The transcription factor Grainyhead primes epithelial enhancers for spatiotemporal activation by displacing nucleosomes

Abstract

Transcriptional enhancers function as docking platforms for combinations of transcription factors to control gene expression. How enhancer sequences determine nucleosome occupancy, transcription factor recruitment, and transcriptional activation in vivo remains unclear. Using ATAC-seq across a panel of Drosophila inbred strains we found that SNPs affecting Grainyhead binding sites causally determine the accessibility of epithelial enhancers. We show that deletion or ectopic expression of Grh causes loss or gain of DNA accessibility, respectively. However, while Grh binding is necessary for enhancer accessibility, it is insufficient to activate enhancers. Finally, we show that human Grh homologs, GRHL1/2/3, function similarly. We conclude that Grh binding is necessary and sufficient for the opening of epithelial enhancers, but not for their activation. Our data support the model that complex spatiotemporal expression patterns are controlled by regulatory hierarchies in which pioneer factors, such as Grh, establish tissue-specific accessible chromatin landscapes upon which other factors can act.

Each cell type in our body expresses a unique set of genes. Deciphering the regulatory programs that govern these transcriptional states requires predictive models that can link the genome sequence with the recruitment of transcription factors and other DNA binding proteins; then link this protein binding to chromatin state, the chromatin state with enhancer function, and the enhancer function with target gene expression. A better understanding of the DNA regulatory code will advance our interpretation of non-coding sequence variation, and may ultimately provide new ways to detect and manipulate cell states, for example in cancer cells or for regenerative medicine.

About a decade ago, the main bottleneck for deciphering gene regulation was to locate all the enhancers involved in the control of a certain cell state. Today that problem is largely solved thanks to advances in epigenomics^1–3. Indeed, high-throughput techniques such as ChIP-seq⁴ against transcription factors, co-factors, and histone modifications; DNaseI-seq⁵ and ATAC-seq⁶ for open chromatin profiling; and DNA methylation profiling⁷, have delivered genome wide enhancer predictions for a multitude of healthy and diseased tissues and cell types^2,3.

Now that we can systematically profile genomes, epigenomes, and transcriptomes, the next challenge is to discover the rules that link DNA sequence with chromatin state and enhancer function. The function of an enhancer is mainly determined by the specific combination of transcription factor binding sites it contains^8,9. Deciphering how motif combinations and their architectures encode a particular output function has been facilitated by Massively Parallel enhancer-Reporter Assays (MPRA)^10–12, which are used to test the function of a library of candidate sequences. Such assays can directly test which sequences yield what output, but are less well suited to include the chromatin state into the equation. Furthermore, only a few studies have addressed, on a genome-wide scale, how enhancers generate spatiotemporal expression patterns in vivo, because this requires the creation of a transgenic animal for each enhancer to be tested^13,14. Most enhancer models, including the billboard and enhanceosome models¹⁵ assume that the cooperative binding of transcription factors causes an enhancer to become accessible and concurrently active^15,16. Active enhancers always have bound TFs and are nucleosome-free¹⁷, however, little is known about the potential different roles of individual transcription factors with respect to nucleosome occupancy versus enhancer output. Some transcription factors can displace nucleosomes more efficiently than others, and these are often referred to as “pioneer factors” or chromatin openers^18–20. Examples of pioneer factors in mammalian genomes include SOX2, PU.1, FOXA1, and TP53^18,21–23. For TP53 it was recently shown that upon binding, it not only outcompetes nucleosomes, but also activates gene expression^22,24. For the other pioneer factors, it is less clear whether they can activate enhancers or whether additional factors (e.g., OCT4 in the case of SOX2)²⁵ are required to activate an enhancer. In Drosophila, the only pioneer factor known thus far is Zelda, which establishes the chromatin landscape in early Drosophila embryos during the maternal-to-zygotic transition^26,27.

Here we set out to study how the three layers, namely sequence information, chromatin accessibility and enhancer function are linked in epithelial cells. To this end we used a combination of chromatin accessibility (bulk, cell-sorted, and single-cell ATAC-seq), natural variation, machine-learning, evolutionary variation and in vivo enhancer-reporter assays in epithelial tissues in Drosophila. Our results provide strong evidence for a hierarchical model of enhancer control that is conserved across Metazoa.

