Polycomb-mediated chromatin loops revealed by a subkilobase-resolution chromatin interaction map

Significance

Chromatin loops, detected by chemical cross-linking and DNA sequencing, are frequently bounded by the polycomb repressive complex 1 in Drosophila. The loops are associated with important developmental genes, often in a repressed state. These results are in contrast with previous studies on mammalian cells, in which chromatin loops are commonly bounded by CTCF protein, and with the generally accepted role of looping in gene activation.

Abstract

The locations of chromatin loops in Drosophila were determined by Hi-C (chemical cross-linking, restriction digestion, ligation, and high-throughput DNA sequencing). Whereas most loop boundaries or “anchors” are associated with CTCF protein in mammals, loop anchors in Drosophila were found most often in association with the polycomb group (PcG) protein Polycomb (Pc), a subunit of polycomb repressive complex 1 (PRC1). Loops were frequently located within domains of PcG-repressed chromatin. Promoters located at PRC1 loop anchors regulate some of the most important developmental genes and are less likely to be expressed than those not at PRC1 loop anchors. Although DNA looping has most commonly been associated with enhancer–promoter communication, our results indicate that loops are also associated with gene repression.

Active and inactive genes are folded differently and located in different regions of the nucleus, but the molecular basis of chromosome folding and nuclear architecture remains to be determined. One folding paradigm is the formation of protein-mediated DNA loops, which are most commonly associated with enhancer–promoter communication (1). On the other hand, gene repression is most often associated with heterochromatin formation and chromosome condensation.

Chromosome folding can be revealed by chemical cross-linking, followed by restriction digestion, ligation, and high-throughput DNA sequencing (2). Such “Hi-C” analysis has revealed intrachromosomal folding on multiple length scales. From lengths of 1 kb, the smallest so far examined (3), to hundreds of kilobases, two features are observed, loops (3) and so-called topologically associating domains, or “TADs” (which we and others have also referred to as A/B domains, physical domains, topological domains, or contact domains) (3–7). Loops bring a pair of loci into close physical proximity; TADs represent genomic intervals in which all pairs of loci exhibit an enhanced frequency of contact, and correspond to stably condensed chromosomal regions (8). On a larger scale, extending to whole chromosomes, TADs interact with one another to form “compartments” (2).

Hi-C, on the smallest length scale, based on the most extensive sequencing, is needed for unambiguous identification of chromatin loops and has been reported thus far only for the mouse and human genomes (3). We have now overcome this limitation in Drosophila by Hi-C analysis at subkilobase resolution. We find an unanticipated correlation with results of ChIP-seq analysis in Drosophila, with an important functional correlate.

Results

Subkilobase-Resolution Drosophila Hi-C.

We performed a variant of Hi-C with improved signal-to-noise ratio (9), termed tethered conformation capture (TCC; hereafter referred to as Hi-C), on embryonic Kc167 (Kc) Drosophila melanogaster cultured cells. We identified 529 million chromosomal contacts (read pairs that remain after exclusion of duplicates, unligated fragments, and reads that align poorly with genome sequences). The resolution of the resulting contact maps was 260 bp or “subkilobase,” based on standard metrics (3) (SI Appendix, SI Materials and Methods, Fig. S1, and Table S1). For this reason, and because 77.1% of the restriction fragments used for the Hi-C analysis were shorter than 500 bp, we plotted the contact maps in units of at least 500 bp. The contact maps were three- to fourfold more dense than those previously reported for human cells at kilobase resolution (3).

Consistent with previously reported Drosophila Hi-C contact maps (4, 8, 10), we observed TADs (apparent in a Hi-C contact map as boxes of enriched contact frequency tiling the diagonal; Fig. 1 A and B and Dataset S1) and noted the presence of compartments (apparent as off-diagonal boxes of alternating enriched or depleted contacts, whose boundaries are aligned with those of the on-diagonal boxes; Fig. 1 A and B). We identified more TADs in Kc cells than were originally reported (10) (2,126 versus 1,110; Dataset S1) because more TAD boundaries could be detected at higher resolution (3). We also identified chromatin loops, which appeared as focal peaks of contact enrichment (Fig. 1C and Dataset S2), and which were previously unobservable in lower-resolution Drosophila maps.

Fig. 1.

