Main

Three-dimensional (3D) genome organization and chromatin dynamics are thought to be crucial for the spatiotemporal control of gene expression. However, little is known about the multiscale dynamics of enhancers and promoters and how this relates to genome organization. In particular, whether chromatin regulators modulate these dynamics remains unclear. To probe the dynamics and organization of regulatory elements at a single-cell level, two complementary methods can be used: live-cell imaging1,2,3,4,5 and single nucleus chromosome conformation capture experiments6,7,8,9,10,11,12 (such as Hi-C) which reveals snapshots of the structure of the dynamic 3D genome in different individual fixed cells6,7,8,9,10,11,12,13,14,15,16,17.

The NuRD complex is a highly conserved 1 MDa multisubunit protein complex that binds to all active enhancers18. NuRD combines two key enzymatic activities: nucleosome remodeling via its helicase-containing ATPase, predominantly CHD4 in mouse embryonic stem (mES) cells; and lysine deacetylation via its HDAC1/2 subunits17,18,19,20,21,22,23. These activities are thought to be present in two subcomplexes: HDAC1/2 associates, along with the histone chaperones RBBP4/7, with the core scaffold proteins MTA1/2/3 to form a stable subcomplex with deacetylase activity24 and the nucleosome remodeler CHD4 interacts with chromatin by itself and also forms a second subcomplex with GATAD2A/B and DOC1 (CDK2AP1)24,25,26. The methyl-CpG DNA binding domain proteins MBD2/3 interact directly with both the deacetylase subcomplex and GATAD2A/B24,27,28,29, and thus has a critical role in linking the CHD4 remodeler and HDAC subcomplexes together to assemble the intact holo-NuRD complex (Fig. 1a).

Fig. 1: Live-cell imaging to study NuRD complex binding kinetics and function.
figure 1

a, Schematic representation of the NuRD complex interacting with chromatin in the presence and absence of MBD3. b, Left, single JF549-HaloTagged molecules in the NuRD complex were tracked in 3D using a double-helix point spread function microscope; two puncta are recorded for each fluorophore with their midpoint providing the lateral x, y position and the angle between them representing the axial position in z relative to the nominal focal plane (see Extended Data Fig. 2 and Supplementary Videos 1–8 for examples of the raw data). Right, examples of extracted single particle trajectories from 20 ms exposure imaging of CHD4 show periods of unconfined and confined diffusion. c, The approach used for segmentation of the single-molecule tracks; the data shown are from the 20 ms exposures of CHD4-HaloTag-JF549. Step 1, Left, a single-molecule trajectory showing an example sliding window (blue). Right, four biophysical parameters are calculated for a sliding window that is moved through the trajectory: α, Dapp, Lc and the norm∥V∥ of the mean velocity, were all estimated from a sliding window of 11 consecutive images. Step 2, Left, several trajectories with example sliding windows (blue). Right, Histograms of the values of the four biophysical parameters extracted in Step 1 from all the sliding windows computed for all the recorded trajectories. Step 3, Left, Based on the values of the four biophysical parameters (producing a four-dimensional feature space) each point in each trajectory is classified as either confined (C) or unconfined (U) using a Gaussian mixture model (GMM). The histograms from Step 2 can then be separated into confined (blue) and unconfined (orange) populations. Step 4, the posterior probability P of the GMM (Step 3) is computed on the four parameters for each sliding window Xi where the index of the trajectory is represented by i = 1,…,N (Xi(k∆t) ∈ C with P(k∆t) > 1 − P(k∆t) (blue); otherwise Xi(k∆t) ∈ U (orange)). The result is a segmented trajectory where each timepoint is assigned as confined or unconfined (see Methods for more details).

Full size image

Assembly of the intact NuRD complex is critical for controlling cell fate transitions. Knockout of Mbd3, which disrupts intact NuRD complex assembly, leads to only moderate up- or downregulation (fine-tuning) of transcription levels, but this modulation prevents mES cell lineage commitment30,31,32,33. Nucleosome remodeling by NuRD regulates transcription factor and RNA polymerase II binding at active enhancers18, but whether this impacts enhancer dynamics or enhancer–promoter interactions has remained unclear. Here, to understand whether enhancer dynamics are regulated by this crucial chromatin remodeler, we combine live-cell single-molecule tracking with Hi-C experiments. Specifically, to explore NuRD function, we exploit the ability to unlink the chromatin remodeling and deacetylase subunits of the intact complex by deleting Mbd3.

Results

An algorithm to segment single-molecule trajectories

To understand how the NuRD complex alters chromatin structure and dynamics, we first set out to understand how its assembly influences chromatin binding. We carried out live-cell 3D single-molecule tracking of NuRD complex subunits in wild-type and Mbd3-knockout (mES cells31. This strategy, which exploits the fact that MBD2 (and thus the MBD2-linked holo-NuRD complex) is expressed at low levels in mES cells and cannot rescue the Mbd3 deletion34,35, allowed us to specifically perturb NuRD complex function. We generated knock-in mES cell lines expressing the endogenous Chd4, Mbd3 and Mta2 genes fused with C-terminal HaloTags, and confirmed that the tags did not prevent NuRD complex assembly (although subtle changes in subunit expression were observed) (Extended Data Fig. 1). We used a double-helix point spread function microscope25 to record 3D tracks of single NuRD-HaloTag-JF549 complexes as they moved through a 4-μm slice of the nucleus (Fig. 1b) at two distinct temporal regimes: 20 ms and 500 ms (Extended Data Fig. 2). Recording at a 20 ms time resolution allows the detection of both freely diffusing and chromatin-bound proteins25, and can thus be used to extract the chromatin binding kinetics of NuRD complexes. In contrast, at a 500 ms time resolution, ‘motion blurring’ substantially reduces the detection of freely diffusing molecules1, allowing us to focus on the slower subdiffusive chromatin-bound NuRD. Videos showing examples of a succession of images tracking both static and moving CHD4 molecules (recorded using either 20 or 500 ms exposures) can be found in Supplementary Videos 1–8.

To extract biophysical parameters, we developed a machine learning method (a Gaussian mixture model (GMM)) to segment the single-molecule tracks into different classes of subtrajectory (confined and unconfined) by studying their behavior over a sliding window of 11 consecutive images (Fig. 1c). From each subtrajectory, we estimated not just the apparent diffusion coefficient, Dapp (as previously used for classifying subtrajectories36,37) but also the anomalous exponent α, the localization length Lc, and the drift magnitude norm∥V∥ (ref. 38). The α value (mean squared displacement ∝ timeα), is particularly informative. Diffusing proteins are characterized by an α close to 1 whereas chromatin-bound (confined) proteins exhibit a lower α (refs. 3,38,39), which represents the condensation state38,40. The Lc of chromatin-bound proteins is also informative as it reflects the spatial scale that the molecules explore within the nucleus. Finally, by computing the magnitude of the drift vector Vi in three dimensions, we can characterize the total displacement of a molecule during the sliding window. Further details of the approach and of the simulations we carried out to test the algorithm can be found in the Supplementary Data and Methods.

Analysis of the 20 ms exposure tracks of single CHD4 molecules using our approach revealed a fast unconfined state that was freely diffusing with an α of 0.94 ± 0.12 and a Dapp of 1.3 ± 0.3 μm2 s–1 (matching previous observations24,25), as well as a confined chromatin-bound state characterized by subdiffusive motion with an α of 0.51 ± 0.02 and a Dapp of 0.43 ± 0.03 μm2 s–1 (Fig. 1c). Similar results were obtained when segmenting the trajectories of two other NuRD complex components, MBD3 and MTA2 (Extended Data Fig. 3c). To demonstrate that our approach can reliably determine differences in the α for diffusing and chromatin-bound molecules, we imaged freely diffusing HaloTag protein. We found that only a small proportion of HaloTag molecules bind to chromatin (as observed41,42), and that many molecules have an α of around 1, which is significantly higher than observed for fixed dye molecules (α of 0.62 ± 0.01) (Extended Data Fig. 3c). We conclude that we can use 20 ms trajectories to distinguish unconfined freely diffusing molecules from confined chromatin-bound proteins. We note, however, that the Dapp of chromatin-bound NuRD molecules can only just be distinguished from those that are stationary when imaging using 20 ms exposures: immobile dye molecules have a median localization error of 60 nm and a Dapp of 0.3 ± 0.2 μm2 s–1, which is quite similar to the 0.43 ± 0.03 μm2 s–1 determined for chromatin-bound molecules (Extended Data Fig. 3c). We also note that shorter exposures are needed for the detection of faster moving smaller molecules (for example, transcription factors).

