Functional Anatomy of Polycomb and Trithorax Chromatin Landscapes in Drosophila Embryos

1Polycomb group (PcG) and trithorax group (trxG) proteins are conserved chromatin factors that maintain, respectively, the memory of inactive or active states of homeotic genes throughout development. They also regulate many other target genes (reviewed in) and misregulation of PcG and trxG genes leads to loss of cell fates, aberrant cell proliferation and tumorigenesis. Moreover, PcG and trxG factors play an important role in diverse epigenetic processes such as stem cell pluripotency and plasticity, genomic imprinting, and X chromosome inactivation. In Drosophila, PcG and trxG proteins are recruited to chromatin by regulatory DNA elements called PcG and trxG response elements (PREs and TREs, respectively). These elements were shown to drive epigenetic inheritance of silent and active chromatin states throughout development. Biochemical studies on PcG proteins revealed that they exist in at least three distinct multiprotein complexes (reviewed in). PRC2-type complexes contain the four core components E(z) (Enhancer of zeste), Esc (Extra sex combs), Su(z)12 (Suppressor of zeste 12), and Nurf-55. The SET domain-containing E(z) subunit trimethylates lysine 27 of histone H3 (H3K27me3). This mark is specifically recognized by the chromo domain of Polycomb (PC), a subunit of the PRC1-type complex. PRC1 contains PC, Polyhomeotic (PH), PSC (Posterior sex combs), and the histone H2A ubiquityltransferase dRing, in addition to several other components, including TBP-associated factors. The PhoRC complexes include the sequence-specific DNA binding proteins Pleiohomeotic (PHO) or its homolog Pleiohomeotic-like (PHOL), as well as the dSfmbt protein (Scm-related gene containing four MBT domains). Several trxG complexes have been identified: TAC1 (Trithorax Acetylation Complex) with the histone methyltransferse Trithorax (TRX), NURF, SWI/SNF, ASH1, and ASH2 (for reviews, see). Interestingly, the human TRX homolog MLL1 has been previously shown to be cleaved at two conserved sites by the Taspase1 enzyme, generating an N-terminal and a C-terminal fragment, which can heterodimerize. However, it is unknown whether the two moieties can have different functions or chromosomal distributions. Additional PcG/trxG proteins have been identified that are not part of the core of these complexes, but are associated with them and, therefore, can be considered as PcG/trxG-associated proteins. These proteins may exist as individual molecules in the cell, but it is also possible that they are part of other protein complexes that contain additional, as yet unidentified PcG/trxG proteins.

2PcG and trxG complexes (except PhoRC) do not bind their target DNA in a sequence-specific manner in vitro, but are recruited to PRE/TRE sequences in vivo. A simple pathway for PcG protein recruitment based on stepwise recruitment of PRC2 proteins by PhoRC, followed by PRC1 recruitment by the H3K27me3 mark deposited by PRC2 has been suggested. However, PcG recruitment seems to be more complex. PHO interacts with PRC2 as well as with the PC and PH subunits of PRC1 in vitro. PHO/PHOL binding sites alone are insufficient to tether PcG proteins to DNA in vivo, and most PcG sites are stained normally in polytene chromosomes in pho/phol double mutants despite lack of detectable PHO and PHOL proteins. However, PcG protein binding is lost at the bxd PRE in pho/phol double-mutant wing discs, suggesting that the role of PHO and possibly PHOL is important. Other factors have been shown to be involved in recruitment, such as GAGA factor (GAF), Pipsqueak (PSQ), Dorsal switch protein (DSP1), Zeste, Grainyhead (GH), and Sp1/KLF (reviewed in). Mutations in the corresponding genes do not have a clear PcG phenotype, and intriguingly, all seem to be involved in activation as well as in repression. In summary, many unresolved questions regarding PcG recruitment still remain, and the current model proposes that a combination of several DNA binding factors, and maybe yet-unknown components, could lead to tethering of PcG proteins to DNA.

3Recently, the distribution of several core components of PcG members and their associated histone modifications has been analyzed in fly as well as mammalian cells. Yet, a comprehensive genome-wide binding map of PcG/trxG recruitment factors and of trxG proteins is still lacking. Here, we have generated high-resolution genome-wide binding maps in Drosophila embryos of two PRC1 components and their associated histone mark H3K27me3, the N- and the C-terminal part of the TRX protein and their associated histone mark H3K4me3 as well as four sequence-specific DNA binding proteins known to be involved in recruitment of Polycomb proteins. Our results show the complementarity between PcG and trxG protein binding in the genome and suggest that multiple DNA binding proteins participate in setting up this PcG and trxG protein distribution.

