Cohesin-Dependent and -Independent Mechanisms Mediate Chromosomal Contacts between Promoters and Enhancers

Summary

It is currently assumed that 3D chromosomal organization plays a central role in transcriptional control. However, depletion of cohesin and CTCF affects the steady-state levels of only a minority of transcripts. Here, we use high-resolution Capture Hi-C to interrogate the dynamics of chromosomal contacts of all annotated human gene promoters upon degradation of cohesin and CTCF. We show that a majority of promoter-anchored contacts are lost in these conditions, but many contacts with distinct properties are maintained, and some new ones are gained. The rewiring of contacts between promoters and active enhancers upon cohesin degradation associates with rapid changes in target gene transcription as detected by SLAM sequencing (SLAM-seq). These results provide a mechanistic explanation for the limited, but consistent, effects of cohesin and CTCF depletion on steady-state transcription and suggest the existence of both cohesin-dependent and -independent mechanisms of enhancer-promoter pairing.

Graphical Abstract

Highlights

• •The majority of promoter-anchored contacts are lost upon rapid degradation of cohesin or CTCF
• •A significant minority of promoter contacts are retained, and some are gained
• •Cohesin-independent promoter contacts preferentially engage active enhancers
• •Loss of cohesin-dependent promoter-enhancer links associates with transcriptional changes

Thiecke et al. use Capture Hi-C to detect the rewiring of promoter-anchored chromosomal contacts upon rapid degradation of the architectural proteins cohesin and CTCF. They find that a significant minority of promoter-enhancer contacts are maintained under these conditions, with implications for long-range transcriptional control.

Introduction

DNA regulatory elements such as enhancers play key roles in transcriptional control, particularly for developmental and stimulus-response genes (Long et al., 2016; Shlyueva et al., 2014). Many enhancers are located large distances (up to megabases) away from their target promoters and are brought into physical proximity with them through DNA looping interactions (Göndör and Ohlsson, 2018). 3D chromosomal architecture and its regulators are, therefore, thought to be important for transcriptional control.

The cohesin complex and regulatory proteins that control cohesin-DNA interactions are critical for shaping chromosomal architecture (Merkenschlager and Nora, 2016). According to the current understanding, cohesin is involved in the extrusion of DNA loops in interphase nuclei (Davidson et al., 2019; Fudenberg et al., 2016; Kim et al., 2019; Sanborn et al., 2015), additionally to its well-characterized role in holding sister chromatids together from DNA replication until mitosis (Uhlmann, 2016). Extruding DNA loops are confined by chromatin boundary elements; in particular, by the binding of zinc-finger protein CTCF to its two recognition motifs on the DNA in convergent orientation (Rao et al., 2014; Vietri Rudan et al., 2015). These loops underpin the formation of topologically associated domains (TADs) (Dixon et al., 2012; Nora et al., 2012; Sexton et al., 2012) and substructures such as insulated neighborhoods (INs) (Hnisz et al., 2016; Nuebler et al., 2018; Szabo et al., 2019). Depletion of cohesin leads to rapid dissolution of TADs, whereas inactivation of CTCF reduces the strength of TAD boundaries (Gassler et al., 2017; Nora et al., 2017; Rao et al., 2017; Schwarzer et al., 2017; Wutz et al., 2017). In contrast, the higher order levels of chromosomal organization, such as A/B compartments that broadly separate active and inactive chromatin, are not reduced upon cohesin and CTCF depletion (Gassler et al., 2017; Nora et al., 2017; Rao et al., 2017; Schwarzer et al., 2017; Seitan et al., 2013; Sofueva et al., 2013; Vian et al., 2018; Wutz et al., 2017; Zuin et al., 2014).

Chromosomal domains such as TADs and INs constrain promoter-enhancer interactions (Bonev and Cavalli, 2016; Sun et al., 2019), albeit incompletely (Freire-Pritchett et al., 2017; Javierre et al., 2016), and perturbations of their boundaries can lead to gene misregulation in development and in cancer (Lupiáñez et al., 2016; Robson et al., 2019). However, genes sharing these domains may have different regulatory wiring and expression patterns (Hnisz et al., 2016; Schoenfelder and Fraser, 2019). In addition, active enhancers might themselves function as TAD boundary elements (Barrington et al., 2019; Bonev et al., 2017), and contacts between superenhancers recover more rapidly upon reversal of cohesin depletion compared with other chromosomal loops (Rao et al., 2017). These observations suggest that factors acting in cis to DNA regulatory elements are involved in establishing their chromosomal interactions. The interplay between higher order domains and specific regulatory chromosomal interactions such as those between enhancers and their target promoters is not fully understood.

Consistent with a role for cohesin in shaping gene regulatory architecture, its depletion prevents adequate activation of inducible genes (Cuartero et al., 2018). Surprisingly however, steady-state gene expression levels are less affected upon architectural protein depletion (Busslinger et al., 2017; Haarhuis et al., 2017; Nora et al., 2017; Rao et al., 2017; Remeseiro et al., 2012; Schwarzer et al., 2017; Seitan et al., 2013; Sofueva et al., 2013; Tedeschi et al., 2013; Zuin et al., 2014). This suggests that gene expression may be maintained by mechanisms independent of these architectural proteins, but whether this involves continued input from enhancers remains unclear.

The effects of architectural protein depletion on 3D chromosomal architecture were typically analyzed using Hi-C. While this is a powerful method for global detection of chromosomal conformation (Lieberman-Aiden et al., 2009), the complexity of Hi-C sequencing libraries limits the coverage and resolution of data obtained using this technology, making the robust analysis of specific enhancer-promoter interactions challenging. Combining Hi-C with sequence capture (Capture Hi-C) makes it possible to mitigate this limitation by selectively enriching Hi-C libraries for interactions involving, at least on one end, regions of interest such as gene promoters (Mifsud et al., 2015; Sahlén et al., 2015; Schoenfelder et al., 2015a). The fact that this approach does not depend on proteins bound to either interaction partner makes it particularly suitable for studying interactions where these proteins are either unknown or ectopically depleted.

Here, we use Capture Hi-C to study the effects of architectural protein depletion on promoter interactions. We show that, while a majority of promoter interactions dissolve upon cohesin and CTCF depletion, large numbers of such contacts remain unaffected, and some are gained in these conditions. Interactions that are lost, gained, and maintained upon cohesin depletion have distinct properties with respect to localization within TADs, interaction distance, and the identity of associated proteins. We further demonstrate that changes in the levels of newly synthesized transcripts of specific genes upon cohesin depletion (as measured by SLAM sequencing [SLAM-seq]) associate with changes in the connectivity of their active enhancers. These results provide a mechanistic explanation for the limited but significant effects of cohesin and CTCF perturbations on gene expression and suggest the existence of alternative mechanisms supporting promoter-enhancer interactions.

Results

Extensive Rewiring of Promoter Interactions upon Rapid Depletion of Architectural Proteins

To study the effects of architectural protein depletion on promoter interactions, we took advantage of HeLa cells, in which all alleles of either cohesin subunit SCC1 or CTCF were tagged with a minimized auxin-inducible degron (AID) (Morawska and Ulrich, 2013) and an mEGFP reporter (mEGFP-SCC1-AID and mEGFP-CTCF-AID cells, respectively). Additionally, these cells stably express Oryza sativa Tir1 protein required for proteasome targeting of AID-tagged proteins (Nishimura et al., 2009). We previously showed that SCC1 and CTCF are rapidly degraded in these cell lines within 20 min of auxin treatment (Wutz et al., 2017).

We performed high-resolution Promoter Capture Hi-C (PCHi-C) in G1-synchronized, auxin-treated mEGFP-SCC1-AID and mEGFP-CTCF-AID cells and auxin-untreated controls. In addition, to compare the effects of depletion of these proteins with cell-cycle effects, we performed PCHi-C in intact HeLa cells synchronized in G1, G2, and mitosis. We also analyzed cells in which the cohesin release factor WAPL was depleted by RNAi (Tedeschi et al., 2013; Wutz et al., 2017). We did not profile Tir1-expressing cells without AID-tagged proteins or the effects of auxin treatment alone, since previous studies had shown that these factors did not significantly contribute to the observed cohesin and CTCF depletion phenotypes (Nora et al., 2017; Rao et al., 2017; Wutz et al., 2017).