HDAC subcomplex requires CHD4 for chromatin binding

Having developed an approach to segment the 20 ms exposure trajectories of the NuRD complex into chromatin-bound and freely diffusing molecules (Fig. 2a,b), we investigated how removal of MBD3, which disrupts the interaction between the HDAC- and CHD4-containing NuRD subcomplexes18,24, affects chromatin binding. To explore whether the two subcomplexes are preassembled before binding to chromatin, we imaged both the CHD4 remodeler and the HDAC-containing subcomplexes. Imaging the HDAC-containing subcomplex using tagged MTA2 revealed a 1.7-fold increase in Dapp for freely diffusing MTA2 in the absence of MBD3 (Fig. 2c and Extended Data Fig. 3c), demonstrating that the deacetylase subcomplex is normally associated with CHD4 in intact NuRD (Fig. 1a). Single-molecule tracking of CHD4, however, revealed only a 1.05-fold increase in the Dapp of freely diffusing CHD4 in the absence of MBD3 (Fig. 2c and Extended Data Fig. 3c). A larger increase might have been expected from the disassembly of the holo-NuRD complex but mES cells also contain CHD4 that is not present in NuRD, both on its own and in the ChAHP complex with ADNP and HP1β,γ43. As a control, we also imaged-tagged MBD3 and showed that both freely diffusing MBD3 and MTA2 molecules have similar diffusion coefficients, consistent with MBD3 linking the two subcomplexes together in intact NuRD18,24,25 (Fig. 1a). Finally, we showed that MBD3 does indeed interact with CHD4 via GATAD2A in vitro using purified GATAD2A in pulldown reconstitution experiments (Extended Data Fig. 3b), and through the knockdown of Gatad2a and Gatad2b, which slightly increased the diffusion coefficient of CHD4 (1.05-fold) (Extended Data Fig. 3d,e). We conclude that the CHD4 and HDAC subcomplexes in NuRD are normally preassembled before binding to chromatin.

Fig. 2: Live-cell single-molecule tracking reveals that the NuRD complex assembles before it binds to chromatin.
figure 2

a, Segmentation of an example 20 ms trajectory of CHD4 into chromatin-bound (C) (blue) and freely diffusing (F) states (orange). b, Percentage of molecules and distribution of Dapp for chromatin-bound and freely diffusing CHD4 molecules. c, Left, boxplot of Dapp for CHD4 and MTA2 molecules in the presence and absence of MBD3; *P = 0.009, **P = 1 × 10–25 (two-sided Kolmogorov–Smirnov test). Center line, median; box limits, upper and lower quartiles; whiskers, 95% confidence interval. Data for MBD3 are shown as a control, and the gray dotted line indicates the upper bound (at the 95% confidence interval) of the Dapp determined for stationary JF549 dye molecules. Right, percentage of CHD4 and MTA2 molecules in the presence and absence of MBD3 (from Gaussian fitting, *P = 1 × 10–6, **P <  1 × 10–27 (two-sided Fisher’s exact test)). The numbers of cells per trajectory used in the analysis were: 30/5,557 (CHD4), 25/2,337 (CHD4–MBD3), 10/336 (MTA2), 10/652 (MTA2–MBD3) and 30/2,224 (MBD3). d, Left, a plot of the confinement probability allows determination of the association TA and dissociation TD times (defined, respectively, as the time a trajectory spends between periods of confined or unconfined motion). Middle, a single exponential curve of rate lambda = 1/TA was then fitted to the distribution of association times. Right, the association times extracted for CHD4 and MTA2 were then compared with those in the absence of MBD3, and with those for the MBD3 control. The number of association times used in the analysis were: 1,436 (CHD4), 668 (CHD4–MBD3), 62 (MTA2), 361 (MTA2–MBD3) and 407 (MBD3). Data are presented as mean values. Error bars show 95% confidence intervals, *P =  1 × 10–47 for CHD4 versus CHD4–MBD3 and **P = 1 × 10–25 for MTA2 versus MTA2–MBD3 (two-way analysis of variance (ANOVA)). e, Schematic representation of a model in which MBD3-dependent assembly of the NuRD complex increases the association rate of the deacetylase subcomplex.

Full size image

We then examined how the NuRD complex interacts with chromatin by comparing the percentage of freely diffusing versus chromatin-bound CHD4 and MTA2 molecules in the presence and absence of MBD3. We observed a 1.1-fold decrease in the percentage of CHD4 molecules bound to chromatin in the absence of MBD3, but there was a much more significant (2.4-fold) decrease in the percentage of chromatin-bound MTA2 molecules upon MBD3 depletion (Fig. 2c). This suggested that CHD4, rather than the deacetylase subcomplex, is primarily responsible for the association of NuRD with chromatin. This finding was supported by in vitro experiments that showed that, in comparison with CHD4, the deacetylase subunit by itself does not bind strongly to nucleosomes (Extended Data Fig. 4a). We conclude that NuRD normally exists as an intact complex in mES cells and that the removal of MBD3 disrupts chromatin binding of the deacetylase complex but not the CHD4 remodeler.

To further investigate the chromatin binding kinetics of the CHD4 remodeler and the MTA2 deacetylase subcomplex in the presence and absence of MBD3, we next determined association times from the time spent freely diffusing between confined chromatin-bound states. The distribution of association times was well approximated by a single exponential, suggesting a Poissonian process. Consistent with our finding that CHD4 is primarily responsible for recruitment of NuRD to chromatin, we found no increase in the association time of CHD4 upon removal of MBD3 (Fig. 2d). (The decrease in the observed association time is consistent with faster diffusion of the smaller CHD4 subcomplex resulting in more frequent collisions with chromatin.) However, we did find a significant (1.3-fold) increase in the association time of MTA2 upon MBD3 depletion (Fig. 2d), consistent with CHD4 recruiting the deacetylase subcomplex to chromatin.

We also attempted to determine dissociation times from the time spent bound to chromatin between unconfined freely diffusing states. Although no changes in dissociation time were observed (Extended Data Fig. 4c), we reasoned that our trajectories would be truncated by photobleaching. We therefore took advantage of ‘motion blurring’ when recording 500 ms trajectories to detect only chromatin-bound proteins1,44, and combined this with time-lapse imaging using different intervals between exposures. To our surprise, this showed that the dissociation times were much longer than we had expected (greater than 100 s for MBD3; Extended Data Fig. 4e), such that it proved impossible to track individual molecules for long enough to determine reliable dissociation rates. We conclude that, once bound to a target site, intact NuRD binds for unexpectedly long times.

Intact NuRD modulates chromatin movement at enhancers

We next studied the dynamics of chromatin-bound NuRD by tracking these slower-moving molecules at a time resolution of 500 ms (Fig. 3a and Extended Data Fig. 5a). Analysis of trajectories that lasted more than 5 s using the GMM (Fig. 3b) revealed two states of chromatin-bound CHD4 (slow and fast moving) with a Dapp of 0.006 ± 0.002 and 0.018 ± 0.006 μm2 s–1 (Fig. 3c and Extended Data Fig. 5b). The slow-moving chromatin-bound NuRD molecules could still only just be distinguished from those that are stationary even when imaging using 500 ms exposures; we found that immobile dye molecules had a median localization error of 34 nm and a Dapp of 0.004 ± 0.003 μm2 s–1, which is again similar to the 0.006 ± 0.002 determined for slow-moving chromatin-bound CHD4 (Extended Data Fig. 5a).

Fig. 3: Assembly of the NuRD complex decondenses chromatin.
figure 3

a, Left, example trajectory of a chromatin-bound CHD4 molecule showing periods of both slow (dark blue) and fast (light blue) subdiffusive motion. Two fast substates (F1 and F2) are observed, with the F2 state showing movement in a defined direction. Right, four biophysical parameters calculated along this trajectory with the fast F2 subtrajectories showing a higher α, increased Lc and increased drift. b, Schematic illustrating the analysis of the 500 ms exposure trajectories of chromatin-bound NuRD complex subunits. c, Histograms of the values of the four biophysical parameters extracted from all the sliding windows computed for all the recorded trajectories to distinguish slow-moving/immobile (dark blue) and fast-moving (light blue) chromatin-bound molecules (Stage 1 of the analysis in b). Gaussian fitting to the distribution of α values (Stage 2 of the analysis in b) identified two values of α for molecules in the fast-moving chromatin state (light blue) (Extended Data Fig. 5). d, Comparison of biophysical parameters for the CHD4 remodeler in the presence and absence of MBD3, and for MBD3 itself. Left, α values resulting from Gaussian fitting (data presented as mean values, error bars show 95% confidence intervals, *P = 1 × 10–4, two-sided Kolmogorov–Smirnov test). Boxplots of (middle) the lengths of confinement and (right) Dapp values (*P = 1 × 10–4 and *P = 1 × 10–32, respectively, two-sided Kolmogorov–Smirnov test). In c and d, the gray dotted lines indicate the upper bounds of the different biophysical parameters (at the 95% confidence interval) determined for stationary JF549 dye molecules. The numbers of cells per trajectory used in the analysis were: 30/3,059 (CHD4), 15/2,111 (CHD4–MBD3) and 30/1,816 (MBD3). Center line, median; box limits, upper and lower quartiles; whiskers, 95% confidence interval. e, Left, percentage of molecules in the slow or fast chromatin-bound states (*P < 0.01, two-way ANOVA). Right, schematic representation of the three states of chromatin-bound NuRD.

