1. Developmental Biology
  2. Genetics and Genomics
Download icon

Genome-wide errant targeting by Hairy

  1. Kurtulus Kok
  2. Ahmet Ay
  3. Li M Li
  4. David N Arnosti  Is a corresponding author
  1. Michigan State University, United States
  2. Colgate University, United States
  3. Janelia Research Campus, Howard Hughes Medical Institute, United States
Research Article
  • Cited 6
  • Views 1,669
  • Annotations
Cite this article as: eLife 2015;4:e06394 doi: 10.7554/eLife.06394

Abstract

Metazoan transcriptional repressors regulate chromatin through diverse histone modifications. Contributions of individual factors to the chromatin landscape in development is difficult to establish, as global surveys reflect multiple changes in regulators. Therefore, we studied the conserved Hairy/Enhancer of Split family repressor Hairy, analyzing histone marks and gene expression in Drosophila embryos. This long-range repressor mediates histone acetylation and methylation in large blocks, with highly context-specific effects on target genes. Most strikingly, Hairy exhibits biochemical activity on many loci that are uncoupled to changes in gene expression. Rather than representing inert binding sites, as suggested for many eukaryotic factors, many regions are targeted errantly by Hairy to modify the chromatin landscape. Our findings emphasize that identification of active cis-regulatory elements must extend beyond the survey of prototypical chromatin marks. We speculate that this errant activity may provide a path for creation of new regulatory elements, facilitating the evolution of novel transcriptional circuits.

https://doi.org/10.7554/eLife.06394.001

eLife digest

The genes encoded in DNA contain the instructions to make proteins and other molecules important for cell behavior. Only a fraction of genes are ‘expressed’ at any particular time; proteins called transcriptional repressors keep many in a silent state. One such repressor in the fruit fly is called Hairy, and its activity is essential for embryos to develop correctly. Similar Hairy-related proteins are crucial regulators of development in mammals.

A central mechanism of controlling gene expression involves the wrapping of DNA around histone proteins to form a structure called chromatin. Attaching chemical tags to histones changes how accessible the genes are within the chromatin—the more accessible the genes are, the more likely they are to be active. Some tags promote gene activation, while other tags block expression. Previous research showed that Hairy reduces gene expression by influencing which tags are added to, or removed from, the chromatin.

Kok et al. have now tracked the effects of the Hairy protein on the entire genome of Drosophila fruit fly embryos. This revealed the genes that Hairy directly targets and the corresponding effects this targeting has on chromatin structure. Hairy altered chromatin chemical tags over large blocks of DNA on silenced genes, with some of the changes being specific to particular genes. However, many areas of chromatin activity were not associated with changes in gene expression. Instead, many genes ignore Hairy-mediated changes in their vicinity, while in other cases chromatin changes occurred on genes that were already silent.

Previous studies have suggested that regulatory factors like Hairy frequently bind to many sites on the genome and have no function. Kok et al. now suggest that these sites—previously regarded as representing ‘inert’ sites—are biochemically very active. Genomic studies that label regulatory sites solely by changes to their chromatin modifications may be fooled by the apparent activity of such ‘errantly targeted’ sites, assuming that they are critical for gene regulation. At the same time, these sites may represent regions that are particularly likely to evolve regulatory properties. Kok et al. therefore propose that errant targeting by Hairy may help new regulatory elements to evolve that could eventually influence how genes are expressed.

https://doi.org/10.7554/eLife.06394.002

Introduction

Metazoan transcriptional circuitry features activation and repression signals that constitute robust regulatory networks important for the unfolding of developmental programs. In the Drosophila embryo, localized transcriptional repressors provide essential patterning information that establishes the primary anterior-posterior and dorsal–ventral axes of the organism. The action of transcriptional repressors is heterogeneous and can exhibit context effects; one of the most striking aspects involves the different classes of repressors that mediate distinct chromatin changes on target genes. Short-range acting proteins Snail and Knirps interfere with transcription only when their cognate binding sites are located within close range of the activator binding sites (Gray and Levine, 1996). These proteins interact with evolutionarily conserved corepressors that possess chromatin modifying activities (Nibu et al., 1998; Payankaulam and Arnosti, 2009). Paradoxically, these same cofactors are also recruited by another class of repressors, the long-range transcriptional repressors, exemplified by the Hairy factor (Paroush et al., 1994; Barolo and Levine, 1997; Poortinga et al., 1998). This protein is a founding member of the Hairy/Enhancer of Split (HES) transcription factors, which play essential roles in animal development, including segmental gene patterning in the early embryo and specification of neuronal differentiation in response to Notch signaling (Kageyama et al., 2007). Thus, elucidation of molecular mechanisms of Hairy activity will shed light on a number of important gene circuits that are prominently represented in key developmental pathways. The biochemical function of Hairy is associated with long-range chromatin modifications, which endow this factor with the ability to interfere with multiple cis-regulatory regions, including activators bound over 1 kb distal to the Hairy binding sites. The long-range effect has been proposed to be due to the recruitment of the corepressor Groucho (Gro), that can oligomerize to spread over large areas of the genome, and colocalization of HDAC to the target genes resulting in deacetylation of specific lysine residues in histones H3 and H4 (Courey and Jia, 2001; Martinez and Arnosti, 2008). In our previous studies, we showed that Hairy induced extensive tracts of deacetylation on ftz, a segmental patterning gene expressed early in embryogenesis (Li and Arnosti, 2011).

While potent in repression potential, Hairy and other long-range repressors are apparently restricted in their ability to exercise transcriptional effects by the local cis-regulatory context in which binding sites are located. Hairy was demonstrated to lack long-range effects on a distal RACE enhancer in the embryonic dorsal ectoderm, when Hairy binding motifs were situated in an element with activators that are restricted to mesoderm/neurectoderm regions. Furthermore, the Dorsal protein, when itself acting as a long-range repressor, is dependent on neighboring Cut and Dri transcription factor motifs to function, indicating that long-range repression complexes may require specific cis-regulatory grammar (Cai et al., 1996; Nibu et al., 2001).

The action of eukaryotic transcriptional repressors involves a number of biochemical activities, including direct antagonism of transcriptional activators and assembly of chromatin-associated factors that are correlated with gene silencing (Perissi et al., 2010). Specific types of covalent histone modifications, such as H3 and H4 deacetylation, H3K9 trimethylation and H3K27 trimethylation are correlated with repressed genes, but there is still no general understanding of how important in a quantitative sense such modifications are for inhibition of transcription at specific genes. Context effects for a particular transcriptional repressor can influence what sort and how much of a response will be generated. At a genome-wide level, specific chromatin features correlate with transcriptionally repressed genes (e.g., H3K9 and 27 methylation, reduced levels of H3 and H4 acetylation, binding of HP1), however these marks are also found within highly active loci (modENCODE Consortium et al., 2010). The epigenetic signature of transcriptional repression is thus context-dependent, consistent with a revised picture of the simple ‘histone code’ hypothesis. In the context of specific transcriptional repressors, we know little about how the context of distinct factors present at cis-regulatory elements shapes their action. Genome-wide information obtained from chromatin immunoprecipitation experiments should provide information about molecular targets and action of transcription factors, however, in addition to bona fide regulatory targets, metazoan transcription factors typically associate with a large number of in vivo binding sites of unknown significance. Recent studies have suggested that these interactions represent off-target genomic interactions, driven by low binding specificity of transcription factors and a general affinity for open chromatin of active enhancers (MacArthur et al., 2009). A survey of possible ‘off target’ binding elements suggested that these tend to be of lower affinity and are transcriptionally inert (Fisher et al., 2012). As noted above, previous studies of Hairy suggested that the protein is unable to mediate transcriptional repression in the absence of other factors co-occupying regulatory elements (Nibu et al., 2001).

Identification of functional properties of Hairy transcends the simple biochemical elucidation of repression; this protein is representative of the regulatory factors comprising conserved gene regulatory networks (GRN) that constitute the basis of animal development. Molecular studies have demonstrated that the acquisition or loss of binding sites or entire regulatory modules appears to drive significant changes in gene expression that initiate critical evolutionary transitions, such as elaboration of novel limb structures (Khila et al., 2009; Pavlopoulos et al., 2009; Tanaka et al., 2011). Significantly, although relatively subtle changes have been linked to such important evolutionary innovations, it appears that functional conservation of gene expression is also compatible with major changes in the structure of transcription control regions (Hare et al., 2008). The constraints for reorganization of existing cis-regulatory elements, or appearance of such elements de novo, are poorly understood; in some cases, the exact placement of multiple transcription factor motifs is essential for transcriptional function, while the composition of other genetic switches appears to be very loosely organized (Arnosti and Kulkarni, 2005). The existence of a large fraction of ‘off-target’ binding sites both complicates the analysis of important functional links, and the interpretation of potential evolutionary changes. Thus, elucidation of the functional targets and chromatin effects of Hairy can provide important insights on the basic substance of evolutionary variation. In this study, we use genetic tools to mediate induction of Hairy on a short time scale, permitting us to identify direct regulatory targets and chromatin effects on a genome-wide level. In addition to identifying common features of Hairy repression mechanisms across many targets, we also show that this protein exerts pervasive biochemical activity to change chromatin states at many loci unlinked to gene expression, revealing a possible pathway to evolution of novel gene regulatory connections.

Results

Genome-wide transcriptional regulation by Hairy

To study transcriptional repression at the genome-wide level at this important developmental stage, we profiled changes in transcriptome, epigenome and RNA polymerase II (Pol II) binding regulated by Hairy in the blastoderm embryo using an inducible system as described previously to capture direct effects with high temporal resolution (Li and Arnosti, 2011) (Figure 1A). Hairy is first expressed in the Drosophila blastoderm embryo in a seven stripe pattern, which is important in controlling downstream pair rule genes that direct segmentation (Ish-Horowicz and Pinchin, 1987). Here, we express Hairy with a brief heatshock, throughout the embryo, which is sufficient to completely repress target genes such as ftz (Figure 1A,B). We treated the control embryos identically to embryos carrying the inducible Hairy transgene to test for possible nonspecific effects of heat shock on gene expression and chromatin marks. In this system, heat shock alone has no effect on the expression patterns of the pair rule and other genes analyzed, and the chromatin marks in heat shocked control embryos were indistinguishable from chromatin patterns previously reported for untreated embryos (Li and Arnosti, 2011 and K Kok, data not shown). In total, we identified 241 down-regulated and 146 up-regulated transcripts in response to induction of Hairy (Figure 1C). Our microarray analysis captured previously identified targets of Hairy, showing downregulation of en, edl, Impl2, and prd, as well as ftz, all of which were previously found to be derepressed in h embryos (Ish-Horowicz and Pinchin, 1987; Bianchi-Frias et al., 2004).

Figure 1 with 1 supplement see all
Global analysis of Hairy regulation.

(A) Schematic expression of Drosophila embryo system used for Hairy repression, with outline of the genome-wide analysis of transcription, chromatin, and RNA polymerase II (Pol II). (B) Repression of ftz, odd, comm and esg revealed by in situ hybridization in wild-type (wt) and Hairy transgenic embryos (hs-hairy) after 20 min induction. Similar repression of 18w, HLHm7 and erm was also observed (not shown). (C) Transcriptionally regulated (red, down; blue, up) and Hairy bound genes identified by microarray and ChIP–chip (MacArthur et al., 2009). A larger fraction of down-regulated genes were physical targets of Hairy than for up-regulated genes (significance: p = 3.8e-95 and p = 3.5e-08 respectively, hyper-geometric test). Differentially expressed genes are selected based on p < 0.05 and fold change >2. (D) Validation of microarray data by RT-qPCR, showing concordance between these methods. Genes are ranked by the fold change from the microarray measurements. Significance was tested by Student's t-test. y-axis values were normalized as described in ‘Materials and methods’.

https://doi.org/10.7554/eLife.06394.003

Differentially regulated genes were compared to those physically bound by Hairy (MacArthur et al., 2009); 70% of down-regulated genes are bound by Hairy, suggesting that most of these are likely to be direct targets (Figure 1C). In contrast, only 30% of up-regulated genes are bound by Hairy, indicating that majority of these genes may be indirect targets. In situ hybridization and RT-qPCR confirmed the repression of a number of target genes we identified (Figure 1B,D). Many of these genes, including odd, comm, comm2, edl, en, Impl2, prd, and 18w, have striped expression patterns complementary to that of Hairy, supporting direct regulation by the repressor. Furthermore, consistent with known biological functions of Hairy, gene ontology analysis showed that categories for down-regulated genes are significantly enriched in transcriptional regulation, cell fate commitment and neurogenesis (p < 3.7e-18). GO categories for the set of upregulated genes were of lower statistical significance, and included reproductive processes (p < 0.03) (Supplementary files 1, 2).