Two biological replicates of each condition were sequenced, aligned, and filtered using the HiCUP pipeline (Wingett et al., 2015) to a median coverage of ~95 M valid read pairs each, with the overall coverage across conditions of ~1.5 billion valid read pairs. Given the ~17-fold enrichment for interactions involving ~22,000 baited promoter fragments compared with conventional Hi-C, the combined coverage of promoter interactions in our dataset is equivalent to that achievable with ~25 billion Hi-C read pairs. Using the CHiCAGO pipeline (Cairns et al., 2016), we detected, on average, ~100,000 significant interactions between promoters and promoter-interacting regions (PIRs) in unperturbed cells (see Figures 1A and 6A for examples). The numbers of significant interactions were, however, markedly lower in architectural-protein-depleted cells compared with controls, with a particularly strong effect for cohesin (118,074 SCC1 control and 61,702 SCC1 depleted, respectively; Figure S1).

Figure 1.

Promoter Interaction Rewiring upon Architectural Protein Depletion

(A) Chromosomal interactions detected for SLC16A6 promoter by PCHi-C. The top four tracks show PCHi-C interaction profiles with interaction arcs for the conditions: SCC1 control, SCC1 depleted, CTCF control, and CTCF depleted. The following five tracks show ChIP-seq pileups in interphase for the targets: SMC3, CTCF, H3K4me1, H3K27ac, and H3K4me3. The bottom two tracks show TAD intervals in cell-cycle stages G1 and G2 and coding sequences (GRCh37).

(B) Clustered PCHiC interaction scores from the following conditions: synchronized interphase cell-cycle stages G1 and G2, AID−auxin control samples for SCC1 and CTCF, AID+auxin depletions of SCC1 and CTCF, RNAi for the cohesin loading factor WAPL and its co-factors PDS5A and PDS5B, and finally synchronized cells in the mitotic prometaphase. A coherent partitioning with 13 K-means clusters was found, which reveals promoter interactions that are dependent on (clusters B, C, and F) as well as independent of (clusters A, E, and J) cohesin and CTCF. Interactions in cluster D are lost upon cohesin, but not CTCF, depletion.

Figure 6.

Perturbation of Enhancers within Cohesin-Dependent PIRs Affect Target Gene Expression

(A) Chromosomal interactions detected for NUAK1 promoter by PCHi-C. The top four tracks show PCHiC interaction profiles with interaction arcs for the following conditions: SCC1 control, SCC1 depleted, CTCF control, and CTCF depleted. The following five tracks show ChIP-seq pileups in interphase for the targets: SMC3, CTCF, H3K4me1, H3K27ac, and H3K4me3. The bottom two tracks show TAD intervals in cell-cycle stages G1 and G2 and coding sequences (GRCh37).

(B) Bar chart showing relative transcript abundance (by qRT-PCR) upon targeting dCAS9-KRAB to the promoter regions of SLC16A6 or NUAK1 (gray bars) and separately to the selected active enhancers (in a pooled setting). Error bars represent ±SD over three biological replicates.

Clustering of interactions based on CHiCAGO scores identified 13 coherent clusters denoted A to M (Figure 1B; see STAR Methods for details). Clusters B, C, and F were characterized by loss of the interaction signal upon cohesin and CTCF depletion, while interactions in cluster D were sensitive to degradation of cohesin but not CTCF. Notably, all of these four clusters also showed loss of promoter interaction signal in mitosis, in which the majority of these proteins are thought to be released from chromosomes. In addition, promoter interactions in these clusters appeared generally stronger in unmodified HeLa cells compared with AID-modified, but auxin-untreated, cells (“control” in Figure 1B). This is presumably due to the baseline residual activity of the degron system toward AID-tagged proteins without the auxin treatment reported previously (Wutz et al., 2017).

Promoter interactions in clusters A, E, and J were maintained upon cohesin and CTCF depletion (Figure 1B). Some of these interactions showed sensitivity to WAPL depletion and were lost in mitosis (cluster A), while others were generally retained in all analyzed conditions (clusters E and J). Finally, distinct subsets of promoter interactions were gained depending on the depleted protein (SCC1, cluster K; CTCF, cluster G).

Jointly, these results point to both the shared and specific effects of cohesin and CTCF depletion on promoter wiring. In addition, they suggest that large numbers of promoter interactions are retained in these conditions, and these interactions appear generally stable throughout interphase.

Guided by the exploratory observations from cluster analysis, we next sought to define subsets of promoter interactions that are lost, maintained, or gained upon cohesin and CTCF depletion formally and with high confidence. For this, we additionally took advantage of our recently developed differential calling pipeline for PCHi-C data, Chicdiff (Cairns et al., 2019). Integrating Chicdiff and clustering results, we defined 36,174 lost, 12,978 maintained, and 2,484 gained interactions upon cohesin depletion (Figures 2A–2C and S1C). The remaining promoter interactions were not assigned to any category at the predefined degree of confidence (see STAR Methods for details). Notably, in an appreciable minority of cases, different interactions of the same regions (either promoters or PIRs) showed different dynamics upon cohesin depletion. For example, 1,259/3,412 (36.9%) of baited promoters whose interactions were lost upon cohesin depletion additionally had interactions that were either maintained or gained in cohesin-depleted cells (Figure 2D). The same was true for 840/9,324 (9%) of PIRs (Figure 2E). Performing the same analysis for CTCF revealed considerably smaller numbers of significantly affected interactions, both overall (17,645 lost, 13,703 maintained, and 1,663 gained) and on average per promoter (Figure S1C).

Figure 2.

Lost, Maintained and Gained Promoter Interactions upon Cohesin Depletion

(A) Overview of the numbers of lost, maintained, and gained promoter interactions.

(B) Median number of lost, maintained, and gained promoter interactions per baited promoter. Error bars indicate the interquartile range.

(C) Illustrative examples of promoters with lost, maintained, and gained interactions upon cohesin depletion. The blue bar indicates the baited restriction fragment containing a promoter. Red circles indicate significant promoter interactions. Red rectangles indicate significantly differential interactions between conditions (SCC1 control and SCC1 depleted). The boundaries of the most proximal TAD are indicated with dotted gray lines, with a shaded area that represents ± 1 SE over four interphase TAD calls.

(D and E) Rewiring of interactions per baited promoter (D) and PIR (E).

See also Figure S1.

The less pronounced effects of CTCF compared with cohesin depletion on promoter interactions are consistent with the hypothesis that cohesin is the primary factor in facilitating long-range interactions, while CTCF is often, but not always, required for cohesin positioning on the chromatin (Busslinger et al., 2017; Hansen et al., 2017; Nora et al., 2017; Schmidt et al., 2010; Wendt et al., 2008; Wendt and Peters, 2009; Wutz et al., 2017). It cannot be ruled out, however, that the observed differences could also be due, at least in part, to the residual levels of CTCF previously detected upon auxin treatment in this system (Wutz et al., 2017). We, therefore, focused on cohesin depletion for the remainder of the study.

TAD Boundaries Anchor Cohesin-Dependent and Constrain Cohesin-Independent Promoter Interactions

Cohesin inactivation rapidly leads to a near-complete loss of TADs (Gassler et al., 2017; Rao et al., 2017; Schwarzer et al., 2017; Wutz et al., 2017). Therefore, we asked how promoter interactions that were lost, maintained, or gained upon cohesin depletion localized with respect to TAD boundaries. To address this, we focused on promoter interactions from either of these classes, for which at least one partner mapped within a TAD (representing 62.1% total). We partitioned these interactions into the following three categories based on the location of one interaction partner (i.e., either a baited promoter or PIR) relative to its nearest TAD boundary: boundary-proximal (mapping within 0%–20% of the TAD length on either side), intermediate (20%–40% TAD length), and mid-TAD (40%–60% TAD length). For each class (lost, maintained, and gained) and category (proximal, intermediate, and mid-TAD), we then obtained the distribution of the locations of the second interaction partner relative to the boundaries of the respective TAD (Figure 3A, summarized in Figure 3B). We performed this analysis for interaction categories defined on the basis of either promoter or PIR locations, with similar results (Figure 3).

