Shadow enhancers modulate distinct transcriptional parameters that differentially effect downstream patterning events

Drosophila blastoderm development occurs rapidly over the course of 3 h. During this time, all of the major tissue types are specified through a burst of intricate transcriptional regulation that culminates in the dramatic morphogenic events of gastrulation (reviewed by Stathopoulos and Newcomb, 2020). This period of development is a powerful system to study transcriptional regulation of developmentally relevant genes. In this study, we explore the conserved phenomenon of ‘shadow enhancers’, first described in Drosophila (Hong et al., 2008). Enhancers are cis-regulatory elements that interact with transcription factors and are capable of producing precise transcriptional outputs by employing a combinatorial logic of bound activators and repressors. Shadow enhancers have overlapping activities; that is, they activate transcription of the same gene in nearly identical patterns and are thought to provide robustness to the system (Frankel et al., 2010; Perry et al., 2010, 2011).

Further experiments have shown that shadow enhancers are widespread in developmentally relevant genes, and appear in multiple organisms, including humans (reviewed by Kvon et al., 2021). Gene editing and transgenic constructs have demonstrated that, despite overlapping activities, RNA production from shadow enhancer pairs can deviate significantly, and multiple modes of enhancer interactions between shadow enhancer pairs have been identified. Shadow enhancers are said to have an additive interaction if the sum of the RNA produced from each individual enhancer matches what is produced from the wild-type enhancer pair. Sub-additive interactions are described as the sum of RNA produced being more than the wild-type RNA, whereas super-additive interactions describe the opposite. Finally, repressive interactions are the result of RNA from one enhancer exceeding the amount from the wild-type pair, suggesting that one of the enhancers is capable of repressing the output of the other (Kvon et al., 2021).

One of the first described shadow enhancer pairs was discovered at the short gastrulation (sog) locus (Hong et al., 2008). When cloned into transgenic expression constructs, both enhancers produce the characteristic lateral stripe of sog expression (Hong et al., 2008; Liberman and Stathopoulos, 2009). A recent report suggested that the one of the two enhancers may have repressive activity (Dunipace et al., 2019), and although particularly interesting at a mechanistic level, it is currently difficult to postulate a biological mechanism for how two enhancers both capable of driving expression can inhibit the total output of RNA. Therefore, we were motivated to quantifiably dissect exactly which features of transcription each enhancer controls, as solely measuring the total RNA produced obscures the multiple mechanistic steps involved in transcription. This would give a better understanding of transcriptional control by enhancers more broadly, as the majority of genes active during early development have been shown to possess shadow enhancers.

To accomplish this, we first created several Drosophila lines with endogenous enhancer deletions to study the developmental consequences of abnormal sog expression. We inserted MS2 tags (see Materials and Methods) into the first intron of sog in all lines, allowing us to compare transcription directly with wild-type alleles in fixed embryos, and to measure transcription in real time to examine how each enhancer modifies the parameters that define transcriptional output. We found that the sog enhancers have distinct but overlapping domains of expression, with individual enhancers capable of modifying different kinetic variables of transcription that combine in a manner that leverages the strength of each individual enhancer. This analysis also revealed that repression, but not activation, appears to be synergistic between the enhancers. Finally, we examined how altered transcription from the loss of individual enhancers leads to idiosyncratic downstream phenotypic consequences that are well explained by differences in the expression profile each enhancer alone generates.

Early sog expression is controlled by two enhancers, originally known as ‘primary’ (or ‘intronic’) and ‘shadow’, but also, and herein, referred to as ‘proximal’ and ‘distal’, respectively (Dunipace et al., 2019; Hong et al., 2008). Fig. 1A shows the location of these two enhancers with respect to the transcription start site (arrow), with the distal enhancer located 20 kb upstream (blue rectangle) and the proximal enhancer located ∼1.5 kb downstream within the first intron (green rectangle).Fig. 1B shows the location of key transcription factor binding sites in each enhancer that are largely responsible for the transcriptional domain of sog. Dorsal (Dl) serves as the primary transcriptional activator across the dorsal/ventral (D/V) axis, whereas Zelda (Zld) potentiates Dl activity down the morphogen gradient and Snail (Sna) represses activity in the mesoderm, resulting in the broad lateral stripes of the sog pattern (Liberman and Stathopoulos, 2009; Foo et al., 2014).

To better understand the individual roles of these enhancers, we created enhancer deletion lines, in which we simultaneously inserted MS2 live-imaging tags within the first intron (Fig. 1C, turquoise rectangle) via CRISPR-Cas9 homology directed repair editing (see Materials and Methods). The size of the enhancer regions removed were 1022 bp in the case of the proximal enhancer and 743 bp in the case of the distal. In order to maintain the spacing between the MS2 loops and the promoter in the proximal enhancer deletion, we adapted a ‘neutral’ DNA sequence of identical size and GC content from Scholes et al. (2019) to replace the proximal enhancer (Fig. 1C, yellow rectangle; see Materials and Methods). In constructing the distal enhancer deletion allele, we included a 3×P3>RFP cassette of ∼1.2 kb that replaced the distal enhancer upon successful integration to aid in screening recombinant flies (see Materials and Methods). The wild-type enhancer allele, proximal enhancer deletion allele and distal enhancer deletion allele will hereafter be referred to as WTsogMS2, ΔPsogMS2 and ΔDsogMS2, respectively, and the double enhancer deletion allele as ΔPΔDsogMS2.