Expression of the majority of genes bound by Hairy did not change (Figure 1C), consistent with previous observations that metazoan transcription factors have apparently many ‘nonfunctional’ interaction sites in the genome (Cusanovich et al., 2014).

Coordinate chromatin transitions mediated by Hairy on diverse genes

Identification of functional and physical Hairy targets allowed us to study gene-specific chromatin changes associated with repression. We performed epigenomic profiling via chromatin immunoprecipitation-high throughput sequencing (ChIP-seq) of chromatin marks that are often correlated with specific features of cis-regulation; H4Ac, H3K27Ac, and H3K4me1 at promoters and enhancers; H3K4me3 at transcription start sites (TSS); H3K36me3 at gene body regions; and H3K9me3 at repressed regions of chromatin (Zhou et al., 2011). The measured signals for specific marks were highly reproducible in separate biological replicates, and Hairy-induced changes in histone marks were consistently observed at specific loci, such as the widespread loss of the H4Ac signal on the ftz locus, with little change to the overall global chromatin landscape (Figure 2—figure supplement 1A). As was apparent from comparison of control chromatin profiles, the induction of Hairy did not cause a global impact on histone marks. In the presence or absence of induced Hairy, the genome features for multiple chromatin marks are virtually identical, except in very discrete regions where there are significant changes (Figure 2—figure supplement 1A–C).

Figure 2 with 1 supplement see all
Examples of coupled, large-scale chromatin changes mediated by Hairy.

Chromatin immunoprecipitation-high throughput sequencing (ChIP-seq) tracks for H4Ac, H3K27Ac, H3K4me1 and H3K4me3 are shown at repressed genes before (−) and after Hairy (+) induction, with gene models below. (AD) Coupled reduction of active histone marks was observed in a wide-spread fashion on ftz, h, 18w and odd genes (scale at top left). (EH) Relatively smaller blocks of chromatin changes were detected on HLHm7, gogo, pros and tup genes. Significantly changed regions (shaded boxes) were identified by the diffReps program. Hairy binding (top track) from MacArthur et al. (2009).

https://doi.org/10.7554/eLife.06394.005

Using single gene techniques, we previously found that Hairy induces a widespread histone H4 deacetylation throughout the entire ftz locus (Li and Arnosti, 2011). To determine if these are general properties of Hairy, we compared all affected loci genome-wide. We observed that on a number of transcriptionally repressed target genes, H4 deacetylation is coupled with loss of the active marks H3K27Ac and H3K4me1. Widespread reduction of these active marks affecting >1 kb blocks was observed on many genes repressed by Hairy, including ftz and other segmentally expressed genes such as h and 18w (Figure 2). Notably, Hairy regulates its own transcription by chromatin alteration, consistent with autoregulatory mechanism of related mammalian HES proteins (Kageyama et al., 2007). In addition to removal of enhancer marks, repression on h and 18w resulted in demethylation of the promoter mark H3K4me3. Furthermore, action of Hairy on another pair rule gene, odd, was limited to removal of acetyl marks on H4 and H3K27; methylation marks on H3K4 are untouched (Figure 2D). These results suggest that Hairy mediates coordinated sets of chromatin transitions. The chromatin changes did however exhibit heterogeneous characteristics; the sizes of altered chromatin domains varied on different repressed genes. For example, changes in levels of H4Ac involved blocks with a range of sizes; generally larger than 1 kb, with the average ∼2.5 kb. Somewhat smaller chromatin blocks were associated with repression of the HLHm7, gogo, pros and tup genes, which showed just as robust regulation of transcription as those genes with large tracts of chromatin modification (Figure 2E–H).

The largest ranges of size in chromatin domains were observed for H4Ac, but similar, although smaller ranges were also seen for H3K27Ac and H3K4me1 marks (Figure 3, Figure 3—figure supplement 1A,B and Supplementary file 3). We found strong correlations between the sizes of the domains of chromatin modification and the direct action of Hairy. Hairy-bound blocks of deacetylation were significantly larger than those not bound by Hairy, and smaller correlations were noted for other modifications, indicating that deacetylation is especially likely to show ‘spreading’ characteristics (Figure 3, Figure 3—figure supplement 1A,B and Supplementary file 3).

Figure 3 with 2 supplements see all
Direct Hairy target genes exhibit broad domains of chromatin effects.

Distribution of genome-averaged ChIP-seq signals before (straight line) and after (dashed line) Hairy induction, showing 4 kb window around affected regions. (A) Distributions of histone H4Ac and H3K27Ac marks of direct Hairy targets were significantly broader than for regions (B) not bound by Hairy (p = 2.55e-92 and p = 5.63e-70 respectively; KM test).

https://doi.org/10.7554/eLife.06394.007

These results suggest that widespread effects found at H4Ac, H3K27Ac and H3K4me1 marks are dependent on presence of Hairy and are consistent with a long-range ‘spreading’ repression mechanism. We saw no correlation between the height or extent of Hairy binding sites and the range of chromatin alteration, suggesting that the effectiveness of this protein is not merely a function of number of binding sites (Figure 3—figure supplement 2A,B). Other local factors may dictate how extensively modifications are propagated on individual genes. Therefore, Hairy induces diverse chromatin transitions associated with gene silencing, indicating that there are gene-specific features dictating how repression is mediated at individual genes.

Global set of chromatin modifications mediated genome-wide

These observations suggest there are context-specific aspects to chromatin modifications directed by Hairy. To determine the nature of changing chromatin states at different genomic loci, we compared the complete set of significant alterations in all measured chromatin marks observed after Hairy induction, regardless of transcriptional effects on the neighboring genes. We observed both loss and gain of these marks on hundreds of regions. Most frequently observed were changes in H4Ac, H3K27Ac, H3K4me1 and H3K36me3; changes in some chromatin marks were much more frequent than in others, indicating that there is some heterogeneity in the impact of Hairy on different regions (Figure 4A). The changes in levels of these marks is not simply due to increased or decreased histone density, as histone H3 levels generally were unchanged (Figure 4A). The roughly equal abundance of regions showing loss or gain of acetylation and methylation would indicate that either secondary effects are common, or that Hairy may exert distinct biochemical activities on different loci. The correlation of Hairy-bound regions with repressed transcripts, as well as the association of Hairy binding with longer-range deacetylations, but not with increased acetylation, supports the idea that indirect effects are common. Indeed, focusing specifically on genes targeted by Hairy, we found that H4 histone deactylation was strongly enriched compared to acetylation gains, suggesting that deacetylations are direct effects (Figure 4B and Figure 4—figure supplement 1A). Further support comes from consideration of the actual Hairy occupancy of the chromatin blocks in question; there was significant correlation between Hairy binding and chromatin blocks exhibiting decreased, but not increased acetylation (Figure 4—figure supplement 2).

Figure 4 with 2 supplements see all
Pervasive genome-wide chromatin effects of Hairy.

(A) All reduced (top) and increased (bottom) chromatin marks in the genome for H4Ac, H3K27Ac, H3K4me1, H3K4me3, H3K36me3, H3K9me3 and H3 shown as heatmaps for 5 kb windows from the center of significantly affected regions before (−) and after (+) Hairy induction. The number of affected regions indicated below each mark. (B) Affected chromatin regions associated with Hairy-bound genes show preferential enrichments for H4Ac, H3K27Ac, and H3K4me1. All affected regions were assigned to closest genes, and those in the vicinity of Hairy-bound genes are shown. (C) Subset of modified regions from (B) that were linked to genes transcriptionally regulated by Hairy. Significance of enrichment for chromatin modifications shown in Figure 4—figure supplement 1A,B.

https://doi.org/10.7554/eLife.06394.010

With respect to another chromatin mark, changes in histone methylation revealed an unexpected and interesting trend. Both decreases and increases in H3K4me1 signals were significantly associated with Hairy-bound genes; decreases were especially found in those regions directly bound by Hairy (Figure 4—figure supplement 1A, Figure 4—figure supplement 2 and Figure 4B). At the same time, about one-quarter of the genes that were transcriptionally silenced by Hairy showed increases in H3K4me1, although these regions of increase did not overlap with Hairy binding. The increase in this mark may represent a reaction of proximal promoter chromatin to distal enhancer silenced by Hairy.

H3K36me3 modification is often associated with active transcription. We found a small fraction of transcriptionally regulated genes that exhibited changes in the mark upon transcriptional repression (Figure 4C and Figure 4—figure supplement 1B). These findings indicate that Hairy repression does not require H3K36me3 changes. Indeed, the direct effect of H3K36me3 on transcription is complex, as has been found for many other histone marks. For example, upregulation of KDM4A histone demethylase target genes in Drosophila occurs without increases in H3K36me3 (Crona et al., 2013). Similar studies with elongation factor Spt6 in Drosophila further indicate that Hsp70 gene expression is not correlated to H3K36me3 levels (Ardehali et al., 2009). In fact, H3K36me3 may in some contexts contribute to gene silencing, due to its presence in heterochromatic domains (Chantalat et al., 2011) and in other cases, removal of H3K36me3 is required to promote transcriptional elongation (Kim and Buratowski, 2007).

A smaller number of H3K9me3 regions were observed to change globally, or on genes that were associated with Hairy (Figure 4A,B). Very few repressed genes showed any alteration in this mark, thus it appears that repression mediated by Hairy does not require changes in such repressive histone modifications (Figure 4C), consistent with our previous report that repression on ftz did not change H3K27me3 levels (Li and Arnosti, 2011). Indeed, other studies have found that these marks are not always simply coupled to repression. For example, only a modest correlation between H3K9me3 and H3K27me3 levels and gene silencing was observed in human cells (Barski et al., 2007; Zhang et al., 2012). In the differentiation of T and B cells, only a small fraction of repressed genes ever acquire H3K27me3 (McManus et al., 2011; Zhang et al., 2012). Interestingly, H3K9me3 was found to be enriched in many active promoters and associated with transcriptional elongation in vertebrates (Vakoc et al., 2005; Squazzo et al., 2006).

Consequently, of the assessed modifications, it appears that Hairy predominantly works to modify acetyl and methyl marks of H4, H3K27 and H3K4 and represses gene expression primarily by eliminating active marks.

Of all chromatin regions impacted by Hairy, only a small number are associated with genes demonstrating measurable transcriptional changes (Figure 4C). Thus, it is striking that the majority of chromatin changes are decoupled from any detectable effect on gene expression (Figure 4—figure supplement 1B). For the many cases where chromatin effect was unlinked to changed gene expression, we observed extensive chromatin alterations associated with both silent and active genes. For example, chromatin transitions occur on transcribed genes not functionally repressed by Hairy, as seen on the pyr gene (Figure 5A). In this case, the gene may remain active because the necessary cis-regulatory elements are located distally and are still able to interact with the promoter and activate it. In other cases, chromatin changes flank silent loci; nht undergoes widespread deacetylation and demethylation even though it is silent during this developmental stage of embryos (Figure 5B). In some cases, binding and changing chromatin near inactive genes by Hairy in the blastoderm embryo may involve the interaction of Hairy with DNA elements that will become active at a later developmental stage, however, this seems unlikely in the case of nht, a testes-specific gene. Here, the physical binding by Hairy and subsequent impact on chromatin may represent ‘errant targeting’. Overall, chromatin changes were observed to correlate with over half of the regions bound by Hairy, suggesting that in most cases, this protein is biochemically active on chromatin, whether or not the changes lead directly to gene repression (Figure 5—figure supplement 1).