Results

Natural variation in chromatin accessibility predicts potential chromatin regulators

To identify DNA features that are essential for establishing chromatin accessibility in a specific tissue, we profiled open chromatin across a cohort of inbred Drosophila strains. Particularly, we performed 30 ATAC-seq⁶ experiments on epithelial tissues (eye-antennal imaginal discs), covering 23 distinct strains from the Drosophila Genetic Reference Panel (DGRP)^28,29 (see Methods). The open chromatin profiles are highly similar (ρ = 0.76-0.96) within this cohort of eye-antennal discs and differ substantially (ρ 0.25-0.50) from the open chromatin profiles of non-epithelial tissues like the adult brain³⁰ (Fig. 1a, Supplementary Fig. 1).

Figure 1. Variation in chromatin accessibility across Drosophila inbred lines.

(a) Accessible chromatin profiles obtained by ATAC-seq on 32 independent tissue samples. The first 30 profiles are from epithelial tissues (eye-antennal discs), the last two from adult brains. (b) Example of two regions with caQTLs: the first has two caQTLs; chr2L:g.20115168G>C and chr2L:g.20115092A>C that correlate with a gain in accessibility (FDR = 2.16*10^-14), in the second region the caQTL chr2R:g.14233638C>A significantly correlates with a loss in accessibility (FDR = 1.1*10^-5) (see Methods). (c) Delta Motif plot, showing the cumulative effect of the 4289 caQTLs on the CRM score of 18832 motifs, plotted on the x-axis (see Methods). The y-axis shows the enrichment of motifs affected by caQTLs compared to control SNPs (one-sided Fisher’s exact test (log transformed)). Significantly enriched motifs are encircled and examples (red or orange) are visualized. (d) AUROC plot, showing the AUROC statistics for motif matching (18832 motifs) between the 2048 accessible and inaccessible sequences. On the x-axis are the AUROC -0.5 values plotted for each motif. The y-axis shows the significance (log transformed pval). Significantly enriched motifs are encircled and examples (red or orange) are visualized.

We identified 30774 accessible regions across the 30 eye-antennal disc samples (see Methods). To link variation in chromatin accessibility with underlying sequence variation (cis-variation), we applied a Generalized Linear Model (GLM) on all 297000 SNPs that were present in accessible regions. This analysis identified 4289 (~1.5%) SNPs that correlated significantly (FDR<0.05) with the accessibility change of their encompassing region (Fig. 1b), termed chromatin accessibility QTLs (caQTLs³¹, see Methods). The 4289 caQTLs were located in 2048 regions with variable accessibility between the different inbred lines (examples Fig. 1b).

One possibility to explain how one or a few SNPs can have such dramatic effects on chromatin accessibility would be that a SNP may break (or create) a recognition sequence of a key transcription factor. To identify such potential factors, we scored every region that contained one or more caQTLs with a curated collection of more than 18 thousand transcription factor binding motifs (see Supplementary Table 1 and Methods). Two independent methods then identified motifs for one transcription factor, namely Grainyhead (Grh), to concordantly change with chromatin accessibility (Fig. 1c and d). Overall, changes in a Grh motif were significantly more associated with caQTLs than with SNPs that have no effect on accessibility (Fisher’s exact test Padj=6.75*10^-22) and could directly explain the variable accessibility of 70 regions (Fig. 1c). These findings thus predict that a Grh binding site can causally determine the in vivo accessibility of an enhancer-size region.

Grainyhead plays a key role in the chromatin landscape of epithelial cells

Interestingly, Grainyhead is a highly conserved transcription factor with essential roles in epithelial cell fate specification and wound healing across Metazoa^32–36. From the 30774 accessible regions of the eye-antennal discs, 10.5% (3246 regions) have at least one Grh binding site. These 3246 potential Grh target regions are located near a large set of 1786 genes that are strongly enriched for GO³⁷ terms like epithelium development (pval=10^-56) (Supplementary Table 2). The Grh target regions contain 22.2% of the mapped reads and are in general the most accessible regions of the epithelial chromatin landscape (Welch’s t-test pval=2.251*10^-166, Fig. 2a). We hypothesised that Grh target regions are the most accessible because Grh stably binds its target sites in a large fraction of cells in the eye-antennal discs. Supporting this hypothesis is the fact that Grh proteins are ubiquitously expressed in basically all cells of the eye-antennal and wing imaginal discs (Fig. 2b-c).