Full size image

We then compared the dynamics of chromatin-bound CHD4 molecules in Mbd3-knockout and wild-type cells. Surprisingly, we found that α, Lc and Dapp of the fast-moving chromatin-bound CHD4 molecules were all higher in wild-type cells (Fig. 3d). The increase in α in wild-type cells unexpectedly suggests that, in the presence of NuRD, chromatin is less condensed, whereas the increased Dapp and Lc show that chromatin-bound CHD4 molecules diffuse more rapidly and explore a larger nuclear volume. We had expected to find that recruitment of the deacetylase by CHD4 would lead to less acetylated chromatin and greater condensation45,46,47,48,49 and that the chromatin-bound CHD4 molecules in wild-type cells would thus explore a smaller nuclear volume.

When we visualized trajectories of the fast-moving chromatin-bound CHD4, we observed a proportion of molecules exhibiting periods of motion in a defined direction, characterized by a high α (>1.0) and high drift (for example, the trajectory in Fig. 3a). This suggested that there may be two types of fast-moving chromatin-bound CHD4. Indeed, when looking at the shape of the distribution of α values extracted from sliding windows of these trajectories, we observed two different populations of fast-moving molecules (Fig. 3c) and used Gaussian fitting to characterize their distributions (Fig. 3b). The two stages of our analysis thus revealed a single slow state (S) with αs of 0.59 ± 0.01 (67% of subtrajectories) and two fast substates (F1 and F2) with different α values: αF1 of 0.60 ± 0.01 (26%) and αF2 of 0.89 ± 0.02 (7%) (Fig. 3c and Extended Data Fig. 5c,d). Molecules in the fast F1 state have the same distribution of α as those in the slow state and they therefore explore the same chromatin environment. However, they diffuse faster and have a larger Lc and thus move further within the nucleus (Fig. 3c). Molecules in the fast F2 state, however, have a higher α and they explore an even larger area of the nucleus (higher Lc) than those in both the slow and the fast F1 states (for example, the trajectory in Fig. 3a). Moreover, they have high drift, indicative of movement in a defined direction; this is also consistent with the higher α.

Having observed both condensed (low α) and decondensed (high α) motion for chromatin-bound CHD4, we carried out a similar analysis in Mbd3-knockout cells. Although chromatin is less condensed in the presence of intact NuRD (see above), we observed a significant decrease in the proportion of CHD4 molecules in the fast decondensed F2 state (7.4% in wild-type cells versus 18% in Mbd3-knockout cells; Fig. 3e and Extended Data Fig. 5c,d). As a control, we also compared the dynamics of chromatin-bound MBD3 with that of CHD4 and found that it too exhibited one slow and two fast states. Both chromatin-bound MBD3 and CHD4 molecules exhibited motion in the fast F1 and F2 states in around 22–26% and 7–8% of trajectories, respectively, confirming that these states are a property of the intact NuRD complex and not just of CHD4 (Fig. 3d,e). Importantly, visualization of individual trajectories identified molecules that switch between the three states: S, F1 and F2 (Fig. 3a and Extended Data Fig. 5e). Thus, they are unlikely to represent either CHD4 forming different complexes or NuRD complex molecules bound in different regions of the nucleus.

The fast F1 and F2 states of chromatin-bound NuRD could result from movement on DNA due to chromatin remodeling or, bearing in mind the long dissociation times we determined for CHD4 (see above), from movement of NuRD-bound enhancers. To distinguish between these possibilities, we targeted sites near active enhancers with dCas9-GFP, either by transfecting a previously studied CARGO vector expressing 36 different gRNAs targeting a Tbx3 enhancer3 or by transfecting a single gRNA that targets DNA repeats near the Nanog gene (Extended Data Fig. 6a,b). Targeting nearby the Nanog enhancer using our single gRNA was confirmed by colocalization of GFP-tagged dCas9 (detected by immunofluorescence) with DNA fluorescence in situ hybridization (FISH) probes (Extended Data Fig. 6c). We carried out these experiments in cells expressing an ER–MBD3–ER (estrogen receptor–MBD3–estrogen receptor) fusion protein in which the nuclear localization of MBD3 (and thus assembly of the intact NuRD complex) is tamoxifen-inducible18. In this system, the intact NuRD complex assembles and remodels chromatin/transcription factor binding following induction; after 24 h, the transcription factor landscape has been reset and transcriptional changes have occurred. This allowed us to study the chromatin environment created by intact NuRD assembly shortly after it had become established (allowing us to distinguish direct from downstream effects). We also imaged cells showing bright undivided foci to exclude data from cells in the S or G2 phases of the cell cycle, which exhibit blurred foci or doublets (Extended Data Fig. 7a). Because of the background fluorescence from freely diffusing dCas9-GFP we had to track the enhancer loci in a single two-dimensional (2D) plane. Although this meant that we could not directly compare the parameters obtained in the 2D (active enhancer) and 3D (NuRD single-molecule) tracking experiments, classification of the subtrajectories once again revealed a slow and a fast-moving chromatin state (Extended Data Fig. 7b), including two subpopulations of fast-moving chromatin (Extended Data Fig. 7c). The longer enhancer locus trajectories (cf. chromatin-bound CHD4) allowed better characterization of the proportions of the different slow and fast states (for example, see Fig. 4a and Supplementary Videos 9–12). As when tracking CHD4 single molecules, addition of MBD3 (and thus the assembly of intact NuRD) significantly increased the Dapp of both enhancers in the fast-diffusing F1 and F2 states (Fig. 4b). Moreover, in the presence of intact NuRD, we again observed a decreased proportion of subtrajectories in the fast decondensed F2 state for both the Tbx3 and Nanog enhancers (Fig. 4c and Extended Data Fig. 7d).

Fig. 4: Assembly of the intact NuRD complex modulates the movement of active enhancers.
figure 4

a, Left, example trajectory of the Nanog enhancer segmented to show periods of slow and fast subdiffusive motion (dark and light blue, respectively). Right, α, Lc, Dapp and norm∥V∥ of the locus extracted from the trajectory shown. b, Boxplots of Dapp calculated for 2D trajectories of loci near the Tbx3 and Nanog enhancers, in the presence and absence of MBD3 (*P = 0.045 and *P = 1 × 10–4 for Tbx3 and Nanog, respectively, two-sided Kolmogorov–Smirnov test). c, Gaussian fitting to the distribution of the α values identifies a single slow and two faster states for enhancer loci (Extended Data Fig. 7). The percentage of subtrajectories of enhancer loci exhibiting slow (S) and fast chromatin motions with either a low (F1) or high (F2) α is shown (*P < 0.01, two-way ANOVA). The numbers of trajectories used in the analysis were 237 (Tbx3 + MBD3) and 287 (Tbx3-MBD3); 546 (Nanog + MBD3) and 229 (Nanog-MBD3). Center line, median; box limits, upper and lower quartiles; whiskers, 95% confidence interval.

Full size image

Previous work has suggested that enhancer dynamics are related to transcription3 and we wondered whether the changes we observe (± intact NuRD) result from altered levels of gene expression. We therefore tracked chromatin-bound CHD4 molecules after adding DRB—a small molecule inhibitor of transcriptional elongation50. Premature termination by DRB led to some reduction in the proportion of bound CHD4 molecules exhibiting the fast F1 motion (from 26% to 19%), but there was no reduction in the proportion of molecules in the fast F2 state or change in the chromatin environment (that is, in α) in the presence of a block on transcriptional elongation (Extended Data Fig. 5c). Finally, we tracked MBD3 molecules while blocking HDAC1/2 deacetylase activity with FK228 (ref. 51). Once again, however, there was no significant change in the proportion of molecules in the fast decondensed F2 state (Extended Data Fig. 5c). We conclude that the changes in enhancer dynamics ± intact NuRD are not due to altered transcription elongation or deacetylation activity.