To evaluate the fitness of our alleles, we performed lethal counts by counting the ratio of unhatched to hatched larvae from homozygous lines over a period of 36 h (Fig. 1C, right). Both enhancer deletion lines showed increases in the number of unhatched larvae, with flies carrying ΔDsogMS2 showing larger losses in viability than those with ΔPsogMS2. ΔPΔDsogMS2 produced neither homozygous adult females nor hemizygous adult males, and therefore was assumed to be embryonic lethal. To evaluate the sog expression domains of these embryos, we performed colorimetric in situ hybridization for sog transcripts (see Materials and Methods). All alleles produced a sog expression pattern of varying intensity with the exception of ΔPΔDsogMS2, which gave no apparent sog expression. We therefore concluded that both enhancers are necessary for sog expression, but a single enhancer is at least sufficient to generate some sog expression. In addition, sog does not appear to contain any other enhancers that drive early blastoderm expression.

Because enhancers regulate gene expression at the level of transcription, we wanted to assess how each of the enhancers contribute to sog transcriptional output. To do this in a quantitative manner, we first turned to single molecule fluorescence in situ hybridization (smFISH), which is capable of producing fluorescence that scales linearly with the amount of RNA stained. We focused on measuring nascent transcripts, which can be seen in nuclei as large foci. To internally control for the dynamic nature of sog expression, particularly when ventral repression sets in during NC14, we crossed our MS2-tagged flies to wild-type flies to create sog heterozygous embryos. This allowed us to directly compare both the level and domain of transcription of wild-type sog to enhancer deletion sog.

As shown in Fig. 2A, the alleles can be discriminated through the use of two probe sets, a sog 5′ exonic-directed probe that labels both alleles in heterozygous flies (magenta), and a second probe set targeting the MS2 region and thus only the MS2 allele (cyan). This enabled us to quantify transcriptional differences at the level of single nuclei (see Fig. 2B schematic of labeled foci). Fig. 2C shows the sog domain of a heterozygous WTsogMS2/wild-type embryo using this double labeling system in combination with an anti-Dl antibody. Note the WTsogMS2 foci are white because of the dual labeling of magenta and cyan probes, whereas the wild-type foci are magenta as they are not labeled with the MS2 probe. For a more detailed explanation of allele discrimination and image analysis, see Fig. S1 and Materials and Methods.

The double labeling assay allows us to internally control for fluorescence of nascent transcripts by examining nuclei that have both alleles active and taking a log ratio between the intensity of the two magenta foci to look for upregulation or downregulation of sog, which will give positive or negative values respectively. To ensure that our WTsogMS2 allele operates identically to our unlabeled wild-type allele, we plotted this ratio across the D/V axis and found minimal fluctuations around 0, indicating that there is no change in sog output with the addition of our MS2 tag (Fig. 2D, orange line). In contrast, ΔPsogMS2 showed significant downregulation in the mesoderm but trended towards wild-type levels in the dorsal portion of the pattern (Fig. 2D, blue line), and ΔDsogMS2 again displayed the opposite trend (Fig. 2D, green line). This suggests that the two enhancers regulate transcriptional output differently across the sog domain.

To further investigate the idea of enhancers having a spatial preference, we assessed the percentage of MS2-expressing nuclei in bins across the sog domain, including in the count not only nuclei with both alleles active, but also those with only the MS2 allele active (monoallelic expression). The WTsogMS2 allele showed robust activation across the sog domain, with major reductions in activity at the ventral and dorsal extremes of the domain (Fig. 2E, orange curve). In contrast, the ΔPsogMS2 allele showed a significant decrease in activation on the ventral end of the pattern (Fig. 2E, blue curve), and the ΔDsogMS2 allele showed the opposite trend, with significant decreases in the dorsal end of the pattern (Fig. 2E, green curve). In addition, neither of the deletion alleles showed as robust an activation as the WTsogMS2 allele in the lateral portion of the domain.

These results suggest that the transcriptional domain displayed by sog is the result of the two separate enhancers summing their individual domains, in a manner that displays simple additivity at the borders of the sog domain, and sub-additivity towards the center. This sub-additivity likely arises from a complete saturation of activation in the center of the pattern, i.e. there are simply no more nuclei to activate, rather than any particular transcriptional mechanism that causes the enhancers to integrate their activation signals in a fundamentally different way across the D/V axis. This simple framework is potentially broadly applicable across multiple shadow enhancer pairs, as it agrees with several previous studies in which shadow enhancers appear to aid in creating robust borders to the transcriptional patterns they give rise to (Perry et al., 2011; El-Sherif and Levine, 2016; Dunipace et al., 2019; Scholes et al., 2019).

We then focused on only the monoallelic-expressing nuclei to determine whether the occurrence of monoallelic expression was differentially influenced by either enhancer at any point in the sog domain. We found a bimodal distribution of MS2 monoallelic expression for the WTsogMS2 allele, with peaks of monoallelic expression on both ends of the sog domain (Fig. 2F, orange curve). This is consistent with earlier reports that have suggested that monoallelic expression occurs more frequently on the border of transcriptional domains (Hoppe et al., 2020). When we examined the enhancer deletion lines, we found a similar trend to the previous experiment, where the dorsal peak of monoallelic expression vanished in our ΔPsogMS2 (Fig. 2F, blue curve), and the ventral peak absent in the ΔDsogMS2 allele (Fig. 2F, green curve). Strikingly, the peak that was not absent in each deletion allele remained at exactly the level we observed in wild type (Fig. 2F, note region of overlap among the three curves). This suggests that monoallelic expression in shadow enhancer pairs may be largely driven by single enhancers acting alone. Taken together, these observations demonstrate that each of the sog enhancers has a preferred domain along the D/V axis, and when combined together create the wild-type pattern.