4Using chromatin immunoprecipitation (ChIP) in 4–12-h-old Drosophila melanogaster embryos coupled with genome-wide high-density tiling arrays, we mapped the distribution of the PRC1 components: PC and PH, the N- and the C-terminal part of the Trithorax protein (TRX-N and TRX-C, respectively), and the histone H3K27me3 and H3K4me3 marks. We also determined the genome-wide binding profile of GAF, PHO, PHOL and DSP1, four DNA binding proteins thought to be involved in PcG recruitment. Reproducibility of biological replicates is shown in Figures S1 and S2. Figure 1 shows an example of the different profiles along part of chromosome 3R including the HOX gene cluster named ANT-C. The statistics on the number and size of regions significantly enriched for various proteins is shown in Figures S3 and S4, and in Table S1. As observed previously, PC and H3K27me3 mark covered over 200 large domains (>5 kb), most of which contain discontinuous subregions with significant p-values for enrichment separated by small intervening subregions that were enriched although their p-values were not significant (see Text S1 for a precise definition of H3K27me3 and PC domains). The number of significantly enriched subregions for PC and H3K27me3 were 2,110 and 2,480, respectively. Nearly all PH binding sites fall into PC- and H3K27me3-bound regions (Figure 2A). The sequence-specific DNA binding proteins PHO, PHOL, DSP1, and GAF are bound to thousands of genomic sites (Table S1). Surprisingly, whereas PcG binding sites strongly predict the presence of one or more of the DNA binding factors, the converse is not true. In fact, the sequence-specific DNA binding proteins are more frequently bound to sites bound by TRX-N and trimethylated on H3K4 (see Figure 2B). Binding of the N-terminal fragment and the C-terminal fragment of TRX (TRX-N and TRX-C, respectively) correlates well at the genome-wide level (Figure S5), but the relative intensities are very different. TRX-N is significantly bound to 4,868 genomic sites, with strong binding correlated to H3K4me3-bound regions (Figure 2; a total of 4,893 regions contained H3K4me3). At most of these sites, TRX-C binding levels are higher than background, but not picked up as significant. Strong binding of TRX-C is only identified at 167 genomic sites, mainly located in PRC1-bound regions (Figure 2C) where TRX-N binds weakly if at all. All the profiles are available at an online browser at the address http://purl.oclc.org/NET/polycomb. This browser also contains data from earlier mapping studies and from transcription profiling of staged embryos. In addition, it contains the annotation of predicted PREs (M. Rehmsmeier, personal communication), whose genomic location can be visualized along with the significantly enriched regions and with the results from our sequence analysis.

5Recent analysis of H3K4me3 and H3K27me3 in mouse and human cells revealed the coexistence of these two marks in a large fraction of the H3K27me3 regions. These regions encompass most of the H3K27 trimethylated sites in embryonic stem (ES) cells and a substantial portion of them in differentiated cells. Although we do frequently observe H3K4me3 occupancy at transcription start sites (TSSs) flanking PH sites, this is almost exclusively observed at the boundary of large H3K27me3 domains (see Text S1). From a total of 4,893 H3K4me3 and 2,480 H3K27me3 regions, only 161 had an overlap, i.e., only 6.5% of the H3K27me3 regions. Considering that most of the genes identified by these regions of overlap are expressed only in a fraction of the embryonic cells, we believe that most of these cases reflect a mixture of cell populations rather than true bivalency. Moreover, the H3K4me3 profile always showed sharp peaks at promoters within large H3K27me3 regions, in contrast to mammalian cells in which bivalent domains often show similar profiles with H3K4me3 and H3K27me3 spread over regions of several kilobases in size. Thus, our data suggest that H3K4me3 and H3K27me3 are generally exclusive in the fly genome. Nevertheless, individual cases of true bivalency may exist in fly embryos or at other developmental stages. A rigorous demonstration of this point will require sequential ChIP with mononucleosomal chromatin and antibodies directed against the H3K4me3 and H3K27me3 marks.

6We sought a comprehensive characterization of the joint distribution of PcG and trxG factors and associated marks. Many of the data tracks are highly correlated among themselves (Figures S5 and S6), and are also tightly associated with other spatial genomic features like TSSs. We therefore developed a new method for dissecting a multivariate genomic profile into a hierarchy of “spatial clusters.” Briefly, “spatial clustering” can be viewed as the genomic analog of gene clustering, since it dissects the genome into clusters that share a common profile across all experimental tracks (detailed information is given in the Text S1). Unlike gene clustering, our model takes into account the genomic layout of the data, and organizes clusters spatially to probabilistically describe the typical genomic order among them. We used the clustering results (Figure 3) as a blueprint for our dataset, validating conclusions by running an independent, supervised data analysis. An example of cluster organization is illustrated in Figure S7. Analysis of the distribution of cluster location with respect to the TSS further demonstrates how the clusters are organized around genes (Figure 3B, note that TSS data were not used by the algorithm to define clusters).