NuRD increases intermediate-range enhancer–promoter contacts

To understand whether the alteration of enhancer dynamics in NuRD affects genome architecture and enhancer–promoter interactions, we next carried out in-nucleus Hi-C experiments. We obtained high-quality contact maps for both wild-type and Mbd3-knockout ES cells after combining our wild-type data with previously published52 (and consistent) datasets (Extended Data Fig. 8a). As previously observed6,8,9,10,11,12,13,14,15,16,17,53, the Hi-C contact maps showed that the genome is segregated into: (1) A and B compartments (regions containing a higher or lower density of genes, respectively); (2) megabase-scale topologically associating domains (TADs), which have a higher frequency of intradomain chromatin interactions; and (3) loops mediated, for example, via CTCF/cohesin binding, where specific genomic regions contact each other more frequently.

Comparison of the Mbd3-knockout and wild-type Hi-C data showed that NuRD leads to an increase of ~30% in the probability of intermediate-range contacts on the scale of TADs (500 kb to 3 Mb) (Fig. 5a and Extended Data Fig. 8b). The genome-wide increase in mean contact length per region binned (Extended Data Fig. 8c) was most noticeable for regions containing NuRD-regulated genes that are within the A compartment in both Mbd3-knockout and wild-type cells (KO-A and WT-A, respectively) (Extended Data Fig. 8d). Indeed, 77% and 17% of NuRD-regulated genes54, respectively, were found in the A and B compartments, with only small proportions (~3%) moving from A to B or vice versa. In addition, NuRD downregulated genes are significantly enriched in the A compartment (P < 1 × 10–10; odds ratio, 1.16), and significantly depleted in the B compartment (P < 1 × 10–10; odds ratio, 0.43) (Extended Data Fig. 8e,f). Chromatin immunoprecipitation followed by high-throughput sequencing (ChIP–seq) and CUT&RUN experiments in the absence of fixation show that NuRD binds predominantly to enhancers with little enrichment at promoters18, suggesting that the MBD3 ChIP–seq signal observed at promoters results from association with NuRD-bound enhancers. We therefore further categorized promoters according to whether or not they bind MBD3, and found that putative ‘NuRD downregulated, NuRD enhancer contacting’ genes are also significantly enriched in the A compartment (P < 1 × 10–10; odds ratio, 1.27) and significantly depleted in the B compartment (P < 1 × 10–10; odds ratio, 0.41) (Extended Data Fig. 8e,f). Thus, analysis of the Hi-C and ChIP–seq data showed that NuRD-regulated genes are predominantly in the A compartment where they may be downregulated through contact with NuRD-bound enhancers.

Fig. 5: Assembly of the NuRD complex increases Mb-range genome and enhancer–promoter interactions.
figure 5

a, log–log plots of contact probability as a function of genomic sequence separation (averaged across the genome), derived from in-nucleus Hi-C experiments of Mbd3-knockout (KO) (red) and wild-type (WT) (black) ES cells, shows a significant (~30%) increase in intermediate-range (~1 Mb) contacts in WT cells (P = 1 × 10–18, two-sided Mann–Whitney U-test; see also Extended Data Fig. 8b). b, Part of the Hi-C contact maps for Chromosome 1; the density of contacts is indicated by the color intensity. TAD boundaries are weakened in wild-type cells, resulting in an increased density of contacts between adjacent TADs, both within and between A/B compartments (red and green arrows); see Extended Data Fig. 8g–i for genome-wide comparisons. c, Boxplots showing intrachromosomal enhancer–promoter link lengths, determined using a modified version of the activity-by-contact algorithm66, present in both KO and WT ES cells (orange), in only KO cells (red) or in only WT cells (green). The number of WT unique links = 7,941; common links = 8,546; KO unique links = 12,932; *P = 1 × 10–10, Bayesian version of t-test. Center line, median; box limits, upper and lower quartiles; whiskers, 95% confidence interval; bars on the left of each plot show all the data. d, Fold-enrichment (upper panel) and (lower panel) chi-squared test (lower) when intergenic enhancer–promoter interactions found in both WT and KO cells, as well as those found uniquely in either KO or WT, are correlated with genes that are up- or downregulated in the presence of intact NuRD. Enriched and depleted types of interaction are colored red and blue, respectively, and significant changes are highlighted using solid and dashed black boxes (see Extended Data Fig. 10c for an example of changes in enhancer–promoter contacts). Interactions where NuRD-bound enhancers can be seen to contact the promoter, which are either found in both KO and WT or uniquely in WT, are enriched at upregulated genes. In contrast, there is a depletion in interactions between intergenic enhancers and promoters of downregulated genes. e, Schematic interpretation of the results of the Hi-C experiments.

Full size image

Comparison of the Hi-C contact maps for Mbd3-knockout and wild-type cells showed that NuRD weakens the boundaries between A/B compartments and TADs (Fig. 5b and Extended Data Fig. 8g), with promoters suspected (see above) of contacting NuRD-bound enhancers having an increased number of cross-compartment contacts (Extended Data Fig. 8h). We also found that NuRD facilitates a significant genome-wide decrease in the insulation between TADs (Extended Data Fig. 8i). Although NuRD-regulated genes do not move between compartments (Extended Data Fig. 8j), these results suggest that the NuRD-mediated increase in enhancer dynamics may facilitate interactions across A/B compartment and TAD boundaries55,56.

The blurring of TAD and A/B compartment boundaries suggested that NuRD might alter CTCF/cohesin binding56,57. We therefore carried out CUT&RUN experiments to study chromatin binding of CTCF and cohesin, which have a key role in loop and TAD formation13,58,59,60,61,62,63. We found that NuRD leads to a redistribution of both CTCF and, more particularly, SMC3 (a subunit of the cohesin complex) (Extended Data Fig. 9a–c,e). Moreover, a significant proportion of NuRD-regulated genes are found near to CTCF/cohesin binding sites; this was most noticeable for genes that are upregulated and whose promoters (we suspect) contact NuRD-bound enhancers, but it was also true for such genes that are downregulated (Extended Data Fig. 9d). We also found that genes that are either up- or downregulated by NuRD tend to be colocated in the linear genome sequence, for example, the Slc6a12, Prmt8 and Htra1 genes (Extended Data Fig. 9e). This clustering of similarly regulated genes suggests that NuRD alters genome structure/dynamics and affects the expression of groups of nearby genes in a similar way.

Because changes in TADs and CTCF-cohesin loops are thought to influence enhancer–promoter proximity60,64,65, we used a modified version of the activity-by-contact algorithm66 to study changes in enhancer–promoter interactions in Mbd3-knockout versus wild-type cells (Extended Data Fig. 10). After defining active enhancers and promoters using H3K27ac and H3K4me3 ChIP–seq profiles18 (Methods), this revealed that enhancer–promoter interactions that occur only in the presence of the intact NuRD complex tend to link together genomic regions separated by longer distances than those that occur in its absence (Fig. 5c). In addition, when we considered enhancer–promoter interactions and changes in transcription54, we found more contacts between intergenic NuRD-bound enhancers and promoters that are upregulated by NuRD, and fewer contacts between intergenic enhancers and promoters that are downregulated by NuRD (Fig. 5d). We also carried out two-color enhancer–promoter DNA-FISH studies of three key pluripotency genes (Bmp4, Sox2 and Tbx3), and showed that the presence of NuRD led to a significant increase in the average distance between their enhancers and promoters (Extended Data Fig. 6d,e). This confirmed (using a different approach to the Hi-C experiments) that NuRD not only modulates chromatin decompaction and enhancer dynamics, but also alters enhancer–promoter interactions.

Discussion

In this work, we have developed a computational approach to analyze 3D trajectories of single molecules and segment them according to their diffusion behavior. Key advantages of our computational approach are that it allows us to (1) measure four different biophysical parameters to explore nuclear dynamics and (2) remove regions of the trajectories where the molecules are essentially not moving—allowing us to estimate biophysical parameters that define the behavior of moving molecules.

Using 20 ms exposure imaging, we were able to distinguish molecules that are freely diffusing from those that were chromatin bound. By comparing wild-type mES cells with a mutant cell line where we had knocked out Mbd3—a gene encoding a protein subunit that links the chromatin remodeling and histone deacetylase activities of NuRD together—we were able to show that the NuRD complex diffuses as an intact entity within the nucleus. We also showed that it associates mainly with chromatin through the CHD4 (chromatin remodeling) subcomplex that serves to recruit the histone deacetylase subcomplex to chromatin. In principle, we can use the approach to study chromatin association/disassociation kinetics. However, for NuRD, although we were able to study the association kinetics at short timescales, we rather unexpectedly found that it disassociates from chromatin too slowly for us to measure the disassociation kinetics (with residence times on the order of a minute or more). Nevertheless, we anticipate that this approach will allow studies of the binding kinetics of proteins that do not interact with chromatin so tightly (for example, many transcription factors), and will become increasingly useful as the field develops more photostable fluorophores.