Figure 5 with 1 supplement see all
Examples of chromatin-modified loci unlinked to changes in gene expression.

(A) pyr is actively transcribed, and not significantly repressed by Hairy, (B) while nht is not expressed at this stage. ChIP-seq tracks for H4Ac, H3K27Ac, H3K4me1 and H3K4m3 are shown before (−) and after Hairy (+) induction.

https://doi.org/10.7554/eLife.06394.013

Hairy coordinates sets of modifications on preferred gene regions

The individual cases described in Figure 2 suggest that Hairy organizes a coordinated set of chromatin changes involving both deacetylation and demethylation of multiple histone residues. To determine if such alterations are a general property of the repressor, we assessed the extent of coordination of modifications on all individual blocks of affected chromatin. Changes in H4Ac, H3K27Ac and H3K4me1 marks were significantly correlated at many loci (Figure 6A). Deacetylation events were also strongly correlated with loss of both H3K4me1 and H3K4me3, indicating that Hairy may form complexes containing both deacetylase and demethylase activities. Indeed, the CtBP cofactor is known to bind both of these classes of enzymes. However, Hairy is not mediating only one average type of transformation; removal of methyl groups from H3K4me1 and H3K4me3 is catalyzed by distinct classes of enzymes; Hairy is likely to interact with both, allowing for removal of H3K4me1 marks on distal sites and H3K4me3 at TSS (Figure 6B). A very similar pattern of correlations between acetylation marks, and between acetylation and methylation marks was observed for regions with increased acetylation and methylation. These elements may represent to a large extent indirect targets of Hairy, as no significant overlap between Hairy binding and these modified regions was found (Figure 4—figure supplement 2).

Coordination in changes of specific chromatin modifications by Hairy.

(A) Very strong overlap between decreases in regions of H4Ac, H3K27Ac, H3K4me1 (heat map, upper left quadrant). Similar coordination between increases of H4Ac, H3K27Ac, H3K4me1 was noted (lower right quadrant). Combined increases and decreases of different marks were rarely observed. (B) Distribution of modified blocks by genomic regions show preferential action of Hairy at a distance from transcription start site (TSS). Affected regions were mapped to intergenic regions, promoter, exon etc. Overall distribution of genomic peaks for measured marks shown at right; the distributions for affected H4Ac and H3K27Ac regions deviated from the genomic averages (left, decreased, and center, increased levels).

https://doi.org/10.7554/eLife.06394.015

Where does Hairy most commonly mediate significant chromatin modifications? We compared the location of individual histone marks genome-wide to those altered by Hairy expression. Although a third of Hairy binding sites are promoter-proximal, where the majority of H4 and H3K27 actetylation occurs, the large majority of affected chromatin sites were found on intergenic and intronic regions, suggesting that successful alterations are targeted to distal sites that may represent transcriptional enhancers (Figure 6B). By contrast, changes in the methylation marks H3K4me1, H3K4me3, H3K36me3, and H3K9me3 are found in the genomic regions where they are naturally enriched (Figure 6B). For instance, H3K4me3 marks are enriched at TSS, as are the bulk of the altered chromatin sites. Hairy may thus have privileged sites on which it is more likely to induce chromatin changes; promoter regions may be in general more resistant to acetylation changes if strong activators are replenishing acetylation marks at these loci. In addition, transcriptional targets of Hairy are enriched in developmentally regulated genes, which typically possess larger cis-regulatory regions with multiple distal enhancers (Supplementary file 1) (Nelson et al., 2004).

RNA Pol II and silencing by Hairy

To directly assess the influence of Hairy on transcriptional machinery, we compared the genome-wide occupancy of RNA Pol II before and after Hairy induction. 75 of 241 repressed genes exhibited changes in Pol II occupancy (Figure 7A). Only three of those are not directly bound by Hairy, indicating a direct regulation by Hairy in loss of Pol II signal. A marked decrease of Pol II occupancy was observed at the ftz promoter, gene body and distal downstream region (Figure 7B). Loss of binding at the promoter, or the body of the gene, or both was detected on other loci (Figure 7C–I). Thus, the loss of Pol II on the promoter and gene body of ftz is not universally associated with transcriptional repression; on other genes, silencing of a distal enhancer may interfere with promoter release without blocking polymerase recruitment to the promoter, consistent with recent studies implicating transcriptional signaling in promoter escape, rather than promoter recruiting (Lagha et al., 2013). As expected, genes with associated chromatin changes without any impact on transcription did not show any change on Pol II occupancy (Figure 7J,K).

Diverse impact on RNA Pol II occupancy by Hairy.

(A) A minority of genes show significant changes in Pol II occupancy after Hairy repression, although a larger proportion of the directly targeted genes have measureable decreases in Pol II. ‘Repressed genes’ shows entire set of transcriptionally downregulated genes, with reduced Pol II occupancy shown in dark gray. Subsets of genes directly bound or not bound by Hairy shown in center and at right. (BI) Pol II occupancy on transcriptionally regulated genes before (−) and after (+) Hairy induction. Pol II occupancy decreases in the promoter and gene body of ftz and odd, only on the gene body of h, 18w and pros, and only at the promoter of HLHm7 and gogo. Pol II signal was not changed significantly on tup. (J, K) Consistent with lack of transcriptional effects on other genes with associated chromatin modifications, Pol II occupancy on pyr is not changed, and absent on nht.

https://doi.org/10.7554/eLife.06394.016

95 repressed genes bound by Hairy did not show any change in Pol II occupancy (Figure 7A). It is possible that Hairy induces a slower transit rate of Pol II without any detectable change in Pol II binding. It has been suggested that repression through elongation control may cause no change in Pol II binding on slp1 and Hsp70 (Adelman et al., 2006; Wang et al., 2007; Ardehali et al., 2009). Our previous analysis of eve repression by short-range repressor Knirps showed similar effects (Li and Arnosti, 2011). Therefore, Hairy may interfere with gene expression at different steps of the transcription cycle, as also suggested for repression by the glucocorticoid receptor, indicating gene specific repression mechanisms (Gupte et al., 2013). An additional consideration is that genes featuring poised polymerase at the promoter in many or most nuclei, but are only expressed in a few nuclei, will have weak signals at the body of the gene. Therefore, the lack of change in Pol II levels on the gene body would reflect the inherently low signal, rather than a distinct biochemical mechanism. This explanation may account for a considerable number of affected genes where no changes in Pol II levels are observed after repression.

Predicting a ‘successful’ repression context

The complexity of chromatin transitions observed genome-wide in the wake of Hairy expression prompted us to ask which features best predict successful repression of a target gene, vs those genes with no chromatin responses or exhibiting errant targeting by Hairy. Here, we alter the expression of only one regulatory factor, rather than the many changes in regulatory factors observed over a developmental time course, therefore our data sets are enriched for direct action of Hairy, potentially simplifying the search space. We sought out correlations between dynamic histone marks, Pol II, Hairy, CtBP and Gro and the repression of targeted genes. Direct inspection reveals that occupancy by Hairy, Gro, and decreases in Pol II are strong predictors of repression, as are several histone marks, compared to genes unaffected or those activated (Figure 8A). However, there are numerous loci that do not fit these simple generalizations. To more systematically assess the connections between these different observed states and transcriptional repression, we applied machine learning to analyze features that may be implicated in the activity of Hairy. We tested 41 features, including the number of observed peaks for Hairy, CtBP, and Gro; the number, width, and magnitude of altered chromatin blocks, and distance to TSS for 583 genes (241 repressed, 146 activated and 196 unaffected genes; activated and unaffected genes were grouped as nonrepressed genes). To identify the most informative features, four different feature selection algorithms were used to rank the information content of the 41 measured properties associated with the genes; the top twenty of these features were then used for predictions (Supplementary file 4). We then tested four classifiers, using 90% of the data for training and 10% for predictions, with 10-fold cross-validation. Overall, each of the classifiers performed better than background, with Random Forests showing superior performance of ∼75% accuracy for repressed and nonrepressed genes (Figure 8B). Three of the feature selection algorithms used with this classifier employed very similar features to achieve this high level of accuracy (Supplementary file 4), indicating that certain features are most informative. The presence and properties of Hairy and Gro peaks are good indicators, although not sufficient information by themselves. RNA Pol II properties, transcript levels, and chromatin modifications, especially H3K4me1 and H4Ac, whether causal or not, are also a close reporter of gene activity. The overall performance differences in these methods are frequently observed in machine learning studies, and likely reflect the underlying data structure and types of features available for analysis. Genes that were correctly predicted as repression targets generally had the most differential features, including binding by Hairy and Gro, and changes in histone modifications. The genes that were least successfully called had one or no differential features, and may represent genes that are expressed in fewer cells and at lower levels where measurement of chromatin changes in a global population is difficult (Figure 8C). The nonrepressed gene pyr was consistently called as ‘repressed’ by the machine learning algorithms, as it exhibited chromatin signatures similar to those found on genes that were actually repressed (Figure 8C). In this case, we propose that the relevant enhancers lie outside of the chromatin regions affected by Hairy. Such genes may represent loci that are poised for capture in the Hairy regulatory network through stepwise acquisition of activator binding sites. Overall, this analysis indicates that from the perspective of Hairy biochemistry, there are intuitive and some non-intuitive combinations of chromatin dynamics that typify this protein's action in the context of transcriptional repression, rather than a ‘practice’ site, but other factors predominate in many instances. The missing information likely relates to the activity of bona fide cis- regulatory elements that are acting on genes in the vicinity of Hairy, which is partially but incompletely known from genome-wide studies (Kvon et al., 2014).

Machine learning reveals complex chromatin code for repression of Hairy target genes.

(A) Changes in histone marks, Pol II occupancy and Hairy, Gro and CtBP binding on repressed (red), activated (green) and unaffected (black) genes upon Hairy induction. Genes were grouped by change in expression, then subgrouped into Hairy bound or unbound, and finally ranked by fold change in gene expression. Activated and unaffected genes were grouped as nonrepressed genes. (B) Relative success rate at calling repressed and nonrepressed genes for four different machine learning models. Background prediction for this entire set is expected to be 58%; Random Forests, Naive Bayes, KNN classifiers had an average success of 75% overall, while the SVM classifier was not better than background. Classifiers were used in conjunction with Information Gain, Symmetrical Uncertainty, Chi Square and Relief feature selection algorithms. The average prediction accuracies of each method are shown in the first column. Expected random success (42%) for repressed genes (middle column) shown on heat map scale bar. (C) Model predictions for subset of repressed genes including those identified in Figure 1; top 19 were successfully predicted by almost all methods. fra, Optix, dib, and onecut were genes with disparate predictions that had few measureable chromatin features. At bottom, uniform false ‘repressed’ calls for pyr, which was not transcriptionally repressed.

https://doi.org/10.7554/eLife.06394.017

Discussion

By testing direct effects of the Hairy repressor in the embryo, we conclude that this protein coordinates a stereotypical set of chromatin modifications, modulated by local context, that underlie its function as a long-range repressor. Most remarkably, these changes on chromatin impact large segments of the genome that are not directly relevant to gene expression in this developmental context. We speculate that these off-target activities may provide an easy entry point for evolution of novel regulatory switches (Figure 9). Our mechanistic analysis of Hairy provides insights into likely mechanisms of related HES factors, as well as other transcriptional repressors that serve as scaffolds for chromatin modifying complexes. Hairy interacts with the widely utilized cofactors Gro, CtBP, and the Sir2 HDAC, and here we provide for the first time a genome-wide picture of the biochemical activities of this archetypal repressor.