Figure 3.

Localization of Promoter Interactions that Are Lost, Maintained, and Gained upon Cohesin Depletion with respect to TAD Boundaries

(A) Promoter interaction frequency profiles. Promoter interactions that are lost, maintained, or gained upon cohesin subunit SCC1 depletion are indicated in the top, middle, and bottom rows, respectively. Vertical gray bars represent the viewpoint window that encompasses the TAD boundary-proximal (peripheral, left column), intermediate (middle column), and central (middle column) positions. Horizontal blue bars represent TAD intervals. Black lines indicate distributions of locations of PIRs whose baited promoters map within the viewpoint window. Blue lines indicate distributions of locations of baited promoters whose PIRs map within the viewpoint window. The shaded area around the profile lines represents ±SEM over four sets of interphase TAD calls.

(B) Bar charts showing the enrichment of baited promoters and PIRs at specific windows within TADs. Error bars represent ±SEM over four sets of interphase TAD calls.

See also Figure S2.

As can be seen in Figure 3 (top row), promoter interactions that were lost upon cohesin depletion tended to have at least one partner located in the vicinity of a TAD boundary, with the other interaction partner confined to the same TAD. Notably, for both the boundary-proximal and intermediate lost interactions, the second interaction partner most commonly localized near the “opposite,” distal boundary as opposed to the nearest one (Figure 3A, top row, left and middle plots). In contrast, promoter interactions maintained upon cohesin depletion, while also constrained by TAD boundaries, showed no preference for the location of either partner within the same TAD (Figure 3, middle row). Notably, unlike the lost promoter interactions, maintained interactions spanned relatively small proportions of the TADs’ lengths, regardless of their location relative to TAD boundaries (Figure 3A, middle row). Finally, promoter interactions that were gained upon cohesin depletion were enriched around TAD boundaries and spanned relatively small proportions of the TADs’ lengths (Figure 3, bottom row). The gained promoter interactions tended to cross the native TAD boundaries, suggesting that they were enabled by the dissolution of TADs upon cohesin depletion. We confirmed the TAD boundary constraint of lost and maintained interactions and the prevalence of TAD boundary-crossing gained interactions formally using logistic regression, accounting for interaction distance (Figure S2). Jointly, these results suggest that cohesin-dependent (“lost”) promoter interactions are anchored to TAD boundaries, while cohesin-independent (“maintained”) promoter interactions are constrained by TADs, and those gained upon cohesin depletion are enabled by TAD dissolution.

Cohesin-Dependent and -Independent Promoter Interactions Have Distinct Properties

We asked what features of promoter interactions were associated with their dependence on or independence of cohesin. First, we found a striking association between longer interaction distances and cohesin dependence (Kruskal-Wallis test, p < 2.2 × 10⁻¹⁶; Figure 4A), which was observed even for the same PIRs involved in both lost and maintained interactions (Wilcoxon test, p = 2.2 × 10⁻¹⁶; Figure 4B). We then assessed the binding of cohesin and CTCF to the interacting fragments (baits and PIRs), using previously published chromatin immunoprecipitation (ChIP) datasets for these proteins (ENCODE Project Consortium, 2012; Wutz et al., 2017). Both baits and PIRs of lost interactions were selectively enriched for the binding of cohesin and CTCF compared with maintained and gained interactions (log odds ratios [LORs] = 1.052 and 1.360 at baits and LORs = 0.526 and 0.766 at PIRs, respectively), demonstrating that architectural proteins likely mediate these interactions via their direct binding to the interacting regions in cis (Figure 4C). We also assessed the presence of the active chromatin marks H3K4me1 (Kuznetsova et al., 2015; Nilson et al., 2017) and H3K4me3 (ENCODE Project Consortium, 2012; Liang et al., 2015) in the same way. Strikingly, we found that these marks were selectively enriched at the PIRs of maintained and gained interactions compared with lost interactions (Figure 4C). Using multinomial logistic regression, we confirmed that the observed enrichment of cohesin-independent interactions for active PIRs is unlikely to be a passenger effect of the underlying compartment signal (Figure S3).

Figure 4.

Chromatin Features of Cohesin-Dependent and -Independent Promoter Interactions

(A) Violin plots showing the log₁₀ genomic distance of lost, maintained, and gained promoter interactions upon SCC1 depletion.

(B) Violin plots showing the log₁₀ genomic distance of lost (cohesin-dependent) and maintained (cohesin-independent) promoter interactions whose PIRs are engaged in the interactions of both types. ^∗p = 2.2 × 10⁻¹⁶, Wilcoxon test

(C) Heatmaps showing log-odds ratios of restriction fragments (baited promoters on the left and PIRs on the right) having strong ChIP-seq signals for active histone modifications and cohesin and CTCF binding. Signals marked with asterisks show significant differences across the lost, gained, and maintained interaction categories (false-discovery-rate [FDR]-corrected Fisher test p values < 0.05).

(D) Bar chart showing the regression coefficients (betas) of important predictors of maintained versus lost interactions identified by LASSO logistic regression.

See also Figures S3, S4, and S5.

Next, focusing on the cohesin-dependent (“lost”) and independent (“maintained”) interactions, we expanded the panel of PIR-bound factors to ChIP datasets for 51 transcription factors (TFs) and 11 histone modifications in HeLa cells available from the Cistrome and ENCODE projects (ENCODE Project Consortium, 2012; Zheng et al., 2019). We processed these data uniformly and devised a strategy to assess the signals at PIRs for each dataset in a consistent manner (see STAR Methods for details). To delineate factors having the strongest association with cohesin dependence, we used LASSO logistic regression, a variable selection approach in which explanatory variables with less important contributions to the outcome are removed from the model by shrinking their regression coefficients toward zero. The binary outcome variable reflected the maintenance versus loss of promoter interactions upon cohesin depletion. As explanatory variables, we used the full set of TFs and histone marks at PIRs. Additionally, we included interaction distance, interaction strength (expressed as CHiCAGO score), and whether an interaction connects two baited promoter regions.

The regression coefficients of the factors showing the strongest effects over a range of values of shrinkage parameter λ (from the most to the least restrictive) are shown in Figure S4, and the barplot in Figure 4D summarizes the effects of factors having significant associations with the maintained versus lost interactions at an optimized level of shrinkage (see STAR Methods for details). Interaction distance showed the strongest association with interactions lost upon cohesin depletion, followed by the binding of cohesin and CTCF (Figure 4D). In turn, the PIRs of cohesin-independent (“maintained”) interactions showed a strong positive association with promoter-promoter contacts and stronger interaction strength, as well as with features of active chromatin, including the histone marks H3K4me1, H3K4me3, H3K27ac, and H3K9ac. These PIRs also preferentially recruited proteins associated with transcriptional activation, such as histone acetyltransferase p300, H3K9 demethylase PHF8, BET family protein BRD4, and LIN9 transcriptional regulator (Figure 4C). We also found an association with the H3K27me3 mark linked with Polycomb repressive complexes (Figure 4C) that have known effects on chromosomal architecture (Cheutin and Cavalli, 2019).

We next asked whether cohesin-independent interactions were stabilized by transcription. To address this, we treated cells with the RNA polymerase II (Pol II) inhibitor triptolide and compared their promoter interaction profiles with untreated cells. Overall, triptolide treatment had only a mild effect on promoter interaction profiles that was somewhat more pronounced for interactions also affected by cohesin depletion compared with cohesin-independent interactions (Figure S5). This result indicates that cohesin-independent interactions are likely maintained by mechanisms that do not directly depend on intact RNA Pol II or transcription.

Taken together, these analyses demonstrate that cohesin-dependent interactions are typically longer range and associate with the binding of cohesin and CTCF in cis, while cohesin-independent interactions are shorter range and associate with features of active promoters and enhancers.

