High-resolution chromatin architecture at key thymocyte genes

To assess at higher resolution whether spatial chromatin domains may be selectively remodeled around genes that are differentially expressed on developmental transitions, we performed Capture Hi-C [43,44] on mouse CD4⁻ CD8⁻ CD44⁻ CD25⁺ (double negative; DN3) and CD4⁺ CD8⁺ (double positive; DP) thymocytes, representing cells just initiating and cells just after the beta-checkpoint for productive T cell receptor beta-chain rearrangement [45]. We used tiled capture oligonucleotides spanning DpnII fragments across 8 approximately 600 kb regions, centered on key thymocyte genes located close to TAD borders, identified from Hi-C maps generated in mouse ESCs [21]: three that are up-regulated on DN3-to-DP transition (Bcl6, Nfatc3, Rag1), three that are down-regulated (Cdh1, Il17rb, Pla2g4a), and two whose expression are relatively unchanged (Cd3, Zap70) (S1A Fig; S1 and S2 Tables). We also performed conventional Hi-C on DN3 and DP cells for genome-wide but lower-resolution assessment of any chromatin topology changes (overview of datasets in this study given in S3 Table). Capture Hi-C was additionally performed on ESCs to provide a reference point and comparison with high-resolution conventional Hi-C data [21]. As expected, the Capture Hi-C provided improved resolution at the target regions compared to previously generated Hi-C maps for thymocytes [46], even though the latter dataset had approximately 4-fold greater sequencing depth (Fig 1A). Biological replicates were highly correlated (Spearman correlation coefficient ≥0.91; S1B and S1C Fig; S4 Table) and were pooled for improved resolution in subsequent analyses. The resultant maps were of sufficient quality to resolve chromatin interactions between promoters and putative enhancers, marked with acetylation of lysine-27 on histone H3 (H3K27ac), and between CTCF-bound sites, further confirmed by 4C-seq analysis (Figs 1B–1D and S1D). Visual inspection of the Capture Hi-C maps showed that thymocytes seem to have better defined TADs and more heterogeneous structures than ESCs (Figs 1B and S2–S7), in line with observations that ESC chromatin is generally more open and “plastic” than differentiated cells [47]. Observing the contact strength decay of the Capture Hi-C datasets with increasing genomic distance reinforces these trends by showing greater variability of contact strength at a given genomic separation within thymocytes (Fig 1E). Further, the cis-decay profiles for both Capture Hi-C and, at lower resolution, the genome-wide Hi-C, show that DP chromatin is globally more compact than the other cell types, in agreement with previous observations [48]. Overall, these results validate the quality of the Capture Hi-C datasets for subsequent in-depth comparisons across cell types.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. High-resolution interrogation of chromatin architecture in thymocytes with Capture Hi-C.

(A) Maps of DN3 chromatin interactions for an approximately 600 kb region around the Bcl6 gene from pooled Hi-C [46] (approximately 1 billion reads; top) or one replicate of Capture Hi-C (approximately 250 million reads; bottom) are shown at 10 kb and 5 kb resolution, respectively. Positions of genes are denoted underneath the maps. (B) Pooled Capture Hi-C interaction maps for an approximately 600 kb region around the Rag1 gene are shown for ESCs, DN3, and DP cells at 5 kb resolution. Positions of genes are denoted underneath the maps. Arrows denote thymocyte-specific interactions. (C) 4C-seq profiles, using the Rag1 promoter as bait, performed in DN3 (blue) and DP (red) cells. Dashed lines denote position of Rag1 promoter and selected CTCF or H3K27ac peaks, which form the bases of the interactions highlighted in (B). (D) Epigenomic profiles at the same genomic location as shown in (B) and (C). Top to bottom: ChIP-seq tracks (normalized as counts per million reads) for CTCF in ESCs (green), DN3 (blue), and DP (red) cells; ChIP-seq tracks (normalized as counts per million reads) for H3K27ac in DN3 and DP cells; RNA-seq (non-strand-specific; normalized as counts per million reads) in DN3 and DP cells. (E) Plot of median normalized interaction score against genomic separation from the pooled Capture Hi-C (top) and Hi-C (bottom) datasets, at 5 kb and 10 kb resolution, respectively, for ESCs (green), DN3 (blue), and DP (red) cells. For the Capture Hi-C data, shading denotes the interquartile range to indicate the variability of these cis-decay distributions. Source data can be found in S1 Data. DN3, double negative; DP, double positive; ESC, embryonic stem cell.

https://doi.org/10.1371/journal.pbio.3002424.g001

Chromatin topology is largely conserved but with cell type–specific differences

Visual inspection of Capture Hi-C maps around the target regions suggested that spatial domain organization was largely conserved across the studied cell types, and particularly in between the 2 thymocyte populations (S2–S7 Figs), in line with lower-resolution studies [46,49]. To compare TAD organization more systematically, we computed insulation scores [50] at 5 kb resolution within the regions targeted by capture oligonucleotides. Compared to other methods that just call the positions of domains [51], insulation has the advantage of giving a quantitative border “score” for all genomic intervals. Since chromosomal domains are known to be nested and hierarchical [52,53], we computed insulation over multiple sliding window widths to alter sensitivity to domains of different size. In agreement with previous observations that TADs appear to be better defined in differentiated cells than in pluripotent cells, insulation scores are more homogeneous in ESCs, whereas more striking insulation score minima (representing strongly insulating borders) and maxima (representing the centers of well-defined folded domains) are apparent in thymocytes (Fig 2A). Over a wide range of window sizes used to compute insulation score, the more homogeneous distribution in ESCs is significantly different to those of the thymocytes (two-sided Kolgorov–Smirnov test; S5 Table). However, the genomic location of TAD border candidates is well conserved across cell types, since insulation scores are overall very well correlated (Spearman correlation coefficient between 0.71 and 0.91; Fig 2B and S5 Table). These observations are supported by analysis of the lower-resolution Hi-C data, both at the specified capture regions and genome-wide (S8A and S8B Fig and S6 Table). Thus, on first analysis, domain border location appears to be well conserved across cell types, but overall, domain strength is higher in differentiated cells than in pluripotent ESCs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. TADs are largely conserved, with cell type–specific differences.