Pervasive biochemical activities on ‘off-target’ loci may represent molecular exaptations that generate novel edges between nodes of a standing gene regulatory networks (GRN).

Functional and nonfunctional interactions of Hairy with chromatin are depicted. (A) Hairy repression of target genes results in loss of active histone marks such as H4Ac, H3K27Ac, and H3K4me1 (dark gray peaks; gene x). Hairy interacts with many other nonfunctional targets where it carries out biochemical activities similar to those seen on transcriptionally controlled loci (gene y). The latter chromatin changes are inconsequential and unlikely to be evolutionarily selected. (B) Gain of activator sites in a region of Hairy-modified chromatin may generate an on/off switch and result in functional targeting. (C) Schematic representation of cooption of Hairy physical interaction into modified GRN.

https://doi.org/10.7554/eLife.06394.018

How is transcription actually controlled by Hairy? The associated chromatin modifications may be effects, rather than direct causes of gene silencing. Our previous studies indicated that Hairy modulated transcription independent of activator occupancy or SAGA co-activator occupancy (Martinez and Arnosti, 2008). These previous observations raised the possibility that Hairy acts through entirely independent pathways from that employed by activators to block transcription. Our work here indicates that Hairy does indeed directly reverse chromatin marks associated with activators, and may therefore work through a dynamic competition with these activators, undoing their positive influence on the chromatin environment that would be necessary for RNA polymerase initiation and/or elongation (Figure 9A). Indeed, Hairy repression is readily reversible, with genes showing reversion to an active state minutes after depletion of the overexpressed repressor (K Kok, unpublished results).

The genome-wide analysis of repression by Hairy revealed an unexpected facet of chromatin activity and highlights the need to consider the activity of ‘off target’ sites in generating novel elements, particularly because for Hairy at least (and likely other factors that employ the same cellular machinery) they are ‘shovel ready’ and not constrained by complex cis-regulatory grammar. Metazoan transcription factors typically interact with thousands of discrete sites in the genome, but only a small subset of these interactions correlate with observable effects on gene expression. In this study, we combined analysis of gene expression and chromatin dynamics in a way that allowed us to attribute effects directly to the induction of Hairy, inferences that would be difficult with a loss-of-function assay due to kinetics of depletion and secondary effects. In contrast, many other genome-wide data sets provide a static snapshot of the extant chromatin landscape or track complex changes through development, which represents the combined contributions of many activators and repressors. Previous studies have noted the presence of detectable but lowly-occupied sites, which have been suggested to reflect non-specific, non-functional interactions that are unavoidable by-products of proteins binding to large genomes (Fisher et al., 2012). Other studies have emphasized that transcription factors may have general nonspecific interaction with HOT sites that represent preferences for open chromatin (Gerstein et al., 2010; modENCODE Consortium et al., 2010). In general, the overall view is that whether or not these interactions are conserved, they may be of little functional consequence, and are not important for activity of GRNs (Cusanovich et al., 2014). Importantly, considering our finding that ‘off-target’ Hairy sites still appear to regulate chromatin structure, we should fundamentally reconsider how we interpret genome-wide data sets. Frequently, an increase in H3K27 acetylation is taken as an indication that the element is an active enhancer, without further functional tests (e.g., Villar et al., 2015). Of course, correlated gene expression measurements indicate that such elements are likely to be enhancers in many cases, but genomic consideration of chromatin marking must not automatically equate changes in certain active marks with enhancers.

Our study provides a new perspective on these previous observations, in that essentially trivial biological interactions may have consequences in evolutionary time. We show that Hairy is engaged apparently in errant targeting of chromatin on many loci during the period when it is expressed, and demonstrate that in many cases, little distinguishes the types of chromatin effects observed on functionally repressed targets compared to ‘non-functional’ interactions on other loci (Figure 9A,B). Thus, unlike an earlier model for Hairy action, in which the protein is active only when embedded in a previously active enhancer (Nibu et al., 2001), our work demonstrates that Hairy is able to mediate biochemical activities in most bound regions, indicating that there is little context necessary for the protein to function. Therefore, Hairy may be relatively nonselective about where it can attract chromatin-modifying agents across the genome. Much molecular biology research has emphasized the high degree of cooperativity necessary for metazoan transcription factors to work well. Enhanceosomes, patterning elements and other enhancers give aberrant readouts if correct stoichiometries and spacings are not respected. These findings suggest that random individual sites are less likely to generate a suitable transcriptional readout. At least for repressors such as Hairy, the demands for generating biochemical activity are lower than anticipated, indicating that enhancers may have a lower threshold for formation that we might have expected. Although some of the targeted genes that are not transcriptionally affected may represent ectopic binding events of the induced Hairy protein, most sites are found in ChIP analysis of endogenous Hairy. The unresponsive genes may in some cases represent later targets of Hairy, may be already repressed by endogenous Hairy, or may have responses too small to measure in this system, however it is likely that there are hundreds of changed chromatin regions that not formally part of the functional Hairy GRN. Thus, a large fraction of the genomic interactions are likely to be with regions that are not strongly selected on an evolutionary timescale. As long as the induced chromatin changes are inconsequential, these effects will not be selected against during genomic evolution. This biochemical activity, however, may provide a unique molecular exaptation to generate novel edges between nodes of a standing GRN (Figure 9B). Most enhancers involve the combined action of transcriptional activators and repressors, thus errant targeting may facilitate formation of new modules with gain of a few activator binding sites (Gould and Vrba, 1982) (Figure 9C).

Materials and methods

Plasmid construction

Request a detailed protocol

The heat-inducible hairy gene was created by introducing a multiple cloning site containing Kozak sequence, initiator ATG and HindIII/BglII sites into the 5′ portion of the hairy ORF in the pCaSpeR-hsh using EcoRI/BstEII sites as described previously (Li and Arnosti, 2011). 400 bp of upstream promoter, 5′ UTR, Kozak sequence, initiator ATG, HindIII/BglII sites, coding sequence and entire hsp70 3′ UTR from the modified pCaSpeR-hsh were amplified using 5′ and 3′ primers with AgeI/KpnI sites and subcloned to the modified pattB vector (Sayal et al., 2011). Oligonucleotides with sequence encoding the double Flag epitope, as described in Zhang and Arnosti (2011), was inserted 5′ of the coding sequence after the ATG using HindII/BglII sites, so that Hairy protein was expressed with the double Flag tag at the N terminus.

Embryo collection, in situ hybridization and antibody staining of Drosophila embryos

Request a detailed protocol

For chromatin analysis 2–3.5 hr embryos were collected and 20 min heat-shock treated for induction of transgenes as described previously (Li and Arnosti, 2011). We treated the wild-type embryos similar to embryos carrying inducible transgene to control for possible nonspecific effects of heat shock. Heat shock alone has no effect on the expression or chromatin patterns (data not shown). For analysis of gene expression by in situ hybridization, embryos were fixed and stained using anti-digoxigenin-UTP-labeled RNA probe for ftz as described previously (Struffi, 2004).

Quantitative reverse transcriptase PCR analysis

Request a detailed protocol

Total RNA from embryos was purified using RNeasy columns (Qiagen), and reversed transcribed using a High Capacity cDNA Reverse Transcription Kit from Invitrogen/Applied Biosystems. The cDNA was then analyzed by real-time PCR using the primer pairs located at transcription units. Data was normalized to act5c. Values for wild-type embryos were set to 1; results represent the average of 2–8 biological replicates. Statistical significance was tested using Student's t-test and p < 0.05. Amplicons were designed using Primer Express and Primer-BLAST.

Expression profiling analysis

Request a detailed protocol

Total RNA from 2–3 hr embryos was purified using RNeasy columns (Qiagen, Valencia, CA). Samples were amplified and labeled using the Quick AMP Labeling kit (Agilent, Santa Clara, CA) and hybridized to 8 × 15K Customized Drosophila Genome Oligo Microarrays (Agilent) according to the manufacturer's instructions. Slide image data was quantified using Agilent's Feature Extraction software. Four biological replicates were performed for each sample. Differential gene expression analysis was performed with the GeneSpring program (Agilent). Functional annotation of down- and up-regulated genes was done using the Database for Annotation, Visualization and Integrated Discovery (Dennis et al., 2003). Differentially regulated gene symbols and their fold changes are listed in Supplementary file 5.

Chromatin immunoprecipitation

Request a detailed protocol

Heat shocks and ChIPs were performed as described previously (Li and Arnosti, 2011), with the exceptions that embryos were sonicated for a total of 20 times using a Branson sonicator in 1 ml of sonication buffer. After precipitation of chromatin-antibody complexes, protein A beads were washed twice with low-salt buffer, once with high-salt buffer, once with LiCl buffer and twice with Tris-EDTA. We used the following antibodies: rabbit IgG (5 μl, Santa Cruz Biotechnology), rabbit anti-H3 (1 μl, Abcam, Cambridge, MA), rabbit anti-acetyl H4 (1 μl, Upstate, EMD Millipore, Billarica, MA), rabbit anti-acetyl H3K27 (1 μl, Abcam), rabbit anti-monomethyl H3K4 (1 μl, Abcam), rabbit anti-trimethyl H3K4 (1 μl, Abcam), rabbit anti-trimethyl H3K36 (2 μl, Abcam), rabbit anti-trimethyl H3K9 (3 μl, Abcam), rabbit anti-Flag (5 μl, Sigma-Aldrich, St. Louis, MO), rabbit anti-Rpb3 (5 μl, gift from Carla Margulies, LMU University of Munich).

ChIP-seq

Libraries

Request a detailed protocol

DNA from chromatin immunoprecipitation (10 ng) was adapter-ligated and PCR amplified (18 cycles) as described in Ford et al. (2014). DNA ligated to the adapter was size selected for 300–500 bp. Illumina HiSeq single-end reads were checked using FastQC and HOMER for sequence quality, base sequence and GC content, sequence duplication, sequence bias, overrepresented sequences and Kmer content. Reads were aligned to genome (BDGP 5.70) with Bowtie version 1.0.0 using--m 1--best parameters. Tags that only mapped uniquely to the genome were considered for further analysis. Summary of tags generated is shown on Supplementary file 6. ChIP-Seq experiments were visualized as custom tracks using Integrative Genomics Viewer (Broad Institute). Total uniquely mapped tags were normalized to 10 million reads to generate tracks. y-axis values shown in all figures use the same scale for an individual measurement of each histone modification in the individual panels. For reasons of clarity, scales can vary between different panels.

Mapping differential regions

Request a detailed protocol

We detected the regions where chromatin states are changed upon induction of Hairy by comparing the level of histone marks at particular genomic locations. Differentially changed genomic regions were identified using the diffReps program (Shen et al., 2013), which uses a sliding window approach to scan the genome and find regions showing read count differences. Default window size with--nsd broad--meth nb parameters was used for the analysis. For downstream analysis, we used regions with p < 0.05 and fold change (log2) > 0.4 or fold change (log2) < −0.4. Input was sequenced from nontransgenic (wt) and Hairy overexpressing embryos and used as background.

Hypergeometric Optimization of Motif EnRichment (HOMER) was used for peak finding and downstream data analysis (Heinz et al., 2010).

Identification of ChIP-seq peaks

Request a detailed protocol

Using HOMER with default settings, peaks for histone marks and Flag tagged Hairy protein were identified using signals from H3 ChIP and input respectively as background.

Annotation of significantly affected regions

Request a detailed protocol

Regions detected by diffReps or peaks called by HOMER were associated with genes by identifying the nearest RefSeq TSS and annotated to a genomic feature such as intergenic, intron, exon etc.

