The spatial organization of the chromosome around a gene is thought to be intimately linked to its expression, providing an important mechanism for gene regulation. At short length scales (<5 kilobase pairs (kbp)), this mechanism might involve nucleosome repositioning and disruption of the chromatin fiber structure1,2, modifying the accessibility of the DNA to proteins. At larger length scales (>5 kbp), this could involve the way chromatin forms loops bringing together promoters and their enhancers. Advances in microscopy3,4, and in next-generation sequencing methods such as chromosome-conformation-capture (3C) and its variants5, have revealed much about gene locus structure. More recently, the development of structural probes operating at the single-cell level has revealed striking variability within a tissue or population of cells that are phenotypically homogeneous3,4,6. The compatibility of such cell-to-cell variability with a robust transcriptional program and emerging phenotype is remarkable and still awaits full explanation. The variability of gene loci was exemplified by our recent modeling work on the locus of Pax6, a highly regulated and highly conserved developmental gene. Our simulations showed that chromatin interaction patterns revealed by CaptureC experiments could be generated by locus conformations that vary widely from cell to cell, and that the level of variation (validated using DNA fluorescence in situ hybridization microscopy (FISH)) also markedly depends on cell type and expression level7.

Although experimental methods probing chromosome organization continue to improve, the majority of studies to date have focused on fixed cells, providing information on only a ‘snapshot’ in time. An understanding of how locus conformations evolve dynamically remains largely elusive. For example, it is unclear whether the structural variation observed across a population is representative of the configurations adopted dynamically within a single cell. Or, does the chromatin in a single cell only visit, or sample, a small part of this ‘configuration space,’ with the observed variability arising only when gathering data from many cells? In other words, if one were to track locus configuration in a live cell, would one observe wide variations or a relatively static picture? The answer to this may give insight into how such variability can still give tight control of expression and phenotype. Although live-cell imaging has advanced markedly in recent years, challenges remain. Super-resolution microscopy allows ever-increasing spatial resolution to be achieved, and high-throughput techniques, such as Hi-D8, have been developed to monitor chromatin diffusivity in vivo. However, it remains difficult to reach high temporal resolution while labeling multiple points of interest simultaneously. Here, we use biophysical modeling and computer simulations9 to study and predict dynamics at the Pax6 locus. This approach provides mechanistic insight as well as new hypotheses and testable predictions, which we hope will stimulate further experiments.

Previously, we developed the HiP-HoP simulation framework to predict structural information on a gene locus at both the single-cell and population level. This was applied to several gene loci, including those of Pax6 in mouse, SOX2 (ref. 7) and CCND1 (ref. 10) in human, and (using an earlier version) mouse alpha- and beta-globin11. Here, we evolved HiP-HoP such that it can be used to analyze the dynamics of the Pax6 locus in three mouse tissue-derived cell lines that express the gene at different levels12,13. We first examined the dynamic properties of simulated chromatin regions at different length scales and found that these vary substantially across the locus. The mobility of a given segment depends both on its biophysical properties (protein binding, local chromatin compaction) and on its surroundings, with local macromolecular crowding playing a pivotal role (we use the term ‘mobility’ to refer to a measure of how far a given segment will move or diffuse within a given time). We then studied the dynamics of the overall locus configuration, analyzing the timescale over which promoter-enhancer interactions change. The results suggest that the locus can sample all of its different configurations within minutes, strongly indicating that wide dynamic variation of structure would be observed within a single cell. Finally, we performed interventional experiments, in which we gave cells treatments either to inhibit transcription or to release topological strain; we made measurements in fixed cells and compared these with dynamic simulations. Interestingly, we found that removing proteins representing polymerase complexes from the simulation did not reproduce the changes observed in the transcription inhibition experiment; this led us to an alternative model scenario.


Modified HiP-HoP framework to explore locus dynamics

We have previously used HiP-HoP7,9,10 to study the structure of the Pax6 locus in three mouse cell lines, denoted Pax6 OFF, ON, and HIGH (indicating the expression state of the gene). The model represents a chromosome region as a chain of beads and combines three key mechanisms that organize the locus (Fig. 1a). First, model proteins (representing complexes of RNA polymerase and transcription factors) that diffuse within the system can interact with the chromatin at specific binding sites; importantly, these proteins are multivalent and can bind multiple sites simultaneously, forming molecular bridges between distal chromatin sites14,15. Second, the loop extrusion mechanism is included16,17,18; this asserts that the cohesin complex can actively push chromatin into loops, which grow until they are stabilized by CTCF proteins bound in a convergent orientation19. Finally, the model includes a ‘heteromorphic polymer’ description of chromatin, meaning that the biophysical properties, for example those determined by the fiber structure, can vary along its length1,2,20. For this, we discriminate between two chromatin states, one with a more compact, thicker fiber, and one with a thinner, less compact (that is, more open, disrupted, or flexible) fiber21 (see Methods and Supplementary Notes). Here, we predominantly consider active chromatin regions, but this model could be extended to account for repressed genomic regions10,22. An important feature is that, in the simulations, proteins tend to come together into clusters; this is driven by a mechanism known as the ‘bridging induced attraction’15,23 and is a consequence of a protein’s ability to form molecular bridges between distal chromatin sites. Clusters form around two or more protein-binding sites on the chromatin and represent the foci or ‘phase-separated droplets’ of transcription-associated proteins observed in recent microscopy studies24,25 (although the mechanism through which this occurs in vivo is not fully understood). In HiP-HoP, proteins continually switch between a chromatin-binding and a non-binding state26, modeling post-translational modifications and resulting in clusters having realistic protein dynamics in terms of the exchange between the cluster and the soluble pool.