Figure 2. Grainyhead plays a key role in the epithelial chromatin landscape.

(a) Box-and-whisker plots visualizing the normalized reads for accessible regions containing a Grh motif (n=3246) and all others accessible regions (n=27528). The median (centre lines), 25^th and 75^th percentiles (box edges), data points within 1.5 times the interquartile range from the edge (whiskers) and outliers (data points) are shown. The Welch two sample t-test (two-sided) was used to evaluate the difference (pval = 2.251*10^-166). (b-c) Grh-GFP protein expression in eye-antennal and wing imaginal discs, using a chimeric Grh-GFP fusion line³⁹ (reproducible GFP pattern in at least 5 discs, scale bar = 100μm) (d) Correlation plot between the accessibility (ATAC) and Grainyhead occupancy (ChIP(mentation)) of the 3246 Grh target regions (Spearman’s ρ = 0.919). (e) Tracks of Grh-ChIP-seq, Grh-ChIPmentation-seq and ATAC-seq in eye-antennal discs. Grh dependent enhancers (black bars), that control the expression of stit⁴⁴ (chr3R:21028800-21037000) and Ddc⁴³ (chr2L:19116600-19121750) upon wounding, are shown. (f) DNA footprint on the Grh motif; the averaged ATAC-seq signal over all 3246 accessible Grh regions, centred on their Grh motif, is shown.

To test whether Grh actually binds to the predicted Grh binding sites inside the accessible regions of the eye-antennal discs, we performed ChIPmentation³⁸ against Grh-GFP³⁹ and also re-analyzed published anti-Grh ChIP-seq data⁴⁰. The same Grh motifs that concordantly changed with chromatin accessibility (Fig 1c,d), were the strongest enriched in the Grh-ChIP peaks (Normalized enrichment score (NES)=12.30)^41,42. Interestingly, the ChIP-seq signal across the 3246 Grh target regions correlated quantitatively (ρ=0.92) with the ATAC-seq signal (Fig. 2d) and covered previously characterised Grh dependent enhancers^43,44 (Fig 2e).

We then evaluated the occupancy of the Grh binding sites by an independent assay by performing an in vivo DNA footprinting analysis^6,45 (see Methods) and found that Grh binding sites had a protection profile that was as strong as nucleosomal or silent DNA (Fig. 2e). These findings indicate that whenever a region with a Grh motif is accessible, Grh is stably bound there. Overall, these data suggest that Grh plays a key role, both in breadth (many target regions) and in depth (the highest peaks) in the accessible chromatin landscape of epithelial cells.

Grainyhead binding sites are essential for enhancer activity

So far, we showed that changes in a Grh binding site could generate or destroy the accessibility of an enhancer-sized region in developing epithelia. To test whether these regions are indeed functional enhancers we cloned four enhancer pairs, each representing different SNP sequences, one with and one without the caQTL that affected a Grh binding site. These regions were individually cloned into GFP reporter vectors⁴⁶ and stably integrated into the same position in the fly genome (see Methods) (Fig. 3a). Notably, the accessibility profile, assayed by ATAC–seq, of these integrated fragments was entirely determined by their sequence, independent of the local 3D chromatin context (Fig. 3a-d).

Figure 3. Activity of Grainyhead enhancers in imaginal discs.

(a) Stable insertion of pairs of cloned region into the genome using the VK37 integration site on chromosome 2L. (a-d) Peak height and shape of each region in their endogenous locus and after genomic insertion into the VK37 integration site, the ATAC-seq experiments were done once for each line. (e-l). GFP-expression (green) driven by the four tested enhancer pairs in eye and wing discs, differentiated photoreceptors are stained by ELAV (red) and nuclei are marked by DAPI (blue) (reproducible GFP pattern in at least 3 discs, scale bar = 100μm). Variable accessibility of the endogenous regions is shown, the accessible ones are each time presented at the top (e,g,i,k), while the inaccessible counterparts are presented underneath (f,h,j,l), the ATAC-seq experiments were done once for each line. A zoom in of the affected sequence is shown (caQTLs are marked by a star) together with the Grh PWM affinity score.