Normalization of ChIP-seq tags for histograms, heatmaps, and scatter plots: We normalized the total number of mapped tags to 10 million for each sample using HOMER so that the read densities were comparable.

Comparison of ChIP-seq experiments using histograms

Request a detailed protocol

ChIP-seq densities of a 4 kb window centered at affected regions detected by diffReps were determined using HOMER. The program normalizes the output histogram such that the resulting units are per bp per region with bin size of 10 bp. Plots were generated using matplotlib (Hunter, 2007).

Comparison of ChIP-seq experiments using heatmaps

Request a detailed protocol

Data matrices were generated using HOMER by counting total tags in a 5 kb window around affected regions or peaks and normalizing to 10 million reads with bin size of 25 bp. Data was visualized using Java Tree View (Saldanha, 2004).

Comparison of ChIP-seq experiments using scatter plots

Request a detailed protocol

Tag densities were calculated by counting the tags at regions defined by peak coordinates of the first experiment (x axis) and compared to the second experiment (y axis). Data was log2 transformed and plotted using matplotlib. Pearson's Correlation Coefficients were calculated to determine the extent of similarity between samples.

Analysis of co-occurrence of differentially changed regions

Request a detailed protocol

mergePeaks program of HOMER was used to find overlapping sites between differentially changed regions of different histone marks upon Hairy induction. These regions were considered as overlapped if changed regions from each experiment share at least 1 bp. Significance of co-occurrence of regions was indicated by natural log p-values using the hypergeometric distribution. Positive values signify divergence.

Linking affected regions to Hairy binding

Request a detailed protocol

Affected regions for chromatin marks were considered as Hairy bound if the nearest gene has at least one Hairy peak. The occupancy of the induced Hairy protein was compared to that of endogenous Hairy binding by conducting ChIP-Seq analysis using the Flag epitope on the inducible protein; a large fraction (40%) of these binding sites were also found in the ChIP–chip study (p = 2.15e-159). Similar Hairy binding motifs were enriched in both data sets, indicating that the induced Hairy protein has similar targeting specificity to the endogenous protein (Figure 1—figure supplement 1).

Machine learning analysis

Request a detailed protocol

We used the differential changes of H4Ac, H3K27Ac, H3K4me1, H3K4me3, H3K36me3, H3K9me3 and Pol II in response to Hairy as features for our analysis here. The genomic blocks detected as significantly altered by diffReps are annotated to closest TSS. We considered four features for each of the ChIP-seq data; number of blocks linked to the same gene, range of blocks, fold change of ChIP-seq signal at blocks, and distance of blocks to closest TSS. Four features from ChIP–chip data sets of Hairy (MacArthur et al., 2009), CtBP, and Gro (Nègre et al., 2011) were used; number of peaks linked to the same gene, width of peaks, peak signal, and distance of peaks to closest TSS. In addition, expression of transcripts in wild-type embryos was included as a feature. In total, these 41 features were collected for 583 genes (241 repressed, 146 activated and 196 unaffected genes; activated and unaffected genes were grouped as nonrepressed genes) in this study. Differentially regulated genes and their fold changes are listed in Supplementary file 5 and randomly selected unaffected genes are listed in Supplementary file 7. Important features were first identified with four feature selection algorithms (Information Gain, Symmetrical Uncertainty, Chi Square and Relief). Then, to predict genes in the repressed and nonrepressed categories, four classifiers (Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Naive Bayes and Random Forests) were employed. To perform this analysis, we wrote Python and Java codes to partition our dataset into 10 parts to perform feature selection and 10-fold cross validation classification utilizing the Weka machine learning software (http://www.cs.waikato.ac.nz/ml/weka/). To increase the robustness of our results we performed 50 iterations of the above procedure and combined the predicted classes for each gene to create a new aggregate predicted class for that gene. Here we took the class that has been predicted more than 50% of the 50 iterations as the predicted class of the gene. We have applied every combination of the four feature selection algorithms and four classification algorithms to the data to obtain the optimal classification methodology for our dataset. The results of our analysis are summarized in the main text.

References

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
    Long-range repression in the Drosophila embryo
    1. HN Cai
    2. DN Arnosti
    3. M Levine
    (1996)
    Proceedings of the National Academy of Sciences of USA 93:9309–9314.
    https://doi.org/10.1073/pnas.93.18.9309
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
  13. 13
  14. 14
  15. 15
  16. 16
    Exaptation-a missing term in the science of form
    1. SJ Gould
    2. ES Vrba
    (1982)
    Paleobiology 8:4–15.
  17. 17
  18. 18
  19. 19
  20. 20
  21. 21
  22. 22
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
  28. 28
  29. 29
  30. 30
  31. 31
  32. 32
  33. 33
  34. 34
  35. 35
  36. 36
  37. 37
  38. 38
  39. 39
  40. 40
  41. 41
  42. 42
  43. 43
  44. 44
  45. 45
  46. 46
  47. 47
  48. 48
  49. 49
  50. 50
  51. 51
  52. 52
  53. 53
  54. 54

Decision letter

  1. Asifa Akhtar
    Reviewing Editor; Max Planck Institute for Immunobiology and Epigenetics, Germany

eLife posts the editorial decision letter and author response on a selection of the published articles (subject to the approval of the authors). An edited version of the letter sent to the authors after peer review is shown, indicating the substantive concerns or comments; minor concerns are not usually shown. Reviewers have the opportunity to discuss the decision before the letter is sent (see review process). Similarly, the author response typically shows only responses to the major concerns raised by the reviewers.

[Editors’ note: this article was originally rejected after discussions between the reviewers, but the article was accepted after an appeal against the decision and further revisions.]

Thank you for choosing to send your work entitled “Genome-wide futile cycling by Hairy repressor suggests mechanism for evolution of gene regulatory networks” for consideration at eLife. Your full submission has been evaluated by Diethard Tautz (Senior editor), a Reviewing editor, and three peer reviewers, one of whom, Michael Eisen, has agreed to share his identity. The decision was reached after discussions between the reviewers. Based on our discussions and the individual reviews below, we regret to inform you that your work will not be considered further for publication in eLife.

This manuscript provides an extensive ChIP-seq data set to study how Hairy regulates gene expression. Using a Hairy over-expression system, authors find that Hairy effects gene expression in a context dependent manner on a subset of target genes. However, on the majority of the bound sites Hairy is able to induce chromatin changes without altering gene expression. Based on these and other observations, authors also speculate about how novel transcription circuits could evolve.

The overall consensus among the reviewers was that although this manuscript does contain some novel observations, it does not provide sufficient mechanistic insights that would explain why only a fraction of Hairy bound sites show a functional correlation and how Hairy induces widespread chromatin changes without changes in gene expression. The fact that the major conclusions are based purely on an over-expression system also poses a major problem. For example, it would be advisable that the authors use other experimental approaches such as depletion experiments to further potentiate their results. The reviewers were also not convinced by evolutionary claims made in the manuscript. The manuscript also lacked a number of controls that made it difficult to assess the quality and strength of the observations.

Reviewer #1:

This is the latest in a series of papers from the Arnosti lab looking at the mechanism of action of the D. melanogaster early embryonic transcriptional repressor Hairy. This paper uses a lot of ChIP-seq data to make a fairly simple point: when Hairy binds to sites where it does not appear to be regulating transcription it nonetheless induces changes in chromatin that are similar to the changes it makes when it does have an effect on transcription. This paper makes a valuable contribution. The main point the authors seem to be trying to make – that Hairy affects chromatin even when it's not affecting expression – is lost in the clutter. I think this could be an eLife paper, but not in its current form.

What they did:

Used a previously generated transgenic line that expresses Hairy ubiquitously in the embryo upon heat shock, and characterized changes in expression in the Hairy-ub line compared to wt. Compared gene expression in induced Hairy-ub line to control (unclear what control was). Carried out ChIP-seq with H3, H4Ac, H3K27Ac, H3K4me1 and H3K4me3 on wt and Hairy-ub embryos. Carried out similar experiments using a mutant form of Hairy that does not interact with CtBP.

What they claim:

Hairy binding is associated with transcriptional repression, histone deacetylation and changes in histone methylation.

The effects of Hairy are mediated to some extent by its interaction with CtBP.

There are significant alterations upon induction of Hairy in regions where the binding of Hairy does not seem to effect transcription.

Genes could come under Hairy control by the acquisition of activator binding to regions already poised for Hairy repression (although they offer no evidence for this claim).

Concerns:

1) Maybe I'm missing something (though I've looked extensively), but as far as I can tell, nowhere in the paper does it say what the control for gene expression and ChIP experiments was. There are at least three possibilities I can think of, and all of them differ in important ways. Wild type embryos grown under normal conditions? Wild type embryos heat shocked in parallel with the transgenic ones? Transgenic embryos that weren't heat shocked?

Li and Arnosti 2011 use heatshocked wild type embryos, but this also wasn't discussed much in the paper. This paper needs a clear statement of what the control samples were and some indication of why that particular control was chosen.

2) A related issue is that the paper doesn't say in the Methods how long the heatshock was, other than that it was short. It's known that heatshock accelerates development. Since gene expression and chromatin marks at this time are known to be highly dynamic, any chance in the stages present with and without heatshock could produce anomalous results. Same could be true for changes that are the result of the induction of h. It would be nice to see some data on the stages represented in the two samples.

3) ChIP-seq data was normalized to 10,000,000 mapped tags for each sample. However it's easy to imagine that altering the levels of an important transcriptional repressor across the whole embryo could have a global impact on the levels of some of these marks. This really calls for some kind of reference standard that would allow a quantitative comparison of overall levels not just their relative distributions, which is what is being done here.

4) There are many claims made associated with Figure 3, and elsewhere, that Hairy binding is associated with the increase/decrease of some measurement. For example, the claim that there is a significant association of Hairy binding with H4 deacetylation as compared with H4 acetylation as compared with the genome-wide changes in H4 acetylation state. This is based on the observation that, genome-wide, there are a roughly equal number of gains vs. losses of H4 acetylation, but there is a strong bias towards deacetylation in Hairy bound regions. However, it matters a lot what the distribution of H4 acetylation is in Hairy bound regions in controls. If Hairy bound regions tend to have relatively high levels of H4 acetylation, then, even if there is no association between Hairy binding and the change in H4 acetylation upon induction of Hairy-Ub, you would still expect to see Hairy-bound regions go down more often than up because of where they started. The lack of conditioning on the starting state of Hairy-bound regions (as well as possible differences in their size, etc.) is a serious problem with the analyses presented in the paper.

5) I didn't really get the point of the machine learning section. I'm not saying it should be removed, just that the authors should try to explain why they did it and what they learned from it a bit better.

6) I'm not a fan of the use of the term “futile cycling” here, since there is no direct observation of cycling here. What the authors have shown is that the ectopic induction of Hairy leads to an alteration of chromatin state in places where Hairy binding doesn't seem functional. However it seems quite possible that what is happening at these sites is that Hairy is preventing the creation of a chromatin state that would otherwise occur in its absence, rather than altering a previously existing chromatin state. And thus it's not really cycling.

7) I'm all for speculation in papers, but the evolutionary model could use some fleshing out – especially what makes it interesting and novel. It seems pretty obvious that you could make a new enhancer by adding an activator to a previously existing repressor. This idea that enhancer evolution is easier if you add an activator to a repressor rather than a repressor to an activator has been around for as long as I can remember.