7As shown in Figure 3, our data reflect two levels of genomic organization. First, the genome is partitioned into three superclusters. Consistent with the mutually exclusive distribution of H3K27me3 and H3K4me3, unsupervised spatial clustering identifies a “H3K27me3-marked” supercluster and “H3K4me3-marked” supercluster, in addition to regions with no particular epigenomic enrichment (“background” supercluster, not shown in Figure 3). Second, each supercluster is subdivided into distinct clusters, and the model identifies the connections between clusters that organize the entire genome (Figure S18). The H3K27me3 superclusters are anchored around clusters characterized by high levels of PH binding (labeled as “PH sites”). These clusters include also strong PHO enrichment, presence of the recruiter factors GAF and DSP1 and TRX-C occupancy. All of the PH site clusters in the BX-C, the ANT-C, the ph, the hh, and the en genes were previously identified as PREs, suggesting that in general, most of the PH clusters are indeed PREs. The H3K27me3 supercluster also included three clusters with lower levels of PC and a general lack of PH and cofactors. We labeled them as “Strong,” “Medium,” and “Weak” PC clusters.

8Similarly, the H3K4me3-marked supercluster was subdivided by the algorithm into four clusters. These clusters reflect clear organization around annotated TSSs, as identified by their TSS enrichment statistics (Figure 3B) and binding preferences (Figure 3D). We denoted the cluster with the most 5′ enrichment as the “K4me3-recruiters” cluster. It is characterized by high levels of GAF, DSP1, and significant, but weaker levels of PHO and PHOL, as well as medium to weak H3K4me3 levels. Enriched exactly at the TSS is the “K4me3-TSS” cluster with high H3K4me3 levels in combination with high levels of TRX-N, PHO and PHOL. The K4me3 cluster has only high levels of H3K4me3 and represent the region downstream the TSS, whereas the “weak K4me3” cluster shows low, but significant levels of H3K4me3 alone and is more weakly enriched around TSSs.

9PC and H3K27me3 were bound in large regions, often greater than 5 kb, with the largest ones spanning several hundred kilobases (see Figures 1 and S4A). Globally, H3K27me3 and PC profiles were very well correlated, facilitating the definition of PC domains (see Text S1), underscoring the significance of the H3K27me3 supercluster (Figure 3) identified by spatial clustering. A similar pattern was observed for PC and H3K27me3 by Schwartz et al. in their genome-wide mapping studies in S2 cells and by Tolhuis et al. who used Kc cells.

10Nearly all PH peaks were specific to PC and H3K27me3 regions (the PH sites; Figure 3) and were present in all the earlier characterized PREs. The average distribution of H3K27me3 around PH peaks takes a dip at the PH sites (Figure 4A), which may be due to nucleosome depletion at the PREs. The distribution of the domain size, number of PH peaks, and genes in H3K27me3 domains is shown in Figure S4 (for an identification of candidate PcG target genes, see Text S1 and Table S2).

11Despite these common features, there are differences in the positions of many of the PcG domains in different biological samples. Although the majority of our 217 H3K27me3 domains also exists in S2 cells, 79 (36%) of them did not overlap any bound regions in S2 cells. These data are corroborated by the analysis of the distribution of the PC protein which, similar to H3K27me3, forms large domains. In general, H3K27me3 differences between embryos and S2 cells paralleled differences in PC binding. The same was observed in a comparison between ChIP on chip binding of PC from embryos and the PC DamID profile obtained previously in Kc cells. Interestingly, a substantial portion of the PC domains in Kc cells differed from those observed both in embryos and in S2 cells. Thus, many common PC domains are identified in various cell types, but a significant subset of them is cell-type specific rather than constitutive. These data are in agreement with previous studies suggesting that part of the PcG binding is cell-type and developmental-stage specific.

12To gain more insight into PRC1 recruitment to chromatin, we examined the distribution of PcG recruitment factors at PH sites that are also bound by PC (PRC1 sites). The combination of different PcG recruitment factors at the PRC1 sites as compared to the genome is listed in Table S3 and shown in Figure 2B. Most PH binding peaks colocalize with the PcG recruitment factor PHO (96.4%) (see Figure 2B and Table 1). DSP1 and GAF were present in about 50% of the PH sites. In contrast, PHOL binding was not common at PH sites, with a frequency (21.1%) comparable to that of TRX-N (26.5%). Surprisingly, only a minority of all recruitment factors binding sites (3.2% to 13.5%) was restricted to PH sites (Table 1). Comparison with previously published Zeste data showed that a moderate 25% of the Zeste sites colocalized with PH peaks. Together these data suggest a correlation gradient between different recruiters and PREs, with PHO > DSP1/GAF > Zeste/PHOL.