Fig. 1: Model schematic and simulation set-up.
figure 1

a, In the HiP-HoP model, a chromosome region is represented by a bead-and-spring polymer, in which each bead represents 1 kbp of chromatin. The model includes diffusing protein complexes represented by spheres, loop extruders represented by additional springs, and a heteromorphic structure in which some polymer regions have additional next-nearest neighbor springs that lead to local crumpling of the chain. Proteins stochastically switch between a binding and a non-binding state; when in the binding state, there is an attractive interaction between the proteins and their binding sites on the chromatin. Further details are given in the Methods and Supplementary Notes. b, A map of the mouse Pax6 locus (mm9 genome build). Positions of CaptureC targets or viewpoints and two sets of FISH probes are indicated (see Supplementary Tables 1 and 2, respectively, for genomic coordinates). c, Schematic showing the simulation set-up. Simulations of a 40-Mbp chromatin fiber (40,000-bead polymer) were performed. Ten copies of a 3-Mbp region around the Pax6 locus were included on each fiber, allowing multiple results to be obtained from each simulation. Copies of the locus used input data from one of the three cell lines; in repeat simulations, versions of the locus from different cell lines were in different positions. In this way, each locus experienced a similar surrounding environment. d, Simulation snapshot showing a typical configuration of the 40-Mbp fiber. One of the ten copies of the locus is shown in orange. In the top right, this same copy of the locus is depicted with the rest of the fiber not visible. In the bottom right, a zoom shows only the region immediately surrounding Pax6 and two nearby regulatory regions (upstream and downstream regulatory regions, URR and DRR). e, Further simulation snapshots show example configurations of the locus (only the 500-kbp region between Rcn1 and Elp4 is shown) in each of three cell lines in which Pax6 is expressed at different levels. Chromatin regions with H3K27ac are shown in yellow, and ATAC-seq peaks (binding sites) in red. The Pax6 gene body is shown in blue.

Three data sets are used as an input to HiP-HoP. First, DNA accessibility data (assay for transposase-accessible chromatin with sequencing (ATAC-seq)) are used to identify protein-binding sites (we use a simplifying assumption that ATAC peaks coincide with binding sites for ‘active’ proteins). Second, chromatin immunoprecipitation (ChIP) data on acetylation of lysine 27 on histone H3 (H3K27ac) are used to identify regions that are in the more open chromatin state: this decision is based on previous studies suggesting that chromatin possessing this mark has a disrupted structure2,27. Finally, ChIP data on CTCF and cohesin (Rad21) are used to identify positions of loop-stabilizing anchors (see Methods and Supplementary Notes).

In our previous study of Pax6, we used HiP-HoP to generate a population of simulated locus structures, which provided good predictions of experimental measures at both the population (validated by CaptureC) and single-cell (validated by FISH) level. CaptureC is a many-to-all 3C method which probes genome-wide interactions at a set of selected target locations28,29.

Obtaining realistic dynamics of locus structure required substantial changes to the simulation set up. We hypothesized that local variation in chromatin density might play a role in variation of dynamics across the simulated region, so we therefore used a more realistic overall chromatin density and simulated a larger chromatin fragment than in previous work. For computational efficiency, a 3-mega-base-pair (Mbp) region around Pax6 (chr2:104,000,000–107,000,000 mm9 genome build) was selected and concatemerized 10 times along a 40-Mbp fiber (that is, one simulation is equivalent to taking measurements of the locus across 10 single cells, Fig. 1c,d). Simulations representing dynamics in 20 single cells from each cell line were performed, and we determined that extracting 400 configurations at regular time intervals provided a good representation of locus motion (see Methods; example configurations are shown in Fig. 1e).

To ensure that the simulations still gave good predictions of locus conformation with this new set up, results were compared with both CaptureC data (Extended Data Fig. 1) and FISH (Extended Data Fig. 2). CaptureC targets were positioned at promoters and CTCF-binding sites across the locus (Fig. 1b, green stars). Three FISH probes were selected to cover the Pax6 promoters and two previously identified enhancers, denoted the upstream and downstream regulatory regions (URR and DRR; Fig. 1b, purple blocks), and were used in three-color imaging experiments to obtain simultaneous measurements of probe separations in single cells. Two quantitative metrics (see Supplementary Notes) confirmed that conformations predicted by this new version of HiP-HoP showed similar levels of agreement with experiments as the original model7 (Extended Data Fig. 3).

Dynamics depend on transcription and fiber structure

To explore the dynamics around Pax6, we used a simulation strategy analogous to live-cell tracking of specific regions of the locus corresponding to the positions of the Pax6, URR, and DRR FISH probes (Fig. 2a). The mean squared displacement (MSD) was calculated as a function of lag time for each probe (Fig. 2b–d). In these simulations, the length-scale of the chromatin (the diameter of the bead (σ)) is 17.6 nm, determined by comparing simulated and experimental FISH measurements. The simulation time unit (τ) is approximately 2.07 ms, determined by comparison with previous motion-tracking experiments30. This gives, for each simulation, an approximate total duration of 27 min (see Methods and Supplementary Notes for details on this mapping).

Fig. 2: Dynamics of individual chromatin regions.
figure 2

a, To quantify the dynamics of a region covered by a FISH probe, the MSD of the center of mass of that region was computed (top). Alternatively, the MSD of a single chromatin bead was computed (bottom). b–d, Plots showing the MSD of simulated fluorescent probes positioned at the Pax6 promoter (c) and the two distal regulatory regions (URR (b) and DRR (d)). These cover the same regions as the FISH probes used in fixed-cell experiments (probe set 1 in Fig. 1b). Simulation lengths and times were mapped to real units, as detailed in the Methods and Supplementary Notes. All results were obtained from 20 independent simulations of the locus; results from three cell lines are shown (lines) and the shaded regions indicate the s.e.m. The inset plots show log(MSD) vs. log(lag time), and the black lines have a slope of 0.5. e, For each cell line, chromatin mobility is plotted as a function of position across the locus (units are ms2). Also shown is the mean local density (units are 10−3 kbp µm−2) (see f). The line indicates values for single polymer beads (each representing 1 kbp of chromatin); the shaded region indicates the s.e.m. (for the density, this is typically similar in size to the line width). Above each plot the colored block shows the input data used for each cell line as indicated: yellow are H3K27ac regions, red are binding sites inferred from ATAC-seq peaks, other regions are gray. Gene positions are indicated above the plots. Purple blocks under the genes indicate the positions of the simulated FISH probes used to obtain the MSDs in bd. f, The local density is determined for each chromatin bead by counting the number of proteins and polymer beads within a radius of 3σ ≈ 53 nm.

MSD curves for different probes, and for a given probe in different cell lines look highly similar, indicating that the mobility of these regions is under similar constraints. For a freely diffusing object, the MSD would grow linearly with time, but as expected for motion of a polymer segment, the MSDs here grew more slowly31 (inset Fig. 2b–d). Some slight differences between cell lines were observed at large times (e.g., longer than 600s); particularly, for the Pax6 probe the MSD grew more slowly for Pax6 ON cells than for the others. Considering the input data, which determine the chromatin properties, Pax6 ON cells have more protein-binding sites (ATAC-seq peaks) than the other two cell lines, suggesting that protein-mediated chromatin looping may lead to reduced mobility.