Next, we wanted to explore how control of transcriptional kinetics differed between the two enhancers. Measuring nascent transcription in fixed tissue confounds two critical variables, the timing of activation and the rate of transcript production. In order to examine these two variables, we turned to live imaging to visualize the number of nascent transcripts produced over time, using our endogenously inserted MS2 tag. Fig. 3A shows how MS2 live imaging operates with the intronically inserted MS2 loops. Co-transcriptional splicing of sog has been experimentally verified in the early Drosophila embryo using intronic FISH probes (Bothma et al., 2011), therefore we expect MCP-GFP fluorescence to be only detectable for transcripts that have not yet been spliced out of the first intron. Imaging was performed on the portion of the embryo that includes the sog expression domain from NC12 to mid-NC14, at which point the defined line of ventral repression in sog becomes apparent (see schematic in Fig. 3B, still images in Fig. 3C and representative Movies 1-3).

In order to characterize the sog transcriptional activity of each line and to validate the fidelity of our live imaging system, we counted the number of foci seen in each nuclear cycle relative to the number of nuclei (Fig. 3C). We classified foci position into four categories across the sog domain: ventral, ventral/lateral, dorsal/lateral and dorsal, with the ventral position encompassing any foci detected in the presumptive mesoderm. Broadly, we observed that WTsogMS2 and ΔDsogMS2 produced similar numbers of MS2 active nuclei, with the exception of the dorsal-most position (see Fig. 3D histograms). This is in contrast to the activity of ΔPsogMS2, which nearly universally underproduced relative to WTsogMS2. This is most striking in NC12, in which barely any transcriptional activity was observed (Fig. 3D, blue bars). Furthermore, in NC14, ΔPsogMS2 produced very little transcription in the ventral-most bin, which suggests that the distal enhancer is more sensitive to Sna-mediated repression. The results of this analysis at NC14 are consistent with our fixed imaging data (Fig. 2D), demonstrating that the MS2 system is faithfully reporting on the transcriptional output of sog.

To better understand how activity differs between the enhancers, we analyzed single nuclei (Fig. 4A,B). Single foci were tracked and their voxel intensity values summed for each time point to produce a trace of MS2 activity over time. Then, several parameters were extracted from these traces: t_on, defined by the time at which a MS2 focus was first observed following the previous nuclear division; RNA polymerase loading rate, which describes the rate of signal increase by fitting a line to values where the GFP signal first increases (Fig. 4B, purple line); and t_off, the time at which the signal is no longer detectable in that nucleus. All parameters were measured for nuclei across the D/V axis. We focused on NC13 and NC14 for this analysis, as these cycles produce far more activity than NC12 and are therefore more relevant to the total transcriptional output of sog.

The t_on times for each genotype at NC13 and NC14 are shown in Fig. 4C. With the exception of the most ventral bins, WTsogMS2 activated transcription at a faster rate than both deletion genotypes. ΔPsogMS2 showed extremely delayed transcription at all positions and times, in line with the results of inefficient activation discussed above. Notably, ΔPsogMS2 embryos exhibited a much broader distribution of t_on times than either WTsogMS2 or ΔDsogMS2 embryos, giving rise to a lack of synchrony of activation across the sog domain (Fig. S2A-C). However, when we examined the loading rates of all lines (shown in Fig. 4D), ΔPsogMS2 outperformed even WTsogMS2 in most cases, and greatly outperformed ΔDsogMS2, which had loading rates that fell severely in the more dorsal bins. All genotypes showed lower loading rates in the ventral bins, likely driven by Sna-mediated repression of sog.

With these apparently opposing enhancer activities for activation and loading rates, we wanted to create a metric that would describe the total transcriptional output of each genotype. To do this, we adapted an approach used by Garcia et al. (2013) that described transcriptional output by combining multiple parameters of transcription (shown in Fig. 4E). For each bin, we multiplied the time active (duration of t_on; Fig. 4B, white area) by the loading rate for each nucleus. The average value obtained in each bin was then multiplied by the fraction of nuclei with detectable transcription to normalize the differences in activation seen across the genotypes. Total output values are plotted in Fig. 4F, showing that WTsogMS2 generates the most activity by this metric. Thus, although at first glance it appeared that the distal enhancer when acting alone drives higher transcriptional activity than when the distal and proximal are combined [Fig. 4D; also shown by Dunipace et al. (2019)], this is not the case when taking into account all transcription variables to determine total output.

Curiously, the only place WTsogMS2 is not highest in this metric is in the most ventral bin in NC14, where ΔDsogMS2 showed higher total output (see Fig. 4F, NC 14, compare orange and green lines in mesoderm). This result indicates a repressive interaction between the two enhancers, as an additive interaction is always indicated by the wild-type enhancer pair showing the highest output.