13The K4me3-recruiter cluster (including strong GAF and DSP1 and medium to weak H3K4me3 levels) is located in a position just upstream to the TSS. The K4me3-TSS cluster (high H3K4me3 levels and strong TRX-N, PHO, and PHOL binding) is usually following it and is almost exclusively observed over the 2 kb around the TSS. Finally, the K4me3 cluster (high H3K4me3 levels without TF occupancy) is enriched 3′ to the TSS. This organization suggests that binding of GAF and DSP1 can promote the activation of a TSS upon binding of TRX-N and the PHO/PHOL factors. Therefore, PH target promoters are strongly bound by PHO and TRX-C and depleted of PHOL and TRX-N (Figure 4B), whereas H3K4me3 promoters are bound by PHO, PHOL, and TRX-N (Figure 4C). Notably, the positions of PHO (and PHOL) in the second class of promoters is right at the TSS, whereas at PH-bound promoters, PHO is colocalized with PH upstream to the TSS (Figure 4D). This different architecture may contribute to PH recruitment or to silencing of PH-bound promoters.

14We further analyzed active promoters and PREs/TREs by analyzing TRX binding. The human TRX homolog MLL1 is cleaved by Taspase1, generating an N-terminal and a C-terminal fragment, which can heterodimerize in vitro. Low levels of TRX-N co-occupied PH binding sites in about 26.5% of cases (Figure 4A; Table1). However, TRX-N is present at thousands of other genomic sites, where no PcG binding can be observed. These genomic sites correspond mainly to annotated 5′ ends of genes carrying H3K4me3 peaks slightly offset towards the body of the gene in comparison to TRX-N (cluster K4me3-TSS; see also Figure 4C). Interestingly, although the TRX-C profile overall looks similar to the TRX-N, its relative binding intensities are different. TRX-C is strongly bound at PcG binding sites, whereas low binding is observed at most promoter regions of non-PcG target genes (Figure 4). These results suggest that whereas the distribution of the N-terminal part of TRX follows a general transcription cofactor role, the C-terminal part is specifically linked to PcG function. PcG proteins might repress transcription by anchoring the C-terminal portion of TRX at PREs. On the other hand, constitutive TRX-C binding at PREs/TREs might allow PcG target genes to switch their state upon strong transcriptional induction.

15In the case of PHO, PHOL, and GAF, sequence-specific DNA binding in vitro has been shown previously. By analyzing the collection of statistically significant bound sites for each of these proteins with the Multiple EM for motif Elicitation (MEME) algorithm, we detected the expected binding sites (Figure 5A and 5B, and Tables S6 and S7), whereas for Dsp1, the results were not conclusive. The “GAAAA” motif was not strongly enriched among the genomic binding sites for this protein, although a degenerated GAAAA motif was found at DSP1-bound as well as at PHO- and PH-bound regions (Figures S8–S11, see Text S1 for a detailed discussion).

16In order to determine whether other sequence features may characterize PREs specifically, we further developed the unsupervised spatial clustering methodology (Figure 3) to allow discovery of sequence motifs that discriminate among clusters or groups of clusters. As shown in Figure 6, we discovered several known and novel motifs that are either shared among clusters or distinguish them. We visualize these results in terms of the affinities (or predicted binding energies) of the inferred position weight matrices (PWMs) in and around each our spatial clusters.

17Two motifs (GAGA and the CA repeat motif) are marking clearly the PH sites and the K4me3-recruiter clusters. Three additional motifs are strongly marking the K4me3-TSS cluster and clearly discriminating it from the spatially coupled K4me3-recruiter cluster sites. Two of them are motifs bound by the Myc, Max, and Mad/Mnt proteins and include the DNA replication element (DRE) TATCGATA, which is also consensus for several other factors including the TRF2n, Cut, and Beaf-32. The third motif (CAGCTG) is an E-box bound by bHLH proteins which, like DREs, are involved in the regulation of many developmental genes. We note that the detected motif enrichments are specific to the K4me3-TSS cluster and not to general TSSs in the genome since general non–H3K4me3-associated TSSs lack these motifs.

18Importantly, we also discovered motifs that discriminate between K4me3-recruiters and PH sites. The CAACAACAA motif is enriched around K4me3-recruiters, but not in and around PH sites (see also Figure S8). On the other hand, the CCGTCGG and the Sp1/KLF-like GGGGTGGG motifs are specific to PH sites and not K4me3-recruiters (see also Figure S11). These motifs constitute candidates to recruit new DNA-binding factors to PREs.

19In addition to these motifs, the consensus sites for PHO/PHOL, DSP1, and GAF are more strongly enriched at the 300-bp core regions around the maximal binding peak of PH than around the other genomic regions bound by the factors without PH (Table S5–S8). Thus, the density of binding sites is specific to PREs, suggesting that cooperative binding may help recruit PcG proteins. Consistent with this idea, the fold enrichment for each of the factors (with the exception of PHOL, see below) is higher at PH-bound regions compared to non–PH-bound regions (Figures S12 and 4).

20Of particular interest is the distribution of the PHO motif around PH sites and the K4me3-TSS clusters. Unlike the GAGA (or CACA) motif, the frequency of motifs with sequence similarity to consensus PHO motifs is high, but these motifs are not well localized at PH sites. High predicted PHO affinities (defined by PWMs; see Text S1) were also present in the strong PcG clusters surrounding PH sites. This pattern matches perfectly with our ChIP data, which also suggest that PHO levels are regionally high around PH sites. In contrast to this pattern, the K4me3-TSS cluster is characterized by weak, but significant peaks of PHO motifs that were localized right at the TSS. This pattern is again matched by the PHO and PHOL ChIP data at the TSS of H3K4me3 associated promoters (Figure 4C).