Reviewer #2:

In this manuscript, Kok and colleagues report the genome-wide effects of misexpression of the transcriptional repressor Hairy on five histone modifications, RNAPolII binding, and gene expression in Drosophila embryos. One of the main findings is that the effects of Hairy binding are context-dependent: some genes that are bound by Hairy are repressed, and many others are not. This is not surprising, in fact it is true of every transcription factor for which regulated genes and genomic binding sites have been compared (see the papers by Fisher and MacArthur cited in the manuscript, or any other ChIP-chip or ChIP-seq study). What is much more surprising is that the H3K36me and H3K9me3 modifications, which are typically strongly associated with actively expressed and silenced genes respectively, were not correlated with Hairy-induced changes in gene expression in this study. The authors suggest that Hairy works through non-canonical mechanisms to regulate gene expression without changing the characteristic markers of gene expression and repression. If so, this is by far the most novel and significant finding of the study, because it overturns the conventional view of the distribution of these histone modifications. However, the authors quickly leave this aspect of the analysis and return to the other modifications, some of whose changes make more conventional sense – assuming that, unlike H3K36me and H3K9me3, their presence or absence can be interpreted in the usual way.

Another unexpected finding is that most genes that are transcriptionally affected by Hairy overexpression (either directly or indirectly) show no change in PolII binding, either at the promoter or in the gene body. A lack of binding at the promoter could be explained by the idea of “promoter escape” mentioned by the authors (subsection “RNA polymerase II and silencing by Hairy”), but the lack of change of PolII in the gene body cannot. It is possible that these targets are regulated post-transcriptionally (and therefore indirectly), so that transcription is unaffected (which would be consistent the lack of change in PolII, H3K36me and H3K9me3) but RNA levels are reduced (which would be consistent with the microarray data)? Or could the result be due to the significant heterogeneity of the cells in the whole-embryo samples?

Another difficulty with the study is the significant problem of direct vs. indirect regulation by Hairy. The authors note that over 91% of Hairy-bound genes are not repressed, and a significant fraction of repressed genes (31%) are not bound by Hairy. They consider the 167 genes that are both bound and repressed to be directly repressed by Hairy, but this is probably a significant overestimate. Since the vast majority of Hairy binding sites are apparently nonfunctional, is it safe to assume that all of the Hairy sites in down-regulated genes are functional, especially since many other genes are regulated indirectly? How many of these binding sites in the 167 regulated genes are located within active enhancers? How many of the known direct target enhancers of Hairy (not the genes, the enhancers) are bound in this experiment? Compounding the uncertainty, the Hairy binding sites shown in the figures are not from the over-expressed hs-hairy used in this study: they are ChIP-chip data from the endogenous protein (MacArthur 2009). Only 40% of peaks are shared between the MacArthur dataset and this study (subsection “ChIP-seq”, last paragraph), suggesting that the distribution of over-expressed Hairy is very different from that of endogenous protein, not unexpectedly since hs-hairy is expressed in all cells of the embryo, most of which do not normally express Hairy. The measurements in this study, unavoidably, are averages across all cells of the embryo, so it is problematic to correlate a change in gene expression with a lack of change in chromatin modifications: these may be occurring in different cells.

Another concern is that the authors explain an up-regulation of HLHm7 by H-mut-CtBP (Figure 6) by proposing “antagonism of the endogenous wild-type Hairy by the mutant” (subsection “CtBP supports Hairy repression activity”, second paragraph). If the authors are correct, this totally nullifies the other conclusions drawn from the experiment.

The attempt to predict repressed vs. non-repressed genes using machine learning does not seem to have produced any useful insights, or at least those insights were not made clear in the text (subsection “Predicting a “successful” repression context”) or in Figure 7.

The Discussion section is far too bold in its claims, relative to the results, especially in its evolutionary speculations. To flatly state that “essentially trivial biological interactions have consequences in evolutionary time” is to go far beyond the data. That statement may be true, it probably is true at least some of the time, but it is not a finding or a reasonable interpretation of this study, and there is nothing in this study to support it. I am also confused by the concept of “futile cycling”: I'm not clear on exactly what is cycling or being cycled. If it is chromatin, is there any evidence for higher chromatin turnover dependent on Hairy binding? Another notable statement is that “little distinguishes the types of chromatin effects observed on repressed genes compared to ‘non-functional’ interactions on other loci” (Discussion, sixth paragraph). That is indeed a finding of the study, but at least for me, it is more likely to reflect a problem with the data than to prove that most of what we know about gene expression, silencing, and their associated chromatin modifications is wrong. If the latter is true, the authors have a tremendously important result on their hands, but that will require a much higher level of proof.

Finally, it is quite easy to propose that non-functional Hairy binding sites are “an off-the shelf module that only needs the addition of a few activator binding sites to emerge as a fully fledged cis-regulatory element.” However, (1) no experimental support for this idea, direct or indirect, is provided here or cited from other studies (mentioning that changes to enhancers can be evolutionarily important does not count as evidence in favor of this particular proposal), and (2) this evolutionary model does not depend on any of the results shown here, since widespread, apparently non-regulatory binding sites have been commonly observed since the first days of ChIP-chip (e.g., Holloway, Genome Inform 2005; Moses, PLOSCB 2006). This is a group of talented, rigorous, and thoughtful investigators, but in this case their large biochemical and evolutionary claims are not well supported by the data.

Reviewer #3:

Kok and co-workers have carried out an extensive genome-wide analysis of the effect of Hairy on the chromatin landscape in an effort to correlate these effects with transcriptional repression by Hairy. The ChIP-seq analysis in comparison with the transcriptome analysis shows that Hairy binds to many more genes than it represses. Analysis of the chromatin landscape shows that repression is largely associated with histone deacetylation and H3K4me1 demethylation in regions that include the Hairy binding sites and spread from these sites. Perhaps surprisingly, however, genes that are not repressed by Hairy, but that bind Hairy, often show similar changes in the chromatin landscape (what the authors call futile cycles of modification). This shows that deaceylation and H3K4me1 demethylation are not sufficient for repression and even suggests that they could be a consequence as opposed to a cause of repression. Various machine learning protocols are used in an effort to determine what features might be most predictive of repression and some moderate success is achieved. The authors propose that these non-functional Hairy binding sites could facilitate the evolution of new cis-regulatory modules through the acquisition of additional transcription factor binding sites. In general, I think this is a well-executed study that brings out some important new ideas. I have a few questions that I would like the authors to address:

1) The authors have carried out the analysis of Hairy targets by comparing Hairy-overexpressing with wild-type embryos. This has both advantages and drawbacks. The advantage, of course, is that it is technically feasible. But the disadvantage is that Hairy is being expressed in cells in which it is never normally present and that this abnormal context could have functional consequences that we can't easily account for. This could potentially account for the lack of correlation between changes in the chromatin landscape and repression. While the ChIP-seq analysis is probably technically out-of-reach in loss-of-function mutants, have the authors considered at least carrying out the RNA-seq analysis in loss-of-function (RNAi knockdown?) embryos. It would be worth knowing how many of the repression targets identified by overexpression are also identified by loss-of-function.

2) I found the lack of correlation between the Pol II ChIP and repression to be puzzling. The authors have not clearly explained how they account for this. Has this lack of correlation between Pol II ChIP and transcription levels been observed in other contexts?

3) I don't fully understand the model in Figure 8. It needs to be explained in the text.

[Editors’ note: what now follows is the decision letter after the authors submitted for further consideration.]

Thank you for resubmitting your work entitled “Genome-wide errant targeting by Hairy” for further consideration at eLife. Your revised article has been favorably evaluated by Diethard Tautz (Senior editor), a Reviewing editor (Asifa Akhtar), and two reviewers. The reviewers appreciated that the manuscript has significantly improved upon revision and therefore we would like to go forward with the manuscript. However, before acceptance we would like you to address the remaining concerns (below) adequately in the text so that the readers are aware of the limitations and the conclusions that can be drawn an over-expression system.

Reviewer #1:

I think the authors did an excellent job addressing our concerns. The revised manuscript is far more direct and readable, and I think paints a more accurate picture of what the data do and do not say. There are aspects of it that I would change if it were my manuscript, but I don't think it's the role of reviewers to shape manuscripts entirely to our tastes, so I will refrain from pointing them out.

I think there one overarching question remains, which is that, to me, none of what they report in the manuscript is surprising. It's very much consistent with lots of data from lots of labs over the past decade. However, what the authors show here has not been demonstrated experimentally yet, and, as such, I think it is an eLife level contribution. But I can see it either way.

Reviewer #3:

The authors have addressed many of the criticisms in the initial review through additional data analysis and reference to the literature. I have two concerns:

1) The original review asked for loss-of-function data to confirm the results of the overexpression studies. In response, the authors cited previous literature showing that hairy loss of function results in derepression of many of the genes that are repressed upon hairy overexpression. However, this does not address the more important point that pertains to the genes that exhibit changes in the chromatin marks and yet show no repression upon hairy overexpression. This is the class of genes upon which the claim to novelty is based and therefore, this is the class of genes whose behavior needs to be verified in loss-of-function studies.

2) I find the machine learning section of the proposal a little puzzling. On the one hand the authors claim that repressed and non-repressed genes often show similar changes in histone marks upon Hairy overexpression. However, the machine learning section seems to suggest that the changes in chromatin marks are not really the same for repressed and non-repressed genes. How do the authors reconcile these apparently disparate claims?

https://doi.org/10.7554/eLife.06394.026

Author response

[Editors’ note: the author responses to the first round of peer review follow.]

Thank you very much for guiding the editorial process of our recent manuscript. We read the reviews carefully, and discussed the main points with two individuals who agreed to be identified. The comments were quite helpful in focusing and clarifying the work; additional analysis on our part was able to address each of the points raised in the review process. […] A major new point that came from responding to questions in this review process was the consideration of how the “errant targeting” and biochemical activity of Hairy, as well as potentially many other factors, calls into question our assumptions in calling enhancers from genomics datasets. This, in addition to consideration of evolution of CREs, will be of great interest to eLife readers.

We address the comments from Reviewers 1, 2 and 3 below.

Reviewer 1 found the presentation complicated and confusing; in part because we presented two parallel narratives. One dealt with the implications of genome-wide chromatin modifications by Hairy, while a separate one focused on the biochemistry of Hairy and role of the CtBP corepressor. We agree with the reviewers that the main focus was on the former story, and accordingly we changed the structure of the paper to reflect this focus, removing data about analysis of the CtBP binding mutant which overly complicated the presentation. To facilitate presentation of our extensive chromatin modification data sets, we now present individual points in separate figures (Figures 2-3, Figures 4-5; previously combined in Figure 2 and 3 respectively). We also extensively revised the model in Figure 8 (now Figure 9) to make it easier to understand.

Reviewer 1 raised a concern about the controls used for inducible gene expression in the ChIP experiments. In the old version of the manuscript, we had cited the controls used for embryo collection and heat-shock treatment (Li and Arnosti 2011). To clarify this point, in the revised version, we explain in the Results section: “We treated the control embryos identically to embryos carrying the inducible Hairy transgene to control for possible nonspecific effects of heat shock on gene expression and chromatin marks. In this system, heat shock alone has no effect on the expression patterns of the pair rule and other genes analyzed, and the chromatin marks in heat shocked control embryos were indistinguishable from chromatin patterns previously reported for untreated embryos (Li and Arnosti 2011; The modENCODE Consortium et al. 2010).” Examples showing how the embryonic chromatin from our samples (heat shock treated) closely resembles that of modENCODE embryos (H3K27Ac, H3K4me1 and me3), and even to some extent from S2 cells (H4Ac) are shown in Author response image 1.

Author response image 1

Comparison of modENCODE data with our data from control embryos not expressing Hairy that were heat shock treated. The high level of similarity for the genome tracks of H4Ac, H3K27Ac, H3K4me1 and H3K4me3 at two loci were representative of the overall similarity across many genomic loci.

https://doi.org/10.7554/eLife.06394.028

We plotted genomic data using reads normalized to the sequencing depth (normalized to 10 million). Reviewer 1 asked whether induction of Hairy might cause a global impact on histone marks, which would be missed by after normalization. We considered this possibility; in fact, we find that whether or not Hairy is induced, the genome features for multiple chromatin marks are virtually identical, except in very discrete regions where there are significant changes, as shown in Figure 2–figure supplement 1), which shows the high similarity in histone marks between control and Hairy induced embryos on a specific locus (A) and globally (B and C). We now emphasize this point in the Results for clarification.