To address this finding, we wondered whether it was possible to construct the observed wild-type transcriptional activation and repression of sog over time using the kinetic parameters gathered from the enhancer deletion lines. By building a model of each individual enhancer's activity over time, we could simulate what would be observed if those enhancers operated in the same nucleus, but did not interact when driving sog transcription. We could then compare the output of this simulation to the transcriptional activity observed in WTsogMS2 embryos, where any significant deviations from the model's prediction and the data could be interpreted as potential synergy between the enhancers.

In order to simulate enhancer activity, we first fit gamma distributions to the t_on and t_off values obtained for each genotype across the D/V axis at NC14. These distributions were then refined by systematically altering the shape and rate parameters of each distribution until the differences between simulated nuclei and the observed activity over time were minimized (see Materials and Methods). During simulation, distributions of t_on and t_off for each nucleus were sampled independently, assuming no correlation between the activation time of a nucleus and the time of loss of signal (see Fig. S3 for validation of this assumption).

Fig. 5A shows the output of simulations based on our model expressed as the percentage of active nuclei over time (solid line) plotted over the data gathered from our NC14 live-imaging experiments (open circles). Note the near perfect overlap for all genotypes, indicating the distributions of t_on and t_off values chosen are sufficient to describe the data. Beside each plot is shown the distributions of t_on (pink) and t_off (blue) values generated from sampling the fit gamma distributions. For a breakdown of distributions and fits for all D/V bins, see Fig. S4.

Having found parameters for all distributions that can accurately describe the data of individual enhancers based on data from our enhancer deletion lines, we created a combined model that simulates the activity of both enhancers in a single nucleus. Nuclei remain ‘on’ if at least one enhancer is simulated to be active based on the values obtained by sampling t_on and t_off distributions obtained from each deletion line. This underlying assumption represents the null hypothesis that there is no interaction between the enhancers, and the activity seen in WTsogMS2 is based purely on the combined activity of the proximal and distal enhancers. Fig. 5B shows conceptually how the model interprets multiple sets of t_on and t_off values sampled from each pair of distributions for the two enhancer deletion genotypes. In this example, the faster-acting proximal enhancer is responsible for the initial activation of sog (green), whereas the slower-acting distal enhancer activates later (blue), with a brief period of overlapping activity of both enhancers (orange) that maintains continuity of transcription.

Using this combined model, we simulated an additional 10,000 nuclei for each bin across the D/V axis, and compared the results with the observed activation kinetics of WTsogMS2. Although in all D/V bins the rate of activation was remarkably well predicted by the model (Fig. 5C, see overlap between initial rise in curves), the t_off values generated by the combined model were much higher than the distribution of t_off values measured by experiment, leading to a dramatic overactivation of the model output compared with the data in the ventral bins (Fig. 5C, note different curve heights). This rigorously demonstrates that the strong repression experienced by the distal enhancer in the mesoderm is somehow influencing the ability of the proximal enhancer to activate transcription in WTsogMS2 embryos. In addition, it identifies the key parameter from which the repressive interaction arises, clearly implicating Sna-mediated repression, not Dl-activation. Understanding this form of crosstalk between enhancer pairs is likely essential for creating a unified model of enhancer biology. For a more detailed look at the implications of this finding and possible underlying mechanisms, see Discussion.

With a better understanding of the kinetic features of sog transcription, we wanted to evaluate the downstream developmental effects that occur due to the loss of a single sog enhancer. To observe developmental consequences of the sog enhancer deletions, we measured the developmental morphogen gradient that Sog protein is directly involved in refining: the dorsally located gradient of phospho-Mothers Against Decapentaplegic (pMAD). Briefly, Sog protein produced in ventro-lateral cells diffuses dorsally, where it inhibits activity of the TGF-β homolog Decapentaplegic (Dpp), resulting in a gradient of Dpp activity (Shimmi et al., 2005; Wang and Ferguson, 2005). Dpp signal transduction leads to the phosphorylation of MAD, and in early NC14 it initially creates a broad region of pMAD. Sog protein binds to Dpp, preventing it from creating high levels of pMAD in the lateral regions of the embryo, and Sog diffusion eventually restricts pMAD to a dorsal stripe four to five nuclei wide (Fig. 6A; Dorfman and Shilo, 2001; Rushlow et al., 2001; Sutherland et al., 2003). By staining embryos with anti-pMAD antibodies, we could visualize any impairments in pMAD-domain formation that may be caused by sog enhancer deletions.

The results of pMAD antibody staining on homozygous WTsogMS2 mid and late NC14 embryos are shown in Fig. 6B. To quantify total pMAD levels, we measured the intensity of pMAD staining and plotted it over the dorsal position centered on the peak of maximum pMAD staining (Fig. 6C). Interestingly, the ΔDsogMS2 and ΔPsogMS2 alleles show peak pMAD intensity nearly identical to that of wild type, suggesting that only a small input of sog activity is required to increase the level of pMAD seen in the dorsal-most cells. This is in contrast to maximum pMAD levels seen in ΔPΔDsogMS2, which completely fail to refine into a narrow peak. However, single enhancer deletions produced an overall broader distribution of pMAD staining. ΔDsogMS2 embryos gave the broadest pMAD domain, which is consistent with the rank order of total output of sog (Fig. 4F).