21PHO and PHOL share sequence homology, were shown to bind the same DNA motif in vitro, and have been proposed to play redundant roles in PcG-mediated silencing (reviewed in). Notably, we observed that PHO and PHOL binding patterns do not always overlap in the genome. In particular, PHO binds much stronger than PHOL at PH sites (Figures 3, 4, S13C, and S14), whereas both proteins bind with similar intensities in K4-recruiter and K4-TSS clusters (Figures 3 and 4). We also noticed that the majority of PHOL sites in the genome colocalized with TRX-N and H3K4me3-bound regions (Figures 3 and S5; Table S4A). To investigate whether PHO and PHOL may fulfill distinct roles in recruitment of PcG and trxG proteins, we computed the genome-wide ratio of PHO/PHOL binding (see Text S1) and plotted it compared to the individual profiles as well as to PH sites. Figure 7A shows that the PHO/PHOL ratio accurately matches the PH distribution profile since the binding of the two proteins at all other sites in the genome cancels out, whereas PHO binding at PREs is much stronger than PHOL. To confirm whether the ratio of PHO/PHOL is linked to the activity state of PRE/TREs, we examined by quantitative ChIP assays the binding levels of PH, PHO, and PHOL at three PcG target genes characterized by ON/OFF expression states in different larval tissues (Figure 7B–7F). Ubx is expressed in haltere/third leg imaginal discs and is repressed in eye imaginal discs (ED). On the contrary, so (sine oculis) and toy (twin of eyeless) have very low expression in haltere/third leg discs and are highly expressed in eye discs (Figure S15A). For Ubx regulation, we analyzed protein binding levels at the bx PRE, bxd PRE, and the Ubx TSS, and for so and toy, we analyzed their TSS, which overlapped with the PH-bound region (Figure S15B). PH, PHO, and PHOL are bound in all the 5′ regions of the genes that we examined in both the ON and OFF state (Figures 7 and S16). However, significant differences in binding levels were noticed. In haltere/third leg discs where Ubx is ON, bx PRE, bxd PRE, and Ubx TSS showed a slight decrease in PH binding (50%) as compared to eye discs. Both so and toy TSS showed higher levels of PH binding in haltere/third leg discs, where these genes are silenced (OFF), as compared to eye imaginal discs (ON). At the Ubx TSS and the bx PRE, levels of PHOL were significantly higher in haltere/third leg discs (ON) as compared to eye discs (OFF). With regards to PHO, stronger binding was observed at the PREs in eye discs (OFF state), whereas at so and toy stronger binding was observed in haltere/third discs (OFF state) compared to eye discs (ON). In summary, a significant decrease in the levels of PH in tissue where target genes are active correlates with a decrease in the PHO/PHOL ratio. On the other hand, increased PH levels at genes that are OFF in a certain tissue correlates well with an increased PHO/PHOL ratio.

22To further examine the function of the PHO/PHOL ratio in Polycomb-dependent gene silencing, we performed quantitative reverse-transcriptase PCR (RT-PCR) on eye, haltere/third leg and wing imaginal discs from wild-type and pho1 homozygous (null mutant allele of PHO) third instar larvae. In wild-type eye discs, the Ubx and Antp genes are repressed, and the detection of their transcripts is limited to few copies. In pho1 mutant larval eye discs, Ubx gene becomes derepressed (5.5-fold), and gene activation is even stronger for the Antp gene (between 10- and 30-fold) (Figure 8A). These results suggest that the loss per se of PHO has an impact on the level of transcription of Polycomb-silenced target genes, and this underscores its fundamental role in setting up Polycomb-mediated silencing. Binding of PHOL to the same sequence motif in the promoter region of these two genes might partially complement for the loss of PHO. Indeed, we detected increased binding levels of PHOL to chromatin in pho1 mutant imaginal discs (unpublished data).

23We then analyzed the effect of the pho1 mutation in haltere/third leg discs where the Ubx gene is transcribed and in wing discs where Antp is active. We detected a consistent, yet slight, decrease of their transcripts (2-fold and 1.5-fold, respectively) (Figure 8B). These results suggest that PHO may also play a role as an activator of homeotic genes, even if this role is weaker than its silencing function.

24Because we found a high colocalization of PHO and PHOL with TRX-N at many gene promoters not related to PcG-mediated silencing, we performed quantitative RT-PCR to check the expression of two constitutively transcribed genes such as Chc and Rp49, which are bound by PHO in wild-type embryos. Again, Chc expression decreased 1.6 times in both eye and haltere/third leg discs and Rp49 1.3 times in eye discs from pho1 mutant larvae (Figure 8C). In contrast, we could not detect major changes in their expression levels in a phol81A null mutant background (unpublished data), pointing to a redundant role of PHOL in gene activation

25These results, together with the recent work of Beisel et al., indicate that PHO is a modulator, not only of PcG-mediated silencing, but also of the active state of many genes.