Using 500 ms exposure imaging, we then studied the movement of chromatin-bound NuRD molecules, where motion blurring prevents the imaging of NuRD molecules that are freely diffusing. We were able to observe two states of chromatin-bound molecules—a slow and a fast state—where the fast state being further divided into two substates: a condensed fast (F1) and decondensed fast (F2) state. Comparison of the movement of chromatin-bound CHD4 molecules in wild-type cells versus the Mbd3-knockout surprisingly suggested that chromatin is less condensed in the presence of NuRD, with chromatin-bound molecules diffusing more rapidly and exploring a larger nuclear volume. Moreover, in wild-type cells, we found that fewer of the chromatin-bound NuRD molecules exhibit movement in a defined direction, suggesting that the intact NuRD complex spends less time in this fast decondensed state. We confirmed that these motions are a property of NuRD-bound chromatin (as opposed to, for example, the movement of NuRD molecules along DNA) by using a dead Cas9-GFP fusion protein to label the genome near specific enhancers and demonstrating that we could observe the same dynamics.

Finally, we used in-nucleus Hi-C and CUT&RUN experiments to ask whether altered enhancer dynamics, mediated by the intact NuRD complex, might affect genome architecture and enhancer–promoter interactions. We showed that the NuRD complex does indeed alter chromosome architecture by increasing the probability of intermediate-range contacts at the scale of TADs, leading to a blurring of the boundaries between A/B compartments and between TADs. We were also able to show that enhancer–promoter interactions that occur only in the presence of intact NuRD tend to link together genomic regions separated by longer distances. Moreover, we found that the NuRD complex leads to a marked redistribution of CTCF and, in particular, cohesin, with a significant proportion of newly formed CTCF/cohesin binding sites being found near to NuRD-regulated genes. We speculate that the NuRD complex promotes an environment with increased chromatin mixing where enhancers and promoters can contact each other over longer distances and where the resetting of enhancer–promoter interactions brought about by the fast decondensed F2 motions is reduced, leading to more stable, long-lived interactions (Fig. 6). This could provide an explanation for the observed increase in transcriptional noise, or low-level inappropriate transcription, observed in both human and mouse ES cells lacking functional NuRD67,68.

Fig. 6: Schematic model of NuRD function.
figure 6

Left, a molecular level view of how assembly of the NuRD complex affects chromatin remodeling18. Right, a chromatin level view of how assembly of the NuRD complex increases the volume explored by an enhancer while at the same time reducing its likelihood of entering the fast F2 state in which the movement of decondensed chromatin might reset enhancer–promoter proximity.

Full size image

Methods

In-nucleus chromosome conformation capture

Data acquisition

In-nucleus Hi-C17 was carried out on E14 wild-type and 7g9 Mbd3-knockout (see Kaji et al.31 for construction and characterization) ES cells; 50 bp paired-end sequencing was carried out on a HiSeq4000 instrument. Our own Hi-C analysis of Mbd3-knockout ES cells cultured in 2i consisted of three experiments: SLX-18035, SLX-7676 and SLX-19611 (which was sequenced twice). Alongside these four datasets, we also collected wild-type 2i data (SLX-7672) and compared this with recently published Hi-C data for ES cells cultured in 2i conditions (https://www.ebi.ac.uk/arrayexpress/experiments/E-MTAB-6591/) (Extended Data Fig. 8a). We used four replicates of this published data (ERR239137, ERR239139, ERR239141 and ERR239143; Supplementary Table 1).

Preprocessing

For each experimental condition and replicate, FASTQ files were first processed and aligned against the mouse GRCm38.p6 reference genome using the nuc_processing package (https://github.com/tjs23/nuc_processing). The number of unique contacts observed varied between replicates and conditions but, overall, a high number of reads were obtained in both conditions. The resulting raw contacts in NCC format were converted into hic and cooler format contact matrices using the juicer Pre (https://github.com/aidenlab/juicer/wiki/Pre) and HiCExplorer hicConvertFormat tools for downstream analysis (https://hicexplorer.readthedocs.io/en/latest/content/tools/hicConvertFormat.html)69,70,71. To assess the correlation between replicates within and between the two experimental conditions, we made use of the HiCExplorer hicPlotdistvscounts tool. Clustering of replicates based on contact distance distributions identified clusters for the two experimental conditions showing good agreement between replicates. We therefore merged the obtained contacts resulting in 705,063,441 and 327,180,884 contacts for the wild-type and Mbd3-knockout conditions, respectively. Owing to the difference in read coverage, we downsampled the merged wild-type dataset to have the same number of contacts as our merged Mbd3-knockout dataset. Finally, we performed Knight-Ruiz normalization of the merged datasets to ensure balanced matrices.

Compartment analysis

To identify A/B chromatin compartments within our merged datasets, we made use of the recently developed CscoreTool (https://github.com/scoutzxb/CscoreTool)72. This tool assigns each genomic window i a score ∈ [–1,1] with 1 assigning a probability \({1}\) that window i is in the A compartment while –1 assigns a probability \({1}\) that window i is in the B compartment. Importantly, unlike eigenvector analysis, c-scores of different samples can be compared directly since they represent probabilities. Binary compartments can also be assigned based on the sign of the c-score.

TAD analysis

TAD calling was performed using Lavaburst, a recently published tool57 that uses an insulation-style metric called the TAD separation score to identify the degree of separation between the up and downstream region at each Hi-C matrix bin. Local minima of the TAD separation score are considered as putative TAD boundaries and assigned q-values for calculation of a false discovery rate. For this analysis, we used parameters gamma = 10 and beta = 50,000. For both datasets, we identified TADs using a binning resolution of 25 kb.

Enhancer–promoter links

We wanted to investigate whether transcriptional misregulation in Mbd3-knockout ES cells was significantly associated with a disruption of cis-regulatory interactions. To identify putative promoter-regulatory element (RE) interactions, we made use of a recent study66 that profiled RE–promoter interactions via thousands of separate CRISPR deletions. In particular, Fulco et al.66 found that a relatively simple activity-by-contact (ABC) model could be used to identify, with reasonable precision and recall, functional RE–promoter interactions. Crucially, the ABC model produced interactions with higher accuracy than those identified using either linear distance along the genome or Hi-C contacts alone.

The ABC model considers the genome in 5 kb bins. For each promoter \({p}\) and regulatory element \(r\) we define:

$${{\rm{ABC}}}_{p,r}=\frac{{A}_{r}\times {C}_{p,r}}{\sum _{s\in N\left(p\right)}{A}_{s}\times {C}_{p,s}}$$

where N(p) is the set of all regulatory elements within 5 Mb of p, Cp,r is the contact frequency between p and r, and Ar is the activity of r. For regions of the genome with poor read coverage, Cp,r is estimated assuming a power-law decay of contact frequencies:

$${C}_{p,r}\propto{d}_{p,r}^{-\gamma }$$

where \({d}_{p,r}\) is the genomic distance between the promoter and regulatory element, and \(\gamma\) is inferred from the Hi-C contact maps. Fulco et al.66 define the activity of RE \(r\) as the geometric mean of read counts of DNase I hypersensitive (DHS) and H3K27ac ChIP–seq at \(r\). We did not have DHS tracks for our Mbd3-knockout ES cells and therefore implemented our own version of the ABC model where we used the read counts in just the H3K27ac ChIP–seq data to score the activity of each regulatory element. Specifically, we identified putative regulatory elements using H3K4me3 and H3K27ac ChIP–seq data from wild-type cells and knockout cells where MBD3 had been depleted18 as follows:

  1. (1)

    Promoter regions were assigned as those regions ±1kb of a transcription start site and overlapping with an H3K4me3 ChIP–seq peak.

  2. (2)

    H3K27ac ChIP–seq peaks from that condition were considered as an initial putative list of condition-specific REs.

  3. (3)

    H3K27ac peaks closer than 500 bp were merged.

  4. (4)

    Peaks with total length <500 bp were discarded.

  5. (5)

    Peaks overlapping with promoter regions were discarded.

  6. (6)

    The union of the nonpromoter peaks and promoter regions was then treated as a master list of putative REs that were each assigned a condition-specific score based on the mean H3K27ac peak strength across that RE.

In particular, REs could be assigned within a single condition as being either ‘intergenic’ or ‘promoter’ associated. Since REs were defined per condition, wild-type-unique, knockout-unique and common overlapping peaks were assigned unique IDs for downstream analysis. The relevant code to perform this analysis and subsequent calculation of ABC scores can be found at https://github.com/dhall1995/Acitivity-By-Contact_Enhancer-Promoter_Link_Prediction.