(A) Violin plots for distributions of insulation scores computed on pooled Capture Hi-C datasets at 5 kb resolution, using a 75-kb (15 bin) window, on ESCs (green), DN3 (blue), and DP (red) cells. The more homogeneous ESC distribution is significantly different to DN3 (p = 1.4 × 10⁻⁵) and DP (p = 0.0041) cells, but the 2 thymocyte distributions are not significantly different to each other (p = 0.12; Kolmogorov–Smirnov test). (B) Scatter plots of insulation scores computed on pooled Capture Hi-C datasets at 5 kb resolution, using a 25-kb (5 bin; left) or 75-kb (15 bin; right) window, comparing DP cells with ESCs (left) or DN3 cells (right). Spearman correlation coefficients (SCC) are denoted. (C) Plot of first 2 principal components for insulation scores computed on biological replicates of Capture Hi-C datasets at 5 kb resolution, using a 25-kb (5 bin) window for ESCs (green), DN3 (blue), and DP (red) cells. Amount of variance accounted for in each principal component is denoted in the axis labels. (D-F) Pooled Capture Hi-C interaction maps for an approximately 600 kb region around the Cd3e gene are shown for (D) ESCs, (E) DN3, and (F) DP cells, above plots showing the positions and scores of called TAD boundaries and ChIP-seq profiles (normalized as counts per million reads) for CTCF in the same cells. Blue and red bars denote the position of a thymocyte-specific TAD, which is accompanied by the gain of boundary at the 5′ end, which is bound by more CTCF (denoted by asterisk). Motif orientations of CTCF sites bordering the new TAD is shown with arrows. Positions of genes are shown below the plots. (G) Venn diagram showing overlapping of called TAD boundaries for the pooled Capture Hi-C datasets of ESCs (green), DN3 (blue), and DP (red) cells. Source data can be found in S1 Data. DN3, double negative; DP, double positive; ESC, embryonic stem cell; TAD, topologically associated domain.

https://doi.org/10.1371/journal.pbio.3002424.g002

To explore further whether there are any cell type–specific differences in TAD organization, we performed principal component analysis on the insulation scores. Across all tested insulation score window sizes, biological replicates clustered closely together, and the 3 cell types were clearly distinct from each other, suggesting that cell type–specific architectural patterns are indeed present (Figs 2C and S8C). Such differences quantified by the principal component analysis could be due to gain/loss of specific borders to generate new domains, altered strengths of preexisting TADs (which has been observed for some genes at early thymocyte transitions [46]), and/or accumulated small but reproducible changes dispersed across the studied genomic regions, which individually make no major difference to the Hi-C maps. Since insulation scores from biological replicates clustered so closely together, we pooled the Capture Hi-C replicates and computed higher-resolution insulation scores over 2 kb genomic bins, finding strong local minima maintained across different sliding window sizes to call the most robust borders. We identified approximately 50 borders within the captured regions for each cell type (92 total; S7 Table). In line with the above observations, many of the strongest borders were conserved throughout development, but cell type–specific/strengthened domains were apparent, in particular when comparing thymocytes to ESCs (Figs 2D–2G and S8D). As may be expected from their developmental proximity, the 2 thymocyte populations had more borders in common (DN3-DP Jaccard index 0.57) than with ESCs (DN3-ESC Jaccard index 0.43; DP-ESC Jaccard index 0.32). Applying an analogous border calling method to the Hi-C data at lower (10 kb) resolution gave similar results genome-wide (S9 Fig and S8 Table), although the separation between thymocytes and ESCs was not as pronounced. This may be due to the focus of the Capture Hi-C around thymocyte-specific genes, which might be expected to have greater architectural differences than more general, housekeeping genes. Overall, we find a general conservation of spatial domain organization between different cell types, although higher-resolution analysis can uncover cell type–specific insulation patterns, sometimes associated with variant borders.