To further quantify the dynamics, we defined a mobility measure (M) as the MSD after a fixed lag time of 104 × τ ≈ 20.7 s. For each cell line, M was calculated for each 1-kbp bead to obtain a ‘mobility profile’ over the locus (Fig. 2e). These profiles revealed large differences in mobility both across the locus and between cell lines. Most strikingly, around the DRR, the mobility of Pax6 HIGH cells was around 1.5 times higher than that of ON or OFF cells. In general, mobility tended to be higher in regions with H3K27ac (Fig. 2e, yellow blocks) and lower at ATAC-seq peaks (red blocks). This is in stark contrast to the FISH probe MSDs, for which little difference between probes or cell lines was observed. This shows that dynamic measurements are highly sensitive to the way that chromatin is probed: local variation in mobility of different 1-kbp beads was largely obscured when we analyzed 40-kbp probes, suggesting that the former depends on the behavior of chromatin over small length scales. Simulating probes of different sizes revealed that both the observed mean and s.d. of mobilities across the locus decreased with increasing probe size (Extended Data Fig. 4a,b).

Measurements of local macromolecular density around each polymer bead (Fig. 2f) revealed that this has a strong effect on diffusion. A lower local density was observed in regions that were more mobile (Fig. 2e): there was a significant negative correlation (Pearson correlation between −0.5 and −0.6, P < 10−10; Fig. 3a). It is also clear that the local density was enriched at protein-binding sites, visualized by the color coding of points in Figure 3a. Beads at ATAC-seq sites (red) had a 5.15% lower mobility, on average, than that of other beads. Reducing the number of model proteins in simulations led to a reduction in local densities and an increase in mobility at binding sites (Extended Data Fig. 4c,d), confirming that protein binding and bridge formation impair motion. By contrast, the mobility of H3K27ac-marked beads (yellow in Fig. 3a) was 7.13% higher, on average, than that of other beads. This is consistent with the more compact (non-H3K27ac) regions being more constrained than H3K27ac regions within the fiber. Increased mobility has previously been associated with histone acetylation in live-cell imaging32; in accordance, there was a positive correlation between the mobility of 20-kbp windows and H3K27ac density in our simulations (Fig. 3b; Pearson’s r = 0.43, P < 10−10).

Fig. 3: Local density, chromatin compaction, and loop extrusion affect dynamics.
figure 3

a, Scatter plots show the relationship between mobility and local density; each point represents a single polymer bead, and data from the whole 3-Mbp simulated region are included. Point color indicates the type of bead (as determined from the input data). b, Plot showing the relationship between mobility and H3K27ac coverage. Values for each quantity were calculated for a 20-kbp window around each 1-kbp bead (the mean value of the mobility within the window, and the fraction of the 20 kbp region covered by the H3K27ac mark are used). To obtain the color map, data were binned on both quantities; the darker the color, the more 20-kbp windows belong to the bin (color bar unit is number of windows). To obtain points, data were binned according to only the H3K27ac level, and the mean mobility was calculated for each bin. Error bars show the s.e.m. The Pearson correlation coefficient and P value (two-sided Student’s t-test, P = 1.5 × 10−13, based on n = 8,973 values) are shown. c, Plot showing the relationship between mobility and extruder occupancy. Extruder occupancy for a given bead was measured as the fraction of time in which it is bound by an extruder, and we used the mobility for each bead. Again, the color map was obtained by binning on both quantities (color bar unit is number of beads), and the points were obtained from binning on extruder occupancy with the mean mobility shown for each bin. Error bars show the s.e.m., and the Pearson correlation coefficient is indicated (two-sided Student’s t-test, P = 0, based on n = 9,000 values).

To examine the effect of loop extrusion on dynamics, the extrusion probability (φe) for each bead was defined as the fraction of time in which it was in the vicinity of an extruder. In HiP-HoP, φe depends only on the positions of and occupancy at CTCF-binding sites, and it shows peaks at high-occupancy CTCF sites (Extended Data Fig. 5). Extrusion does affect dynamics: there was a relatively weak negative correlation between extrusion probability and mobility (Fig. 3c). To understand how extrusion affects dynamics more generally, we performed simulations in which we varied the extrusion rate and the density of CTCF sites (by adding new sites in random positions; Extended Data Figs. 6 and 7 and Supplementary Notes). Intriguingly, changes to mobility across the locus were subtle and difficult to predict a priori. The largest changes occurred at CTCF sites, which typically showed decreased mobility when either the extrusion rate or the number of CTCF sites was increased (Extended Data Figs. 6c and 7c). Both changes led to extruders reaching CTCF sites more quickly, with CTCF–CTCF loops forming more readily. Because extruder-mediated loops can stabilize (or disrupt) nearby protein-mediated loops, changing the pattern of CTCF looping leads to changes across the locus (see Supplementary Notes for further discussion).

A concern from this analysis is that the dynamics of a given chromatin bead are simply a reflection of the properties with which it is endowed in the simulation scheme. However, beads that do not overlap ATAC, H3K27ac, or CTCF peaks (that is, ‘unmarked’ beads) exhibited a broad range of mobility values (Extended Data Fig. 4d), indicating that the dynamics of a given bead emerge both from its own properties and those of its local environment33.

Dynamics show correlation with interaction locality

We hypothesized that the chromatin properties giving rise to the variation in mobility might lead to differences in chromatin interactions. A common measure extracted from 3C methods is the ratio between the amount of local and long-range chromatin interactions at a given site34. Although the interpretation of long-range interactions within a 3-Mbp locus is limited, a measure of interaction ‘localness’ can be defined as the ratio between the number of interactions with regions within 100 kbp of the target and the number of interactions with regions farther than 100 kbp (Fig. 4a; see Supplementary Notes for details); 100 kbp was chosen because it is slightly smaller than typical promoter-enhancer loop sizes, so cis-regulatory loops will be counted as long-ranged (but different thresholds might be more informative if a larger locus is being considered).