Because we found large differences in the onset of transcription in our enhancer deletions, we were interested to see if this influenced the timing of pMAD refinement. To test this, we plotted the width of the pMAD domain of both mid and late NC14 embryos for all genotypes (Fig. 6D). As expected, WTsogMS2 embryos refine their pMAD domain over these two time points. ΔPsogMS2 embryos carry out a more extreme refinement, initially showing a far larger pMAD domain. In contrast, ΔDsogMS2 embryos initially show a modestly expanded pMAD domain less so than ΔPsogMS2 embryos; however, this undergoes no appreciable change in late NC14. Finally, ΔPΔDsogMS2 displays an incredible expansion of the pMAD domain, which, in the absence of any sog production, does not undergo any subsequent retraction.

These results are well explained by our MS2 data, which showed significant delays in the onset of transcription of sog in ΔPsogMS2 embryos. However, the high loading rates achieved by the distal enhancer allow ΔPsogMS2 embryos to eventually produce enough Sog protein to refine the pMAD gradient. The lack of refinement of pMAD in ΔDsogMS2 is likely due to the inability of the primary enhancer alone to continuously produce sog transcripts late into NC14.

To determine whether the changes in the pMAD gradient impact the expression domains of pMAD target genes, we performed colorimetric in situ hybridization for two representative pMAD targets; u-shaped (ush), thought to be an ‘early’ pMAD target, and hindsight (hnt; also known as peb), thought to be a ‘late’ pMAD target (Hoppe et al., 2020) (Fig. 6E). In ΔPsogMS2 embryos, the ush pattern is much wider than in wild type, whereas hnt is only slightly wider than in wild type (Fig. 6E), matching the changes in pMAD over time (Fig. 6D). In ΔDsogMS2 embryos, both ush and hnt are somewhat wider than in wild type, matching the unchanging pMAD domain over time, but not as wide as ush in ΔPsogMS2 embryos. These results suggest that the changes observed in pMAD stainings functionally impact the subsequent patterning steps, and that changes in the onset and rate of transcription of sog have specific and defined consequences for downstream signaling events.

In this study we sought to understand how two shadow enhancers collectively contribute to the output of a gene. We used fixed and live imaging techniques to characterize the position, timing and rate of transcription of each enhancer separately. Far from being redundant, we found that these enhancers contributed to different aspects of transcription, and loss of enhancers produced different downstream consequences for development in terms of altered tissue patterning and embryo survivability. In addition, by separating out different key features of transcription, we have shown that enhancer additivity functions differently at particular steps in transcriptional activation and repression.

Our fixed imaging experiments demonstrated that the proximal and distal enhancers contribute to the ventral and dorsal locations of the sog transcriptional pattern, respectively, with the highest overlapping activity located in the lateral region of the pattern (Fig. 2). Higher rates of monoallelic expression were seen on both edges of the sog pattern in WTsogMS2 embryos, which are presumably the result of reduction in the frequency of activation the farther away a given nucleus is from the target region of sog expression encoded by the enhancers. This is supported by the observation that the peak of monoallelic expression found at either end of the pattern disappears when the enhancer that has a preference for that position is lost. However, it is unclear whether monoallelic expression represents a complete loss of activity from a single allele, or whether a small amount of activity remains, but has dipped below our detection threshold for nascent transcription.

Modeling of enhancer activity found that the collective action of the two enhancers complement each other in mostly an additive fashion, that is, the action of the two enhancers together can be adequately explained by assuming that there is no mechanistic interaction between them. However, this is not the case in the ventral portion of the D/V axis where Sna acts to repress transcription of sog. Instead, there appears to be enhanced repression by the proximal enhancer in the presence of the distal enhancer, as seen in Fig. 5C, in which the prediction of our model deviates from the observed WTsogMS2 data, indicating interaction between the two enhancers.

The cause of this effect is unknown, but a plausible mechanism can be postulated based on the current understanding of Sna-mediated repression. Sna works to repress transcription in the early embryo by the recruitment of the co-repressor CtBP, which is thought to operate at small genomic distances less than 200 bp (Keller et al., 2000). In the classic example of the short-range repressive effect of CtBP, Krüppel is responsible for repressing the activity of the eve stripe 2 enhancer to create the sharp posterior border of stripe 2. Located just 1.7 kb away is the eve stripe 3+7 enhancer, which does not experience any repressive effects despite eve stripe 3 being found in the domain where Krüppel is most active in the blastoderm embryo (Nibu et al., 1998). Importantly, the portion of the enhancer that drives stripe 3 is locally depleted for Krüppel binding sites (Vincent et al., 2018).

However, this lack of a shared repressor responsible for recruiting CtBP is not the case for the enhancers of sog, where both enhancers contain binding sites for Sna (see Fig. 1B). Efficient recruitment of the co-repressor by high occupancy of Sna at the distal enhancer may amplify the action of Sna at the proximal enhancer by increasing the local concentration of CtBP in the microenvironment of the sog locus, thereby allowing Sna at the proximal enhancer to recruit CtBP more efficiently. A modeling based approach that attempted to derive how enhancer sequence changes transcriptional output based on the binding characteristics of recruited transcription factors in Drosophila embryos found that Sna repression required uniquely high levels of homotypic cooperativity in the context of a single enhancer compared with all other repressors examined by the study (Fakhouri et al., 2010). It is unknown whether this cooperativity could scale to larger genomic distances, but repressive factors, including the Ciona Sna homolog, have been shown to form condensates that may extend the range of repressive activity (Treen et al., 2021).