Next, by examining the expression pattern of the GFP reporter gene in eye and wing imaginal discs, we tested the direct effect of the caQTLs on the activity of these potential enhancers. Strikingly, the accessible sequences with an intact Grh binding site drove GFP expression in specific and reproducible patterns for all four fragments (Fig. 3e-g-i-k). Their counterparts on the other hand, lacking Grh sites, were predominantly inactive (Fig. 3f-h-j-l). Thus, the Grh binding sites are necessary for both enhancer accessibility and activity.

Grainyhead regulates enhancer accessibility but not activity

Our transgenic reporters showed a clear correlation between the presence of a Grh binding site and GFP expression. The four expression patterns were however distinct from each other and not ubiquitous (Fig 3e-g-i-k). Taking into account that Grh binding nevertheless determined chromatin accessibility and that Grh is ubiquitously expressed in the discs, we hypothesized that Grh binding opens or “primes” its target enhancers, without necessarily activating them. Regions that did not yield any noticeable transcriptional activity despite of being accessible or bound by TFs had also been reported in other studies^10,47.

To test whether Grh binds to enhancers without activating them, we measured the reporter activity of 21 additional Grh target enhancers (Supplementary Table 3), of which 15 showed activity in the eye-antennal disc (Fig. 4a). Interestingly, there was no correlation (ρ = 0.05) between the accessibility of these Grh bound enhancers and the number of cells in which the GFP reporter was active (Fig. 4b). Furthermore, using single-cell ATAC-seq, we found that the actual number of single cells, where the Grh bound enhancers were accessible, did correlate with their bulk ATAC-seq signals (ρ = 0.82, Fig. 4c, supplementary Fig. 2). These results already suggest that Grh binding opens its target enhancers, without necessarily activating them.

Figure 4. Accessibility of Grainyhead enhancers does not imply activity.

(a) Confocal images showing the activity (GFP, green) of 15 Grh bound Janelia-Gal4 enhancers¹⁴ in the eye-antennal discs (reproducible GFP pattern in at least 4 discs per shown image, scale bar = 100μm). (b) Plot visualizing the relative activity (GFP% of encompassing eye-antennal disc) and accessibility (ATAC-seq averaged over 30 independent eye-antennal disc samples) of the 15 Grh bound enhancers. (c) Correlation plot showing the accessibility of the 3246 Grainyhead regions (normalized ATAC-reads) on the x-axis versus the actual number of single cells where the region is accessible (Spearman’s ρ = 0.82). (d) Confocal image showing the reproducible (5 discs, scale bar = 100μm) expression pattern of the Optix2/3 enhancer⁴⁸ (GFP, green; cell nuclei DAPI, blue; and differentiated photoreceptors (ELAV, red)) in the eye-antennal disc. (e) Plot visualizing the relative activity (GFP%) and accessibility (ATAC-seq, circle: entire eye-antennal discs, triangle: Optix positive subpopulation) of the 15 Grh bound enhancers. (f) Accessible chromatin and Grh-ChIP profiles of two Grh bound enhancers, 40436 and 47530, both are accessible in the Optix positive subpopulation, but show no activity there (2 biological replicates for each sample).

Next, to demonstrate that these enhancers can indeed be accessible in cells where they are not active, we FAC-sorted and performed ATAC-seq on a specific subpopulation of cells from the eye-antennal disc⁴⁸ (Fig. 4d, supplementary note 1). This revealed that Grh-bound enhancers were similarly accessible in the sorted subpopulation compared to the rest of the tissue (Fig 4e) and that multiple enhancers, like 40436 and 47530, were accessible in the subpopulation without having GFP reporter activity in that domain (Fig. 4f). These findings indicate that Grh binding opens or “primes” its target enhancers, without activating them.