Pause-released polymerase may broaden spatial domain boundaries

We next further investigated the cell type–specific chromatin architectures uncovered by the Capture Hi-C approach. When comparing ESCs to thymocytes, we found cases of thymocyte-specific subdomains arising from establishment or strengthening of boundaries near the promoter of genes expressed specifically in thymocytes, both in Capture Hi-C (e.g., the Cd3 cluster; Figs 2D–2F and S5) and lower-resolution Hi-C (e.g., Runx1; S9C Fig) maps. In these described examples, the new boundaries were accompanied by increased binding of CTCF near the promoter during thymocyte development. The downstream boundary appeared to be maintained across the studied cell types, corresponding with a conserved CTCF binding site. Based on the loop extrusion model, these developmental architectural changes between ESCs and thymocytes could be explained solely by differential CTCF binding. Whether underlying gene activation could be a cause or consequence of 3D genome structure, if functionally linked at all, thus remains unclear from these cases. However, interrogating differences between DN3 and DP chromatin structures, which have very similar CTCF binding profiles, uncovered potentially more direct links between transcription and chromatin architecture. Firstly, a strong, conserved border at the Bcl6 promoter is altered in DP cells where the gene is highly expressed: An additional weak “stripe” extending the upstream spatial domain indicates the presence of a secondary broader border, adjacent to the major sharp boundary at the promoter (see rectangle and arrow in Figs 3A and S7). Notably, the domain “extension” corresponds with the Bcl6 coding region, and ChIP-seq data show that RNA polymerase II is paused at the promoter/major boundary in DN3 cells while being released to fully transcribe Bcl6 in DP cells (Fig 3B and 3C). These Capture Hi-C data are thus consistent with a model whereby bound and/or engaging RNA polymerase can act as a topological boundary itself, perhaps by stalling loop extrusion processes, as has been reported in bacteria [23] and proposed in a parallel study for mammalian cells [24]. In DN3 cells, the accumulation of RNA polymerase at the paused promoter could define a clear boundary, whereas the exact position of the elongating polymerase varies within each cell of a fixed DP population, generating a broader, more blurred boundary in the average map. While an attractive model, RNA polymerase II is absent from the inactive Bcl6 gene in ESCs, where a relatively strong but somewhat broader border is nevertheless maintained at the promoter (S7 Fig), so other mechanisms must also contribute to architectural maintenance at this locus. Conserved CTCF binding at the promoter would be expected to be such a mechanism, although homozygous deletion of the major binding site had mild effects on locus architecture in ESCs (S10 Fig), in line with other studies suggesting that TADs are built up cooperatively from multiple elements [19,54,55]. We note also that the Bcl6 gene locus contains multiple CTCF sites that are conserved in position in ESCs and thymocytes but do vary slightly in binding strength. Since clusters of CTCF are associated with “transition zones” between TADs [56], this may also influence architectural differences across cell type.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Spatial domain border width at Bcl6 gene tracks with RNA polymerase pause release.

(A) Pooled Capture Hi-C interaction maps for an approximately 600 kb region around the Bcl6 gene are shown for DN3 and DP cells, above the differential heat map, where DP-enriched interactions are displayed in red and DN3-enriched interactions in blue. The white dotted lines represent the conserved strong border, with the arrows on the Hi-C maps and black rectangle on the differential map highlighting the broadening of the domain boundary in DP cells. Positions of genes are shown below the plots. (B) Insulation scores for the same genomic region as in (A), computed at 2 kb resolution with a 100-kb (50 bin) window for DN3 (blue) and DP (red) cells. (C) ChIP-seq profiles (normalized as counts per million reads) for the same genomic region as in (A), for DN3 (blue) and DP (red) cells. Top to bottom: RNA polymerase II; CTCF; H3K27ac. Red stripe corresponds to the broadened domain boundary in DP cells, with corresponding extension of lower insulation score, RNA polymerase II binding, and H3K27ac loading into the gene body of Bcl6 indicated by an asterisk. (D) Pooled Capture Hi-C interaction maps for the same genomic region as in (A) are shown for untreated ESCs and ESCs with ectopic activation of Bcl6, above the differential heat map. The white dotted lines representing the strong thymocyte domain border, and the arrows on the Hi-C maps and black rectangle on the differential map representing the broadening of the domain boundary in DP cells, have been included for reference. Positions of genes are shown below the plots. (E) Insulation scores for the same genomic region as in (A), computed at 2 kb resolution with a 100-kb (50 bin) window for untreated (green) and Bcl6-induced (black) ESCs. (F) Bar chart showing normalized Bcl6 expression levels in untreated and CRISPRa-treated cells determined by qRT-PCR (2 biological replicates). (G) Virtual 4C plots derived from pooled Capture Hi-C datasets, using an upstream putative DP enhancer/H3K27ac peak as bait (position given by dashed line), for DN3 (blue), DP (red), untreated ESCs (green), and CRISPRa-treated ESCs (black). Red stripe corresponds to broadened domain boundary in DP cells, with corresponding increased interaction of the Bcl6 gene body with the enhancer highlighted with an asterisk. Source data can be found in S1 Data. CRISPRa, CRISPR activation; DN3, double negative; DP, double positive; ESC, embryonic stem cell.

https://doi.org/10.1371/journal.pbio.3002424.g003

To directly test the effect of transcription on spatial domain border extension, we performed CRISPRa [42] to recruit the activator VP64-p65-Rta (VPR; [57]) specifically to the Bcl6 promoter in ESCs. Despite a greater than 50-fold induction of the gene, which we estimate to elevate expression to comparable levels as DP thymocytes (S2 Table), there were negligible effects on the insulation score profile and, actually, a slightly reduced interaction of the Bcl6 gene body with upstream regions (Fig 3D–3F). Transcription of the gene therefore does not appear to directly affect topological insulation at this gene. Alternatively, distal enhancer interactions may also play a role in defining spatial chromatin domains [19], and they have previously been reported to track from the promoter into the gene body, presumably accompanying the progress of the engaged RNA polymerase [58]. In line with this, virtual 4C from a putative interacting enhancer located approximately 230 kb upstream of Bcl6 (denoted by an H3K27ac peak) shows thymocyte-specific interactions with the Bcl6 promoter, with DP-specific persistence of interaction into the downstream coding region (Fig 3G). The enhancer is inactive and noninteracting in ESCs, so it does not affect topological organization regardless of Bcl6 expression. It is worth noting that DN3 and DP cells have an identical positioning of the local insulation score minimum, at the Bcl6 promoter, with an equivalent score in both thymocyte populations. Instead, the insulation score stays consistently lower in the Bcl6 coding sequence in DP cells, which tracks well with both the DP-specific extension of RNA polymerase II and the extended region interacting with the upstream enhancer (* in Fig 3B, 3 and 3G). Such subtle insulation changes are missed from conventional TAD boundary calling approaches, so it is unclear whether the border broadening we observed at the Bcl6 locus is a common occurrence in the genome. We did not observe this in any other region interrogated by the Capture Hi-C, and the resolution of our Hi-C data was insufficient to detect such subtle changes at the Bcl6 gene or other loci in thymocytes.