Cohesin-Dependent Promoter Interactions Participate in Transcriptional Control

To address the immediate effects of rapid depletion of cohesin on gene expression, we generated the profiles of newly synthesized mRNA using SLAM-seq (Herzog et al., 2017; Muhar et al., 2018) in SCC1-AID cells treated with auxin for 60 min and in untreated controls (see STAR Methods for details). Since the median half-life of mRNA in highly proliferating mammalian cells exceeds 4 h (Herzog et al., 2017), using SLAM-seq as opposed to the conventional RNA sequencing (RNA-seq) made it possible to mitigate the confounding effects of RNA stability on detecting transcriptional changes in this system.

The majority of genes were relatively unperturbed upon cohesin depletion, consistent with previous findings in other cell types in steady-state conditions (Busslinger et al., 2017; Haarhuis et al., 2017; Nora et al., 2017; Rao et al., 2017; Schwarzer et al., 2017; Seitan et al., 2013; Sofueva et al., 2013; Tedeschi et al., 2013). However, we detected 421 and 266 genes whose levels of newly synthesized mRNA were significantly upregulated and downregulated in cohesin-depleted cells, respectively (adjusted p value < 0.05; absolute log₂ fold change > 0.1). We also defined a set of 1,197 “constant” genes showing strong expression levels (top 25% RNA-seq signal) and no significant changes in the levels of newly synthesized transcripts upon cohesin depletion (see STAR Methods for details; Figure 5A). Gene Ontology (GO) term enrichment analysis revealed that downregulated genes show involvement in translational processes, mRNA-to-endoplasmic reticulum (ER) targeting, and negative regulation of gene expression. The upregulated genes show involvement in cell-cycle processes and negative regulation of gene expression (Table S1).

Figure 5.

Transcriptional Response upon Cohesin Depletion Associates with Rewiring of Promoter Interactions with Active PIRs

(A) Volcano plot of SLAM-seq-detected changes in the levels of newly synthesized transcripts upon SCC1 depletion. In total, transcripts were detected for 7,424 genes, of which 266 were categorized as downregulated and 421 as upregulated; additionally, a subset of 1,197 highly transcribed constant genes was identified and used in further analyses (see STAR Methods for details).

(B) Effect plot of ordinal logistic regression analysis probing an association of transcriptional response to SCC1 depletion with rewiring of promoter interactions with active enhancers for the respective genes. The response variable describes gene category with respect to the transcriptional response to SCC1 depletion (see A). The predictor variable describes the difference between the number of gained and lost promoter interactions per bait as a proportion of all interactions per bait (see STAR Methods for details).

(C) Same as (B), but using the rewiring of PIRs not containing active enhancer annotations as the response variable.

(D) Violin plots showing a log₂ fold change of the SLAM-seq signal upon SCC1 depletion for a small subset of genes whose promoters contact a shared active PIR in unperturbed cells but either lose or maintain interactions with it upon SCC1 depletion. ^∗p = 0.048, one-sided t test.

See also Figure S6.

We asked whether the observed changes in transcription upon cohesin depletion were consistent with the dynamics of enhancer-promoter interactions in these conditions. We devised an ordinal logistic regression model, using the transcriptional response of a gene (“upregulated,” “constant,” and “downregulated”) as the outcome variable and the relative change in the numbers of connected active enhancers as the explanatory variable (see STAR Methods for details). The direction of transcriptional change significantly associated with relative changes in the numbers of connected active enhancers (p = 0.02, effect size = 0.74; Figures 5B and S6). Notably, the same was not true when PIRs without active enhancer annotations were used instead of active PIRs in the same regression framework (p = 0.59; Figure 5C). Next, we focused on 15 active enhancers that engaged both in lost (n = 15) and maintained (n = 19) promoter interactions with different genes upon cohesin depletion. Transcription of genes that lost interactions with these enhancers was downregulated in cohesin-depleted cells compared with the genes that maintained interactions with the same enhancers (one-sided t test, p = 0.048, Figure 5D). Jointly, these results suggest that rewiring of connections between active promoters and enhancers contributes toward changes in gene expression upon cohesin depletion.

Finally, to validate the causal effect of enhancers with cohesin-dependent promoter connections on target gene expression, we selected two genes, NUAK1 and SLC16A6, that show abundant loss of enhancer-promoter interactions upon cohesin depletion (Figures 6A and 1A, respectively). Previous studies have established that recruitment of the KRAB-repressor domain using dCas9 can inhibit short- and long-range enhancer function (Gilbert et al., 2013; Thakore et al., 2015). Inhibiting the enhancers of these genes located at cohesin-dependent PIRs by targeting dCas9-KRAB to them (in a pooled setting) resulted in a significant downregulation of both of these genes (Figure 6B). These observations support the model that a subset of cohesin-dependent interactions connect promoters to functionally relevant enhancers.

Discussion

In this study, we have combined rapid depletion of cohesin and CTCF with high-resolution detection of promoter interactions by Capture Hi-C and the sequencing of recently synthesized transcripts by SLAM-seq to study the immediate roles of these architectural proteins in gene control. We report extensive rewiring of promoter contacts upon perturbation of these architectural proteins, which associate with changes in target gene transcription. The loss of large numbers of promoter-enhancer interactions upon cohesin depletion observed in our study is consistent with the superenhancer rewiring observed with Hi-C in HCT-116 cells (Rao et al., 2017), as well as with recent observations from a PCHi-C experiment in that cell line (El Khattabi et al., 2019).

We find that cohesin-dependent promoter interactions commonly invovle (at least on one end) TAD boundary-proximal regions and are characterized by an extended interaction distance and the binding of architectural proteins in cis to the interacting regions. These observations confirm and extend previous reports of cohesin and CTCF binding to promoters and enhancers (Dorsett and Merkenschlager, 2013; Yan et al., 2018). To what extent cohesin-dependent promoter interactions are a cause or effect of TAD organization is a complex question, since many boundaries contain and potentially arise from active regulatory elements (Barrington et al., 2019; Bonev et al., 2017). Furthermore, the role of architectural proteins in maintaining enhancer-promoter interactions in cis is not limited to TAD boundaries (Ren et al., 2017). However, the relative depletion of active PIRs among cohesin-dependent interactions compared with other interaction categories suggests that these interactions are not restricted to gene-regulatory contacts, consistent with the structural roles of architectural proteins and TADs in processes such as genome replication and maintenance (Marnef and Legube, 2017; Pope et al., 2014; Racko et al., 2018; Yu et al., 2016).

Importantly, we also observe that a significant number of promoter interactions are maintained upon architectural protein depletion. These findings are consistent with observations that cohesin depletion abrogates TADs, but not all chromosomal interactions (El Khattabi et al., 2019; Gassler et al., 2017; Rao et al., 2017; Schwarzer et al., 2017; Wutz et al., 2017), and are supported by recent microscopic evidence of cohesin-independent fine-scale chromosomal topology (Bintu et al., 2018; Luppino et al., 2019). Notably, the maintained promoter interactions are generally constrained by TAD boundaries, reinforcing the established role of these elements in delimiting chromosomal contacts (Robson et al., 2019).

We additionally detect a smaller subset of generally short-range promoter interactions that emerge de novo upon architectural protein depletion. These gained interactions often cross native TAD boundaries, consistent with the promiscuous rewiring of promoter-enhancer contacts observed upon genetic perturbations of TADs (Lupiáñez et al., 2015; Narendra et al., 2016). Collectively, these results indicate that architectural proteins have two important and separable functions in gene regulation: first, they directly facilitate long-range interactions between promoters and regulatory elements; and second, they insulate promoters from potentially inappropriate enhancers via TAD organization.

We show that cohesin-independent interactions often connect active promoters with active promoters and enhancers. At the same time, these interactions do not depend on RNA Pol II activity and transcription, consistent with the conclusions of a recent study using a different RNA Pol II inhibitor (El Khattabi et al., 2019). We further demonstrate that the observed enrichment of cohesin-independent contacts for active enhancers cannot be explained by the underlying A/B compartment signal that is also independent of architectural protein activity (Nora et al., 2017; Nuebler et al., 2018; Rao et al., 2017; Wutz et al., 2017). Jointly, these results suggest that cohesin-independent enhancer-promoter contacts do not merely represent the passenger effects of compartment structure or transcription. It cannot be ruled out, however, that active enhancer-promoter contacts contribute to shaping up compartment organization and its strengthening upon architectural protein depletion (Nora et al., 2017; Nuebler et al., 2018; Rao et al., 2017; Wutz et al., 2017) alongside the hypothesized role of heterochromatic contacts in this process (Falk et al., 2019; Nuebler et al., 2018).