Using calculated ABC scores, wild-type and knockout links were assigned unique link IDs based on the promoter and enhancer pair in question as well as the promoter genomic position in the case of a gene with multiple possible promoter regions. The top 10% of all identified links (wild-type and knockout) when ranked by ABC score were then selected as ‘strong’ links. In our data, this corresponded to an ABC threshold of ~0.12. Based on this threshold, unique link IDs could be assigned as wild-type-unique, knockout-unique or common depending on the conditions in which we observed that link. These thresholds were chosen to maximize the precision of identified links while retaining a large number of links to analyze although we acknowledge that maximizing precision comes at the expense of recall using the ABC model. Finally, after identification of wild-type-unique, knockout-unique and common links, enrichment analysis was performed by associating each link with a promoter and performing \({\chi }^{2}\) analysis using the statsmodels Python module.

CUT&RUN

CUT&RUN experiments were carried out according to Meers et al.73 using a Drosophila genomic DNA spike-in and ‘input’ controls consisting of samples processed in parallel with the CUT&RUN samples but with untethered MNase. Western blots of nuclear lysates were also carried out67 to measure the relative levels of CTCF and SMC3 in wild-type and Mbd3-knockout ES cells (Extended Data Fig. 1a). For the antibodies used see Supplementary Table 2.

We carried out 50 bp paired-end sequencing on a Novaseq instrument, with three biological replicates per sample obtaining 8–16 million mapped reads per replicate, respectively, whilst the input samples had 8–23 million mapped reads per replicate, respectively. All CUT&RUN data was trimmed using trim_galore (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/) and then aligned using Bowtie2 (ref. 74) with standard parameters to the GRCm38.p6 Mus musculus reference genome (https://www.ncbi.nlm.nih.gov/assembly/GCF_000001635.26/). Heatmaps of CUT&RUN enrichment were made using Deeptools v.2.5.0 (ref. 75). CUT&RUN bigwig tracks were calculated using BedTools and bedGraphToBigWig with standard parameters. The coverage was calculated with computeMatrix reference-point with options–binSize 10. The heatmap of standardized signal was then plotted using plotHeatmap. Peaks were called using MACS2 (ref. 76) so as to give a false discovery rate of 1% and above fivefold enrichment. All Venn diagrams were plotted using the matplotlib-venn library in Python v.3.6.

mES cell line generation

mES cell lines were cultured in 2iL conditions77 (50% DMEM/F-12 medium (Gibco catalog no. 21041025) and 50% Neurobasal Medium (Gibco catalog no. 12348017)) supplemented with 1× N2 to a final concentration of 2.5 µg ml–1 insulin (provided in-house by the Cambridge Stem Cell Institute), 0.5× B-27 Supplement (Gibco catalog no. 17504044), 1× minimum essential medium nonessential amino acids supplement (Sigma-Aldrich, catalog no. M7145), 2 mM l-glutamine (Life Tech, catalog no. 25030024), and 0.1 mM 2-mercaptoethanol (Life Tech, catalog 21985023), 2i inhibitors (1 µM PD0325901, 3 µM CHIR99021) and 10 ng ml–1 mouse leukemia inhibitory factor (mLIF provided by the Biochemistry Department, University of Cambridge). Cells were passaged every 2 days by washing in PBS (Sigma-Aldrich, catalog no. D8537), adding Trypsin-EDTA 0.25% (Life Tech, catalog no. 25200072) to detach the cells, and then washing in medium before replating in fresh medium. To help the cells attach to the surface, plates were incubated for 15 min at room temperature in PBS containing 0.1% gelatin (Sigma-Aldrich, catalog no. G1890). All cell lines were screened routinely for mycoplasma contamination at least twice yearly and tested negative.

ESC cells expressing CHD4 tagged at the C terminus with HaloTag were generated in the presence and absence of MBD3 (refs. 17,24,25). Briefly, this was achieved by CRISPR–Cas9 based knock-in of a cassette containing mEos3.2-HaloTag-Flag and a puromycin selection gene into one CHD4 allele. The puromycin cassette was then removed using Dre recombinase to generate the CHD4 allele with a C-terminal HaloTag fusion. Since knockout of CHD4 is lethal, we used cell viability assays and the ability to immunoprecipitate NuRD component proteins (Extended Data Fig. 1c) to verify that the function of the tagged CHD4 was not impaired. We similarly generated knock-in ES cells in an E14Tg2a (XY) background expressing MBD3 tagged at the C terminus with HaloTag (Extended Data Fig. 1b). MTA2-HaloTag knock-in cell lines were generated in MBD3-inducible ES cells18 (Extended Data Fig. 1d), in which MBD3 is fused to the estrogen receptor at both the N and C termini so that it initially localizes at the cytoplasm but then translocates to the nucleus when induced with 4-hydroxytamoxifen added directly to the culture medium to a final concentration of 0.4 nM. Western blots were carried out using nuclear lysates18, to confirm the expression and assembly of the NuRD complex (Extended Data Fig. 1a). Immunoprecipitations were carried out using antibodies to CHD4 or MTA2, or Halo-Trap beads (ChromoTek) in the case of the CHD4-Halo line. For the antibodies used, see Supplementary Table 3.

Live-cell 3D single-molecule imaging

ESCs expressing HaloTag-tagged CHD4, MBD3 and MTA2 were passaged 2 days before imaging onto 35 mm glass-bottom dishes No 1.0 (MatTek Corporation P35G-1.0-14-C Case) in serum/LIF imaging medium: Fluorobrite DMEM (Thermo Fisher Scientific, catalog no. A1896701) containing 100 mM 2-mercaptoethanol (Life Tech, catalog no. 21985023), 1× minimum essential medium nonessential amino acids (Sigma-Aldrich, catalog no. M7145), 2 mM l-glutamine (Life Tech, catalog no.25030024), 1 mM sodium pyruvate (Sigma-Aldrich, catalog no. S8636-100ML), 10% fetal bovine serum (HyClone FBS, catalog no. (lot no.) SZB20006, GE Healthcare catalog no. SV30180.03) and 10 ng ml–1 mLIF (provided by the Biochemistry Department, University of Cambridge). Glass-bottom dishes had been prepared by incubation in 0.01% poly-l-ornithine (Sigma-Aldrich catalog no. P4957) for 30 min, followed by two rinses in PBS and incubation in PBS containing 10 µg ml–1 laminin (Sigma-Aldrich catalog no. L2020) for at least 4 h. Just before single-molecule imaging experiments, cells were labeled with 0.5 nM HaloTag-JF549 ligand for 15 min, followed by two washes in PBS and a 30 min incubation at 37 °C in imaging medium, before imaging the cells in fresh serum/LIF imaging medium. Cells were underlabeled to prevent overlap of fluorophores during single-molecule tracking experiments. The HaloTag dyes were a kind gift from L.D. Lavis (Howard Hughes Medical Institute).

For the HaloTag-NLS control, the pEF-HaloTag-NLS vector was generated by replacing the HP1 sequence in a HaloTag-HP1 expression vector78 with a SV40 NLS sequence. 1 µg of the pEF-HaloTag-NLS vector was transfected into ES cells using Lipofectamine 2000 Transfection Reagent (Thermo Fisher Scientific, catalog no. 11668019) during passaging onto 35 mm glass-bottom dishes No 1.0 (MatTek Corporation P35G-1.0-14-C Case). Media was changed after 24 h and samples were both labeled as above and imaged the following day.

Transcription elongation was inhibited using 100 μM 5,6-dichloro-1-β-d-ribofuranosylbenzimidazole (DRB) and deacetylase activity using 10 nM FK228 (TOCRIS Bioscience) both for 2 h before imaging51,79.

A custom-made double-helix point spread function (DHPSF) microscope was then used for 3D single-molecule tracking25. The setup incorporates an index-matched 1.2 numerical aperture (NA) water immersion objective lens (Plan Apo VC ×60, Nikon) to facilitate imaging above the coverslip surface. The DHPSF transformation was achieved by the use of a 580 nm optimized double-helix phase mask (PM) (DoubleHelix) placed in the Fourier domain of the emission path of a fluorescence microscope (Eclipse Ti-U, Nikon). The objective lens was mounted onto a scanning piezo stage (P-726 PIFOC, PI) to calibrate the rotation rate of the DHPSF. A 4f system of lenses placed at the image plane relayed the image onto an EMCCD detector (Evolve Delta 512, Photometrics). Excitation and activation illumination was provided by 561 nm (200 mW, Cobolt Jive 100, Cobolt) and 405 nm (120 mW, iBeam smart-405-s, Toptica) lasers, respectively, that were circularly polarized, collimated and focused to the back focal plane of the objective lens. Oblique-angle illumination imaging was achieved by aligning the laser off axis such that the emergent beam at the sample interface was near-collimated and incident at an angle less than the critical angle (θc ~ 67°) for a glass–water interface. The fluorescence signal was then separated from the excitation beams into the emission path by a quad-band dichroic mirror (Di01-R405/488/561/635-25 ×36, Semrock) before being focused into the image plane by a tube lens. Finally, long-pass and band-pass filters (BLP02-561R-25 and FF01-580/14-25, respectively; Semrock) placed immediately before the camera isolated the fluorescence emission. Using 561 nm excitation, fluorescence images were collected as videos of 60,000 frames at 20 ms or 4,000 frames at 500 ms exposure. A continuous 561 nm excitation beam at ∼1 kW cm–2 was used for 20 ms exposure imaging and at ~40 W cm–2 for 500 ms exposure imaging. Each experiment was carried out with at least three biological replicates (three fields of view, each containing around three cells).