26The genome-wide mapping of PcG factors, TRX, their associated histone marks, and potential PcG recruiter proteins in Drosophila embryos revealed several important features. First, similar to the PcG distribution in Drosophila cell lines, PcG proteins strongly colocalize and form large domains containing multiple binding sites. Second, the N-terminal and C-terminal fragments of TRX show different binding affinities to repressed and active chromatin. The N-terminal fragment of TRX has low affinity to PcG binding sites but is strongly bound to thousands of active promoter regions that are trimethylated on H3K4, whereas the C-terminal fragment of TRX only showed high binding affinity to PcG binding sites. Third, the majority of PcG recruiter binding sites are associated with H3K4me3 and TRX-N foci and not with PH binding. The binding ratio between the PHO protein and its homolog PHOL is a major predictive feature of PcG versus TRX recruitment. Finally, supervised and unsupervised sequence analysis methods led to the identification of sequence motifs that discriminate between most of the PcG and TRX binding sites, but these motifs are likely to be working jointly, and none of them seems to drive recruitment by itself.

27To date, PREs have been only characterized in Drosophila. These elements are not defined by a conserved sequence, but include several conserved motifs, which are recognized by known DNA binding proteins like GAGA factor (GAF), Pipsqueak (PSQ), Pleiohomeotic and Pleiohomeotic-(like) (PHO and PHOL), dorsal switch protein (DSP1), Zeste, Grainyhead (GH), and SP1/KLF. Our genomic profiles provide a comprehensive view on the potential role of these factors in the establishment of PcG domains.

28The presence of PHO at all PREs indicates that PHO is a crucial determinant of PcG-mediated silencing, consistent with earlier analysis on one particular PRE. On the other hand, PHOL and Zeste were bound at a small subset of PREs. Zeste was previously shown to be necessary for maintaining active chromatin states at the Fab-7 (Frontabdominal-7) PRE/TRE. Therefore, Zeste and PHOL may primarily assist transcription rather than PcG-mediated silencing. GAF and DSP1 resemble PHO as they bind to many (albeit less than PHO) PREs as well as to active promoters. Supervised DNA motif analysis indicated a higher density of GAF, DSP1, and PHO binding sites at PREs as compared to other bound regions at non-PH sites. This suggests that cooperative binding of these proteins may provide a platform for PcG protein binding. Moreover, GAF may act by inducing chromatin remodeling to remove nucleosomes, since the regions bound by PcG proteins show a characteristic dip in H3K27me3 signal that has been attributed to the absence of nucleosomes in those regions. These nucleosome depletion sites are the places wherein histone H3 to H3.3 replacement takes place. Indeed, several of the Zeste-bound regions and GAGA binding sequences were shown to localize to peaks of H3.3, suggesting the possibility that GAF may recruit PcG components to PHO-site–containing PREs as well as recruit TRX to promoters via nucleosome disruption.

29In addition to an increased density of motifs for GAF, PHO, and PHOL, unsupervised spatial cluster analysis identified specific motifs that distinguish the PH sites from the K4me3 cluster. Although the identity of the factors binding to these motifs is unknown, this suggests that the DNA sequence of PREs contains much of the information needed to recruit PcG proteins and to define silent or active chromatin states. With this distinction, it may be possible to develop an algorithm to faithfully predict the genomic location of PREs. Earlier attempts to predict PREs in the fly genome have made progress toward this goal, but they are still far from reaching the required sensitivity and specificity (see also Tables S9 to S11). The use of a sequence analysis pipeline that is not dependent on prior knowledge was demonstrated here to generate new discriminative motifs with a potential predictive power. The unique genomic organization of PcG domains may suggest that the genome is using, not only local sequence (high-affinity transcription factor binding sites located at the binding peaks) information to determine PREs, but also integration of regional sequence information (stronger affinity on 5 kb surrounding PREs). Using such regional information to predict PREs may break the current specificity and sensitivity barriers.

30Our ChIP on chip data showed that PHO binding comes in two distinct flavors. In one class of target sites, PHO binding coincides with PH sites within PC domains, whereas outside these domains, it is largely colocalized with PHOL, TRX-N, and H3K4me3 (Table S4). PHOL binding was weaker at PH sites and was mainly present along with marks associated with gene activation. Quantitative ChIP assays (Figure 7) revealed that PH, PHO, and PHOL were bound in PREs/TSS of their target genes in both ON and OFF states, but the ON state was marked by a decrease in PH binding and a corresponding increase in PHOL levels, whereas the OFF state was characterized by an increase in both PH and PHO binding levels.