Mechanisms that support promoter interactions independently of cohesin remain largely unknown, and the broad range of chromatin co-factors that we detect as enriched at these interactions points to the potential diversity of these mechanisms. Such co-factors include p300 acetyltransferase (Heintzman et al., 2007; Visel et al., 2009) and BET family protein BRD4 (Kanno et al., 2014; Lovén et al., 2013; Zhang et al., 2012), commonly found at active enhancers. BRD4 was previously shown to play a role in chromatin insulation, while depletion of BRD4 led to widespread chromatin decompaction (Floyd et al., 2013; Wang et al., 2012). A role in chromatin insulation was also previously reported for another BET family protein, BRD2, albeit notably in a CTCF-dependent manner (Hsu et al., 2017). We also find an enrichment for the H3K27me3 histone mark associated with Polycomb repressive complexes at cohesin-independent PIRs, consistent with the findings of a recent study in pluripotent stem cells (Rhodes et al., 2020). The effects of Polycomb in our system, however, appear to be less pronounced than in pluripotent cells, in line with the particular importance of Polycomb in shaping up genome architecture and transcriptional patterns in early development (Denholtz et al., 2013; Mas and Di Croce, 2016; Schoenfelder et al., 2015b). Other factors that appear associated with cohesin-independent interactions include chromatin regulatory proteins such as NuRD-related factor PWWP2A, DREAM complex subunit LIN-9, and histone demethylase PHF8, which are known to be involved in the transcriptional control of the cell cycle (Liu et al., 2010; Pünzeler et al., 2017; Sadasivam and DeCaprio, 2013). Both PWWP2A and LIN-9 associate with the H2A.Z histone variant (Latorre et al., 2015; Pünzeler et al., 2017), while interestingly, the DREAM complex was found to possess insulator activity in Drosophila (Bohla et al., 2014; Korenjak et al., 2014). Notably, the enrichment of chromatin co-factors at the PIRs of cohesin-independent interactions is unlikely to be fully explained by their general association with active enhancers, since these factors were retained as predictors by LASSO feature selection, alongside the classic enhancer-associated histone marks.

Although the binding of cohesin and CTCF is not restricted to enhancers, we confirm, using regression modeling and CRISPR interference (CRISPRi) perturbations, that these proteins mediate subsets of promoter-enhancer interactions that are relevant for gene expression, consistent with previous reports (Merkenschlager and Nora, 2016; Seitan et al., 2013; Xu et al., 2016). Our regression model is based on the relative change in the number of promoter-connected active enhancers and does not incorporate other important information such as the levels of enhancer activity and contact probability (Fulco et al., 2019), enhancer redundancy (Cannavò et al., 2016; Osterwalder et al., 2018; Spivakov, 2014), and the potential effects of initiator enhancers discussed below. Despite these limitations, however, this model is sufficient to demonstrate that the relatively mild effects of architectural protein depletion on steady-state transcription can be explained by the retention of many interactions between promoters and active enhancers and the gain of only a small number of promoter-enhancer contacts in these conditions.

How can the modest transcriptional changes upon cohesin depletion be reconciled with the requirement for cohesin for gene induction (Cuartero et al., 2018) and lineage-specific gene regulation (Liu et al., 2019)? One possibility is that, in non-transcriptionally permissive chromatin contexts, cohesin/CTCF-independent promoter contacts are unable to form in the absence of some prior activation events, such as chromatin decompaction of the locus or establishment of appropriate chromosomal conformation. It is possible that these events require cohesin-dependent enhancer interactions to act as “pioneers” (Darbellay and Duboule, 2016). This model is broadly consistent with the findings that the first enhancer in chains of enhancers is often distal and contains CTCF motifs (Song et al., 2019). It may also explain the functional distinction between the potential initiator and maintenance enhancers (Sen and Grosschedl, 2010; Song and Ovcharenko, 2018), such as when a weak distal enhancer, but not a strong proximal one, is required for developmental activation of sex-determining gene Sox9 (Gonen et al., 2018). Elucidating the interplay of cohesin-dependent and -independent mechanisms of long-range gene control is an important research direction for future studies, with broad implications for health and disease.

STAR★Methods

Key Resources Table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rabbit anti-RNA polymerase II phospho-Ser2	Abcam	Cat#ab5095; RRID:AB_304749
Rabbit anti-RNA polymerase II phospho-Ser5	Abcam	Cat#ab5131; RRID:AB_449369
Mouse anti-tubulin	Sigma-Aldrich	Cat#T-5168; RRID:AB_477579
Chemicals, Peptides, and Recombinant Proteins
Auxin (indole-3-acetic acid sodium salt)	Sigma-Aldrich	Cat#I5148
Lipofectamine RNAiMAX Transfection Reagent	ThermoFisher Scientific	Cat#LMRNA015
Fugene 6 Transfection Reagent	Promega	Cat#E2691
Triptolide	Sigma-Aldrich	Cat#T3652
4-Thiouridine	Carbosynth	Cat#NT06186
Critical Commercial Assays
SureSelectXT Custom 3-5.9Mb library	Agilent Technologies	Cat#5190-4831 (probe sequences available on request)
QuantSeq 3′ mRNA-Seq Library Prep Kit FWD for Illumina	Lexogen	Cat#015.24
Deposited Data
Raw Promoter Capture Hi-C and SLAM-seq data generated in this study	Deposited to GEO database	GEO:GSE145736
Processed data generated in this study	Deposited to OSF	https://osf.io/brzuc
Public ChIP-seq data	Cistrome project / SRA and ENCODE project	See Table S4
Experimental Models: Cell Lines
HeLa Kyoto cells	Peters lab, IMP	Cellosaurus:CVCL_1922
HeLa SCC1-mEGFP-AID	Wutz et al., 2017	N/A
HeLa CTCF-mEGFP-AID	Wutz et al., 2017	N/A
HeLa SCC1-mEGFP-AID / dCas9-KRAB	This study	N/A
293HT packaging cells	ThermoFisher Scientific	Cat#R70007
Lenti-X 293T packaging cells	Takara Bio	Cat#632180
Oligonucleotides
WAPL/PDS5A/PDS5B triple knockdown siRNAs	Wutz et al., 2017	See Table S2
gDNA probes for CRISPRi experiment	This study	See Table S3A
qPCR primers for gene expression quantitation following CRISPRi	This study	See Table S3B
Recombinant DNA
pHR-SFFV-KRAB-dCas9-P2A-mCher	Addgene	Cat#60954
pCMV-VSV-G	Addgene	Cat#8454
pCMVR8.74	Addgene	Cat#22036
pETN	Michlits et al., 2020	N/A
Software and Algorithms
HOMER	Heinz et al., 2010	http://homer.salk.edu/homer/
HiCUP	Wingett et al., 2015	https://www.bioinformatics.babraham.ac.uk/projects/hicup/
Chicago	Cairns et al., 2016	http://functionalgenecontrol.group/chicago
Chicdiff	Cairns et al., 2019	http://functionalgenecontrol.group/chicdiff
FastQC		https://www.bioinformatics.babraham.ac.uk/projects/fastqc/
FastQScreen	Wingett and Andrews, 2018	https://www.bioinformatics.babraham.ac.uk/projects/fastq_screen/
TrimGalore		https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/
Bowtie2	Langmead and Salzberg, 2012	https://github.com/BenLangmead/bowtie2
htseq-count	Anders et al., 2015	https://pypi.org/project/HTSeq/
normalizeQuantiles	Bengtsson et al., 2004	http://bioconductor.org/packages/release/bioc/html/aroma.light.html
Varistran (Anscombe transformation)	Harrison, 2017	https://github.com/MonashBioinformaticsPlatform/varistran
glmnet (LASSO logistic regression)	Friedman et al., 2010	https://cran.r-project.org/web/packages/glmnet/index.html
fixedLassoInf (significance of LASSO regression coefficients)	Lee et al., 2016	https://cran.r-project.org/web/packages/selectiveInference/
SlamDunk	Neumann et al., 2019	https://t-neumann.github.io/slamdunk/
LAGO	Boyle et al., 2004	https://go.princeton.edu/cgi-bin/LAGO

Resources Availability

Lead Contact

Further information and requests for resources should be directed to the Lead Contact, Mikhail Spivakov (mikhail.spivakov@lms.mrc.ac.uk), and will be fulfilled by the Lead Contact or study co-authors.