Transcription can directly remodel chromatin architecture

An additional type of topological change we observed when comparing thymocytes was the establishment of subdomains encompassing single genes, specifically in the cell type(s) where the gene is more highly expressed and contains most bound RNA polymerase II. In the regions targeted by Capture Hi-C, we observed a DP-specific spatial domain at the Nfatc3 gene (Figs 4A–4C and S4) and a DN3-specific domain at the Tmem131 gene (S2 and S11A–S11C Figs), neither of which are accompanied by altered CTCF binding. To test whether transcription can be a direct driver of such chromatin architecture, we used CRISPRa to induce Nfatc3 approximately 4-fold in ESCs where the subdomain is absent, bringing expression to approximately 80% of the level measured in DP cells (S2 Table) [59]. On induction, a topological domain encompassing the Nfatc3 gene was identified in the Capture Hi-C map (Fig 4D, 4E and 4I), providing, to our knowledge, the first direct evidence that transcription can be instructive in spatial chromatin domain formation. Notably, the borders of the Nfatc3 domain are also local minima of insulation score in conditions where the gene is silent, and either developmental (DN3-to-DP transition) or ectopic (CRISPRa in ESCs) induction of the gene does not greatly change insulation scores at these “proto-boundaries.” Instead, the overall intradomain insulation score is increased, suggesting that transcriptional induction does not necessarily impede chromatin interactions between domains per se but rather reinforces intradomain contacts or compaction. As for Bcl6, these structural changes are largely overlooked by conventional insulation score analysis but can be detected by the alternative Hi-C analysis tool, CHESS (Comparison of Hi-C Experiments using Structural Similarity) [60]. This method successfully identified cell type–specific subdomains (alongside apparent false negatives) in the Capture Hi-C maps, but missed the widened border at Bcl6 (S12 Fig), and was unable to find differential subdomains in the lower-resolution Hi-C maps. For Tmem131, the insulation scores of the boundaries did differ somewhat between thymocyte stages, but the increase in intradomain contacts was much more pronounced. Trials of multiple gRNAs failed to induce Tmem131 in ESCs in CRISPRa trials, so we were unable to assess whether this domain was also directly remodeled by transcription.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Transcription can directly remodel chromatin architecture around activated genes.

(A) Pooled Capture Hi-C interaction maps for an approximately 600 kb region around the Nfatc3 gene are shown for DN3 and DP cells, above the differential heat map, where DP-enriched interactions are displayed in red and DN3-enriched interactions in blue. The arrows indicate the DP-specific domain. Positions of genes are shown below the plots. (B) Insulation scores for the same genomic region as in (A), computed at 2 kb resolution with an 80-kb (40 bin) window for DN3 (blue) and DP (red) cells. (C) ChIP-seq profiles (normalized as counts per million reads) for the same genomic region as in (A), for DN3 (blue) and DP (red) cells. Top to bottom: RNA polymerase II, CTCF. Red stripe indicates DP-specific domain, with increased insulation score maximum, increased RNA polymerase II loading, and negligible changes in CTCF binding. (D) Pooled Capture Hi-C interaction maps for the same genomic region as in (A) are shown for untreated ESCs and ESCs with ectopic activation of Nfatc3, above the differential heat map. Arrow indicates position of strengthened domain brought about by Nfatc3 induction. (E) Insulation scores for the same genomic region as in (A), computed at 2 kb resolution with an 80-kb (40 bin) window for untreated (green) and Nfatc3-induced (black) ESCs. Dark stripe indicates position of reinforced domain with corresponding increase in insulation score maximum. (F) Pooled Capture Hi-C interaction maps for an approximately 600 kb region around the Il17rb gene are shown for ESCs and DN3 cells, above the differential heat map, where DN3-enriched interactions are displayed in red and ESC-enriched interactions in blue. The arrows indicate the thymocyte-specific domain, encompassing Il17rb and Chdh. Positions of genes are shown below the plots. (G) Insulation scores for the same genomic region as in (F), computed at 2 kb resolution with an 80-kb (40 bin) window for ESCs (green), DN3 (blue) and DP (red) cells. (H) CTCF ChIP-seq profiles (normalized as counts per million reads) for the same genomic region as in (F), for ESCs (green) and DN3 cells (blue). Yellow stripe indicates thymocyte-specific domain. (I) Bar charts showing normalized Nfatc3 and Il17rb expression levels in untreated and CRISPRa-treated cells determined by qRT-PCR (2 biological replicates). (J) Pooled Capture Hi-C interaction maps for the same genomic region as in (F) are shown for untreated ESCs and ESCs with ectopic activation of Il17rb, above the differential heat map. Arrow indicates position of strengthened domain, and rectangle shows reduction in interdomain interactions brought about by Il17rb induction. (K) Insulation scores for the same genomic region as in (F), computed at 2 kb resolution with an 80-kb (40 bin) window for untreated (green) and Il17rb-induced (black) ESCs. Yellow stripe indicates position of new domain, containing new insulation score minima flanking an insulation score local maximum. Source data can be found in S1 Data. CRISPRa, CRISPR activation; DN3, double negative; DP, double positive; ESC, embryonic stem cell.