Subkilobase-resolution Hi-C identifies Drosophila chromatin loops. (A) Intrachromosome 3R Hi-C contact map at 25-kb resolution shows off-diagonal boxes of alternating enriched or depleted contacts indicative of compartmentation. (B) Hi-C contact map at 500-bp resolution of a region of chromosome 3L shows TADs (yellow outlines) as boxes of enriched contact frequency tiling the diagonal. On-diagonal boxes (TADs) align with off-diagonal boxes indicative of compartmentation. (C) Hi-C contact maps at 500-bp resolution reveal the presence of chromatin loops, identified as focal peaks of contact enrichment (cyan circles). TADs are indicated by yellow outlines, and the corner of the TAD closest to a loop is indicated by a black arrowhead. (D) Cumulative distributions of the 2D Euclidean distance between a loop and the closest TAD corner for Drosophila loops (red), human loops (blue), and the mean of 10,000 shuffled sets of Drosophila loops (black), with 95% confidence interval (gray shading). Drosophila loops are farther from TAD corners than human loops [P = 3.79 × 10⁻¹¹; two-sided Kolmogorov–Smirnov (KS) test], but closer to TAD corners than the shuffled control (P = 5.89 × 10⁻⁹; two-sided KS test). Distance is normalized for each loop by the size of the closest TAD in the respective species because TAD sizes differ between flies and humans.

Drosophila Loops Are Unrelated to TADs.

We identified 120 chromatin loops, far fewer than the number of TADs and also far fewer than the 9,448 loops identified in human GM12878 B-lymphoblastoid cells (3), even after accounting for the difference in genome size.

Drosophila loops differed from mammalian loops in their relationship to TADs (Fig. 1D). Mammalian loop anchors are frequently (38% of loops) located at TAD corners and therefore coincide with TAD boundaries. In contrast, the vast majority (82.5%) of Drosophila loops did not appear at TAD corners (Fig. 1 C and D), and, conversely, 99.1% of Drosophila TADs did not have focal peaks at their corners. Evidently, chromatin condensation revealed by TADs (8) is not due in Drosophila to loops detectable by our methods. A close relationship between TADs and loops detectable by Hi-C is not conserved throughout metazoans.

Lack of CTCF at Drosophila Loop Anchors.

In humans, 86% of loop anchors are associated with CTCF, and the CTCF-binding motifs are in a convergent orientation (3, 11). In contrast, Drosophila loop anchors did not tend to align with CTCF ChIP-seq signals (12) (Fig. 2 A and B and SI Appendix, Fig. S2), and only 28.2% of loop anchors overlapped CTCF-binding sites genome-wide (Fig. 2C). Loop anchors were less likely to occur at the strongest CTCF ChIP-seq peaks than at the weakest peaks (Fig. 2C, Left). Conversely, the strongest CTCF ChIP-seq peaks were less likely to occur at loop anchors than at the weakest peaks (Fig. 2C, Right). At a minimum ChIP enrichment of 16-fold, loop anchors are 9.55-fold enriched at CTCF sites, but this accounts for only 13.2% of loop anchors and 1.66% of CTCF sites. Evidently, CTCF is rarely, if ever, significantly associated with chromatin loops, and the vast majority of CTCF sites are unrelated to the chromatin loops we identify in Drosophila. Similar results were obtained for the other Drosophila insulator proteins, BEAF-32, Su(Hw), and CP190 (SI Appendix, Figs. S2 and S3). In contrast to the results for CTCF and in keeping with reports for human loop anchors, the majority (72.8%) of Drosophila loop anchors were associated with the cohesin subunit, Rad21. Moreover, stronger Rad21 ChIP peaks were more likely to overlap with loop anchors (Fig. 2 A, B, and D and SI Appendix, Fig. S2).

Fig. 2.

Drosophila loop anchors are bound by PRC1. (A) Hi-C contact map at 500-bp resolution of a region of chromosome X shows chromatin loops (cyan circles) that align with Pc ChIP-seq peaks. CTCF, histone H3K27me3, Rad21, and Pc ChIP-seq profiles are aligned above the map. TADs are indicated by yellow outlines. (B) Hi-C contact map at 500-bp resolution of a region of chromosome 2L shows a small network of chromatin loops (cyan circles) that align with Pc ChIP-seq peaks. CTCF, histone H3K27me3, Rad21, and Pc ChIP-seq profiles are aligned above the map. TADs are indicated by yellow outlines. (C, Left) Fold enrichment of loop anchors at CTCF ChIP peaks (red) and percentage of loop anchors bound by CTCF (blue) at the CTCF minimum ChIP enrichment indicated on the abscissa. (Right) Fold enrichment of CTCF ChIP peaks at loop anchors (red) and percentage of CTCF ChIP peaks at loop anchors (blue) at the CTCF minimum ChIP enrichment indicated on the abscissa. Red shading indicates 95% confidence interval. (D) Same as C except for Rad21. (E, Left) Z-score of the enrichment/depletion of entire loops within five functional classes (colors) of chromatin relative to 10,000 random shuffle controls shows that entire loops are strongly enriched in PcG-repressed chromatin. (Right) Z-score of enrichment/depletion of entire loop anchors within five functional classes (colors) of chromatin relative to 10,000 random shuffle controls shows that entire loop anchors are strongly enriched in PcG-repressed chromatin. Enrichment is given by a positive Z-score, and depletion is given by a negative Z-score. (F) Same as C except for Pc.