Materials Availability

All unique/stable reagents generated in this study are available on request with a completed Materials Transfer Agreement.

Data and Code Availability

All raw sequencing data for this project is made available online through Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/) under accession number GEO: GSE145736. Processed datasets and code supporting this study have been uploaded to Open Science Framework (https://osf.io/brzuc/).

Experimental Model and Subject Details

HeLa Kyoto cells (female) were used as a model system for all experimental analyses. Cells were cultured at 37°C in suspension in DMEM supplemented with 10% FCS, glutamate and penicillin-streptavidin.

Method Details

Cell cycle synchronization

HeLa cells were synchronized at the G1/S-phase transition by two consecutive cell cycle arrest phases using 2mM thymidine and released into fresh medium for 6h (G2-phase), or 15h (G1-phase). For mitotic cells, Nocodazole (100 ng/ml) was added 8h after release from double thymidine block, to arrest the cells in prometaphase. Post-mitotic cells were removed by shake off after five hours.

WAPL/PDS5A/PDS5B RNA interference

HeLa cells were transfected with siRNAs as described in van der Lelij et al., 2014 followed by addition of thymidine. The siRNA sequences directed against WAPL, PDS5A and PDS5B (Table S2) were obtained from Ambion. Transfection was performed by incubating duplex siRNA with RNAi-MAX reagent (100 nM) in growth medium lacking antibiotics. Cells were harvested after 72h of RNAi treatment.

Auxin induced degradation of SCC1 and CTCF

Auxin-inducible degron systems for the analysis were generated as described in Wutz et al. (2017). Briefly, C-terminal tagging of SCC1 and CTCF with an Auxin induced degron (AID) was performed using CRISPR/Cas9-mediated genome editing with a double-nicking approach (Ran et al., 2013). The C-termini were extended with monomeric EGFP (L221K) and the IAA1771- 114 (AID^∗) mini-degron from Arabidopsis thaliana (Morawska and Ulrich, 2013). Single clones were selected using flow cytometry. PCR was used to confirm that all alleles were successfully modified. Cells stably expressing Tir1 for auxin-inducible protein degradation were generated by transduction of SCC1-AID and CTCF-AID knock-in cells with the lentiviral vector SFFV-OsTir1-3xMyc-T2A-Puro (SOP) (Wutz et al., 2017) and puromycin selection (2 μg/ml, Invitrogen).

Promoter Capture Hi-C (PCHi-C)

PCHi-C was performed on SCC1-mEGFP-AID and CTCF-mEGFP-AID HeLa cells synchronized in G1 and treated with auxin for either 0 or 120 minutes, on WAPL/PDS5A/PDS5B RNAi knockdown HeLa cells, as well as on unmodified HeLa cells synchronized in G1, G2 or mitosis. HindIII-based Hi-C libraries from our previous study (Wutz et al., 2017) were captured using SureSelect target enrichment system (Agilent Technologies) according to manufacturer’s instructions, using custom-designed RNA bait library and paired-end blockers. Promoter-based ligated products were isolated and amplified as previously described (Schoenfelder et al., 2018) and paired-end sequenced using Illumina HiSeq 2500 at the Babraham Institute Sequencing Facility.

SLAM-seq

Transcriptional responses to SCC1-degradation were profiled by SLAM-seq as described previously (Muhar et al., 2018) with minor modifications. HeLa SCC1-AID cells were synchronized in G1 by double thymidine block as for PC-HiC. SCC1 degradation was induced by addition of 500 μM auxin (indole-3-acetic acid sodium salt, Sigma-Aldrich). Newly synthesized RNA was labeled by addition of 4-Thiouridine (4SU, Carbosynth) at a final concentration of 100 μM for 60 minutes. Cells were snap-frozen, and extracted total RNA was subjected to alkylation by iodoacetamide (Sigma-Aldrich) for 15 minutes at room temperature in an appropriate buffer. Alkylation was stopped by addition of dithiothreitol (DTT, GE-Healthcare) and alkylated RNA purified by ethanol precipitation. 3′ end mRNA sequencing libraries were generated from 500 ng of alkylated RNA using the QuantSeq 3′ mRNA-Seq Library Prep Kit FWD for Illumina (Lexogen). Single-read sequencing was performed by the Vienna Biocenter Core Facilities (VBCF) for 100 cycles on a HiSeq2500 sequencer (Illumina).

Triptolide treatment and western blot

Scc1-mEGFP-AID cells were treated with 300 nM Triptolide (Sigma, T3652) for 3 hours, and cells were collected for western blot or Hi-C. For western blot, cells were resuspended in RIPA buffer (50 mM Tris pH 7.5, 150 mM NaCl, 1 mM EDTA, 1% NP-40, 0.5% Na-deoxycholate and 0.1% SDS), which additionally contained pepstatin, leupeptin and chymostatin (10 μg/ml each) and PMSF (1 mM). The protein concentration was determined with the Bradford Protein Assay (Bio-Rad Laboratories). SDS-PAGE and standard western blot technique were applied to detect individual proteins with specific antibodies: rabbit anti-RNA polymerase II phospho-Ser2 (Abcam, ab5095), rabbit anti-RNA polymerase II phospho-Ser5 (Abcam, ab5131) and mouse anti-tubulin (Sigma, T-5168).

CRISPRi using dCas9-KRAB

HeLa SCC1-mEGFP-AID cells stably expressing dCas9-KRAB for CRISPRi-mediated repression of enhancer or promoter activity were generated by transduction with the lentiviral vector pHR-SFFV-KRAB-dCas9-P2A-mCherry (Gilbert et al., 2014) (Addgene plasmid #60954) and subsequent sorting for mCherry-positive cells using a FACS Aria III cell sorter (BD Lifesciences). Lentiviral packaging was performed in Lenti-X packaging cells (Takara Bio) transfected with lentiviral transfer plasmids and packaging plasmids helper plasmids pCMV-VSV-G (Addgene plasmid #8454) and pCMVR8.74 (Addgene plasmid #22036) using standard procedures. gRNA oligos (sequences in Table S3A) were cloned into pETN expression vector (Michlits et al., 2020) and transfected into packaging cells (293FT, Thermo Fisher, R70007) using Fugene 6 (Promega, E2691). Virus was then harvested and applied to Scc1-EGFP-AID/dCas9-KRAB cells. Infected cells were selected by G418 (GIBCO, 108321-42-2) and used for RT-qPCR.

RT-qPCR

RNA was prepared using TRIzol (Thermo Fisher, 15596026) according to the manufacturer’s instructions. The cDNA was generated by reverse transcription using Random hexamers (Thermo Fisher, N8080127) and SuperScript II Reverse Transcriptase (Thermo Fisher, 18064014) according to manufacturer’s instructions. Different transcripts were then compared by qPCR using GoTaq® qPCR Master Mix (Promega, A6001) and primers specific for selected genes and GAPDH (see Table S3B for primers). GAPDH levels were used for normalization of the selected genes and data were presented as fold change compared to the control gRNA.

Quantification and Statistical Analysis

Definition of topologically associating domains

Quality-controlled, aligned and filtered Hi-C datasets (Wutz et al., 2017) were processed with HOMER (Heinz et al., 2010). Hi-C reads were binned at a 40kb resolution and normalized using iterative correction (Imakaev et al., 2012). Directionality index (DI) scores (Dixon et al., 2012) were calculated using a 5kb step size, a 25kb window size and a 1Mb upstream and downstream window size, and were subsequently smoothed using 25kb kernel density smoothing. Topologically associating domain (TAD) boundaries were called as local extrema in DI transitions from negative to positive were detected. TADs were calculated by setting minIndex = 1 and minDelta = 2.