Reviewer 1 noted that the tendency for Hairy-bound regions to show loss rather than gain of acetylation/methylation marks may be influenced by the starting level of these chromatin marks. If Hairy is overrepresented in regions highly enriched for these marks, then there might be a bias. However, this is not the case. For H4Ac, the majority of Hairy reduced regions were not in the ∼2700 regions called as peaks by HOMER, suggesting that Hairy is not preferentially acting in “hot spots” for this modification (Figure 2–figure supplement 1B). Furthermore, for differentially reduced levels of H4Ac, the average level of this modification in regions not bound by Hairy was actually higher than the average level for regions on which Hairy specifically acts (Figure 3–figure supplement 1A-B, blue plots). Similar relationships hold for H3K27Ac and H3K4me. These results indicate that the tendency for Hairy-bound regions to show a loss in histone marks is not simply due to a biased high starting state.

Reviewer 1 asked that we better explain the purpose of carrying out machine learning on our chromatin data, and what we learned from it. The variety of chromatin alterations induced by Hairy did not contain a simple pattern, such as magnitude of change or width of affected area, predictive of transcriptional repression. Clearly, the ability to repress or not might hinge on where activators are bound; we didn’t measure the location of all activators, as this is not currently available knowledge. However, using machine learning, we did test the hypothesis that certain combinations of factors relating to position, size of chromatin block, magnitude of changes, occupancy of Pol II and other factors would tend to differentiate “real” regulatory events from “simulations”. The “Random Forests” method, shown in Figure 8B, was able to perform substantially better than random guessing for calling both “repressed” and “nonrepressed” genes (e.g. for repressed, 75% success, vs. 40% for random guesses), indicating that some special combination of certain features – as discussed in the Results – do provide a partial indication of where chromatin alterations are more likely to be associated with changes in gene expression. The overall success rate is still well below 100%, however. We changed the figure legend in Figure 8B to show on the heat map which level exceeds the performance of random guessing. Other machine learning approaches were able to make useful predictions, but were less successful; such differences in performance are routinely found for different sorts of datasets, and usually machine learning studies do not dwell on the weaker algorithms. We have revised this section to emphasize that there are at least two kinds of information relevant for interpreting “real” from “background” chromatin signatures, one of which is directly accessible from this sort of study. The importance of this point is that many genome-wide analyses simply assume that a particular set of modifications proves that there is a regulatory element at a locus – this is an erroneous and misleading way to approach genome-wide information. We address this this point with revised text in Results and Discussion.

Reviewer 1 was not in favor of using the term “futile cycling” for the biochemical changes observed on non-functional Hairy-bound sites. Although we noted in the manuscript that the changes are indeed impermanent “Hairy repression is readily reversible, with chromatin reverting to its previous state minutes after depletion of the overexpressed repressor” we agree that this term may be misleading. Therefore we use “genome-wide errant targeting by Hairy” as now employed in the revised title and text.

Reviewer 1 urged us to flesh out the evolutionary model, noting that the idea of evolutionary appearance of binding motifs to make new cis-regulatory elements is not new. We agree that this part of the argument was weak; removing the CtBP – related data allows us to focus directly on this point. We now present in the Discussion (and provide relevant background information in the Introduction) two major points; first, previous studies had suggested that Hairy as a repressor only functions in the context of an intact and active enhancer. Our work demonstrates that this protein is able to mediate its biochemical activity in most bound regions, indicating that there is little context necessary for the protein to go to work. Second, much molecular biology research has emphasized the high degree of cooperativity necessary for metazoan transcription factors to work well. Enhanceosomes, patterning elements and other enhancers give aberrant readouts if correct stoichiometries and spacings are not respected. These findings suggest that random individual sites are less likely to generate a suitable transcriptional readout. At least for the repressor side of the equation, we see here that the demands for generating biochemical activity are lower than anticipated – our work challenges our colleagues in the field to determine the generality of previous assumptions, and measure the independent potential for many activators binding to “off target” sites.

Reviewer 2 notes that one of the main findings presented is that Hairy interacts with many “off target” sites, with relatively few linked to transcriptional function. The reviewer notes that such observations have been made previously, and this finding is hardly surprising. We completely agree, and in fact cited previous studies discussed this point (MacArthur et al. 2009; Fisher et al. 2012; Cusanovich et al. 2014). As emphasized in our manuscript, it is not the presence of off-target binding that is novel, but the realization that these events are linked with extensive biochemical activity, in contrast to previous assumptions, and the biological implications of this finding – namely, 1) enhancers may have a lower threshold for formation that we might have expected, and 2) genome-wide studies that blithely assume a change in chromatin mark “proves” you have an enhancer at some locus are making an unwarranted assumption.

Reviewer 2 was surprised that genes silenced by Hairy lacked certain canonical chromatin changes previously suggested to be linked to repression, namely, loss of H3K36me3 and gain of H3K9me3. The reviewer notes that this would overturn the conventional view of how these histone modifications work, and asked why we didn’t comment on this point. Indeed, our initial draft did not comment on this apparent discrepancy, an omission that we have now remediated in the revised Results. Although transcriptional review articles have pointed out general correlations between individual marks and gene expression, the actual literature clearly shows that such correlations are not strong, and there is considerable heterogeneity in how they are involved in gene expression. We now cite some of these studies; for example, upregulation of KDM4A target genes in Drosophila occurs without increases in H3K36me3 (Crona et al. 2013). Similar studies with Spt6 in Drosophila further indicate that H3K36me3 levels do not correlate with Hsp70 gene expression (Ardehali et al. 2009). In fact, H3K36me3 may in some contexts contribute to gene silencing due to its presence in heterochromatic domains (Chantalat et al. 2011) and in other cases, removal of H3K36me3 is required to promote transcriptional elongation (Kim and Buratowski 2007). Similarly, repressive marks such as H3K9me3 and H3K27me3 are not always simply coupled to repression. For example, only a modest correlation between H3K9me3 and H3K27me3 levels and gene silencing was observed (Barski et al. 2007; Zhang et al. 2012). In differentiation of T and B cells, a small fraction of repressed genes ever acquired H3K27me3 (Zhang et al. 2012; McManus et al. 2011). Interestingly, H3K9me3 was found to be enriched in many active promoters and associated with transcriptional elongation (Squazzo et al. 2006; Vakoc et al. 2005). In summary, the simple correlations previously noted mask much more complex pictures that are revealed by analysis of genome-wide datasets, such as ours. We think that these complexities are themselves intriguing, and now have expanded our Discussion to provide more context.

Reviewer 2 was intrigued by the lack of change in Pol II occupancy on the bodies of many repressed genes, and offered suggestions on why this may be so, including a possible post-transcriptional regulation that would identify genes as “repressed” when in fact it was enhanced mRNA turnover at work. The short time span from Hairy induction to collection of mRNA makes it less likely that posttranscriptional effects mediated through Hairy targets would account for this effect, however. We favor the second hypothesis offered by the reviewer, namely, that we measure transcription in embryos with heterogeneous nuclei. Genes that feature poised polymerase at the promoter in many or most nuclei, but are only expressed in a few nuclei will have strong Pol II promoter signals but weak signals at the body of the gene, as observed on the gogo, pros, and tup genes shown in Figure 7. Therefore, in this view, the lack of change in Pol II levels on the gene body reflects the inherently low signal, rather than a biochemical mechanism. This explanation accounts for a considerable number of affected genes. In addition, Hairy is associated with many repressed genes with higher gene body signals where Pol II levels do decrease (Figure 7A). However, Hairy is also associated with many repressed genes where Pol II levels can be measured on the gene body, and these levels are unchanged even as mRNA levels drop. These different observations suggest that depending on context, repression may differentially impact Pol II properties. Such mechanistic heterogeneity has been previously observed; in the revised Results, we cite several examples where reduction in gene expression does not result in loss of Pol II from the gene body (Wang et al. 2007; Adelman et al. 2006; Ardehali et al. 2009; Li and Arnosti 2011). We now note that Hairy may interfere with gene expression at different steps of the transcription cycle as suggested for repression by GR, indicating gene specific repression mechanisms (Gupte et al. 2013).

Reviewer 2 raised a point about whether target genes that were repressed and bound by Hairy in vivo were in fact direct functional targets of the protein, since some repressed genes were not bound by Hairy, and many genes bound by Hairy are not repressed (consistent with our findings that Hairy has many nonfunctional sites). 1. The reviewer asked whether it is safe to assume that all of the Hairy sites in downregulated genes are functional. Indeed, to assume that every single bound site must be functional is unlikely, however, importantly, our study does not rest on that assumption, only that the large majority of target genes are bona fide Hairy targets based on the criteria below. 2. The reviewer asked how many of the Hairy binding sites on the 167 genes scored as direct targets overlap with active enhancers. High-quality data about validated enhancers remains to be determined for most of these genes, but if we use ChIP occupancy of transcription factors as a proxy for “active enhancer” (which is an admittedly soft proposition), and map Hairy sites within these elements, we find that 163 of these 167 genes has Hairy associated with a putative regulatory module (data from MacArthur et al. 2009; Kok unpublished observations).

We emphasize that the current state of knowledge in this field is still too fragmentary to decisively answer this question. 3. The reviewer asked how many known Hairy target enhancers are bound in this data set. Most functional studies addressing enhancer interactions with Hairy have been performed with synthetic cis elements, however, Hairy direct regulation of ftz 5’ and 3’ enhancers has been reported, and we find that these enhancers are indeed bound by Hairy (Li and Arnosti 2011). We also identify the Sxl promoter region as a direct target bound by Hairy; this gene has previously been characterized as a deadpan target, which has a similar bHLH DNA binding domain to Hairy, and Hairy is shown to regulate this gene (Dawson et al. 1995). There are no other reports of Hairy interacting with natural enhancers, where the role of Hairy has been described at a molecular level; our study provides the first elucidation of other possible targets. 4. The reviewer asked whether the use of ChIP data for non-overexpressed Hairy was suitable, as only 43% of our Flag epitope signals for the overexpressed Hairy protein overlapped with the ChIP-chip data. We selected the ChIP-chip data for our analysis because the Flag epitope gave low signals overall, although high-confidence functional targets such as ftz, Impl2, odd, h, 18w, wg, tup, pros, nht, and en were found. Practitioners of the trade know that there are often discrepancies in the exact sets of genes captured by ChIP studies, but the bulk of evidence from this study strongly supports the set of genes as a whole being likely Hairy targets; first, the overlap between repressed genes and genes bound in vivo is highly significant (p=3.8e-95), second, down-regulated genes are significantly enriched in classes consistent with known Hairy function (Supplemental file 1), including transcriptional regulation, cell fate commitment and neurogenesis (p<3.7e-18), third, many of the predicted target genes including odd, comm, comm2, edl, en, Impl2, prd, and 18w have striped expression patterns complementary to Hairy’s, and fourth, several of these have been shown to be derepressed in h mutant embryos (Ish-Horowicz and Pinchin 1987; Bianchi-Frias et al. 2004). In sum, we have good evidence that down-regulated genes are likely to be direct Hairy targets, and based our conclusions not on the activity of any individual gene but on the groups as a whole. Uncertainty about small numbers of genes would not significantly change the conclusions. We revised the Results to include information about the correspondence of the genetically confirmed Hairy targets and our molecular results.

Reviewer 2 raised a question about derepression of HLHm7 upon induction of the CtBP mutant version of Hairy. Since we removed this section from the manuscript as described above to streamline the presentation, it is not a part of the Discussion anymore.