Loops Are Found Within Polycomb-Repressed Chromatin.

For a more comprehensive assessment of the relationship between loops, loop anchors, and chromatin-bound proteins, we compared our loop annotation to the five functional classes of chromatin that were previously identified on the basis of genome-wide nonhistone protein localization in Drosophila Kc cells: PcG-repressed chromatin, HP1 heterochromatin, another type of repressed chromatin (“null/inactive”), and two types of active chromatin (“H3K36me3-enriched or -depleted”) (13). We found that 17.5% of loops were entirely contained within a region of PcG-repressed chromatin, and such enrichment of entire loops within PcG-repressed chromatin was far greater than expected on a random basis (10.9-fold, P = 1.82 × 10⁻³⁵) (Fig. 2E). Few loops were entirely located within H3K36me3-enriched or -depleted active chromatin (no loops) or HP1 heterochromatin (0.8% of loops), and loops were neither significantly enriched nor depleted from these classes (Fig. 2E). Loops were depleted from null/inactive chromatin (2.5% of loops, a 5.1-fold depletion, P = 5.07 × 10⁻⁴). The low percentage of loops entirely located within a single functional chromatin class indicates that the majority of loops must span more than one class.

Many loop anchors (39.6%) were located within PcG-repressed chromatin (larger than the 17.5% of loops contained entirely within PcG-repressed chromatin because sometimes one anchor of a loop was located within PcG-repressed chromatin but not the other anchor). Enrichment of loop anchors within PcG-repressed chromatin was greater than expected on a random basis (2.35-fold, P = 1.79 × 10⁻¹⁶) (Fig. 2E). A smaller fraction of loop anchors was found in H3K36me3-depleted active chromatin (18.8% of anchors, 2.76-fold enriched, P = 6.01 × 10⁻¹⁰) (Fig. 2E). Loop anchors were significantly depleted from HP1 heterochromatin, null/inactive chromatin, and H3K36me3-enriched active chromatin (HP1: 0.1% of anchors, 3.32-fold depletion, P = 0.0436; null/inactive: 17.8%, 2.47-fold depletion, P = 1.19 × 10⁻¹⁴; H3K36me3-enriched active: 8.42%, 1.82-fold depletion, P = 7.07 × 10⁻³) (Fig. 2E).

Polycomb Protein Is Found at Loop Anchors.

PcG proteins are involved in gene repression during development and are components of two evolutionarily conserved complexes, polycomb repressive complex 1 (PRC1) and PRC2. PRC2 trimethylates histone H3 at lysine 27, which demarcates PcG-repressed chromatin and is bound by the chromodomain of the Polycomb (Pc) subunit of PRC1 (14). Loops were readily identified within H3K27me3-enriched regions (15) (Figs. 2 A and B and 3 A and B), and loop anchors tended to coincide with Pc ChIP-seq signals (16) from Kc cells (Figs. 2 A and B and 3 A and B and SI Appendix, Fig. S2). Occasionally, loop anchors that aligned with Pc ChIP signals formed networks of interactions (Fig. 2B). The enrichment of loop anchors at Pc ChIP peaks increased with the strength of the Pc ChIP peak, and enrichment of Pc ChIP peaks at loop anchors continually increased with the strength of the Pc ChIP peak as well (Fig. 2F). The percentage of Pc ChIP peaks at loop anchors increased 18.8-fold from 2.1 to 39.5% as the strength of the Pc peak increased from zerofold to 100-fold minimum ChIP enrichment. This is in contrast with CTCF, BEAF-32, Su(Hw), and CP190, none of which exhibited a monotonic relationship between peak strength and the frequency of overlap with loop anchors (Fig. 2C and SI Appendix, Fig. S3). To ensure that these results reflected the presence of Pc, rather than off-target binding by the antibody that was used for ChIP, we analyzed ChIP-seq data from a second antibody against Pc (16) and obtained essentially the same results (SI Appendix, Figs. S2 and S4).