Fig. 4: Dynamics are correlated with interaction locality.
figure 4

a, Localness is defined for a given chromatin bead as the number of interactions with regions closer than 100 kbp genomically, divided by the number of interactions with regions farther away than 100 kbp. An interaction was defined as two chromatin beads being closer together than 3.5σ ≈ 62 nm. b, Plots showing how localness of interactions varies across the locus in each cell type in simulations (colored lines; a value is shown for each 1-kbp chromatin bead). log2 values are shown. For comparison, gray lines show the mobility as in Figure 2e (a different vertical scale is used). Above each plot, the gray, yellow, and red bars indicate unmarked chromatin regions, chromatin regions marked with H3K27ac (open chromatin), or ATAC-seq peaks (binding sites), respectively, in simulations. c. Plot showing the relationship between mobility and localness of interactions. To obtain the color map, each chromatin bead was put into a bin on the basis of its localness and mobility; the darker the color, the more beads in the bin (color bar unit is number of beads). Points were obtained by binning each chromatin bead only according to its mobility; the mean and s.e.m. for each bin are shown. The Pearson correlation coefficient and P value (two-sided Student’s t-test, P = 8.3 × 10−14, based on n = 9,000 values) are indicated.

Plotting localness for each 1-kbp bead across the simulated region (Fig. 4b) revealed a pronounced reduction in localness at protein-binding sites (ATAC-seq peaks, red bars), consistent with these sites being involved in bridging interactions with distant (>100 kbp) regions. Some, but not all, H3K27ac-marked regions (open chromatin, yellow shading) also displayed reduced interaction localness. There was a small, statistically significant positive correlation between localness and mobility (Fig. 4c). The trend was clear for beads with lower mobilities but did not continue for those with mobilities larger than 0.05 µm−2 (which had a larger spread of localness values, and possibly a non-monotonic relationship, but these constitutes <9% of the beads), consistent with protein-binding sites having low localness values, being in higher density regions, and having lower mobility.

Locus conformations change on a timescale of minutes

The dynamics of the locus conformation as a whole can be accessed from a simulation by tracking in time the relative positions of all beads (N) in the chain. A more tractable approach (closer to what might be realized experimentally) is to track a smaller number of points. Here, we considered the same three FISH probe regions as above (Pax6, URR, and DRR) and took the three pairwise separations of these probes as a description of the locus configuration. These separations are represented by a point in a three-dimensional space (Fig. 5a): we can write a vector X = (xUP,xUD,xPD), where the components xUP, xUD, and xPD are the separations of the URR and Pax6, the URR and DRR, and the Pax6 and DRR probes, respectively. In Figure 5b, each point represents a single instant in time in 1 of our 20 simulations of the locus in Pax6 ON cells. The volume of the cloud of points represents the range of structures that the locus can adopt (Fig. 5d), providing a metric for ‘locus variability.’ The trajectory of X as it moves through this ‘configuration space’ during a single simulation can be overlaid on the scatter plot (Fig. 5c; see Extended Data Fig. 8 for examples from Pax6 HIGH and OFF cells). Examination of these trajectories suggests that the whole volume can be explored within a single simulation of duration 7.96 × 105τ, equivalent to roughly 27 min—considerably shorter than a typical cell cycle (see Supplementary Notes for details). This indicates that much of the structural variability of a locus can be exhibited within a single cell.

Fig. 5: Dynamics of locus conformation.
figure 5

a, Locus conformation can be characterized using the three distances between three regions covered by FISH probes (Pax6, URR, and DRR). These can be considered as a point in a three-dimensional configuration space. As the configuration changes in time, this point moves in configuration space. b, Scatter plot showing configurations adopted in Pax6 ON cells. Each point represents a single time point; data are shown for 20 repetitions of the simulation. Units are nm, and all axes show the same range. c, The same scatter plot as in b, but a trajectory from a single representative simulation is overlaid. The time interval between points along the path is equivalent to 4 s; the total simulation duration is 27 min. d, The size of the cloud of points represents the variability of the locus. e, The trajectory of a point through configuration space (X(t)) can be used to quantify how quickly locus conformation changes via a shape-change parameter. f, The shape-change parameter as a function of lag time. Error bars show s.e.m. (for each point, the number of independent measurements is at least n = 3,570). g, The same data as in f are shown on a logarithmic scale (error bars show s.e.m.). The black line has a slope of 0.5. h, Scatter plot of configurations for Pax6 ON from simulations with a reduced protein switching rate (black points show ksw ≈ 0.097 s−1; gray points show ksw ≈ 0.49 s−1, as in b). Axis ranges are the same as in b. i, Bar plot showing how locus variability changes with switching rate. j, The shape-change parameter is shown for Pax6 ON from simulations with different switching rates. Error bars show s.e.m. (the number of independent measurements is at least n = 1,850). k, Similar plot to that in h, but showing simulations with 100 proteins (black) and 1,000 proteins as in a (gray); ksw ≈ 0.49 s−1 in both cases. l, Bar plot showing how the number of proteins affects locus variability (ksw ≈ 0.49 s−1). m, The shape-change parameter is shown for Pax6 ON from simulations with different numbers of proteins. Error bars show the s.e.m. (the number of independent measurements is at least n = 1,850).

To examine this quantitatively, a shape-change parameter (S(t)) was defined, which tracks the mean change in locus configuration over a lag time (t). Mathematically, this is the MSD of X as it moves through configuration space (Fig. 5e and Supplementary Notes). As expected, S(t) reaches a plateau at large times, because the locus explores a finite volume within configuration space (Fig. 5f). This occurs within about 100–200 s, implying that all of the configurations are explored within this time frame. The behavior at short times, where S grows as some power α of time (i.e., Stα), gives information about the dynamics of the locus structure: all three cell lines show α < 1, characteristic of sub-diffusive behavior (Fig. 5g). Pax6 ON and HIGH have α ≈ 0.5; HIGH cells generally have a larger S(t), indicating that the configuration changes more quickly. Sub-diffusion is expected, owing to the polymeric nature of the chromosome; loops being transiently stabilized by protein bridges or loop extruders likely act as dynamical traps, further inhibiting motion. The Pax6 OFF S(t) curve shows a smaller α, perhaps reflecting the very different pattern of protein-binding sites in those cells.

