Genome-wide errant targeting by Hairy

[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

Download asset Open asset

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

Download asset Open asset

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).