31Papp and Muller analyzed chromatin at the Ubx TSS, the bx PRE, and the bxd PRE (the same primers were used in our study) by comparing haltere/third leg imaginal discs (ON state) with wing imaginal discs (OFF state). They found a 50% reduction of PH binding levels at the bx PRE, a minor decrease at bxd, and no change in the Ubx TSS. Our ChIP experiments demonstrated a 50% decrease in PH levels at bx PRE and at the Ubx TSS and a minor decrease at bxd PRE when comparing haltere/third leg imaginal discs to eye imaginal discs. We also observed a slight decrease in the levels of PHO in haltere/third leg disc (ON state) as compared to eye imaginal discs (OFF state) at the bx and bxd PRE, whereas Papp and Muller did not see differences in the levels of PHO. The most likely explanation for these discrepancies is that the peripodal membrane cells of the wing imaginal discs express Ubx, whereas all cells silence this gene in eye imaginal discs.

32In pho1 mutant eye discs, the absence of PHO causes derepression of the homeotic genes Ubx and Antp. However, the expression levels in pho1 mutants are still much weaker compared to tissues where these genes are normally expressed. This low degree of activation could be explained by compensatory binding of PHOL to the PHO sites in order to maintain PcG-mediated silencing, even if the PHOL-dependent rescue function is incomplete as pho1 mutants die as pharate adults. PHO and PHOL have indeed been described as redundant in their role in PcG-mediated silencing since they bind to the same DNA sequence motif in vitro. However, out of the 1,757 places wherein both PHO and PHOL were significantly bound, only 807 shared the same local maxima (46%). Another 559 (32%) peaks were within 250 bp of each other. This suggests that, in vivo, these two proteins prefer slightly different sequences, with PHO more strongly attracted to PREs, whereas PHOL binds better to promoters. Moreover, PHO interacts directly with PC and PH, as well as with the PRC2 components E(z) and Esc, whereas PHOL only interacts with Esc in yeast two-hybrid assays. Stronger interactions between PHO and PcG components may stabilize PHO binding at PREs, favoring it over the binding of PHOL. It is thus possible that the primary function of PHOL is as a transcription cofactor, and that its recruitment to PREs is subsidiary to PHO.

33Here, we report for the first time, to our knowledge, the genome-wide distribution of TRX. This protein has been proposed to counteract PcG-mediated silencing. Petruk et al. demonstrated that TRX colocalizes with Polymerase II and elongation factors in Drosophila polytene chromosomes. They then showed that PcG and TRX proteins bind to a PRE mutually exclusively in salivary gland chromosomes. In contrast, two other studies found binding of TRX at discrete sites at PREs and promoter regions of HOX genes, and suggested that TRX coexists with PRC1 components at silent genes. We postulated that these differences might be explained by the use of different TRX antibodies, one against the N-terminal domain and one against the C-terminal domain of TRX. Notably, the TRX protein is proteolytically cleaved into an N-terminal and a C-terminal domain, but the fate of the two moieties after cleavage has never been addressed in vivo.

34Our genome-wide mapping studies using the same antibody against the N-terminal fragment (TRX-N) as used by Petruk et al., showed that the binding affinity of the N-terminal fragment to PREs is rather weak, whereas TRX-N binds thousands of promoter regions trimethylated on H3K4, indicating a general role of TRX-N in gene activation. In contrast, ChIP on chip profiling using an antibody against the C-terminal TRX fragment showed high binding levels at PRE/TREs, whereas binding to promoter regions (where the TRX N-terminal fragment is strongly bound) is rather weak. The strong quantitative correlation between the binding intensities of PH and TRX-C suggests that TRX-C can indeed bind to silent PcG target genes. These data are confirmed by the colocalization of PH and TRX-C at inactive Hox genes in salivary gland polytene chromosomes and in diploid cell nuclei (as seen in a combination of DNA fluorescent in situ hybridization (FISH) and immunostaining; unpublished data). Thus, PcG silencing may involve locking the C-terminal portion of TRX in an inactive state that perturbs transcription activation events. The fact that TRX is recognized by two different antibodies that recognize PREs (H3K4me3-depleted regions) or TSSs suggests that these antibodies reflect the activity state of the protein and thus represent a powerful tool to study the switching of genes between silencing and activation.

35Similar to mapping studies in Drosophila cell lines, H3K27me3 also forms large domains in Drosophila embryos. These large PcG domains could provide the basis of a robust epigenetic memory to maintain gene expression states during mitosis. As previously suggested, stably bound PcG complexes at PREs may loop out and form transient contacts with neighboring chromatin, which become trimethylated on H3K27. H3K27me3 might then attract the chromodomain of the PC protein, which may be occasionally trapped at these remote sites by cross-linking mediated by the chromodomain of PC. Alternatively, PcG subcomplexes missing some of the subunits might spread from the PRE into flanking genomic regions containing H3K27me3 histones.