How well do simulations reproduce the dynamics of real locus structures in vivo? It is likely that the dynamics depend heavily on the model parameters, for example the number of proteins, the rate at which proteins switch between binding and non-binding states, the number of extruders, or extrusion speed35. So far, we have used biologically reasonable parameters that were optimized to best predict static (fixed-cell) measurements7. However, it is possible that there are distinct sets of parameters that give similar static predictions, but different dynamic behavior. Although experimental determination of the parameters remains challenging, we can use the simulations to examine the effect of varying them.

First, simulations with different rates at which proteins switch between binding and non-binding states (the protein switching rate, ksw) were performed for Pax6 ON cells (Fig. 5h): the rate was reduced from ksw = 10−3τ−1 ≈ 0.48 s−1 (gray points) to ksw = 2 × 10−4 τ−1 ≈ 0.097 s−1 (black points). Less frequent switching led to a smaller volume in the configuration space being explored (reduced variability, Fig. 5i): configurations in which one or more of the probe pair separations is small became favored over extended configurations (reduced mean locus size). The locus also changed configurations more slowly (revealed by examining S(t), Fig. 5j). These effects arise because reducing the protein switching rate leads to an increase in the size and longevity of protein clusters26: protein-stabilized loops are more likely to form for extended times, slowing the dynamics and favoring more compact configurations (including multi-loop ‘rosette’ structures).

If the switching rate is kept constant, but instead the number of proteins is decreased, there is a smaller decrease in variability (Fig. 5l), and configurations in which the probe pair separations are small are disfavored (mean locus size increases, Fig. 5k). If there are fewer proteins, there will be a lower likelihood of protein-stabilized loops. Unlike when altering ksw, there was no change in configurational dynamics in this scenario (Fig. 5m). These results further support the idea that there is non-trivial interplay between different model ingredients. A decrease in locus variability can be accompanied by either an increase or a decrease in mean locus size and does not necessarily lead to a change in configurational dynamics. We note that the changes to the parameters led to poorer agreement with the fixed-cell experiments (CaptureC and FISH), but nevertheless provide insight into the underlying biophysical mechanisms. Interestingly, there was very little effect on S(t) when loop extrusion parameters were varied (Extended Data Figs. 6d and 7d), likely owing to the specific pattern of CTCF binding around Pax6 (loop extrusion does not seem to play a major role in promoter-enhancer interactions in this locus7). Other recent work using HiP-HoP suggests that loop extrusion plays an important role in cell-to-cell variability of expression22; it would be interesting in the future to study configurational dynamics in other loci where, for example, enhancers and promoters sit at opposite ends of a CTCF loop domain.

Finally in this section, we note that for Pax6, the choice to consider three probes is natural because there are two distal enhancers. One could consider more probes and examine how the system moves through a higher-dimensional configuration space. That might be more relevant for larger, more complex loci with many regulatory elements, but the general conclusions are unlikely to be affected. A complementary approach for characterizing locus dynamics is to measure time intervals between enhancer-promoter collisions and the duration of their interaction; we include such an analysis in Supplementary Notes (also Supplementary Fig. 1). A surprising result is that interaction durations are at most twofold longer in ON cells than in OFF cells; this suggests that the stabilization of loops by protein clusters is modest.

Perturbing gene expression weakly affects locus conformation

Given the apparent links between locus configuration, gene activity, and dynamics, we performed new CaptureC and FISH experiments on Pax6 HIGH cells after treatment with drugs to perturb either transcription or topology. Alpha amanitin was used to inhibit transcription through selective degradation of elongating polymerases36,37. Surprisingly, but as observed previously38, CaptureC profiles looked very similar to those from untreated cells, with similar, if slightly higher, interaction peaks (Fig. 6a). Separations of FISH probes located at the Pax6 promoters and two neighboring genes also showed no significant change after treatment (Fig. 6b). Together, these findings suggest that, at least at Pax6, inhibiting transcription per se does not greatly affect structure, but frequently occurring interactions (CaptureC peaks) are enhanced.

Fig. 6: Perturbing transcription weakly affects locus conformation.
figure 6

a, CaptureC data from Pax6 HIGH cells in an experiment involving alpha amanitin (aa) treatment are shown (yellow bars) alongside data from untreated cells (green line). b, FISH data for alpha-amanitin-treated and untreated Pax6 HIGH cells. Probes were positioned over Pax6 and the promoters of the two adjacent genes Rcn1 and Elp4 (purple bars in a and probe set 2 in Fig. 1b). Box plots show the distributions of each separation (from left to right, the number of measurements in each case are n = 147, 142, 125, 131, 130, 131). Here and in later panels, box limits give the interquartile range with whiskers extending by a factor of 1.5 and the center line giving the median. A Mann–Whitney U test did not reject the null hypothesis that treated and non-treated distances were drawn from the same distribution (P > 0.05 for all three pairs of probes; from left to right, P = 0.87, P = 0.44, and P = 0.53). c, In one possible model of alpha amanitin treatment, all proteins are removed, because these could represent complexes containing RNA polymerase (see text). d, Simulated CaptureC for Pax6 HIGH cells from simulations in which all protein-binding sites were removed (lime green bars). Equivalent results without the simulated treatment are also shown (green line). e, In an alternative model, at the point of alpha amanitin treatment, all protein-chromatin bonds are fixed (see text and Supplementary Notes). f, Simulated CaptureC for Pax6 HIGH cells from simulations in which all protein-chromatin bonds were fixed at the point of treatment (yellow bars). g, Plot showing simulated FISH measurements for the two models of alpha amanitin treatment, as well as the untreated case (the number of measurements in each case is n = 7,980). h, The MSD of the Pax6 promoter FISH probe is shown for the three cases; shaded bands show the s.e.m. i, The mobility is calculated for each chromatin bead within the locus. The distributions of mobility values are shown as violin plots, with overlaid boxes showing the median and interquartile range, with whiskers extending by a factor of 1.5.

To interpret these results, we considered how transcription inhibition could be modeled within the simulations. A loss of transcription could be implemented as a loss of protein binding, because our model proteins represent complexes of polymerase and transcription factors (Fig. 6c). However, a simulation in which all binding sites were switched off showed a dramatic loss of many interaction peaks in the simulated CaptureC, and a general decrease in interactions, an effect opposite to that observed experimentally (Fig. 6d).