PCHi-C processing and analysis

Promoter interactions were called using the CHiCAGO pipeline (version 1.1.5) (Cairns et al., 2016). CHiCAGO models expected promoter interaction frequencies based on the Delaporte distribution, which has a negative binomial component for random Brownian collisions and a Poisson component for technical noise. CHiCAGO corrects for multiple testing by means of a p value weighting procedure based on the expected true positive rate at a given interaction distance estimated on the basis of consistency between biological replicates. CHiCAGO scores represent soft-thresholded -log weighted p values. Enrichment of multiple chromatin marks is typically maximized at promoter interacting regions (PIRs) at a CHiCAGO score of 5 (Cairns et al., 2016) and this score cutoff was used to call significant promoter interactions.

Clustering of promoter interactions

K-means clustering was used to partition promoter interaction scores across samples. The function kmeans from the base R package stats was called with the Hartigan-Wong clustering algorithm, a maximum of 1000 iterations and 25 initial random configurations. Prior to clustering, CHiCAGO scores were asinh-transformed and capped at median + 3 MAD. The number of clusters k was determined by iteratively performing k-means clustering with k-values ranging from 2 to 16 and choosing a partitioning with a low k as well as a low total within-cluster sum of squares value. Additionally, partitionings that resulted in multiple clusters with similar centroid values were avoided. Using these criteria, a partitioning with k = 13 was selected.

Differential PCHi-C interaction calling

Chicdiff (Cairns et al., 2019) was used to detect differential promoter interactions between SCC1-AID control (Aux-) and depleted (Aux+) samples, and between CTCF-AID control (Aux-) and depleted (Aux+) samples. Chicdiff uses DESeq2 (Love et al., 2014) differential interaction calling with custom feature selection, normalization and multiple testing correction procedures. Testing was performed on promoter interactions exceeding a Chicago score cutoff of 5 detected on merged-replicate data for either of the two conditions. Chicdiff v0.2 was used, using the normalization procedure available in the later versions as norm = “fullmean”. The difference in the mean asinh-transformed Chicago scores between conditions above 1 was used to prioritise the potential “driver” differential PIRs within the tested aggregated fragment bins, which is the default setting.

Defining promoter interaction rewiring categories upon SCC1 depletion

We constructed consensus sets of interactions that are lost, maintained and gained upon SCC1 depletion based on results from Chicdiff analysis, Chicago promoter interaction calling for each condition and K-means clustering. The criteria were as follows. “Lost”: Chicdiff adjusted weighted p value ≤ 0.01; log2 fold-change < 0; K-means clusters B, C, D or F; “gained”: Chicdiff adjusted weighted p value ≤ 0.01; log2 fold-change > 0; K-means cluster K; “maintained”: Chicdiff adjusted weighted p value > 0.01; Chicago score ≥ 5 in the merged-replicate samples for each condition; difference in the mean asinh-transformed Chicago scores between conditions not exceeding 1; K-means clusters A, E or J). The absolute majority (> 87%) of Chicdiff-detected differential interactions had consistent K-means cluster assignment (36,174/40,666 lost and 2,484/3,653 gained, respectively).

Analysis of the distribution of promoter interactions within TADs

To create the diagrams in Figure 3A, the linear genomic distance between a given restriction fragment (RF; either the baited promoter fragment or PIR) and the nearest TAD boundary was normalized to the length of the TAD:

$$R{F}_n^{\text{'}}=\frac{\left(B_1-R{F}_n\right)}{B_2-B_1},$$

where RF’_n is the relative location of the n^th restriction fragment with respect to the nearest TAD boundary; B₁ and B₂ are the start and end of the nearest TAD, respectively; and RF_n is the center coordinate of the n^th restriction fragment. RF’_n takes the value of 0 when the fragment is located at the TAD start boundary and the value of 1 when the RF is located at the TAD end boundary.

RF’_n was calculated for each RF, and the frequency distribution of RF_1-N was then calculated and visualized for the interval: RF’_n = [-2, 3]. Note that this interval corresponds to the x-axes of the panels in Figure 3A, showing the values [-2, −1, S, E, 1, 2]. The frequency distributions for multiple TAD partitionings were combined by calculating the mean and the standard error per bin.

RF’_n distributions were plotted separately depending on the relative location of the second interaction partner (i.e., either the baited promoter or PIR) within a pentile interval with respect to TAD length (TAD boundary-proximal: 0%–20% and 80%–100%; intermediate: 20%–40% and 60%–80%; mid-TAD: 40%–60%). The results for the boundary-proximal and intermediate windows were combined by taking the reflection along the TAD center of the frequency distributions.

TAD window enrichment analysis

RF enrichment within TAD windows (Figure 3B) was performed by dividing the observed proportion of RFs in a pentile window (peripheral, intermediate and central) by the expected proportion according to the uniform distribution. For the central window the expected proportion is 0.2 (one fifth), while for peripheral windows, as well as for intermediate windows, the expected proportion is 0.4 (since there are two of each of these windows). Enrichment is then defined as:

$$En{r}_w=lo{g}_{10}\left(\frac{O_w/O_t}{E_w}\right),$$

where Enr_w is the enrichment in window w (peripheral, intermediate or central), O_w is the observed number of interacting RFs in window w, O_t is the observed total number of interacting RFs in the TAD, and E_w is the expected proportion of RFs under the uniform distribution for window w (0.4, 0.4, or 0.2 respectively). The enrichment scores are calculated for TAD partitionings obtained on each of the four datasets (two replicates of G1- and G2-synchronized HeLa cells, respectively) and the mean ± 1SE are shown in Figure 3B.

Association between cohesin dependence and TAD boundary crossing of promoter interactions

For the analysis in Figure S2, logistic regression was performed at the level of individual promoter interactions, where the Boolean dependent variable was set to 1 if a promoter interaction crosses a TAD boundary in all four TAD partitionings (generated from two replicates of G1- and G2-synchronized HeLa cells, respectively), and to 0 otherwise. Promoter interactions crossing TAD boundaries defined in a subset of partitioning were removed from the analysis. The independent variables were the log₁₀(bp) promoter interaction distance and the rewiring category (lost, maintained or gained). A ‘sum to zero’ contrast matrix was used. The R function allEffects from the effects package (Fox, 2003) was used for visualization.

ChIP-seq data sources and processing

HeLa-specific ChIP-Seq data were obtained from two sources: the ENCODE project (ENCODE Project Consortium, 2012) and the Gene Expression Omnibus (GEO) (Edgar et al., 2002). ENCODE files were downloaded manually as BAM files, aligned against GRCh37. Dataset and publication IDs are listed in Table S4. Quality control on ENCODE files was performed using the “Read coverage audits” QC metrics provided on the ENCODE website (https://www.encodeproject.org/data-standards/audits/). Only entries with replicate data and with a read length of ≥ 36 were included and files with read-depth limitations or complexity limitations were excluded (i.e., the dataset must have ≥ 10 million usable fragments and at least “moderate library complexity” is required). Separately, unaligned data were downloaded from GEO as fastq files using the sra_fqdump command in Cluster Flow (Ewels et al., 2016). Read quality was tested using FastQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/), and FastQscreen (Wingett and Andrews, 2018) was used to test for contaminants and to ascertain that the sequencing material indeed originated from the human genome. Overall, 19 ChIP-Seq targets were excluded from further analysis. Datasets passing the QC were processed with Trim Galore (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/) to remove adapters. Reads were then aligned against GRCh37 using Bowtie2 (Langmead and Salzberg, 2012)

Definition of ChIP-seq signal enrichment per restriction fragment

Read counts per restriction fragment (RF) were obtained by using htseq-count (Anders et al., 2015), requiring reads to have a minimum mapping score of 10 and with the -m flag set to ‘union’. RFs that were shorter than 50bp or longer than 50kbp were excluded from further analysis. The Anscombe variance-stabilizing transformation was applied to RF-level read counts (Harrison, 2017), with the dispersion parameter set to 0.4. Following that, between-chromosome quantile normalization was applied (separately for each ChIP-Seq dataset) using the normalizeQuantile function from the aroma.light R package (Bengtsson et al., 2004). Linear regression was then used to define the relationship between the Anscombe-transformed, quantile-normalized RF-level read counts (response variable) and the RFs’ log₁₀-length (predictor variable). Replicates were merged by taking the mean residual score per RF. A cutoff on the studentized residuals from these regression models was then applied to define RFs with enriched signals in each dataset. The cutoff was taken to be 1.6449, corresponding to Student distribution p value = 0.05 at df = n-k-1, where n = 801363 (the number of RFs in the model) and k = 1 (the number of explanatory variables in each model).