https://doi.org/10.1371/journal.pbio.3002424.g004

Another gene within our own capture strategy, Il17rb, may have also been expected to comprise a DN3-specific domain: a cell type–specific transcribed region flanked by local insulation score minima (Figs 4G and S6). Despite only negligible local interaction differences between DN3 and DP at this locus, closer inspection revealed that the reduced expression of Il17rb in DP cells is nonetheless approximately 40-fold higher than in ESCs (S2 Table). Differential Capture Hi-C heat maps revealed the presence of a thymocyte-specific contact domain, comprising Il17rb and the neighboring gene Chdh, again correlating with increased transcription of both genes (Fig 4F–4H). We used CRISPRa to induce Il17rb to comparable total expression levels as was obtained for Nfatc3, but this time representing an approximately 5,000-fold induction of the gene compared to almost completely silent wild-type ESCs, which we estimate to be approximately 40-fold greater than the level in wild-type DN3 thymocytes (and approximately 130-fold greater than wild-type DP thymocytes; Fig 4I and S2 Table). As may be expected, this overexpression had minimal effects on chromatin architecture around the neighboring Chdh gene but remodeled a new spatial domain at and just downstream of the induced Il17rb gene (Fig 4J and 4K), providing a second demonstration that transcription can directly remodel chromatin architecture. Compared to Nfatc3, which is an approximately 6-fold longer gene, the induced Il17rb domain was less apparent on the differential contact map, but the insulation score profile demonstrated a clear local maximum at the center of the novel domain, indicative of a similar intradomain compaction/interaction increase. Unlike on Nfatc3 induction, where local insulation minima already demarcated the domain in wild-type ESCs, new domain boundaries were generated as shown by local insulation minima. The border downstream of Il17rb is at the same location as that found in thymocytes, and which is missing from wild-type ESCs. More strikingly, the border induced just upstream of the Il17rb promoter, also clearly visible as a “stripe” of reduced interdomain interactions (rectangle in Fig 4J), is unique to induced ESCs. The spatial domain extends further 3′ in thymocytes, presumably due to incorporation of the highly expressed Chdh gene, and is not apparent at all in wild-type ESCs.

Overall, these findings are reminiscent of previous observations of triptolide-sensitive single-gene domains in Drosophila [37] and of a correlation between RNA polymerase occupancy and intragene contacts in ESCs [35] in ultrahigh-resolution Hi-C or Micro-C maps, although neither study had assessed whether domains were directly formed as a consequence of transcription. Analysis of the thymocyte Hi-C datasets, restricted to sufficiently long genes to accommodate for the resolution limit, also showed an overall positive correlation between intragene contact strength and RNA polymerase occupancy, as well as other epigenetic marks of gene activity, such as H3K27ac and monomethylation of lysine-4 on histone H3 (H3K4me1) (Fig 5A). The RNA polymerase correlation was quantitatively very similar to what was observed in ESCs (Spearman correlation coefficient of 0.60 (35), compared to DN3 (0.62) and DP (0.61) in this study), and presence of the repressive histone modification, trimethylation of lysine-27 on histone H3 (H3K27me3), weakly anticorrelated with intragene contact strength. When clustering genes into 4 groups based on transcriptional output, metagene analysis further confirms the gradual increase of intragene compaction by augmenting gene expression (Fig 5B). Additionally, the most highly expressed thymocyte genes also demonstrated stronger interactions between transcription start sites and transcription termination sites in metagene analysis (rightmost column in Fig 5B), in line with previous suggestions of active gene looping events [39], although the resolution is insufficient to distinguish between point-to-point interactions between gene termini and general compaction of the entire gene body to a more homogeneous domain. These and the previously reported correlations apply to comparisons of gene sets within the same cell type. An advantage of our experimental setup is that we can also compare architectures for the same genes in different, but developmentally very close, cell types. When comparing intragene compaction changes between DN3 and DP cells, we also find a significant but weaker positive correlation with changes in RNA output, polymerase occupancy, and active histone modifications (Spearman correlation coefficient 0.21 to 0.30; p-value < 1 × 10⁻²⁴) (Fig 5C). This suggests that, globally, gene up-regulation can remodel local spatial chromatin domains but is not necessarily a universal occurrence, with domains having different sensitivities. This is mirrored in our Capture Hi-C results when comparing spatial chromatin domains at Il17rb and Tmem131. Even though both genes form induced spatial domains, the apparent approximately 4-fold reduction in Il17rb expression on DN3-to-DP transition is insufficient to significantly disrupt the domain, whereas a smaller (approximately 2.5-fold) reduction in Tmem131 expression causes major loss of architecture. Overall, our results strongly support that direct contact domain remodeling by transcription can occur, but with an apparent context dependence, which is hidden from previous global analyses.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Genome-wide intragene domain contacts correlate with hallmarks of transcription and upregulation.