In HiP-HoP, the bridging complexes continually switch between a binding and a non-binding state26 to model chemical reactions (for example, post-translational modifications) and provide a realistic turnover of proteins in foci. Another possible effect of alpha amanitin is that it might interrupt such reactions, for example, in the polymerase transcription cycle, and polymerases could become stuck in an initiation state. A modified simulation was used to test this: all protein-chromatin bonds were ‘fixed,’ so that from the point of simulated alpha amanitin treatment, any protein bound to a chromatin bead remained bound (Fig. 6e and Supplementary Notes). This led to an increase in interactions at many peaks in the simulated CaptureC (Fig. 6f), a situation closer to the experimental observations than in the scenario in which proteins were removed.

Simulated FISH measurements showed different trends for each of the two models (Fig. 6g). Removing proteins led to increased probe pair separations, whereas fixing protein-chromatin bonds led to a decrease. Locus dynamics also changed for both scenarios (Fig. 6h,i): protein removal led to faster dynamics, whereas protein fixing led to slowed dynamics. Recent particle-tracking experiments in which histone proteins were labeled fluorescently have indicated that chromatin is typically more dynamic after alpha amanitin treatment39.

Since neither model fully predicts both FISH and CaptureC data, we suggest that alpha amanitin treatment actually results in a mixture of these effects: loss of some bridging (for example, mediated by complexes involving elongating polymerase) results in increased mobility and local decompaction, but some other enhancer-promoter bridging (for example, involving paused polymerases) may be stabilized, resulting in more pronounced CaptureC peaks. A proper representation of the action of alpha amanitin in simulations would require a more explicit inclusion of transcription than is currently possible within HiP-HoP.

In a complementary experiment, Pax6 HIGH cells were treated with bleomycin to perturb the local chromatin topology (DNA is nicked, releasing superhelical tension so that it becomes torsionally relaxed37). After perturbation, CaptureC profiles looked very similar to those from untreated cells, whereas FISH revealed no significant change (Extended Data Fig. 9). This suggests that at this scale of analysis, superhelical tension, which has dramatic effects on larger scale chromatin structure37, does not affect chromatin architecture within the locus.


In this work, we extended the HiP-HoP chromatin modeling scheme to study gene locus dynamics. Chromatin mobility can be determined by extracting MSDs from simulations. The observed dynamics depend strongly on the size of the chromatin segment being tracked. Chromatin neighborhoods (35–40-kbp regions) at different points within a locus, or in different cell lines, showed highly similar dynamics; by contrast, different 1-kbp regions could have very different mobilities. This was dependent on two factors: local chromatin fiber disruption (marked by H3K27ac), and bridging interactions between regulatory elements. Typically, disrupted chromatin20 was more dynamic than compact chromatin, whereas regions enriched in protein-mediated loops, often marked by ATAC-seq peaks, were less dynamic. However, active binding sites tend to be embedded within H3K27ac regions, leading to a balance between the bridging-mediated slowing and the disruption-mediated speeding up of motion; the dynamics of a segment depend on the surrounding environment as well as its own properties.

The prediction that the position and size of a probe determine the observed dynamics will be critical when designing future live-cell imaging experiments, and could explain some previous contradictory results. In live-cell imaging in which single enhancers and promoters were labeled using CRISPR–dCas9 with guide RNAs spanning 2 kbp, these became more mobile when active40. Conversely, experiments using an ANCHOR/ParB DNA-labeling system to track a promoter showed motion becoming constrained upon activation41. Several studies tracking multiple points across the genome (for example, by labeling histones) showed that dynamics tend to increase as activity is reduced (by transcription inhibition, serum starvation, or RNA polymerase II (Pol II) degradation8,32,39). These observations can be reconciled if, as our model suggests, gene activation leads to chromatin decompaction (increasing mobility) in some regions, but additional looping (decreasing mobility) in others.

The concept of promoter-enhancer interactions often leads to a notion of somewhat static gene locus configurations; by contrast, our study introduces the concept of sampling between regulatory elements, where interactions continually change over time. Our results indicate that, for a typical locus, many possible interaction events can be sampled within tens of minutes (Fig. 5), suggesting that although promoter-enhancer contacts revealed by 3C are enriched above a background, there is continuous sampling. Importantly, we found that, within a simulation equivalent to 30 min, the locus was able to explore all of its configuration space. Dynamics depend on the model parameters, but our chosen parameter set is validated by comparison to fixed-cell experiments. Varying the number of active proteins and the rate at which these switched between binding and non-binding states26 had subtle effects. Intriguingly, we found that different changes to the parameters that have the same effect on static properties do not necessarily have the same effect on dynamics. We note that configurational dynamics will also depend strongly on locus size: we would expect, for example, a 1-Mbp TAD to take longer to re-organize than the 200-kbp region studied here. This is consistent with previous modeling based on fitting to 5C data42 that suggested that multiple TAD configuration states should be dynamically accessible during a cell cycle.

We found that inhibiting transcription using alpha amanitin did not greatly alter Pax6 locus structure, leading only to small increases in CaptureC interaction peak heights. Our experiments stand in contrast with recent Hi-C, HiChIP, and OCEAN-C data showing that acute degradation of Pol II led to a small decrease in looping interactions43; another study using MicroC in mice found that transcription inhibition using triptolide or flavopiridol did not greatly affect promoter-enhancer interactions, but other features in the data were affected44. Other work showed that alpha amanitin did not disrupt enhancer interaction hubs38, and that elongation inhibition via DRD treatment did not disrupt Pol II foci45. Together, these results suggest that there is a complicated relationship between transcription and chromatin contacts and point to subtleties in the action of different inhibitors, which are not understood.

These simulations of the Pax6 locus give insight into the factors affecting chromatin dynamics and suggest that for a locus of this size, in a given time, a single cell would show the same level of variability as that observed from cell to cell across a population. We expect that a similar order of magnitude for the timescale for locus rearrangement would be found in other polymer models for chromatin (for example, in simpler models with fewer ‘mechanistic ingredients’11,14,16); however, our results suggest that differences across a locus or between loci depend on details of the model (that is, on which local chromatin features are included). In the future, it would be interesting to simulate larger regions or to include additional model ingredients that might shed light on other observations made in live cells (for example, correlated chromatin motion8,32, dependence on nuclear structural proteins like lamins46, and gel-like features of chromatin47,48,49). HiP-HoP could be extended to include, for example, repressive proteins that compact DNA, nuclear lamina interactions, or proteins such as SAF-A that are thought to form an RNA-dependent gel constraining chromatin motion47.