36Although genome-wide PcG profiles in Drosophila embryos correlate well with profiles from Drosophila cell lines, it has recently been shown that PcG protein binding profiles are partially remodeled during development. Comparison of our PcG target genes (Figure S19 and Tables S14–S16) with Schwartz et al. showed that 40% of our targets were unique (Figure S17). The fact that a consistent number of targets are only found in one or two of the samples indicates tissue specific PcG occupancy. Thus, although PcG proteins have been often invoked as epigenetic gatekeepers of cellular memory processes, they may be involved as well in dynamic gene regulation during fly development, similar to their function in mammalian cells.

37All antibodies used in this study are listed in Table S12.

38ChIP assays were performed on 4–12-h-old embryos of the Oregon-R w1118 line of Drosophila melanogaster. The complete experimental details of the ChIP experiments are available in Text S1. Briefly, ChIP samples were amplified by ligation-mediated (LM) PCR, as described previously, and hybridized to whole-genome tiling arrays manufactured by NimbleGen Systems (the array design is described in Text S1). A list of all significantly enriched regions (p-value < 0.0001) for all profiles are shown in Table S17.

39Spatial clustering was performed by training a Hidden Markov Model (HMM) to fit the available genomic profiles using a small set of clusters. The HMM represents both the relations between clusters and the joint profile distribution emitted from each cluster. We developed a hierarchical version of the algorithm so that the two layers of genomic organization in the data can be characterized (for details, see Text S1). We further enhanced the spatial clustering framework to search for motifs that discriminate among clusters. We also used the MEME and Motif Alignment and Search Tool (MAST) programs to search for enriched motifs directly (a detailed description can be found in Text S1).

40ChIP assays of imaginal discs were performed as described for embryos with the following modifications: third instar larval eye discs and haltere/third leg discs were dissected in SS M3 insect medium and kept on ice during dissection. A hundred discs were used per immunoprecipitation (IP). Discs were pelleted by centrifugation at 4.000 g for 5 min, resuspended in 1 ml of Buffer A1, and then cross-linked for 15 min in the presence of 1.8% formaldehyde by homogenization in a Tenbroeck homogenizer. Chromatin was sonicated using a Bioruptor (Diagenode) for 12 min (settings 30 s on, 30 s off, high power). Sheared chromatin had an average length of 500 to 1,000 bp. Antibodies used for IP (PHO, PHOL, and PH) were diluted 1:100 (PH and PHO) or 1:20 (PHOL). Enrichment of specific DNA fragments was analyzed by real-time PCR, using Roche Light Cycler equipment and accessories as described in Comet et al.. Enrichment in specific IPs was determined by normalizing the amount of DNA obtained in each reaction by the amount of a negative control fragment from the robo3 gene. Primer sequences are listed in Table S13.

41pho1 homozygous larvae were collected from a stock ey-GAL4/ey-GAL4; pho1/GS15194 kindly provided by R. Paro's lab. Wild-type and pho1/pho1 mutant larvae were dissected in PBS, and 40 eye or haltere/third leg discs were taken for RNA isolation using TRIzol reagent (Invitrogen). RT-PCR was performed using Superscript III First Strand Synthesis Kit from Invitrogen following the manufacturer's instructions. Reverse transcription was primed using hexamer primers. Quantitative polymerase chain reaction (qPCR) analysis was done as described for ChIP experiments. The copy number for each investigated gene was normalized to the copy number of the 18S RNA gene. Primer sequences are listed in Table S13.

42Experiment, first part (combined replicates; K27, PC, PH, PHO, DSP1, PHOL, GAF, TRX-N, and K4): E-MEXP-1708.

43Additionally, all employed microarray designs have their own accessions: PhysicalArrayDesign name: 2005-08-08_Henikoff_Dros_ChIP_1, ArrayExpress accession: A-MEXP-1251; PhysicalArrayDesign name: 2005-08-08_Henikoff_Dros_ChIP_2, ArrayExpress accession: A-MEXP-1252; PhysicalArrayDesign name: 2005-08-08_Henikoff_Dros_ChIP_3, ArrayExpress accession: A-MEXP-1253; PhysicalArrayDesign name: 2007-03-13_Henikoff_Dros_ChIP_1, ArrayExpress accession: A-MEXP-1254; PhysicalArrayDesign name: 2007-03-13_Henikoff_Dros_ChIP_2, ArrayExpress accession: A-MEXP-1255; PhysicalArrayDesign name: 2007-03-13_Henikoff_Dros_ChIP_3, ArrayExpress accession: A-MEXP-1256; and PhysicalArrayDesign name: Cavalli_Dmel_1_tiling, ArrayExpress accession: A-MEXP-1257.

44Gene accession numbers: Antp: FBgn0000095; ato: FBgn0010433; cad: FBgn0000251; Chc: FBgn0000319; Dll: FBgn0000157; dsx: FBgn0000504; grn: FBgn0001138; hb: FBgn0001180; robo3: FBgn0041097; Rp49: FBgn0002626; so: FBgn0003460; toy: FBgn0019650; and ubx: FBgn0003944.