In the case of activation, our data do not support any mechanism of super-additivity. Activation rates of sog are well predicted by a model that assumes enhancers act independently. Decreases in measured t_on values seen in WTsogMS2 embryos are likely accounted for by the wide distribution of t_on times measured in ΔPsogMS2 embryos (Fig. 4C). Activation of sog by the distal enhancer occasionally precedes the proximal enhancer, thus modestly lowering the average t_on values in WTsogMS2 embryos. However, in most cases, the proximal enhancer will activate first, and the later t_on value contributed by the distal enhancer will be ‘masked’ and will therefore not contribute to raising the average t_on value. This form of additivity is similar to the additivity seen in the domain of activation from our fixed experiments, in which saturation of activation occurs in the central region of the sog domain.

Although we do see evidence that RNA Pol II loading rates are diminished in the WTsogMS2 embryos when compared with ΔPsogMS2 embryos (Fig. 4D), potentially suggestive of so-called ‘enhancer interference’ (Fukaya, 2021), we believe that this result is well explained by the initial activation of transcription being performed by the proximal enhancer in the majority of nuclei, which appears to drive much lower rates of transcription. This confounds our loading rate measurement, as the rise in signal intensity in WTsogMS2 embryos is likely a composite of the two enhancers acting sequentially. Techniques that attempt to estimate the promoter state at any given time using an MS2 trace may be able to dissect out the individual contributions each enhancer makes; however, we believe that this analysis is not required to explain our data.

Our study has uncovered the primary biologically relevant transcriptional parameters responsible for the phenotypic differences in the downstream signaling pathway of sog. The slower activating distal enhancer drives insufficient levels of sog to achieve the early refinement of the pMAD gradient. However, the high loading rates achieved by the distal enhancer enable enough build-up of Sog in the later stages of NC14 to eventually reach near wild-type restriction of the pMAD domain. In contrast, the faster acting proximal enhancer is capable of achieving an early contraction of the pMAD domain but fails to drive sustained expression of sog at high enough levels to continue this contraction. The expansion of the expression domain of the ‘early’ pMAD target gene ush, but not the ‘late’ pMAD target gene hnt seen in ΔPsogMS2 embryos, while the opposite is seen in ΔDsogMS2 embryos, gives good indication of the validity of this model.

With this in mind the question naturally arises: why have two enhancers at all, if it is possible to achieve this result with only one? Based on our previous work on the distal enhancer in reporter constructs, we know that placement of the distal enhancer directly upstream of a promoter is capable of driving fast transcriptional activation at high levels (Yamada et al., 2019). Beyond increasing the robustness of transcription as proposed by previous studies of shadow enhancers (Frankel et al., 2010; Perry et al., 2010; Tsai et al., 2019), we believe our data elaborates on the original hypothesis that shadow enhancers act as a source for evolutionary novelty (Hong et al., 2008). In its original conception, the de novo creation of a shadow enhancer allows one of these enhancers to drift, potentially adding new functionality without disturbing the core role of the original transcriptional program.

An alternative view to this interpretation is that selection may favor the creation of enhancers that allow for the tuning of individual transcriptional parameters. In our study, loss of a single enhancer produced defined and unique differences in phenotypic outcomes based on the parameter that enhancer was principally responsible for controlling, either the activation speed in the case of the proximal enhancer, or loading rate in the case of the distal enhancer. By keeping these activities separate, mutations in either enhancer will create smaller, but more precise, changes in the downstream patterning events, reducing potential pleiotropy that would be present if sog was driven by a single enhancer. Overall, this partitioning of enhancer activity would allow for a more defined exploration of the landscape of potential phenotypes during periods of increased selective pressure.

All flies were grown on standard fly (Drosophila melanogaster) cornmeal-molasses-yeast media. Fly stocks used in this study were: y[1]w[1118] (used as wild-type flies) and y[1] sog[S6]/FM7c, sn[+] [used as a sog null allele; Bloomington Drosophila Stock Center (BDSC), stock number 2497]. zld⁻ embryos were made using UAS-zld shmir lines and the Gal4 driver, MTD as previously described (Sun et al., 2015). Flies of the genotype y[1] w*; P{His2Av-mRFP1}II.2; P{nos-MCP.EGFP}2 (BDSC, stock number 60340) carried two transgenes, one on chromosome 3, P{nos-MCP.EGFP}2, which expresses the MS2 coat protein (MCP) fused to EGFP under the control of the nanos promoter active in oogenesis, and the other on chromosome 2, P{His2Av-mRFP1}II.2, which expresses RFP-tagged His2Av in all cells under the control of His2Av. Embryos from these and CRISPR-engineered flies (see below) were collected on yeasted grape juice agar plates, aged, and either fixed or live imaged (see below). Both fixed and live experiments were performed using female embryos, with the MS2 allele always coming from the male flies. Females were selected for either by imaging embryos carrying two sog alleles in the fixed experiments or by looking for MS2 expression in the live experiments.