(A) Scatter plots correlating intragene contacts as measured by Hi-C in DN3 (top) and DP (bottom) cells with median ChIP-seq signal on the gene for marks (from left to right: RNA polymerase II, H3K27ac, H3K4me1, H3K27me3). Spearman correlation coefficients are denoted on the plots. (B) Pileup Hi-C maps at 10 kb resolution showing the cumulative interaction between TSSs and TTSs in DN3 (top) and DP (bottom) cells with genes stratified according to expression levels, from the lowest quartile (Q1; left) to the highest quartile (Q4; right). (C) Scatter plots for the ratios of intragene contact score relative to the ratio of median ChIP-seq gene signals between DN3 and DP cells. Spearman correlation coefficients and p-values are denoted on the graphs. Source data can be found in S1 Data. DN3, double negative; DP, double positive; TSS, transcription start site; TTS, transcription termination site.

https://doi.org/10.1371/journal.pbio.3002424.g005

Experimental design

Capture Hi-C and Hi-C experiments were performed in biological duplicates on purified mouse thymocytes or cultured ESCs before and after manipulations such as CRISPRa for specific genes or homozygous deletion of candidate CTCF binding sites. Detailed below, biological replicates were first assessed for reproducibility, before pooling the data for higher-resolution analysis. TAD analyses via computation of insulation scores was performed over multiple window sizes to ensure robustness of subsequent findings. 4C-seq experiments were performed as single experiments. qRT-PCR verification of CRISPRa gene induction was performed in 2 biological replicates, once for each induction experiment.

Thymocyte isolation

Thymocytes were obtained from 4-week-old c57/Bl6 mouse (Charles River; equal numbers of males and females) thymus by FACS. Thymuses were dissected, gently homogenized, and filtered through a 70-μm strainer in cold PBE (0.5% BSA and 2 mM EDTA in 1X PBS). To purify DP cells, approximately 600 million cells were incubated with anti-mouse CD4-PE, anti-mouse CD8a-FITC, and anti-mouse CD3e-APC antibodies (eBioScience), before washing with cold PBE, adding DAPI, and purifying viable (DAPI⁻) PE⁺ FITC⁺ APC⁺ cells by FACS. To purify DN3 cells, approximately 4 billion cells were first depleted of DP cells by incubating with rat anti-CD4 and rat anti-CD8 anti-sera (kind gift of Susan Chan), washing with cold PBE, and binding the DP cells to Dynabeads-sheep anti-rat IgG (Invitrogen). The supernatant was centrifuged to harvest the cells, which were then incubated with anti-mouse CD4-FITC, anti-mouse CD8a-FITC, anti-mouse CD3e-FITC, anti-mouse B220-FITC, anti-mouse CD11b-FITC, anti-mouse Ly-6G(Gr-1)-FITC, anti-mouse NK1.1-FITC, anti-mouse CD44-APC, and anti-mouse CD25-PE antibodies (eBioScience) and washed with PBE. DAPI was added and viable PE⁺APC⁻FITC⁻ cells were purified by FACS.

ESC culture

Mus musculus J1 ESCs (ATCC SCRC-1010) were grown on gamma-irradiated mouse embryonic fibroblast cells under standard conditions (4.5 g/L glucose-DMEN, 15% FCS, 0.1 mM nonessential amino acids, 0.1 mM beta-mercaptoethanol, 1 mM glutamine, 500 U/mL LIF, gentamicin) and then passaged onto feeder-free 0.2% gelatin-coated plates for at least 2 passages to remove feeder cells before treatments.

CRISPRa

Twenty million cell batches of J1 ESCs were transfected each with 20 μg dCas9-VPR vector (Addgene #63798) and 20 μg plasmid constructed by the IGBMC molecular biology platform (available on request), containing optimized scaffold for 4 gRNAs alongside a puromycin resistance gene and mCherry reporter (gRNA sequences are given in S9 Table), with Lipofectamine 2000 according to the manufacturer’s instructions. After 24 h, transfected cells were treated with 3 μg/mL puromycin and 1 mg/mL G418, and after another 24 h, mCherry-positive cells were sorted by FACS before immediate processing.

Deletion of major Bcl6 CTCF site

One million J1 ESCs were transfected with 1 μg of a plasmid constructed by the IGBMC molecular biology platform (available on request), containing optimized scaffold for 2 gRNAs alongside a Cas9-HF-EGFP construct and a puromycin resistance gene, with Lipofectamine 2000 according to the manufacturer’s instructions. After 3 days, transfected cells were treated with 3 μg/mL puromycin for 24 h, then with 1 μg/mL puromycin for a further 24 h, before sorting individual GFP-positive cells into 96-well plates, pre-prepared with feeder cells, with FACS. Colonies were amplified and clones with homozygous deletions were identified by genomic DNA extraction and PCR screening with different primer pairs flanking and inside the expected deletion region. Sequences of the deletion were confirmed by TA-cloning and sequencing the PCR products. This ESC line is available on request.

RNA isolation and gene expression analysis by RT-qPCR

Total RNA was extracted using the Nucleospin RNA kit (Macherey Nagel), and cDNA was prepared using SuperScript IV (Invitrogen) with random hexamers according to the manufacturer’s instructions. Quantitative PCR was performed in a LightCycler 480 (Roche) using SYBR Green I master mix according to the manufacturer’s instructions. Expression was normalized to Gapdh (primer sequences are given in S10 Table).