Evolutionary comparison of Grainyhead enhancers identifies candidate activators

The tested Grh-bound enhancers showed a large variety of expression patterns, suggesting that multiple other TFs can cooperate with Grh. To identify candidate co-TFs, we performed a cross-species motif analysis, since this is a powerful approach to identify functional TF binding sites⁴⁹. We thus performed ATAC-seq in the eye-antennal disc of 10 additional Drosophila species, and selected conserved enhancers that have a Grh motif and are accessible in the other species (example Fig. 5b). These two criteria were nearly 100% predictive of enhancers that are actually bound by Grh (Fig. 2d,e). Next, we scanned these enhancers for conserved motifs from our library of 18832 TF PWMs, using a Branch Length Score (BLS)⁴⁹ (Fig. 5a) and identified several co-conserved TF binding motifs (Fig. 5c, Supplementary table 5). Interestingly, one of the top co-conserved motifs is (CANNTG), an E-box motif for which the best candidate activator is Atonal, a bHLH factor that is active in a subset of cells in imaginal discs⁴⁶. These E-boxes were conserved in 92 Grh enhancers (Fig. 5d*, example Fig. 5b) which were located near Ato-induced genes⁴⁶ (Supplementary Fig. 3), suggesting that the [Grh+Ato] enhancers are functional. Furthermore, from twenty experimentally validated Atonal target enhancers⁴⁶, we found six to be co-bound by Grh and Ato (Supplementary Fig. 3), and identified four new Grh target enhancers that were active in Atonal expressing cells (Fig. 4a, BL48037, BL46823, BL38727 and BL50129). We conclude from these results that the activity of enhancers primed through Grh binding requires the recruitment of additional factors like Atonal.

Figure 5. Identification of co-regulator motifs through evolutionary conservation.

(a) Branch Length Scores for the 11 Drosophila species. (b) ATAC signal from eye-antennal discs of 11 Drosophila species on a conserved Grh-Atonal enhancer (two eye-antennal disc ATAC-seq replicates per specie). The Grh motifs are shown in red, Atonal in purple and repeats in green. (c) Histogram of the cumulative Branch Length Scores for the 18k motifs. The top Grh motif (red X) has a score of 3371, the top Atonal motif (purple X) has a score of 878. (d) Heatmap showing the CRM scores for the Atonal motif across the conserved Grh enhancers ordered by BLS score. The top 480 enhancers are shown, with the 92 most conserved Grh-Ato enhancers (BLS > 3) marked by *.

Loss-of-function and gain-of-function experiments demonstrate that Grainyhead is a pioneer factor

The data presented so far suggests that Grh is a pioneer TF for epithelial cells. The main function attributed to pioneer TFs is that they can bind their recognition sites within closed chromatin, making their target regions accessible and available for other proteins to bind^18–20. This means that removing a pioneer factor should directly reduce the accessibility of its target regions and vice versa that these regions should become accessible upon ectopic expression of the factor.

We tested both predictions and first investigated how lack of Grh proteins impacts the open chromatin landscape of epithelial cells. We thus performed bulk ATAC-seq on eye-antennal discs that were largely mutant for grh (grh^IM mutant clones)⁵⁰ (Fig. 6a, Supplementary Fig. 5). Comparing the accessible chromatin landscape of grh^IM mutant discs with matching controls identified 1076 regions with reduced accessibility (logFC<-0.5, pval<0.1) in the grh^IM mutants. Interestingly, only Grh motifs (from our 18K motifs) were strongly enriched in these regions (i-cisTarget⁴² NES=21). Known Grh target enhancers near epithelial genes, like Cad99C or jar, lost accessibility in the grh mutant discs (Fig. 6b). Overall, the 3246 Grh target regions were significantly less accessible in the Grh mutant tissue (Fig. 6c, Welch’s t-test pval=3.28*10^-8). In summary, these results demonstrate that loss of Grh directly reduces the accessibility of its target regions.

Figure 6. The pioneering role of Grh.

(a) Confocal images of control (top) and discs largely mutant for Grh (bottom), Grh C-terminal antibody⁶⁸ staining (red), nuclei DAPI (blue) (reproducible results for 5 discs, scale bar = 100μm). (b) Tracks visualizing ATAC of control (black) and Grh mutant (grey) eye-antennal disc and Grh-ChIP (dark-red) in eye-antennal disc (2 biological replicates each). Grh target regions near epithelial genes Cad99C and jar lose accessibility. (c) Volcano plot showing the change in accessibility (DESeq2 log2FC and –log10 pval) of the 30774 regions in control versus Grh mutant eye-antennal discs, red dots mark the 3246 Grh regions. (d) Ectopic grh expression in the neurons of developing larvae. (e) Volcano plot showing the change in accessibility (DESeq2 log2FC and –log10 pval) of 37668 accessible regions in control larval brains versus larval brain with ectopic grh expression, red dots mark the 3246 Grh regions. (f) Tracks visualizing ATAC on control larval brains (black), larval brains with 18 and 24h grh expression (purple) (ATAC-seq experiments were done once for each time point) and eye-antennal discs (dark grey, 30 biological replicates), and Grh-ChIP on eye-antennal discs (dark-red, 3 biological replicates). Genomic loci near CG34057 and scab are shown with Grh target region (red).