Definition of PIRs containing active enhancers

To define PIRs containing active enhancers, we used the studentized residuals from the regression analysis described above to determine ChIP-seq signal enrichment. PIRs with top 5% studentized residuals for at least two out of the three marks H3K4me1, H3K4me3 and H3K27ac were considered as containing active enhancers.

Association of interaction rewiring with compartment signal and the presence of active enhancers at PIRs

For the analysis in Figure S3, we performed multinomial logistic regression using the multinom function from the nnet R package. The interaction rewiring category was used as the response variable (lost, maintained or gained; interactions that were not assigned to either category were not considered in the analysis). As explanatory variables, we used the presence of active enhancers at PIRs (a binary flag denoted ActivePIR and defined as described in “Definition of PIRs containing active enhancers” above) and the compartment signals at both baited promoter fragments and PIRs defined in 250kb bins on basis of Hi-C data in control HeLa cells in our previous study (Wutz et al., 2017). To rule out collinearity between ActivePIR and the compartment signal, we assessed variance inflation factors calculated on the basis of our multinomial model using the vif function from the car R package. ActivePIR did not show collinearity with the other factors (vif = 1.07), while bait and PIR compartment signals showed moderate collinearity (vif = 6.0 and 5.8, respectively) that was expected given that > 90% of promoter interactions spanned fewer than 500kb. To verify that the collinearity of compartment signal variables did not influence our conclusions, we reran the same model using only ActivePIR and the PIR compartment signal as explanatory variables (vif = 1.06 and 2.32, respectively), which confirmed a strong association of ActivePIR with maintained and gained interactions (not shown).

Features of PIRs associated with maintained versus lost interactions

The analysis was performed using LASSO logistic regression, with the rewiring category (lost versus maintained in the absence of cohesin) as the response variable. The predictor variables were the binary ChIP-seq signal enrichment scores for each analyzed dataset computed as above, asinh-transformed Chicago interaction scores, log10-transformed linear distance between interaction partners, and a flag indicating promoter-promoter interactions. The distance and interaction score parameters were scaled and centered to mean = 0 and SD = 1. Since a PIR can have multiple interactions, we performed regressions separately using the lowest, mean (Figure 4D) and highest score per PIR and the distance of the respective interaction, which produced comparable results (data not shown). PIRs associated with interactions from both categories (i.e., lost as well as maintained) were excluded from this analysis.

The glmnet function from the glmnet R package (Friedman et al., 2010) was used to perform LASSO logistic regression, with the alpha flag set to 1 (indicating LASSO regression) and the thresh parameter set to 10⁻¹². Cross-validation was performed using the function cv.glmnet. Significance was tested at a λ value that minimizes the number of parameters while remaining within 1SE of the cross-validated error (λ 1SE). To infer p values on the regression coefficients, the function fixedLassoInf from the R package selectiveInference (Lee et al., 2016) was used with parameters tol.beta = 10⁻³, and tol.kkt = 0.3.

Association between the rewiring of promoter-enhancer interactions and the transcriptional dynamics upon cohesin depletion

Ordinal logistic regression approach

Dynamics ofnewly synthesized mRNA levels based on SLAM-seq data was used as the outcome variable and the relative change in the number of promoter interactions upon cohesin depletion was used as the predictor variable. SLAM-seq data were processed using the SLAM-DUNK pipeline (Neumann et al., 2019).

The relative change in the number of interactions per baited promoter was computed as follows:

$$R=\frac{nGained-nLost}{nLost+nMaintained+nGained+1},$$

where nGained, nMaintained and nLost represent the numbers of gained, lost and maintained promoter interactions per baited fragment, respectively, and a pseudocount of 1 is added to avoid division by 0. This quantity was computed separately for promoter interactions with PIRs (R_active) containing active enhancers (defined as described above) and for promoter interactions with other PIRs (R_non-active) and used in the analyses shown in Figures 5B and 5C, respectively. To construct the outcome variable, genes were assigned into three categories (downregulated, constant, upregulated; Figure 5A), based on the log2 fold-change (LFC) in SLAM-seq read count and DESeq2 FDR-adjusted p value between the control and SCC1-depleted samples (downregulated: LFC ≤ −0.1 and p value ≤ 0.05; constant: |LFC| ≥ 0.1, p value > 0.05, RNA-seq RPM > 0.21 (75th percentile); upregulated: LFC ≥ 0.1, p value ≤ 0.05). Genes not meeting the criteria for either of the three categories were removed from the analysis.

Ordinal logistic regression was then performed using the polr function from the MASS package (Venables and Ripley, 2002). The proportional odds assumption was tested by performing a graphical parallel slopes test (data not shown).

Permutation-based approach

Contingency tables for association between transcriptional regulation and promoter interaction rewiring upon SCC1 depletion were constructed for lost, maintained, and gained promoter interactions separately. Promoter interaction rewiring and transcriptional response were then expressed as a log-odds ratio:

$$LOR=ln\left(\frac{\left(nLost\&downRegulated/n\neg Lost\&downRegulated\right)}{\left(nLost\&\neg downRegulated/n\neg Lost\&\neg downRegulated\right)}\right).$$

LOR is > 0 if genes with lost promoter interactions tend to be more strongly downregulated than genes with maintained or gained promoter interactions. Log-odds ratios were calculated for all combinations of rewiring category and regulatory category (i.e., lost & upregulated, lost & non-regulated, maintained & downregulated, etc). To construct regulatory categories (downregulated, constant and upregulated), a more stringent log₂-fold change cutoff was applied (downregulated: LFC ≤ 0.1 and p value ≤ 0.05; constant: |LFC| ≥ 0.1 and p value > 0.05, upregulated: LFC ≥ 0.1, p value ≤ 0.05). Note that in this analysis, there was no requirement for PIRs to bear active marks.

An empirical null distribution of log-odds ratios was then computed for each of these combinations, by constructing the contingency tables on randomly sampled baits (sampled from the set of baits that were categorised as described above) over 10,000 iterations. Subsequently, empirical p values were calculated by dividing the number of times the values in the expected distribution exceeded the observed value by the total number of values in the expected distribution. The p values were converted to z-scores (shown in Figure S6) using the following calculation:

$$z_{ij}=\frac{LO{R}_{ij}^{obs}-LO{R}_{ij}^{exp}}{S{D}_{ij}^{exp}},$$

where z_ij is the z-score for the i^th rewiring category (lost, maintained or gained) and the j^th regulatory category (downregulated, constant or upregulated), LOR^obs is the observed log₂ odds ratio, LOR^exp is the mean expected log₂ odds ratio and SD^exp is the standard deviation of the distribution of expected log odds ratios.

Gene Ontology enrichment analysis

The online tool LAGO (https://go.princeton.edu/cgi-bin/LAGO) (Boyle et al., 2004) was used to determine GO term enrichment for genes that showed a transcriptional response upon SLAM-seq analysis. Two analyses were performed by providing LAGO with a list of gene IDs of the upregulated genes upon SCC1 depletion and separately by providing a list of gene IDs of the downregulated genes upon SCC1 depletion. The results were merged and similar GO terms were grouped manually.

Published: July 21, 2020

Supplemental Information

Table S1. GO Term Enrichment Analysis of Genes Showing Differential Recent Transcription upon SCC1 Depletion, Related to STAR Methods

(A) Unmerged terms. (B) Merged terms.

Table S3. gDNA Probes and qPCR Primers Used in the CRISPRi Experiment, Related to STAR Methods

(A) gDNA probes. (B) qPCR primers.

Table S4. ChIP-Seq Datasets Used in the Analysis, Related to STAR Methods

(A) Datasets downloaded from Cistrome / SRA. (B) Datasets from the ENCODE project.