Fig. 3.

Developmentally regulated promoters are found at PRC1 loop anchors. (A) Hi-C contact map at 1-kb resolution shows chromatin loops (cyan circles) at the ANT-C Hox gene complex. RNA-seq, histone H3K27me3, Rad21, and Pc ChIP-Seq profiles are aligned above the map. TADs are indicated by yellow outlines. (B) Hi-C contact map at 500-bp resolution shows chromatin loops (cyan circles) at the inv and en promoters. RNA-seq, histone H3K27me3, Rad21, and Pc ChIP-Seq profiles are aligned above the map. TADs are indicated by yellow outlines. (C) GO term P value chart indicates that PRC1 loop anchor promoters are enriched for genes that regulate transcription and development. Logarithm is base 10. (D) Promoters at PRC1 loop anchors are less likely to be expressed than promoters at loop anchors not bound by PRC1 or promoters not at loop anchors and not bound by PRC1. Expression levels indicated by violin plots that are overlaid with Tukey box plots (black) and the median expression level (white circle). Logarithm is base 10. *P = 1.03 × 10⁻³; one-sided Mann–Whitney U test; **P = 8.34 × 10⁻⁴; one-sided Mann–Whitney U test; N.S., not significant.

The presence of strong Pc ChIP sites that are not associated with loop anchors may reflect the presence of chromatin loops that we cannot detect. Especially in the case of strong Pc ChIP sites separated by only a few kilobases, it is difficult to identify focal peaks in the Hi-C map due to the strong signal along the diagonal.

PRC1 Loops Are Associated with Gene Repression.

We found that Pc ChIP sites enriched at least 30-fold were not only particularly large (in terms of 1D extent along the DNA sequence; see, for instance, Pc ChIP track in Fig. 2 A and B), but the majority of these sites (64.2% versus 21.8% for random shuffle control; SI Appendix, Fig. S5) contained both GAGA and PHO motifs, which are characteristic of polycomb response elements bound by PRC1 (17). We therefore refer to the anchors found at such sites as “PRC1 loop anchors.” Loop anchors were 11.5-fold enriched relative to random shuffle control at these sites, which accounted for 26% of all loop anchors and 26.2% of PRC1 sites. Twenty-five percent of all loops had a PRC1 peak at both anchors (51.4-fold enrichment versus random shuffle control).

A large fraction (32.7%) of PRC1 loop anchors were found at promoters of important developmentally regulated genes, including those for the homeotic PcG target genes Antennapedia (Antp) and sex combs reduced (Scr) at the Antennapedia Hox gene complex (ANT-C) (Fig. 3A), as well as invected (inv) and engrailed (en) (Fig. 3B). Gene ontology (GO) term analysis indicated that genes involved in regulating transcription and development were particularly enriched at PRC1 loop anchor promoters (Fig. 3C and Dataset S3).

PRC1 loop anchor promoters were less likely to be expressed than promoters at loop anchors not bound by PRC1 (0 compared with 3 median RPKM expression level, P = 8.32 × 10⁻⁴; one-sided Mann–Whitney U test) or promoters not bound by PRC1 and not at loop anchors (0 compared with 3 median RPKM expression level, P = 1.03 × 10⁻³; one-sided Mann–Whitney U test) (Fig. 3D). Maximum expression levels of promoters at PRC1 loop anchors were less than that of promoters bound by PRC1 but not at loop anchors, even though median expression levels did not significantly differ. A similar result was obtained for promoters at loop anchors not bound by PRC1 compared with promoters not at loop anchors and not bound by PRC1 (Fig. 3D). Taken together, these results suggest a strong relationship between gene repression and chromatin looping, but whether loops cause gene repression or form as a consequence of gene inactivity remains to be determined.

Discussion

The locations of loop anchors in Drosophila determined here are notable both for correlations with ChIP-seq data and for the lack thereof. The lack of correlation with locations of CTCF protein was unexpected, inasmuch as most loop anchors in mammals are associated with CTCF protein, apparently bound to CTCF sequence motifs in a convergent orientation (3, 11). There are evidently multiple patterns of protein association with loop anchors in metazoans. The association of loop anchors in Drosophila with Pc protein is noteworthy because it points to a role of looping not only in gene activation, as widely observed in the past, but in gene repression as well.