HiP-HoP simulations

The HiP-HoP model was used as previously described7. In brief, a chain of beads connected by springs represented a chromosome region, with 1 kbp of DNA per bead. Diffusing spheres represent complexes of proteins which bind at active chromatin sites, and these switch back and forward between a binding and a non-binding state with rate ksw = 10−3 τ−1 (unless otherwise stated), where τ is the simulation time unit. DNA accessibility data (ATAC-seq) were used to identify binding sites. Loop extrusion was modeled by introducing additional springs between adjacent beads in the chain, which were then moved at regular time intervals to extrude a loop. Extruders were initiated at random positions, moved at a rate kex = 2 bp τ−1, and were removed with rate koff = 2.5 × 10−5 τ−1. ChIP followed by microarray hybridization (ChIP–chip) data for CTCF and Rad21 binding were used to identify loop anchor sites, and extrusion was halted at loop anchor sites in a direction-dependent manner. In some regions of the bead chain, additional springs were added to ‘crumple’ it into a more compact structure. Histone modification data (H3K27ac ChIP–chip) were used to identify regions that do not have the crumpling springs. The dynamics were evolved through a Langevin dynamics scheme (implicit solvent) using the LAMMPS molecular dynamics software50. In our previous work7, the simulations matched the average diffusion properties of the model chromatin to experimental measurements, but in that case only a short region of the genome was simulated, so spatial variations in chromatin density were not accurately represented. In this work, a much larger chromatin fragment (40 Mbp) was simulated to better match the overall density found in vivo. For efficiency, 10 instances of a 3-Mbp region around Pax6 (chr2:104,000,000–107,000,000 mm9 reference genome) were placed along this 40-Mbp (40,000 bead) fragment. Periodic boundary conditions were used, and the system size was chosen to give a roughly realistic chromatin density. We used data from 3 mouse cell lines, and in a single simulation, a mixture of loci from the 3 were included among the 10 instances. In repeat simulations, the positions of the loci in different cell lines were randomized (meaning that, on average, each copy of the locus was embedded in a similar chromatin context). As noted above, to ensure that this new set up still gave good predictions of locus conformation, results were compared with both CaptureC data and FISH (Extended Data Figs. 1 and 2).

In the simulations, time is evolved in discrete steps, with a simulation ‘time unit’ τ, which can be mapped to a real time by comparison with experiments. Typically, each individual simulation was run through 1.5 million simulation time units. We initialized the fiber in a configuration resembling a mitotic chromosome; during the first part of the simulation, the system relaxes to a realistic configuration representative of the in vivo locus in interphase. We therefore discarded the first 7 × 105 τ of each simulation, after which time the dynamics had reached a steady state (see Supplementary Notes). The locus configuration varied dynamically through the remaining 800,000 τ, and we determined that extracting 400 configurations at regular time intervals provided a good representation of the locus motion over this time (see Supplementary Notes). We generated 20 trajectories (representing single cells) for each cell line, extracting a total of 8,000 configurations per cell line. Full details of the simulations, interaction potentials, and all parameters are given in the Supplementary Notes. The input data were previously published7 and are available at GEO: GSE119660, GSE119656, GSE119659, GSE119658, GSE120665, and GSE120666. Full details of the data analysis are given in Supplementary Notes.

Mapping simulation length and time scales to real units

Simulation length and time units were mapped to physical units by comparison with experimental data. To estimate the length unit, for a given pair of FISH probes, a distribution of separations was obtained from simulations and experiments7, and compared using the Kolmogorov–Smirnov statistic (the smaller this statistic the closer the two distributions). We obtained nine such distributions (three distances between Pax6, URR, and DRR probes in each of three cell lines), and used the mapping that minimized the average of the nine Kolmogorov–Smirnov statistics. This gave an estimate for the length unit of σ ≈ 17.6 nm. To map the simulation time unit, we calculated an MSD for every chromatin bead in every simulation, and obtained an average. This was compared with data from motion-tracking experiments in ref. 30, where MSDs were obtained for several chromosome regions in Saccharomyces cerevisiae. We used a linear fit to find the mapping that minimized the difference between the simulated and experimental MSD curves. This led to an estimate for the simulation time unit of τ ≈ 2.07 × 10−3 s.

Cell culture

Pax6 HIGH cells (also known as β-TC3 cells, obtained from DSMZ, cat. no. ACC-324, RRID: CVCL_0172) were isolated from a mouse insulinoma51 and were cultured in Dulbecco’s Modified Eagle Medium (DMEM) (Thermo Fisher) supplemented with 10% fetal calf serum and 1% penicillin–streptomycin at 37 °C in 5% CO2. No cell authentication was performed, and the sex of the cell line is not known.

Alpha amanitin and bleomycin treatment experiments

To block transcription, Pax6 HIGH cells were plated in a 6-cm dish and, once at 80–90% confluence, treated with 100 µM alpha amanitin (Sigma A2263-1MG) for 7 h or with mock treatment (PBS alone). Inhibition was assessed by real-time quantitative reverse transcription PCR (qRT–PCR); samples were washed with PBS, and RNA was extracted using RNeasy mini-Kit (Qiagen). All RNA samples were treated with on-column DNase1 treatment (Qiagen); the RNA concentration was corrected across the sample and reverse transcribed to complementary DNA (cDNA) using SuperScript II (Thermo Fisher) standard first-strand synthesis protocol with oligo(dT) primers (Promega), with two biological replicates per condition. Real-time qPCR on cDNA was performed using a LightCycler 480 II (Roche) and SYBR Select Master Mix (Thermo Fisher), using the standard manufacturer’s protocol. Gene-specific primers were designed for intron-exon junctions to assay nascent RNA, and the fold change against mock treatment control was calculated using the 2–∆∆CT method, normalizing against 18S as a housekeeping gene.

For bleomycin treatment, Pax6 HIGH cells at 80–90% confluency were trypsinized to a single cell suspension; 1 × 106 cells were treated with 250 µM of bleomycin (Cayman Chemical) for 15 min in PBS, centrifuged to remove the bleomycin, and washed with PBS. Samples were processed using a standard protocol for genomic DNA preparation from a Pure link genomic DNA extraction kit (Invitrogen). To assay DNA nicking, a comparison between 250 µM bleomycin and PBS treatment alone was performed on 1 µg extracted genomic DNA and run on an alkaline agarose gel under denaturing conditions, before staining with ethidium bromide and imaging, using a standard alkaline gel protocol52.