Next, we tested whether Grh was able to bind to nucleosomal DNA, making its target regions accessible, as this is a crucial feature of pioneer factors^18,19. To evaluate this, we analyzed larval brains, which mainly comprises of neurons and glial cells that do not express Grh, and where we found that the majority (74.3%) of Grh target regions were inaccessible. If Grh is a true pioneer factor of the epithelial epigenome, it should be able to bind to its target regions within this closed chromatin, making them accessible. To test this, we ectopically expressed the epithelial isoform of grh (grhN)³⁴ for 18 and 24 hours in all neurons of 3^rd instar Drosophila larval brains (Fig. 6d), using a pan-neuronal ELAV driver⁵¹, performed ATAC-seq, and compared their accessible chromatin landscape with that of wild-type brains. Remarkably, basically all 3246 Grh target regions, identified in the eye-antennal discs, specifically gained accessibility in response to ectopic Grh expression (Fig. 6e, Welch’s t-test pval=1.9*10^-10), while the neuronal cistrome did not change (supplementary note 2). Furthermore, only Grh motifs were significantly enriched (NES=26.6) in the 1774 regions with the strongest accessibility gain (logFC > 0.5, pval < 0.05) (Fig. 6f shows two examples). Potential target genes near these regions were enriched for GO terms that are normally not present in brain tissues, like embryonic dorsal epidermis (pval=9.6*10^-50) and epithelium development (pval=9.3*10^-29). Thus, consistent with our caQTL analysis and in vivo enhancer reporters, these results demonstrate that Grh is a pioneer factor, sufficient to directly and specifically open its target regions.

Functional Grainyhead motifs are embedded in a specific DNA context

Having established that Grh is a pioneer factor, capable of binding and opening its target regions in multiple tissues, we sought to investigate which sequence elements determine in vivo Grh binding. We first assessed whether the genomic Grh binding pattern was simply determined by the affinity for its binding sites. We collected the 10,000 highest scoring Grh motif matches in the genome, and ordered them according to their Grh occupancy (ChIP-seq meta-analysis, see Methods). The Grh-ChIPmentation and ATAC signal in wing imaginal discs clearly follows this ranking, indicating that despite their different cell fates, Grh binds to the same target regions in both epithelial tissues (Fig. 7a). From these 10,000 highest scoring Grh motif matches, we identified a set of 1,300 regions with recurrent Grh binding, and a set of 4,000 regions with non-functional (not bound nor accessible) Grh motifs (Fig. 7a). By combining five different Grh PWMs into a Random Forest model⁵² (Supplementary Fig. 5), we could distinguish functional from non-functional sites relatively well (AUROC=0.76, Padj<0.001), suggesting that the affinity for a Grh motif is already predictive for Grh binding.

Figure 7. Discovery of additional sequence features instructive for the in vivo binding of Grainyhead.

(a) Seqminer⁶⁹ plots visualising the Grh-ChIP(mentation) and ATAC-seq signals on the 10k highest scoring Grh binding sites genome wide (ranked using order statistics see Methods). Similar de novo Grh motifs were found on the top 1300 bound regions and bottom 4000 unbound regions. (b) Average GC fraction of the 1300 bound regions and 4000 unbound regions, showing 200 bp up and downstream of the Grh motif. (c) Repeat sequences of variable length (n) enriched around (±300bp) the 1300 functional Grh motifs. (d) Random Forest score for the encompassing Grh regions (smoothed over 15 regions). (e) Receiver Operating Characteristic curves assessing the performance of a Random Forest classifier to discriminate between 1300 Grh bound and 4000 unbound sequences. (f) Predicted nucleosome affinity⁵⁵ for the 1300 functional (orange) and 4000 non-functional (black) Grh regions.