Capture oligonucleotide design

The mouse genome (mm10) was digested in silico with DpnII, and fragments were filtered to include those ≥140 bp and with GC content between 20% and 80%. To assess target mappability, the genome was split in silico into 50 bp fragments and remapped using Bowtie [67] with a filter to only include uniquely mapping fragments. The DpnII fragments were further filtered to only include those where ≥80% of the sequence is covered by this uniquely mapping reference point. Fragments within 600 kb of the target genes (S1 Table) were retained, and the 120 nt directly adjacent to the DpnII sites were used for capture oligonucleotides (full capture design is given in S11 Table), synthesized as a SureSelect library (Agilent).

Capture Hi-C

Capture Hi-C was essentially performed as in [68]. Five million aliquots of cells were fixed in 2% formaldehyde for 10 min, quenched with 125 mM cold glycine, then collected by centrifugation and washed with PBS. Cells were lysed in lysis buffer (10 mM Tris-HCl (pH 8), 100 mM NaCl, 0.2% NP-40, protease inhibitor cocktail) on ice for 30 min and nuclei collected by centrifugation and resuspended in DpnII restriction buffer (NEBuffer 2 for HindIII Hi-C). ESCs were permeabilized by treatment with 0.4% SDS for 1 h at 37°C, followed by 1.6% Triton X-100 for 1 h at 37°C. Thymocytes were permeabilized by treatment with 0.8% SDS for 20 min at 65°C, followed by 40 min at 37°C, then with 3.3% Triton X-100 for 1 h at 37°C. Nuclei aliquots were digested overnight at 37°C with 1,500 U DpnII (2,000 U HindIII for HindIII Hi-C), then cohesive ends were filled in with biotin tags by incubating for 90 min at 37°C with 15 μM dATP, 15 μM dTTP, 15 μM dGTP, 15 μM biotin-14-dCTP, and 25 U DNA polymerase I Klenow fragment. Nuclei were collected by centrifugation, and in situ ligation was performed overnight at 16°C in 500 μL volumes of 1X ligase buffer (NEB) with 20,000 U T4 DNA ligase, before de-crosslinking overnight at 65°C with 0.75 mg/mL proteinase K. DNA was purified by RNase A treatment, phenol/chloroform extraction and isopropanol precipitation. Around 5 μg aliquots of DNA were sonicated to a fragment size of approximately 100 to 400 bp in a Covaris sonicator E220 and bound to Steptavidin MyOne T1 beads, following the manufacturer’s instructions. The DNA was then end-repaired by incubating for 1 h at 20°C with 0.8 mM dNTPs, 0.3 U/μL T4 DNA polymerase, 1 U/μL T4 polynucleotide kinase, and 0.1 U/μL DNA polymerase I Klenow fragment in ligase buffer, then A-tailed by incubating for 1 h at 37°C with 200 μM dATP and 0.2 U/μL Klenow fragment (3′-5′ exo⁻) in NEBuffer 2 (NEB), then Illumina PE adapter was added by incubating overnight at 20°C with 15 μM PE adapter and 2,000 U T4 DNA ligase in ligase buffer, washing the beads with appropriate buffers in between each treatment. Hi-C material was then amplified from the beads with 7 to 9 cycles of PCR using Herculase II Fusion DNA polymerase (Agilent) according to the manufacturer’s instructions. Material at this stage was quantified on a BioAnalyzer (Agilent), and DpnII and HindIII Hi-C material was sequenced (2 × 50 nt) on a HiSeq 4000 (Illumina) following the manufacturer’s instructions). For Capture Hi-C, 750 ng DpnII Hi-C material was concentrated in a SpeedVac vacuum concentrator and then target capture was performed using the SureSelect XT Target Enrichment system (Agilent) with the custom-designed library (S11 Table), following the manufacturer’s instructions, and the library was then sequenced (2 × 50 nt) on a HiSeq 4000 (Illumina).

4C-seq

3C was performed on ESCs and thymocytes essentially as for the first steps of the Hi-C described above (up to DNA purification before sonication), except that the biotin fill-in step was omitted. The DNA was digested with 5 U/μg Csp6I at 37°C overnight, then repurified by phenol/chloroform extraction and isopropanol precipitation. The DNA was then circularized by ligation with 200 U/μg T4 DNA ligase under dilute conditions (3 ng/μL DNA) and purified by phenol/chloroform extraction and isopropanol precipitation, before PCR amplification with primers containing Illumina adapter sequence (sequences given in S12 Table). Excess primers were removed with SPRI beads, and the PCR products were assessed on a BioAnalyzer (Agilent), then sequenced (1 × 50 nt) with a HiSeq4000 (Illumina).

Hi-C analysis

For convenience, Hi-C data were processed using the FAN-C suite of tools [69], entailing read mapping to the mm10 genome assembly, mating pairs, attribution to restriction fragments with filtering of self-ligated fragments and PCR duplicates, matrix balancing, and computing insulation scores. For compatibility with Capture Hi-C plots, the data for the relevant regions were extracted using the dump tool and visualized as for the Capture Hi-C matrices.