All engineered fly lines were created through CRISPR-Cas9-mediated homology directed repair. sog enhancer sequences that were deleted are listed below. Transgenic Cas9 flies were co-injected with pCDF5 plasmids (Addgene, plasmid #73914; Port and Bullock, 2016) encoding guides targeting relevant genomic targets and pGEM-T vectors (Promega, A3600) containing homology repair templates. All injections were performed by BestGene. pGEMT donor DNA vectors were generated from fragments obtained through genomic PCR for homology arms, and sequences subcloned or PCR amplified from existing plasmids. All 24× MS2 loops containing plasmids used the MS2 sequence found in the MS2v5(-TAG) vector (Yamada et al., 2019). The neutral spacer DNA in the primary deletion plasmid was generated using the spacer sequence found in Scholes et al. (2019) and was generated as an IDT (Integrated DNA Technologies) gene block. The 3×3P-RFP sequence (Berghammer et al., 1999; Sheng et al., 1997) for the distal deletion (ΔD) plasmid was a generous gift from the Desplan Lab (NYU Biology, NY, USA). For screening of ΔD crispants 3×3P-RFP was used as a marker; there was no apparent RFP fluorescence in early embryos of flies carrying this allele. Moreover, this allele behaved transcriptionally identically to a straight deletion of the enhancer in colorimetric in situ hybridization assays (see Fig. S5). Plasmids were assembled using a combination of restriction enzyme digest and ligation, and Gibson assembly cloning (Gibson et al., 2009). Primers used to create donor vectors for each fly line are listed in the supplementary Materials and Methods, along with the guide sequences associated with each injection. Plasmid sequences and maps can be found at https://rushlowlab.bio.nyu.edu/resources/.

Proximal+distal deletion (ΔPΔD) flies were generated by injecting the ΔD guide plasmid and donor plasmid on the background of the proximal deletion (ΔP) fly line homozygous for transgenic Cas9. Flies expected to contain 3×3P-RFP cassettes were screened for red fluorescence, all other lines were screened via PCR using primers that spanned the MS2 insertion: Fwd, 5′-tgacgtttgattagccaccagttggg; Rev, 5′-gccaacctcaacttccaatctccg.

Embryos were collected and aged to be 1-3 h old at room temperature and dechorionated in Clorox for 2 min. They were then fixed in 4% formaldehyde (1× PBS) and an equal volume of heptane for 25 min while shaking vigorously. Devitellinization was performed by pipetting off the bottom fixative phase and adding 4 ml of methanol and shaking vigorously for 30 s. Embryos were rinsed in methanol and transferred to ethanol for storage at −20°C. Hybridization of fixed embryos used a standard in situ hybridization protocol and DIG-labeled sog cDNA or lacZ RNA antisense probes (hybridized at 55°C overnight). Visualization of the labeled probe was carried out using anti-DIG-AP (alkaline phosphatase) antibodies (Roche Biochemicals) followed by histochemical enzymatic staining reagents (Roche Biochemicals). Embryos were mounted on slides with Aqua-Polymount (Polysciences) using 1.5 coverslips (Fisher Scientific) and imaged with Zeiss Axiophot DIC optics and a Zeiss Cam and ZEN2012 software.

Probe sets for smFISH were generated using the online Stellaris (LGC Biosearch Technologies) probe designer. sog probes were ordered to be conjugated to Atto-670 and MS2 probes were ordered to be conjugated to Atto-570. Embryos were fixed in the same manner as outlined above, and stained following the Drosophila whole embryo staining protocol found on the Stellaris website (https://www.biosearchtech.com/support/resources/stellaris-protocols). After in situ staining, embryos were washed 3× with PBS-Tris, and stained overnight at 4°C with anti-Dorsal antibodies (see below) followed by staining with fluorescently labeled secondary antibodies for 1.5 h at room temperature (see below).

Antibody staining was performed at 4°C for 16 h followed by three 20 min washes in PBS+0.1% Tris (pH 7.0). Anti-Dl antibody (Dl_7A4) was obtained from the Developmental Studies Hybridoma Bank and used at 1:50 dilution. Anti-pMAD antibodies were obtained from Cell Signaling Technology and used at a 1:300 dilution. Embryos were then stained with secondary antibodies: Alexa Fluor 488 anti-mouse (Thermo Fisher Scientific, A-21202; 1:500) or Alexa Fluor 488 anti-rabbit (Thermo Fisher Scientific, A-21206; 1:500) for 1.5 h at room temperature and washed in the same manner. After DAPI (D9542, Sigma-Aldrich) staining for 20 min, embryos were mounted on microscope slides using ProLong™ Diamond Antifade Mountant (Thermo Fisher Scientific) and 1.5 glass coverslips (Fisher Scientific). Embryos were imaged using a Zeiss 880 with Airyscan confocal microscope.

All confocal images were captured on an LSM Zeiss 880 microscope. Images for the pMAD experiments were captured using a 20× objective with 1.1 Digital Zoom and a 2000×800 scanning area. Images all contained ∼20 z-planes. Laser power was set at 0.5% (408 nm) and 3% (488 nm), with gain set at 750. Images for all smFISH experiments were captured using the Airyscan module, and processed using the suggested Airyscan Processing strength. These images were captured using a 40× objective with 1.0 Digital Zoom and a 2000×1500 scanning area. Images all contained ∼50 z-planes. Laser power was set at 0.5% (405 nm), 5% (488 nm), 7% (561 nm) and 20% (633 nm), with gain set at 750.