Next, we examined the local sequence context around the Grh motifs (±300bp) to identify additional features that may affect Grh binding. We found that functional regions had a higher GC content (Fig. 7b), which may increase the flexibility of the DNA⁵³. Additionally, di- and tri-nucleotide repeat sequences (GA, AA, CAA and CAa/g) were enriched around functional Grh motifs (Fig. 7c). Combining all these features in a Random Forest classifier⁵² allowed us to quantitatively predict functional Grh binding sites (Fig. 7d) (AUROC=0.870, Padj<0.001, Fig. 7e, Precision-Recall Supplementary Fig. 5). Hence, a suitable Grh motif in a local “favourable” sequence context, is predictive for its in vivo binding.

Since nucleosome positioning is also strongly affected by the local sequence context⁵⁴, we investigated whether there is a difference in nucleosome affinity between the functional and non-functional motifs. We found that the predicted nucleosome occupancy profiles⁵⁵ differed significantly, with a pronounced dip in nucleosome occupancy at non-functional motifs, and a wider increase around functional Grh motifs (Fig. 7f). This indicates that Grh preferentially binds to DNA sites in regions that have a high intrinsic affinity for nucleosomes (i.e., likely to be bound by nucleosomes in the absence of Grh), similar to other pioneer factors like PU.1, FOXA, SOX2 and TP53^{18,20–22,24,56}. Thus, in vivo Grh binding is highly predictive and associated with nucleosome displacement.

Grainyhead homologs play similar roles in human

We have identified Grh as the principal pioneer factor of the epithelial accessible chromatin landscape in Drosophila. The mammalian homologs are three Grainyhead-like transcription factors (GRHL1, GRHL2, GRHL3)⁵⁷ with known and conserved cell adhesion target genes^58–60. Since the DNA-binding domain of Grainyhead proteins is highly conserved across Metazoa^32,33, we investigated whether Grh could have retained its pioneering functions in mammals.

First, using publically available ChIP-seq datasets for GRHL2^58,59, we confirmed that the DNA binding motif for Grh is conserved between Drosophila and mammals (Fig. 8a). We then identified a core set of regions that are recurrently bound by GRHL2 (Fig. 8b, ChIP-meta-analysis, see Methods) and compared the predicted nucleosome occupancy profiles⁵⁵ of functional versus non-functional GRHL2 motifs. We found that GRHL2 binding occurs more frequently in regions with a high preference for nucleosomes (Fig. 8c), very similar to the Grh binding profiles observed in Drosophila (Fig. 7f). Additionally, the crucial amino acids for GRHL1 to specifically engage the DNA⁶¹ were all conserved in the DNA-binding domain of Drosophila Grh (Supplementary Fig. 6). These findings suggest that part of the interaction mechanisms between Grh proteins and the DNA is conserved from Drosophila to human.

Figure 8. Human GRHLs have similar properties as their Drosophila homolog.

(a) Top enriched motifs found in the GRHL2-ChIP peaks. (b) Seqminer plots visualizing the GRHL2-ChIP signal of six ChIP experiments on OVCAR3, OVCA429, PEO1⁵⁹ and airway epithelium d29, d36 and d44⁵⁸ respectively, regions are ordered using a ChIP-meta-analysis (see Methods). (c) Predicted nucleosome preference of regions with functional GRHL motifs (red) and regions with non-functional GRHL motifs (black), centred on the GRHL motif. (d) Seqminer plots visualizing the accessible chromatin of MCF7 on the GRHL bound regions, public DHS data on non-treated cells, omni-ATAC on MCF7 cells 48h after non-targeting siRNAs, omni-ATAC on MCF7 cells 48h after mix of GRHL targeting siRNAs. (e) Omni-ATAC-seq tracks of the MCF7 cell line, control (black) and 48h after treatment with GRHL targeting shRNAs (Grey) (Omni-ATAC-seq experiments were done once for each sample). Predicted GRHL target regions (red) near two epithelial genes, PCDH1 and SPINT1 are shown.