Capture Hi-C preprocessing

To ensure that the captured regions were treated independently of the noncaptured regions, the datasets were processed with custom scripts, as entailed in [70]. The initial steps are essentially the same as FAN-C, entailing read mapping with Bowtie [67] to the mm10 genome assembly, mating pairs, and attribution to restriction fragments with filtering of self-ligated fragments and PCR duplicates. Custom perl scripts filter the data to include only those where both ends of a paired read correspond to regions targeted by capture oligonucleotides and to bin these submatrices to fixed genomic bins. Each captured subregion was treated separately for matrix balancing with the Knight–Ruiz method [11]; R code kindly shared by Aleksandra Pękowska [71]. All plots in this study visualizing (Capture)-HiC matrices alongside ChIP-seq tracks and insulation scores were generated by custom R scripts.

4C-seq analysis

4C-seq data were mapped to the mm10, processed, normalized, and visualized using 4See [72], as in [19].

(Capture) Hi-C analysis

The cis-decay plot for Fig 1E was obtained from 10 kb Hi-C data with the expected function of FAN-C. For the 5-kb Capture Hi-C data, the cis-decay plot was derived directly from the balanced submatrices at the captured regions, obtaining the distribution of normalized interaction scores at different genomic separations, and plotting the median values ± the interquartile range.

Insulation scores were essentially computed as in [50], determined by counting interactions along sliding windows off the Hi-C diagonal. Lower counts/weaker interactions imply greater topological insulation and, thus, a greater likelihood of the existence of a TAD border. This has the additional advantage of giving a TAD border “score” for all genomic intervals, rather than simply calling TADs using a more complex set of (often arbitrary) parameters. The only tunable parameter required in computing insulation is the sliding window size, altering the sensitivity to TADs of different sizes within the folded hierarchy [53]. Local insulation score minima are candidate TAD boundaries, and their boundary scores are computed as the difference in insulation score between the minima at the boundary and the adjacent bins (the delta in [50]), with higher scores indicating sharper, stronger TAD borders. For 10 kb Hi-C data, insulation score was computed with the insulation tool of FAN-C, using sliding windows of 70, 100, and 150 kb (7, 10, and 15 bins). Local minima were subsequently identified with the FAN-C tool boundaries, using a window of 3 bins for computing the boundary score and with no initial threshold filter for minimal boundary score. For Capture Hi-C, insulation score was computed with custom R scripts on the balanced submatrices at 5 kb (for individual replicates) and 2 kb (for pooled data) resolutions, using sliding windows of 15, 25, 35, 50, and 75 kb (3, 5, 7, 10, and 15 bins), and 30, 40, 50, 80, and 100 kb (15, 20, 25, 40, and 50 bins), respectively. Windows of 3 bins were used for computing boundary scores for each of these insulation profiles. To identify the most robust TAD boundaries (e.g., as shown in Fig 2D), insulation score minima with a boundary score ≥0.1 in at least 2 different window sizes were maintained. Adjacent boundaries were merged and the overall boundary score for each boundary was calculated as the mean of the component scores, which were ≥0.1. These final called borders were intersected using the GenomicRanges R package to derive the Venn diagrams (e.g., Fig 2G). Distributions of insulation score were plotted as violin plots (e.g., Fig 2A) using the ggplot2 package of R and were compared across cell type by two-tailed Kolgorov–Smirnov tests.

Biological replicates were compared by Spearman correlation coefficients of normalized interaction scores (for Capture Hi-C) and insulation scores (for Hi-C and Capture Hi-C). Principal component analysis was performed on the biological replicates of Capture Hi-C insulation scores using the prcomp function of R.

To derive the differential interaction maps (e.g., Fig 3A) the normalized scores of one matrix were subtracted from the other, and the difference was expressed as a z-score:

Virtual 4C plots were derived as one row from the relevant Capture Hi-C normalized interaction score submatrix, applying smoothing with a running mean over 3 bins.

CHESS [60] was applied to each captured region independently after first converting the normalized submatrices to FAN-C objects. The sim function of CHESS was applied on sliding 40-bin (200 kb) windows, with steps of 1 bin (5 kb). The results of the submatrices were pooled, and regions with structural differences were filtered as having a signal-to-noise threshold >0.5 and a z-score for ssim, computed for the grouped dataset, <−1, as per recommendations in the original article [60].

Metagene analysis

For all mouse genes of size ≥100 kb (≥10 bins in Hi-C), the intragenic interaction for each gene in each thymocyte population was computed as the median observed/expected score (defined as the Hi-C interaction normalized by the median Hi-C interaction level at the same genomic separation computed over the whole genome; see cis-decay curve in Fig 1E) for all intragenic pairwise bin-to-bin combinations. For the same gene set, the median ChIP signal within the gene body (taken over 10 kb bins) was also calculated and used to generate the scatter plots in Fig 5A. Fig 5B was made like in [35,37]. Genes ≥100 kb were grouped into 4 quartiles based on their expression levels, and, for each group, the median heatmap of observed/expected interaction scores is computed, by scaling genes to a pseudo-size of 10 bins to align all the transcription start and termination sites for the intragene part of heatmap and by considering extensions of 400 kb (40 bins) beyond the genes (upstream and downstream) for the rest. In Fig 5C, for each gene, the log2 ratio of the intragenic interaction score in DP and DN3 is plotted versus the corresponding log2 ratio of median ChIP-seq signal within the gene body.

Epigenomic datasets

Publicly available thymocyte and ESC RNA-seq and ChIP-seq data were obtained from the Genome Expression Omnibus (see S13 Table). Fastq files were mapped to mm10 with Bowtie2, and then normalized BigWig tracks (counts per million reads) were generated with DeepTools [73]. The tracks were plotted in R using the same y-axis scale when comparing across tissues.