Virgin females maternally expressing MCP-GFP and H2Av-RFP were crossed with males of the MS2 reporter lines. Embryos at 0-1 h-old were collected, dechorionated and transferred onto a breathable membrane (Lumox Film, Sarstedt AG & Co.) in the middle of a plastic microscope slide (3D printed on Ender 3 Pro, Creality). Live imaging was performed using a LSM Zeiss 880 63× objective lens with the following settings: optical sections: 1024×1024 pixels, 20 z-stacks 0.7 μm apart, 12 bit; zoom: 1.0; time resolution: 25 s per frame. Laser power was set at 0.6% (488 nm) and 0.4% (561 nm), with gain set at 800. Embryos were imaged for ∼1 h, typically from NC12 to late NC14.

Processing for images followed a pipeline, starting with feature extraction using standard tools in Imaris, then data exported to .csv files for organization, further processing and plotting. For pMAD experiments, nuclei positions were obtained using the ‘spots’ function with an estimated diameter of 4 μm, and a z-axis diameter of 7 μm, with background subtraction enabled. Spot positions were restricted to an area of interest ∼75%-25% of egg length. Fluorescence intensity from the pMAD channel at all spot positions was extracted and processed using the ‘pMAD_quant.R’ script. This script aligned, plotted and extracted gradient widths from all pMAD gradients measured.

Nuclei with more than two assigned foci were excluded from the analysis, and represented less than 1% of the data. During the creation of this pipeline, accuracy was additionally assessed by visual inspection. Each nucleus and its assigned foci were plotted on 2D graph with their x and y coordinates and given a unique color, making it easy to spot any errors not picked up by the filtering step.

Live imaging analysis was performed on Imaris by tracking nuclei using the ‘spots’ function with an estimated diameter of 4 µm and a z-axis diameter of 6 µm. Tracking was performed using the ‘retrograde motion’, with a maximum allowable gap of 1, and a maximum allowable displacement of 10 µm. Foci were also tracked using the ‘spots’ function with an estimated diameter of 1.3 µm and a z-axis diameter of 2 µm. Tracking was performed using the ‘retrograde motion’, with a maximum allowable gap of 0, and a maximum allowable displacement of 2.5 µm. Spots were filtered by inclusion of foci with ‘Quality’ scores greater than 33.0, median RFP fluorescence greater than 200 AU, mean GFP fluorescence greater than 250 AU and a distance from the xy-border greater than 1 µm. All tracking data, including position and mean GFP fluorescence, was exported to .csv files for further analysis in R.

Foci were assigned to nuclei by finding minimum distance between foci and nuclei.

Subsequently, any nuclei that came within 3 μm of the xy-border were filtered out to reduce edge effects. Nuclear cycle times and D/V axis relative positions to the mesoderm were annotated manually and stored in a separate .csv file. Nuclei were assigned into positional bins by taking the difference between the annotated mesoderm y-coordinate and the average position of the nucleus for each nuclear cycle. t_on values for NC13 and NC14 were obtained by subtracting the time GFP foci were first detected from the annotated cycle time of the respective nuclear cycle. Loading rates were estimated by fitting a linear model to the first five time points of the GFP foci intensity. Negative values were discarded, and represented less than 5% of the data. Total output values were calculated by multiplying the loading rate of each nucleus by the total time that foci were detected, with a maximum allowable time of 25 min in NC14 to account for differences in imaging time between each movie. These values were then averaged for each positional bin and multiplied by the percentage of active nuclei in the corresponding bin.

All plots were generated using base R plotting functions. All error bars were computed using the standard error of the mean (s.e.m.).

Models of activation were constructed by fitting gamma distributions to measured t_on and t_off values using the function fitdist() included in the ‘fitdistrplus’ library. Fits were achieved via maximum likelihood estimation. The shape and rate of each distribution was extracted and used to construct new distributions of values which were sampled independently to generate simulated t_on and t_off values. Distribution parameters were subsequently refined by comparing simulated nuclei to measured activation traces for each bin. New sets of potential shapes and rates for each distribution were generated by allowing each parameter to vary by up to 20%, and selecting new shape and rate values based on which parameters minimized the residuals between the prediction generated by the model and data.

During simulation, each nucleus was assigned a t_on value and t_off value generated from the corresponding distribution. At each time point, the number of nuclei that had a t_on value less than the current time, and a t_off value greater than the current time were considered ‘on’. The number of nuclei ‘on’ was divided by the total number of nuclei in the simulation, generating the value of the proportion of nuclei active for that time point. If the assigned t_off value was less than the assigned t_on value, the nucleus was considered ‘off’ at every time point. This allowed us to account for nuclei which never activate transcription without skewing the distribution of t_on, which was essential to accurately simulate the ventral bins.

For the combined model, nuclei were assigned two t_on and t_off values each sampled from the two different enhancer deletion distributions. Nuclei were evaluated in the same manner as described above, but only required one enhancer's values to meet the criteria of ‘on’ to be considered as such. All simulations were carried out using a set of 10,000 nuclei, which represented a compromise between accuracy of prediction and computing power.

The authors thank Sevinc Ercan for many insightful suggestions over the course of this study and Enrique Rojas for helpful discussions of our model. The authors also thank Michelle Pollak Lanis, Alicia Lina Zhu, Yandel Morel, Lauren Stafford, Caitlin O'Neil and Dhruvansh Shah for their tireless help with embryo collection and PCR screening for crispants, Caichen Duan for help with the live imaging, and Ignatius Tan (NYU Biology Microscopy Core) and Michael Cammer (NYU Langone Imaging Core) for advice on imaging.

The peer review history is available online at https://journals.biologists.com/dev/lookup/doi/10.1242/dev.200940.reviewer-comments.pdf.