Residence time analysis from time-lapse 500 ms exposure imaging

Since photobleaching is related to the number of exposures, and the residence time is related to the time a molecule spends bound to chromatin, it is possible to change the time-lapse between exposures and use the data to extract both the residence time1 and photobleaching rate. However, when we imaged at time intervals of 0.5 s, 2.5 s, 8 s and 32 s, we discovered that, at the longest time-lapse (32 s), we could see no decrease in the mean number of frames imaged before photobleaching, implying that the residence time had no impact on the measurement, which was thus dominated by photobleaching (Extended Data Fig. 4e). To estimate the residence time would probably require imaging at much longer time-lapses but, because chromosomes and the cell itself move during periods longer than this, it becomes unreliable to track individual chromatin-bound NuRD complex subunits.

3D single-molecule image processing, generation of trajectories and determination of experimental precision

Single molecules were localized from 3D videos using the easy-DHPSF software80 with a relative localization threshold of 100 for all six angles for the 20 ms data and relative thresholds of 116, 127, 119, 99, 73 and 92 for the 500 ms data. The trajectories of individual molecules were then assembled using custom Python code for connecting localizations in subsequent frames if they were within 800 nm for 20 ms trajectories and within 500 nm for 500 ms trajectories (https://github.com/wb104/trajectory-analysis). This code also outputs average signal intensity per trajectory and trajectory lengths (OPTION -savePositionsFramesIntensities) and a summary of these data is reported in Extended Data Fig. 4b.

Our precision values (measured for fixed dye molecules on the coverslip and calculated using the approach described by Endesfelder et al.81) were 60 nm and 34 nm for the 20 and 500 ms tracking experiments, respectively. The lower limits of the effective diffusion coefficient (Deff) one can measure are dependent on the precision values. The Deff is equal to the displacement squared over time. Thus, if the upper limit of the precision is 60 nm, then the lower limit of Deff that we can measure for 20 ms imaging is 0.06 × 0.06/0.02 = 0.18 m2 s−1. For 500 ms imaging on the other hand, the upper limit of our precision was 34 nm, corresponding to a lower limit of Deff that we can measure of 0.034 × 0.034/0.5 = 0.002 m2 s−1. Consistently, when we measured the Dapp values for dye molecules attached to a coverslip we determined values of 0.3 ± 0.2 μm2 s–1 and 0.004 ± 0.003 μm2 s–1, respectively (Extended Data Figs. 3c and 5a). Any measured diffusion coefficients below these values do not have a biophysical interpretation.

Single-molecule trajectory analysis

The development of the algorithm to classify subtrajectories into confined and unconfined states based on four biophysical parameters using a GMM, and the use of this classification algorithm to analyze the single particle trajectories is described in the Supplementary Data and Methods.

In vitro biochemical assays of the NuRD complex with and without nucleosomes

Drosophila PMMR and Human CHD4 were expressed in insect Sf21 cells and purified as described24. Sf21 cells expressing Human GATAD2A-MBP were resuspended in 50 mM Tris-HCl pH 7.5, 1 M NaCl, 5 mM DTT and 1× complete EDTA-free protease inhibitor cocktail (Roche), lysed by sonication and cleared by centrifugation at 50,000g for 1 h. The supernatant was applied to amylose resin pre-equilibrated with lysis buffer and incubated for 2 h with rotation at 4 °C. The resin was washed with 20 column volumes of lysis buffer and then eluted with 10 mM maltose in lysis buffer. Fractions containing hGATAD2A-MBP protein were concentrated and further purified by size exclusion chromatography using a Superose 6 Increase 3.2/300 column (GE Healthcare) equilibrated with 50 mM Tris-HCl pH 7.5, 150 mM NaCl, 5% glycerol and 1 mM DTT.

For pulldown experiments, purified protein was immobilized on MBP-Trap resin (ChromoTek) pre-equilibrated in pulldown buffer (50 mM HEPES pH 7.5, 300 mM NaCl, 1 mM DTT and 5% v/v glycerol) followed by incubation for 1 h with rotation at 4 °C. A sample of the 6% protein:bead mixture was retained as ‘Input’. The resin was washed three times with pulldown buffer, then a washed ‘beads’ sample was retained for analysis on a 4–12% NuPAGE gel (Invitrogen).

Electrophoretic mobility shift assays were performed with n3-Widom-78bp DNA or recombinant nucleosomes made with this template82,83 in 10 µl of binding buffer (20 mM Hepes pH 7.5, 2 mM MgCl2, 5% glycerol and 1 mM TCEP) with varying concentrations of the indicated proteins. The reaction mixtures were incubated at 30 °C for 30 min followed by centrifugation at 1,000g. The resulting reaction mixtures were loaded onto 5% native polyacrylamide gels and run in 0.2× TBE. Gels were stained with SYBR Gold (Invitrogen) and imaged using a Typhoon FLA 9000 (GE Healthcare).

dCas9-GFP imaging of enhancer loci

ESCs expressing dCas9 tagged with GFP were generated3. Briefly, MBD3-inducible ES cells were transfected with the PB-TRE3G-dCas9-eGFP-WPRE-ubcp-rtTA-IRES-puroR vector containing a dual promoter backbone, with a TRE3G (Tet-on) promoter expressing GFP-tagged inactive dCas9 and the ubiquitin C promoter expressing the reverse tetracycline-controlled transactivator, rtTA, and a puromycin cassette via an IRES sequence. Puromycin-resistant ES cells were then selected for 7 days and doxycycline was added for 24 h to induce expression of dCas9-GFP (through activation of the rtTA). Stable transfectants were then FACS sorted for low levels of GFP expression to select cells where only a few copies of the plasmid were integrated stably into the genome.

Before imaging, 1 µg ml–1 of doxycycline was added to ES cells for 24 h to induce expression of low levels of dCas9-GFP. For imaging of Tbx3 enhancer loci, three CARGO vectors in total expressing 36 gRNAs targeting the Tbx3 enhancer were then transfected using lipofectamine 2000 (Invitrogen). The CARGO and dCas9-GFP expressing plasmids were gifts from the J. Wysocka laboratory. For imaging of Nanog enhancer loci, a custom designed gRNA was annealed with SygRNA Cas9 Synthetic Modified tracrRNA (Sigma-Aldrich catalog no. TRACRRNA05N). The gRNA was designed such that a single gRNA sequence could be used to uniquely target a repetitive sequence close to the relevant enhancer (Extended Data Fig. 6b). Cells were transfected with the tracr:crRNA complex using Lipofectamine 2000 (Invitrogen catalog no. 11668019). In all cases, cells were transfected during passaging straight onto imaging dishes in Fluorobrite imaging medium as described above. After 24 h, the medium was replaced with fresh medium and, for +MBD3 samples, 4-hydrooxytamoxifen was added to a final concentration of 0.4 nM. All samples were then imaged after a further 24 h.

2D tracking of genomic loci was carried out using oblique-angle illumination on a custom built 2D single-molecule tracking microscope84. Briefly, an IX73 Olympus inverted microscope was used with circularly polarized laser beams aligned and focused at the back aperture of an Olympus 1.40 NA ×100 oil objective (Universal Plan Super Apochromat, catalog no. UPLSAPO100XO/1.4). A 561 nm laser was used as a continuous wavelength diode laser light source. Oblique-angle illumination imaging was achieved by aligning the laser off axis such that the emergent beam at the sample interface was near-collimated and incident at an angle less than the critical angle (θc ~ 67°) for a glass/water interface. This generated a diameter excitation footprint of ~50 μm. The power of the collimated 488 nm beam at the back aperture of the microscope was 100 W cm–2. The lasers were reflected by dichroic mirrors that also separated the collected fluorescence emission from the TIR beam (Semrock, Di01- R405/488/561/635). The fluorescence emission was collected through the same objective and then further filtered using a combination of long-pass and band-pass filters (BLP01-561R and FF01-587/35). The emission signal was projected onto an EMCCD (Photometrics, Evolve 512 Delta) with an electron multiplication gain of 250 ADU per photon operating in a frame transfer mode. The instrument was automated using the open-source software micro-manager (https://www.micro-manager.org) and the data were displayed using the ImageJ software85,86.