As did Reviewer 1, Reviewer 2 asked about the purpose of the machine learning study, which we address above.

Reviewer 2 raised questions about how we handled the discussion of evolutionary implications of this study, noting that we do not show examples of how these “off target” biochemical interactions may facilitate the generation of novel cis-regulatory elements. We agree that, buoyed by the novelty of the findings, our description did indeed veer heavily into a speculative mode. We have made significant changes to the Abstract and Discussion to tone things down a bit and point out what we know and what the field should be investigating as possible leads. At this point, we do not have in hand an example where a Hairy bound “off target” site was subsumed into a novel cis-regulatory element. Studies that identify new enhancers have been carried out successfully in flies only to extent of showing how new genes (in male reproduction, for example) can pick up cis-regulatory elements from a variety of exapted ancestral or novel DNA elements. Alternatively, very beautiful individual cases have been described showing how subfunctionalization or neofunctionalization of cis-regulatory regions plays a role in evolution of gene expression. These studies fail to provide much information about which factors are present on individual elements, or how their activities may be relevant to generating the new element. Our study highlights the need to consider the activity of “off target” sites in generating novel elements, particularly because for Hairy at least (and likely other factors that employ the same cellular machinery) they are “shovel ready” and not constrained by complex cis-regulatory grammar. The new version of this manuscript provides a more balanced discussion of what we know and what we don’t know, and now develops a second theme that was stimulated by considering your comments. That is, genomic consideration of chromatin marking must not equate changes in some active marks with enhancers – there appears to be a significant possibility for false positives if one relies on the simple correlations that are widely employed.

The reviewer was confused by the term “futile cycling”. We recognize that this was not a helpful designation, and have altered this, as described above.

“H” was used in Figure 1 to denote Hairy protein; the reviewer suggested we use hs-hairy to distinguish this from Hairless, which is another gene. We have modified the figure accordingly.

Reviewer 3 noted that induction of Hairy in cells in which it is never normally present could induce artifacts, accounting for the low correlation between changes in the chromatin landscape and repression. We agree that it is possible that some loci identified are not real Hairy targets, but only when the protein is overexpressed. However, for hundreds of the loci considered where chromatin changes are taking place, ChIP data from both Biggin as well as Kevin White confirm that Hairy can be found binding to these genomic regions. The early point in development selected for our study means that there are fewer cell-specific modifications, Polycomb repressed regions, and other markers of differentiated or senescent cells; thus the heterogeneity question is of lesser concern. We now do make a remark about this consideration in the Discussion, however.

Reviewer 3 recognized that the overexpression system allows us to carry out manipulations that would be otherwise impossible, but asked if we can also use knockdown approaches. The inducible system allows us to capture direct effects with good temporal resolution, which would not be possible in loss of function assays such as RNAi knockdown. In response to this suggestion, in the current version we note h loss of function mutants cause derepression of genes such as ftz, en, edl, Impl2, and prd as shown in previous studies, and these same genes are down-regulated upon Hairy induction in our hands (Ish-Horowicz and Pinchin 1987; Bianchi-Frias et al. 2004).

Reviewer 3 points to the lack of correlation between the Pol II ChIP and repression and asks if it has been observed in other contexts. We addressed this question in our response to Reviewer 2.

Reviewer 3 asked that Figure 8 (now Figure 9) be better explained in the text. We made a new figure for the model and clarified the Discussion.

[Editors’ note: the author responses to the re-review follow.]

We appreciate the reviewers’ careful reading and helpful suggestions. We address the comments from Reviewers 1 and 3 below.

Reviewer 1 noted the revised manuscript is more direct and readable; we thank the reviewers for their feedback that aided us in improving this paper. The reviewer noted that data from many groups are consistent with our findings, but this study provides for the first time direct experimental evidence for “errant targeting” that results in direct biochemical modification of chromatin regions, not necessarily linked to changes in expression. We agree with the reviewer that our findings indeed have a bearing on numerous ChIP-seq studies.

In the first round of review, Reviewer 3 noted that overexpression used here means Hairy will be expressed in (interstripe) cells not normally containing the protein. There, the Reviewer asked if knockdown studies might indicate “how many of the repression targets identified by overexpression are also identified by loss of function”. We interpreted this point as asking whether the repressed genes found in our study were corroborated by other methods, and cited previous literature supporting the identification of a number of Hairy targets. In response, in the second round the reviewer noted that “this does not address the more important point that pertains to the genes that exhibit changes in the chromatin marks and yet show no repression upon hairy overexpression”. Based on this comment, I think that we misinterpreted the thrust of the question in the first place, so here we address the question of the unaffected genes experiencing chromatin modification by Hairy. We have previously discussed data indicating that many of such elements are known to be occupied by endogenous Hairy. The reviewer asked that the behavior of this class be verified in loss-of-function studies. As previously addressed, the ChIP-seq/chromatin change studies require the quick kinetics that overexpression but not knockdown allow, but what might depletion of Hairy reveal? We thought about the possible mechanistic models possible in this case. 1) A gene normally bound by Hairy in the nuclei within the blastoderm stripes whose global expression is unaffected despite induction of Hairy may have separate enhancers that are far from Hairy sites. Even when Hairy binds in the interstripe nuclei, the enhancers would still escape regulation. 2) Alternatively, the gene may be silent, in which case Hairy binding to the gene in interstripe nuclei would not alter its silence. 3) A third possibility is that within the normal Hairy stripe regions, there are special activators that are only present in these stripes, but the endogenous Hairy is at high enough levels so that expressing more Hairy has no effect on transcription – the observed chromatin changes would come only from interstripe nuclei. Only in this last case would we mistakenly “call” a gene as “not regulated transcriptionally” but still having chromatin modifications. We think that there are few genes that fit this third category; known activators at this stage tend to be more broadly expressed, and loss of function hairy mutants show widened expression of target genes (stripes expand), consistent with broadly expressed activators. These scenarios are shown in Author response image 1. We know that hundreds of genes we consider are broadly expressed, thus the conclusions drawn are based on this more general picture. To be frank, we had not considered all of the scenarios before this point was raised, and it is a useful one. We elaborated this point in the Discussion so that the readers are aware of these issues.

Author response image 2

A) Three scenarios explaining why genes might be unresponsive to Hairy induction in interstripes and stripes even though chromatin changes were detected. 1. Enhancers are far from Hairy sites. 2. Gene is generally inactive. 3. Stripe-specific activators are present within Hairy stripe region, and are partially or completely repressed by the endogenous Hairy, so that ectopic Hairy has no effect on transcription. Promoters are shown in Hairy stripe and interstripe areas, before and after Hairy induction. B) Diagrams show expression pattern of Hairy (red) and Type 3 unresponsive genes (blue) to Hairy overexpression in wild-type (wt) embryos, loss-of-function and overexpression assays.

https://doi.org/10.7554/eLife.06394.029

Reviewer 3 pointed out the apparent contradiction between our assertions that Hairy action on repressed and non-repressed loci was similar, yet machine learning methods can differentiate promoters that are transcriptionally repressed from others.

The two settings for these conclusions are somewhat different. For machine learning, we classified the gene sets as repressed and nonrepressed (activated and unaffected), whether or not they are bound by Hairy or affected at the chromatin level in a particular manner. The machine learning algorithms predict gene expression with up to 75% accuracy, not perfect but much better than random guessing. Separately, in the Discussion, we pointed out the similarity of chromatin changes on targets bound by Hairy, whether or not they were transcriptionally regulated. We address these differences in the revised manuscript, where we clarify these points in the text and Figure 8 (subsection “Predicting a “successful” repression context” and modified labeling in Figure 8).

https://doi.org/10.7554/eLife.06394.027

Article and author information

Author details

  1. Kurtulus Kok

    Genetics Program, Michigan State University, East Lansing, United States
    Contribution
    KK, Conception and design, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article
    Competing interests
    The authors declare that no competing interests exist.
  2. Ahmet Ay

    Departments of Biology and Mathematics, Colgate University, Hamilton, United States
    Contribution
    AA, Performed machine learning analysis
    Competing interests
    The authors declare that no competing interests exist.
  3. Li M Li

    Janelia Research Campus, Howard Hughes Medical Institute, Ashburn, United States
    Contribution
    LML, Designed and acquired microarray data, Analysis and interpretation of data
    Competing interests
    The authors declare that no competing interests exist.
  4. David N Arnosti

    1. Genetics Program, Michigan State University, East Lansing, United States
    2. Department of Biochemistry and Molecular Biology, Michigan State University, East Lansing, United States
    Contribution
    DNA, Conception and design, Analysis and interpretation of data, Drafting or revising the article
    For correspondence
    arnosti@msu.edu
    Competing interests
    The authors declare that no competing interests exist.

Funding

National Institutes of Health (NIH) (GM056976)

  • David N Arnosti

The funder had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank Carla Margulies for the Rpb3 antisera, Shin-Han Shiu, Monique Floer, and George I Mias for critical reading of the manuscript, John Johnston and the MSU Institute for Cyber-Enabled Research (iCER) for help with the High Performance Computing Center. We thank Samuel Daulton (Colgate University) for his help in the Machine Learning analysis of the data. We also thank the Arnosti laboratory members for useful discussions, and Sandhya Payankaulam for assistance with establishing library preparation protocols.

Reviewing Editor

  1. Asifa Akhtar, Max Planck Institute for Immunobiology and Epigenetics, Germany

Publication history

  1. Received: January 9, 2015
  2. Accepted: July 24, 2015
  3. Version of Record published: August 25, 2015 (version 1)

Copyright

© 2015, Kok et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 1,669
    Page views
  • 282
    Downloads
  • 6
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Developmental Biology
    Aynur Kaya-Çopur et al.
    Research Article Updated

    Skeletal muscles are composed of gigantic cells called muscle fibers, packed with force-producing myofibrils. During development, the size of individual muscle fibers must dramatically enlarge to match with skeletal growth. How muscle growth is coordinated with growth of the contractile apparatus is not understood. Here, we use the large Drosophila flight muscles to mechanistically decipher how muscle fiber growth is controlled. We find that regulated activity of core members of the Hippo pathway is required to support flight muscle growth. Interestingly, we identify Dlg5 and Slmap as regulators of the STRIPAK phosphatase, which negatively regulates Hippo to enable post-mitotic muscle growth. Mechanistically, we show that the Hippo pathway controls timing and levels of sarcomeric gene expression during development and thus regulates the key components that physically mediate muscle growth. Since Dlg5, STRIPAK and the Hippo pathway are conserved a similar mechanism may contribute to muscle or cardiomyocyte growth in humans.

    1. Developmental Biology
    Sweta Parab et al.
    Research Article

    Vascular endothelial cells (vECs) in the brain exhibit structural and functional heterogeneity. Fenestrated, permeable brain vasculature mediates neuroendocrine function, body-fluid regulation, and neural immune responses, however its vascular formation remains poorly understood. Here we show that specific combinations of vascular endothelial growth factors (Vegfs) are required to selectively drive fenestrated vessel formation in the zebrafish myelencephalic choroid plexus (mCP). We found that the combined, but not individual, loss of Vegfab, Vegfc, and Vegfd causes severely impaired mCP vascularization with little effect on neighboring non-fenestrated brain vessel formation, demonstrating fenestrated-vEC-specific angiogenic requirements. This Vegfs-mediated vessel-selective patterning also involves Ccbe1. Expression analyses, cell-type-specific ablation, and paracrine activity-deficient vegfc mutant characterization reveal that vEC-autonomous Vegfc and meningeal fibroblast-derived Vegfab and Vegfd are critical for mCP vascularization. These results define molecular cues and cell types essential for directing fenestrated CP vascularization and indicate that vECs' distinct molecular requirements for angiogenesis underlie brain vessel heterogeneity.