plots showing loop distance distributions (kb) between RAD21-lost promoters, and FLI1 loop anchors (bottom). E. Stacked bar plots showing genomic annotation (Distal Intergenic, Exon, Intron, Promoter) of FLI1 target genes in control (left) and RAD21-depleted iEcs (right), colored by expression change upon RAD21 loss (upregulated: red, and downregulated: blue). F. (Left) Mean log₂fold change (± SE) of gene expression grouped by changes in FLI1 binding (increased, unchanged, decreased) following RAD21 depletion, with statistical comparison between increased and unchanged groups (left). Enrichment plot showing the distribution of FLI1 binding categories (Increased: red, Unchanged: yellow, Decreased: blue) across ranked gene expression changes, with enrichment scores (ES) and area under the curve (AUC) indicated.

We hypothesized that RAD21 loss may alter genome-wide FLI1 occupancy. CUT&Tag profiling confirmed widespread FLI1 redistribution, with an increased number of binding sites in RAD21-depleted iEcs (Figure 5(B), Venn diagram), particularly in intronic and distal intergenic regions, indicative of potential enhancer activity (Figure 5(B), bar chart). To determine whether FLI1 redistribution is preferentially linked to RAD21 binding dynamics, we compared FLI1 enrichment at RAD21 lost vs. preserved sites upon its depletion in iEcs. The strongest FLI1 signal was observed at sites of RAD21 loss, suggesting cohesin may restrain FLI1 genomic occupancy (Figure 5(C)).

We next predicted the number of chromatin loops linking RAD21-lost promoters to distal FLI1-bound sites using control iEC Micro-C data [13]. Although a similar number of these loops were detected before (22) and after (24) RAD21 loss (Figure 5(D), bar plot), they showed a limited overlap (Figure 5(D), green bar), and an increased, more variable anchor distance (Figure 5(D), violin plot). This indicates that RAD21 loss may promote the formation of new, weakened enhancer-promoter contacts rather than maintaining existing loops, consistent with RAD21's role in maintaining 3D genome structure.

To understand the functional consequences of FLI1 redistribution, we annotated FLI1 binding sites in both control and RAD21-depleted iEcs by genomic location (promoter, exon, intron, distal-intergenic), and identified associated target genes. We compared these genes with those that were differentially expressed following RAD21 depletion. We detected that FLI1 target genes were more frequently associated with the downregulated rather than the upregulated genes (Figure 5(E), red vs blue bar), particularly for FLI1 binding sites at intronic and intergenic regions. This provides a global overview of how FLI1 target genes are affected by RAD21 loss. It suggests that loss of RAD21 may disrupt chromatin looping, potentially impairing FLI1's enhancer-mediated regulation, and leading to gene repression.

To further quantify the relationship, we grouped RAD21-responsive genes based on whether FLI1 binding increased, remained unchanged, or decreased after RAD21 depletion. As shown in Figure 5(F) left panel, genes with increased FLI1 binding showed stronger downregulation on average compared to other groups. Consistently, the enrichment analysis (Figure 5(F), right) displays that these genes are preferentially detected among the most downregulated genes (AUC = −0.202), pointing to a genome-wide association between increased FLI1 binding and transcriptional repression.

To determine whether this FLI1-RAD21 association is conserved in other endothelial cell types, we downregulated RAD21 in human umbilical vein endothelial cells (HUVEC) using the same shRAD21 lentiviral constructs. We achieved ~50% knockdown of RAD21 at the protein level (Fig. S8A). We then assessed the expression of FLI1, and its target genes. Similar to iEcs, RAD21 depletion did not affect FLI1 expression, but significantly reduced the expression of its target genes, including ENG, TEK, and VWF, compared with control HUVECs (Fig. S8B). These findings further strengthen the idea that RAD21 regulates FLI1-dependent endothelial gene expression without altering FLI1 transcription.

Together, these observations lead us to conclude that RAD21 supports FLI1-driven gene expression in iEcs not by regulating its expression, but by stabilizing its chromatin binding. RAD21 loss leads to genome-wide redistribution of FLI1, particularly to intronic and intergenic regions, disrupting enhancer-promoter communication, and repressing target genes. This underscores a key cohesin-FLI1 axis in maintaining EC-specific gene expression through coordinated regulation of chromatin architecture.