Regions of PcG-repressed chromatin (“PcG domains”) that are separated by hundreds of kilobases to megabases are known to be in enhanced spatial proximity (4, 18, 19), but details of their internal organization have only been investigated by averaging over many PcG domains (17). The high resolution of our Hi-C contact maps revealed chromatin loops within individual PcG domains, giving insight into their internal organization. PRC1 is known to compact nucleosome arrays in vitro (20). Knockdown of the PRC1 subunit, Polyhomeotic (Ph), in vivo decompacts PcG-repressed chromatin (21), and Ph that is unable to polymerize impairs the ability of PRC1 to form clusters (22). Together with these findings, our results suggest that PRC1-bound chromatin loops within PcG-repressed domains either establish or maintain a condensed state.

Previous analyses by 3C have pointed to associations of PcG proteins with chromatin loops for the Bithorax complex (BX-C) in S2 cells (23); for inv and en in BG3 and Sg4 cells (24); and for an embryonic, pupae, and adult transgenic reporter system (25). Our Hi-C data are, however, at higher resolution and genome-wide. Higher resolution allowed more comprehensive analysis, such as the unambiguous identification of loops and the segmentation of ANT-C into a series of TADs with one or two homeotic gene promoters per TAD (Fig. 3A). Genome-wide analysis revealed both the pervasive nature of Pc protein association and the absence of significant CTCF protein association, despite conservation of CTCF from Drosophila to human (26).

A report by Cubeñas-Potts et al. on Drosophila chromatin loops in Kc cells appeared while our manuscript was in preparation (27). Cubeñas-Potts et al. noted an enrichment of cohesin but a lack of Drosophila CTCF at loop anchors, consistent with our observations. Cubeñas-Potts et al. did not mention Pc, but our analysis of their data revealed an enrichment of loop anchors at Pc ChIP peaks and an enrichment of Pc ChIP peaks at loop anchors (SI Appendix, Fig. S6). We find a likelihood of repression of promoters at Pc-bound loop anchors, especially for developmental genes; Cubeñas-Potts et al. observed an enrichment of active developmental enhancers at loop anchors, possibly because these are among the many loop anchors not bound by Pc, or because the chromatin at these loop anchors is bivalent, bound by nonhistone proteins and histone posttranslational modifications associated with both gene activation and repression (28).

The occurrence of PRC1 at loop anchors could reflect a role in loop formation similar to that proposed for CTCF in mammals, wherein cohesion complexes extrude loops, in a process halted upon reaching bound CTCF (11, 29, 30). Consistent with this model, a large majority (72.8%) of Drosophila loop anchors are bound by the Rad21 subunit of cohesin. Regardless of whether PRC1 performs such a role, additional proteins must be involved, because PRC1 is present at only 26% of Drosophila loop anchors.

Materials and Methods

Hi-C.

Hi-C was performed using the tethered conformation capture approach (8, 9), which improves the signal-to-noise ratio needed for detection of chromatin loops. In brief, D. melanogaster Kc cultured cells were fixed with 1% EM grade paraformaldehyde. Cells were lysed, cross-linked proteins were biotinylated at cysteine residues, and the DNA was digested with DpnII. Digested chromatin was bound to streptavidin beads, thoroughly washed to remove uncross-linked DNA, DNA ends were filled in with biotin-14-dATP, and free DNA ends were ligated together. DNA–protein cross-links were reversed, DNA purified, biotinylated nucleotides marking unligated ends removed, and the DNA sheared to ∼500 bp. The biotinylated DNA was pulled down with streptavidin beads and prepared for and subjected to high-throughput Illumina DNA sequencing. Further details provided in SI Appendix, SI Materials and Methods.

Hi-C Analysis.

Hi-C data were analyzed using the Juicer pipeline as previously described (3, 31). In brief, Hi-C reads were mapped to the dm3 reference genome using BWA-MEM. Aligned reads were assigned to restriction fragments, duplicates were removed, reads with a MAPQ < 30 were removed, and intrafragment reads were removed. The genome was then divided into equally spaced bins, and the number of contacts was counted in each pair of bins. Hi-C contact maps were normalized by matrix balancing (3, 31).

TADs were identified using the previously described Arrowhead algorithm (3, 31), and loops were identified by visual inspection and manual annotation using Juicebox (32). Further details provided in SI Appendix, SI Materials and Methods.

External datasets used in this study can be found in SI Appendix, Table S2.