CaptureC experiments

For Pax6 HIGH cells that had been treated with alpha amanitin or bleomycin, NG CaptureC was performed as previously described28,29, but with the following alterations. Two replicates of 5 × 106 cells were processed for each case; cells were fixed with 2% formaldehyde and lysed with standard 3C lysis buffer for 15 min before being snap frozen. Cells were further lysed by re-suspension in water and then in 0.5% SDS for 10 min at 62 °C. Each replicate was split between three tubes, re-suspended in 800 µL 1× DpnII buffer (NEB) with 1.6% Triton X-100, and digested with 3 sequential additions of 750 units DpnII enzyme at 37 °C with 1,200 r.p.m. shaking over 24 h. Samples were heat inactivated at 65 °C for 20 min, and 3 samples from each replicate were combined into 7 mL with 1× T4 DNA Ligase Buffer (NEB), with 1% Triton X-100 and 12,000 units of T4 DNA ligase at 16 °C overnight. Samples were treated with Proteinase K overnight at 65 °C and RNase A/T1 (Thermo Fisher) for 1 hour at 37 °C, before a standard phenol–chloroform extraction and ethanol precipitation was performed. Complete digestion and ligation were assessed by gel electrophoresis.

Purified 3C DNA from each sample was sonicated to 200–400 bp with a Soniprep 150 probe sonicator at 4 °C and purified with a standard Ampure XP Bead protocol (Beckman Coulter) using a 1/1.5 DNA to bead ratio. Two Illumina sequencing libraries were prepared per capture pool replicate, with 6 µg of starting DNA in each, and generated using NEBNext DNA Library Prep Kit (NEB). Samples were indexed with unique barcodes using NEBNext Multiplex Oligos for Illumina (NEB). Two separate capture pools were designed to the following Pax6 locus elements, as in ref. 7 (a list of targeted restriction enzyme fragments is given in Supplementary Table 1). Capture oligonucleotides were designed for each end of the targeted DpnII fragments29, and each was synthesized in a separate 4 nM synthesis, with a 5′ biotin label on a 120-bp Ultramer (IDT). Capture oligonucleotides from each of the two pools were mixed at equimolar amounts and pooled to a final concentration of 13 pmol in a volume of 4.5 µL per sequence capture. Libraries were sized and quality controlled on a D1000 Tapestation tape (Agilent).

NG CaptureC sequence capture was performed using SeqCap EZ HE-Oligo Kit A or B (dependent on the multiplex barcode) and SeqCap EZ Accessory Kit (Nimblegen)29, using each of the two capture pools, with 1.5–2 µg 3C library DNA per hybridization reaction. Each hybridization reaction was performed on a thermocycler at 47 °C and incubated for between 66 and 72 h. Each hybridization reaction was then bound to streptavidin beads from SeqCap EZ Pure Capture Bead Kit and washed with SeqCap EZ Hybridization and Wash Kit (Nimblegen), following the manufacturer’s protocol. Hybridization reactions were split into two and libraries were re-amplified using Post LM-PCR oligonucleotides (Nimblegen) and Q5 High-Fidelity DNA polymerase (NEB) directly from the beads, and then the DNA was purified using Ampure XP Beads, with a 1/1.8 DNA to bead ratio. A second hybridization reaction was performed as above on the re-amplified 3C libraries with 2 reactions pooled together (1 µg in each) and incubated for 22–24 h. Washed and re-amplified double-captured libraries were sized and quality-controlled on a D1000 Tapestation tape (Agilent), and paired-end sequenced on an Illumina Hi-seq 2500 or Hi-seq 4000.

CaptureC data were analyzed using the capC-MAP software53 (see Supplementary Notes for further details).

Three-dimensional DNA fluorescence in situ hybridization

Cells were grown overnight on glass slides. Slides were rinsed with PBS and fixed in 4% paraformaldehyde for 10 min. Slides were rinsed with PBS and cells were permeabilized for 10 min on ice with PBS supplemented with 0.2% Triton X-100. After rinsing, slides were stored in 70% ethanol at 4 °C.

For processing, slides were dehydrated through an ethanol series and incubated with 2× SSC supplemented with 100 µg ml−1 RNase A (Invitrogen) at 37 °C for 60 min. Slides were then rinsed briefly with 2× SSC, dehydrated through an ethanol series, and air dried. Slides were warmed by incubation in a 70 °C oven for 5 min before denaturation for 1 min in 70% formamide in 2× SSC, pH 7.5, at 70 °C. Slides were then transferred to 70% ethanol on ice, dehydrated through an ethanol series, and air dried before overnight hybridization at 37 °C with pairs of fosmid probes (listed in Supplementary Table 2). Probes (BacPac resources) were labeled in green-500-dUTP (ENZO life sciences), digoxigenin-11-UTP (Roche), or biotin-16-dUTP (Roche). Then, 150 ng of each labeled probe was hybridized with 5 µg salmon sperm and 10 µg human Cot1 DNA. Slides were washed 4 times for 3 min in 2× SSC at 45 °C and 4 times for 3 min in 0.1× SSC at 60 °C before being transferred to 4× SSC with 0.1% Tween 20 at room temperature. Digoxigenin-labeled probes were detected using one layer of rhodamine-conjugated sheep anti-digoxigenin and a second layer of Texas-red-conjugated anti-sheep (Vector Laboratories). Biotin-labeled probes were detected using one layer of FITC-conjugated streptavidin followed by a layer of biotin-conjugated anti-avidin and a second layer of FITC-conjugated streptavidin (Vector Laboratories). Slides were counter-stained with 0.5 µg ml−1 DAPI.

Image capture and analysis

Three-color 3D DNA FISH slides were imaged using a Hamamatsu Orca AG CCD camera (Hamamatsu Photonics) Zeiss Axioplan II fluorescence microscope with Plan-Neofluar objectives, a 100-W Hg source (Carl Zeiss), and a Chroma 83000 triple band-pass filter with single excitation filters installed in motorized filter wheels (Prior Scientific Instruments). Image capture and analysis were done using in-house scripts written for Iola Spectrum (Scanalytics). For FISH, images were collected from at least 50 randomly selected nuclei for each experiment and then analyzed using custom Iola scripts that calculate the distance between two probe signals.

Reporting summary

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