CTCF and RAD21 binding dynamics across enhancer states

Given that enhancer–promoter communication is fundamentally mediated by CTCF- and cohesin-dependent chromatin loops, we next examined how CTCF and RAD21 binding is distributed across distinct enhancer states in iEcs. CTCF and cohesin regulate transcription by facilitating promoter-enhancer communica tions through chromatin loops [47]. Enhancer activity is defined by distinct chromatin marks [48,49], allowing classification into active (H3K4me1-H3K27ac), poised (H3K4me1-H3K27me3), and primed (H3K4me1-only) [49,50]. We first analyzed the genome-wide unique and shared binding regions of H3K4me1, H3K27me3, and H3K27ac in iEcs. H3K4me1 showed substantially greater overlap with H3K27ac (23706) than H3K27me3 (2865) (Figure 6(A)). In addition, H3K4me1-H3K27me3 marks were enriched at promoters (Figure 6(B)), consistent with silent or bivalent gene promoters [51]. In contrast, H3K4me1-H3K27ac and H3K4me1-only marks were predominantly found in intronic and intergenic regions, reflecting distal regulatory roles (Figure 6(B)).

Figure 6.

CTCF/RAD21-specific enhancer peaks.

A. Venn diagram showing shared and unique genomic occupancy of three histone marks: H3K27me3, H3K4me1, and H3K27ac in iEC. B, and C. Bar plots depicting the genome-wide distribution of all (B) and non-promoter (C) binding sites for three enhancer classes: H3K4me1-only (primed), H3K27me3-H3K4me1 (poised), and H3K27ac-H3K4me1 (active). D. Line graph presenting the number of super-enhancers (red) defined by the ranked H3K27ac occupancy signal at the non-promoter region (TSS ±2 kb) in iEcs by ROSE. E, and F. Bar plots showing the percentage of lost and preserved CTCF (D) and RAD21 (E) binding sites within each enhancer class. G. I.G.V.

Browser view of theSHB locus in iEcs. The top five tracks show control chromatin features: super-enhancer calls (black), H3K27ac (grey), H3K4me1 (violet), H3K27me3 (maroon), and H3K4me3 (green), with both signal and peak tracks displayed for each. The bottom four tracks show CTCF, and RAD21 occupancy (signals and peaks) in control versus CTCF-depleted or RAD21-depleted iEcs, respectively. Red boxes highlight regions of CTF/RAD21 co-occupancy overlapping active enhancer marks (H3K27ac/H3K4me1) within the SHB super-enhancer domain.

Focusing on non-promoter loci, introns were preferentially enriched for active and primed enhancers (Figure 6(C)), consistent with prior reports that tissue-specific enhancers often reside within introns [52]. In contrast, poised enhancers were more frequently located in distal intergenic regions (Figure 6(C)), potentially serving as a reservoir of developmentally regulated elements, inactive in iEcs, but primed for future activation. In addition, we also detected 384 iEC-specific super-enhancers by using H3K27Ac non-promoter binding sites by ROSE (Figure 6(D)).

To assess whether CTCF and RAD21 influence enhancer organization, we analyzed their overlap with enhancer classes at lost and preserved binding sites. Both factors showed increased overlap with super-enhancers, followed by active and poised enhancers, at lost sites compared to preserved ones (Figure 6(E,F)), indicating that these enhancer states are sensitive to disruption of CTCF/RAD21-mediated chromatin organization. Specifically, ~6% and ~27% of CTCF-lost sites overlapped with active and super-enhancers, respectively, compared to 2.5% and 5.2% at preserved sites (Figure 6(E)). RAD21 loss showed a stronger effect, with 47% and 15% overlap with active and super-enhancers, respectively (Figure 6(F)), consistent with a more direct role in stabilizing enhancer–promoter loops. Notably, RAD21 retained more preserved binding sites across all enhancer classes than CTCF, suggesting a more stable association with EC-specific enhancer architecture. Taken together, CTCF and RAD21 preferentially occupy and stabilize the iEC-specific enhancers, and their loss may disrupt these regulatory regions, highlighting a critical role in maintaining enhancer-promoter chromatin architecture.