For image processing, PeakFit87 was used to localize genomic loci from the images using the filter settings: ‘shiftFactor’:1.0, ‘signalStrength’:5.0, ‘minPhotons’:30.0, ‘precisionThreshold’:40.0, ‘minWidthFactor’:0.5, ‘maxWidthFactor’:0.5 and ‘precisionMethod’:‘MORTENSEN’. Trajectories were then tracked in 2D using custom Python code for connecting foci in subsequent frames if they were within 500 nm (https://github.com/wb104/trajectory-analysis). Trajectories were classified as for single molecules using a GMM (see Single-molecule trajectory analysis in the Supplementary Data and Methods).

Enhancer–promoter DNA-FISH

FISH probes were prepared from mouse BAC library clones (Source Biosciences)88,89,90,91,92,93,94,95,96,97,98,99,100,101. BAC vector DNA was purified using the Qiagen Large Construct Kit. BAC DNA was labeled using Cy3 and Alexa Fluor 647 Nick Translation Labeling Kits (Jena Bioscience, catalog nos. PP-305S-CY3, PP-305S-AF647) and purified17. The BAC probes generated are shown in Supplementary Table 7.

For two-color DNA-FISH, 2 × 104 cells were seeded per well on microscope slides with removable eight-well silicone chambers (Ibidi catalog no. 80841). Cells were fixed in 4% formaldehyde in PBS (Pierce catalog no. 28906, Thermo Fisher Scientific) for 10 min at room temperature, followed by permeabilization in 0.3% Triton X-100 (Sigma-Aldrich X100) in PBS for 15 min at room temperature. After three washes in PBS, cells were incubated in prewarmed 2× saline-sodium citrate (SSC) with 100 µg ml–1 RNAse A (Qiagen, catalog no. 158922) for 1 h at 37 °C. Chambers were removed and slides washed in 2× SSC at room temperature. Slides were then dehydrated using 70% ethanol, 90% ethanol and 100% ethanol for 2 min each and left to air dry. Cells were denatured in 70% deionized formamide (Sigma-Aldrich catalog no. S4117) in 2× SSC at 80 °C for 15 min. Slides were then again dehydrated quickly through ice-cold 70% ethanol, 90% ethanol at room temperature and 100% ethanol at room temperature for 2 min each and again left to air dry.

For each sample slide, 150 ng of Cy3-labeled BAC probe and 150 ng of AF647-labeled BAC probe were precipitated with 5 µg of salmon sperm DNA (Invitrogen, catalog no. 15632011) using 0.1 volumes of 3 M sodium acetate (pH 5.2) and 2.5 volumes of 100% ethanol. Precipitated DNA was pelleted through centrifugation at 15,000g for 20 min and resuspended in 50 µl hybridization buffer (50% formamide, 10% dextran sulfate (Sigma-Aldrich, catalog no. 42867), 0.1% SDS, 2× SSC) through incubation for 1 h at 37 °C. Probes were denatured at 80 °C for 10 min and then transferred to 37 °C. Sample slides were overlaid with 50 µl hybridization solution and covered with Parafilm, after which hybridization was allowed to occur at 37 °C overnight in a humidity chamber. The following day, the coverslip and hybridization solution were removed, and the slides washed four times in 2× SSC at 40 °C for 3 min each, then four times in 0.1× SSC at 60 °C for 3 min. Slides were cooled by washing in 4× SSC at room temperature. After removing all the wash solution, cells were mounted in VECTASHIELD Antifade Mounting Medium with 4,6-diamidino-2-phenylindole (DAPI) (Vector Laboratories, catalog no. H-1200-10).

Sequential immunofluorescence for dCas9-GFP and DNA-FISH

ESCs expressing dCas9 tagged with GFP were transfected with either CARGO plasmids or a gRNA as described above (dCas9-GFP imaging of enhancer loci). Fresh medium was added after 24 h and cells were fixed the following day in PBS containing 4% formaldehyde (Pierce, catalog no. 28908, Thermo Fisher Scientific) for 10 min at room temperature. Cells were permeabilized in PBS containing 0.5% Triton X-100 (Sigma-Aldrich X100) for 5 min, washed three times with PBS and then treated with blocking buffer (4% bovine serum albumin (Sigma-Aldrich, catalog no. A9418) in 0.1% Triton X-100 in PBS) for 30 min. Cells were incubated with GFP-Booster Alexa Fluor 488 nanobody (ChromoTek, catalog no. gb2AF488) in blocking buffer (1:1,000) through incubation for 1 h at room temperature. Samples were washed three times in PBS, each for 5 min.

Cells were postfixed in PBS containing 3% formaldehyde (Pierce, Thermo Fisher Scientific, catalog no. 28908) for 10 min at room temperature, followed by repermeabilization in 0.1 M HCl in 0.7% Triton X-100 in PBS for 10 min on ice. After two washes in 2× SSC for 5 min each, cells were incubated in prewarmed 2× SSC with 10 U ml–1 RNAse A (Qiagen, catalog no. 158922) for 1 h at 37 °C. Slides were then equilibrated in 20% glycerol in PBS for 1 h, followed by three consecutive freeze–thaw cycles using liquid nitrogen. After incubation for 1 h in denaturing solution (50% formamide in 2× SSC) at room temperature, slides were denatured at 70 °C for 5 min and then washed several times in ice-cold 2× SSC. Probes were denatured at 70 °C for 10 min and then placed on ice to cool. Hybridization solution was prepared with 50 ng of BAC probe and 10 µg salmon sperm DNA (Invitrogen, catalog no. 15632011) per 100 µl of hybridization buffer (50% formamide, 10% dextran sulfate, 1 mg ml–1 BSA and 2× SSC). Sample slides were overlaid with 25 µl hybridization solution per well and covered with Parafilm. After denaturation at 70 °C for 5 min on a heat block, the slide was gradually cooled to 37 °C and hybridization allowed to occur at 37 °C overnight in a humidity chamber. The following day, the coverslip and hybridization solution were removed, and the slides washed three times in 2× SSC at 40 °C for 5 min each, then three times in 2× SSC at room temperature for 5 min. After removing all the wash solution, cells were mounted in VECTASHIELD Antifade Mounting Medium with DAPI (Vector Laboratories).

Immunofluorescence and DNA-FISH image acquisition and analysis

Imaging was carried out using a Zeiss LSM Airy Scan 2 super-resolution microscope set for imaging of DAPI (405 nm laser, 0.8%), Cy3 (514 nm laser, 10%) and AF647 (639 nm laser, 90%). Three stacks of horizontal plane images (38,04 × 3,804 pixels corresponding to 136.24 × 136.24 μm2) with a z-step of 150 nm were acquired for each field of view. CZI image files were then imported into IMARIS v.9.6 (Bitplane) for 3D modeling. Quantitative analysis of interprobe distances within nuclei was carried out using the Surfaces and Spots modules of Imaris v.9.6.

Software and code

The following tools were used for data collection: microscope image acquisition, Micro-manager (https://www.micro-manager.org); ImageJ software85,86.

The following tools and methods were used: for Hi-C analysis - NucProcess (https://github.com/tjs23/nuc_processing), NucTools (https://github.com/tjs23/nuc_tools), Juicer (https://github.com/aidenlab/juicer), Cooler (https://cooler.readthedocs.io/en/latest/), CscoreTool (https://github.com/scoutzxb/CscoreTool) and Lavaburst (https://github.com/nvictus/lavaburst); for enhancer–promoter analysis (https://github.com/dhall1995/Acitivity-By-Contact_Enhancer-Promoter_Link_Prediction); for CUT&RUN analysis - Trim galore (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/), Bowtie2 (ref. 74), Deeptools v.2.5.0 (ref. 75) and MACS2 (ref. 76) (data were processed using the GRCm38.p6 mouse reference genome (https://www.ncbi.nlm.nih.gov/assembly/GCF_000001635.26/); for 2D single-molecule peak fitting - PeakFit87; for 3D single-molecule peak fitting - easy-DHPSF80; for trajectory analysis (https://github.com/wb104/trajectory-analysis); for trajectory overlay visualization - TrackMate; for GMM classification (https://zenodo.org/record/6497411#.YmlGFy8w3q0); for 3D DNA-FISH analysis - Imaris v.9.6.

Biological materials

All constructs and cell lines are available upon request. The Chd4, Klf4, Mbd3 and Mta2 Eos-Halo targeting constructs have also been deposited with Addgene.

Reporting summary

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