Discussion

In this study, we compared the roles of CTCF and RAD21 in undifferentiated iPSCs and lineage-committed iEcs to understand how chromatin architectural proteins contribute to cell-type-specific gene regulation. Although CTCF and RAD21 are classically viewed as structural organizers of chromatin loops, our findings extend this framework by highlighting their distinct contributions to lineage-specific transcription in iEcs. Notably, RAD21 depletion had a stronger impact on the endothelial transcriptome than CTCF loss (Figure 1), suggesting that cohesin supports lineage-specific regulatory programs independently of CTCF. This cell-type-specific regulation is consistent with genome-wide chromatin interaction studies, showing that while CTCF/cohesin co-occupancy demarcates structural domain boundaries [53], only a subset of these sites contributes to tissue-specific gene expression [54,55]. These studies have shown that cohesin can operate independently of CTCF by co-localizing with tissue-specific TFs, such as estrogen receptor α (ERα) in MCF-7 cells or hepatocyte nuclear factors in HepG2 cells [54]. These findings emphasize that cohesin’s recruitment and function are highly context-dependent, diverging from that of CTCF. These findings, along with other chromatin interaction studies, suggest that CTCF often acts as a static boundary element, while cohesin plays a more dynamic, lineage-specific regulatory role through TF interaction and enhancer-promoter coordination [9,56]. Moreover, recent high-resolution mapping of cohesin dynamics reveals that cohesin facilitates TF binding to chromatin to regulate transcription, particularly in differentiated or lineage-committed cells [57].

By leveraging partial depletion of each factor, a biologically relevant model given the lethality of complete loss, and the clinical relevance of haploinsufficiency [58,59], we uncovered striking context-dependent functions that emerge upon endothelial differentiation. In contrast to iPSCs, iEcs displayed a strong dependence on CTCF- and RAD21-mediated promoter architecture (Figure 2), aligning with reports that only a subset of CTCF/cohesin-bound sites drives tissue-specific gene expression despite their widespread role in defining chromatin domains [53–55]. Supporting this, studies of human ESC differentiation into pancreatic islet cells have demonstrated that CTCF loops are dynamically established during lineage commitment, either through recruitment to new genomic sites or functional activation of pre-existing sites, facilitating enhancer–promoter interactions and transcriptional activation [60]. Similarly, in human monocyte differentiation into macrophages or dendritic cells, cohesin depletion disrupts chromatin loops and impairs the activation of differentiation-specific genes, highlighting its essential role in lineage-associated transcriptional programs [61]. These findings suggest that the dependency on CTCF and cohesin for promoter architecture and gene regulation becomes more pronounced as cells acquire lineage-specific identities. However, our data uniquely demonstrate this dependency in iEcs and reveal a promoter-centric sensitivity that is not observed in undifferentiated iPSCs, highlighting a cell-type-specific regulatory function of CTCF and RAD21 in vascular differentiation.

Consistent with this specialization, CTCF and RAD21 displayed both shared and distinct genomic distributions in iEcs, supporting their context-dependent regulatory functions (Figure 3). A recent study in mESC highlighted that the stability and recruitment of CTCF-independent cohesin at promoters and enhancers are maintained by common TFs like SP1, NFYA, and by a chromatin modifier, WDR5 [5]. Depletion of these factors led to the loss of cohesin on chromatin at those co-bound sites. While depletion of these factors disrupts cohesin occupancy at co-bound sites, loss of cohesin does not, in turn, displace these TFs or chromatin modifiers, highlighting a unidirectional dependence [5].

Together, these studies position cohesin as a dynamic regulator of lineage-specific gene expression, in contrast to the more static boundary-defining role often attributed to CTCF [9,57]. In line with this model, our chromatin-loop analyses suggest that RAD21 primarily supports promoter-regulatory elements (such as promoters, intronic, and distal intergenic) connectivity, whereas CTCF contributes more selectively to promoter–promoter interactions (Figure 4), reinforcing models in which cohesin enables flexible chromatin connectivity required for stable lineage-specific transcriptional programs [62,63]. Despite their distinct effects on cell-type-specific gene regulation, CTCF and cohesin act cooperatively to maintain chromatin‑loop architecture at loop anchors (Figure 4). Notably, partial depletion of either factor had a greater impact on gene expression than on global loop stability, indicating that modest reductions in CTCF and cohesin primarily impair transcriptional regulation rather than the overall structural integrity of chromatin loops.

A key mechanistic insight from our study is the critical association between RAD21 and the endothelial transcription factor FLI1 (Figure 5). Despite unchanged FLI1 expression, RAD21 depletion redistributed FLI1 occupancy, and attenuated transcription of FLI1 target genes, indicating that cohesin may play a role in regulating FLI1 activity at the level of chromatin binding. Importantly, we validated this finding in primary HUVECs, where RAD21 depletion similarly preserved FLI1 expression, while reducing the expression of FLI1 target genes. Together, these findings support a model in which cohesin stabilizes the chromatin interactions required for productive FLI1-dependent enhancer-promoter and endothelial gene regulation. This mechanism echoes cohesin’s broader role in transcriptional regulation across disease contexts. In Ewing sarcoma and TMPRSS2-ERG-positive prostate cancer, RAD21 stabilizes chromatin contacts at FLI1- or ERG-bound enhancers, mitigating oncogene-induced stress and sustaining transcriptional output [44,45]. Likewise, loss of STAG2, another cohesin subunit, disrupts looping between EWSR1-FLI1-bound enhancers and their gene targets, impairing transcription in Ewing sarcoma [64]. Further extending this framework, including our group’s work in an inv(16) acute myeloid leukemia (AML) mouse model, demonstrates that cohesin haploinsufficiency accelerates leukemogenesis by enhancing Fli1 activity in hematopoietic stem and progenitor cells (HSPCs) [65]. Another study in hematopoiesis also demonstrates that STAG2 deficiency leads to aberrantly increased chromatin accessibility at FLI1 targets in both mouse and human HSPCs [66]. In the context of hematopoietic cells, cohesin depletion or genetic haploinsufficiency has been shown previously to cause an increase in accessibility for FLI1/ERG motifs in chromatin accessibility data [67–69]. In contrast to cohesin loss in the hematopoietic system, where enhanced FLI1 chromatin accessibility is correlated with transcriptional activation, facilitating leukemic transformation, our study showed that RAD21 depletion in iEcs leads to FLI1 re-distribution, accompanied by transcriptional repression of vascular gene programs. These observations highlight a potential interplay between RAD21 and FLI1 that may contribute to lineage-specific transcriptional regulation. However, further investigation is needed to clarify the underlying mechanisms, including whether RAD21 directly modulates FLI1 binding or the observed effects are secondary to broader changes in chromatin organization, and to clarify how this interplay contributes to vascular identity. Our findings also suggest potential contributions of CTCF and cohesin in maintaining the chromatin architecture of enhancers (specifically active and super-enhancers), which may play a crucial role in establishing iEC-specific enhancer-promoter communication (Figure 6).

Conclusions

In summary, our study demonstrates that while CTCF and RAD21 both contribute to global chromatin organization, they play distinct and context-dependent roles in shaping endothelial gene expression. Notably, our findings suggest that RAD21 contributes to the regulation of FLI1-dependent transcriptional programs in iEcs, potentially through stabilizing functional chromatin interactions rather than altering FLI1 expression. These findings advance our understanding of how chromatin architecture integrates with TF networks to establish and maintain endothelial identity and may have implications for vascular disease states associated with cohesin dysfunction.

Funding Statement

This work was supported by the National Heart, Lung, and Blood Institute [P01 HL149620 to M.L., S. R., A.S.G., A.W.C., and A.M.G.], National Institute of Diabetes and Digestive and Kidney Diseases [R01 DK134064 to S.R.], National Institute of Allergy and Infectious Diseases [R01 AI183571 to S.R.], and National Institutes of Health [HL173778, HL144098, K24HL152143, R38HL167238 to M.E.W.].

Disclosure statement

No potential conflict of interest was reported by the author(s).

During the preparation of this work, the author(s) used CURSOR to write the code, and ChatGPT to improve the manuscript writing.

Data availability statement

The authors confirm that the data supporting the findings of this study are available within the article. Raw and processed data files for all CUT&Tag and RNA-seq data are available from GEO under accession numbers: GSE300316 (CUT&Tag data; confidential reviewer token is yholeiqujruttsz), GSE283789 (iEC RNA-seq data; confidential reviewer token is ufsfmwukjhidrmn), GSE300315 (iPSC RNA-seq data; confidential reviewer token is wvsvikoixjqzjgf), and GSE283789 (MicroC data; confidential reviewer token is ufsfmwukjhidrmn).

All manuscript codes are available on GitHub via the following link https://github.com/Raolab-VBRI/Chromatin-Rewiring-in-ECs-Upon-Cohesin-Reduction-Redistributes-Genome-Wide-FLI1-Binding.git

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/15592294.2026.2749250