To determine whether the GRHL factors also play a key role in setting up the mammalian epithelial chromatin landscape, we analysed the accessible chromatin landscape of the epithelial-like breast cancer cell line MCF7, using DHS data from ENCODE⁶². The entire accessible chromatin landscape of MCF7 cells was significantly enriched for GRHL motifs (NES=3.7), which were present in 22.8% of the accessible regions. These accessible GRHL target regions are located near 1337 expressed genes that were strongly enriched for GO terms like epithelium development (FDR=1.43*10^-65) and cell junction (FDR=4.21*10^-49). Combined with previous studies^59,58,36, these results indicate that the GRHL transcription factors play a key role in the mammalian epithelial accessible chromatin landscape.

Finally, we investigated whether the GRHL transcription factors also have a direct impact on chromatin accessibility. The GRHL1 and GRHL2 genes are highly expressed in MCF7 cells⁶³ and we had identified a set of putative GRHL target regions. To test whether the accessibility of these regions is due to GRHL binding, we performed omni-ATAC-seq on control MCF7 cells and MCF7 cells in which all three GRHL genes were knocked down for 48 hours using siRNAs (Fig. 8d, see Methods). We found that the regions with reduced accessibility upon GRHL knockdown were strongly enriched for GRHL target regions (Fisher’s Exact Test pval=2.11*10^-47) and GRHL binding sites (NES=12.8). These regions included GRHL regulated enhancers near target genes coding for proteins involved in cell adhesion, such as PCDH1 and SPINT1 (Fig. 8e). These findings suggest that the function of Grh as pioneer factor of the epithelial chromatin landscape is conserved between Drosophila and mammals.

Discussion

We performed to our knowledge the first in vivo caQTL study that links sequence variation with changes in chromatin accessibility and enhancer activation. Previous caQTL studies have been performed in human lympoblastoid or iPS cells^31,64. Performing a caQTL study in a complex, developing tissue, with a diversity of spatiotemporal expression patterns, enabled us to link regulatory variation with developmental enhancers. We discovered that SNPs that change a Grainyhead binding site can causally determine the in vivo accessibility of an enhancer-size region. Even though these enhancers also contain binding sites for other TFs, their accessibility can be attributed to Grh binding. Indeed, accessibility to epithelial cell enhancers is instigated when Grh is ectopically introduced in non-epithelial cells and abolished when Grh is removed from the epithelial tissue. We found that functional Grh binding sites are generally located in regions with a high affinity for nucleosomes, suggesting that in the absence of Grh its target enhancers are repressed due to strongly bound nucleosomes. Such an elegant scenario, where pioneer target enhancers are kept “off” in other cell types by nucleosome binding, was recently proposed as the “default off” model²⁰.

Grh is ubiquitously expressed in the imaginal disc, and its stable binding across multiple cell types is in agreement with a recent study in the Drosophila embryo³⁵, where Grh was found stably bound to its target genes throughout embryonic development. Even though Grh is bound to its targets in all cells of a tissue, the tested Grh bound enhancers were active in different patterns or at different stages. This observation is reminiscent of Zelda in the embryo^19,26,27 where the spatiotemporal enhancer activity of Zelda targets is induced by binding of combinations of other factors, such as the maternal activator Bicoid and the gap genes Hunchback, Giant, and Kruppel⁶⁵. Also for Grh we discovered many potential co-regulatory transcription factors, like for example Atonal, which is expressed and active in well-defined subparts of the eye-antennal disc⁴⁶. On the other hand, there are some Grh target genes, like coracle³⁴, Fasciclin3³⁴ and grh itself, that are ubiquitously expressed in epithelia. Identifying the co-factors or additional mechanisms behind this ubiquitous expression could be an interesting future challenge.

Importantly, only a subset of the enhancers that are active in imaginal discs are dependent on Grh. Many eye-specific enhancers such as Optix or p53 target enhancers (see Supplementary note 1 and 3), likely become accessible due to binding of the respective factor. Thus, multiple regulatory programs run in parallel within the same tissue, each with their array of target enhancers that are nucleosome-free either due to the binding of pioneer/lineage factors, or through the cooperative binding of multiple TFs¹⁶. The possibility that multiple regulatory layers are simultaneously active in the same cell also became apparent in neurons upon ectopic expression of Grh, which triggered ectopic accessibility of the entire epithelial cistrome, without affecting the neuronal cistrome (supplementary note 2). This further suggests that pioneer factors work in parallel, where each pioneer finds and opens its specific target regions, jointly establishing the accessible chromatin landscape upon which other factors can act.

A consequence of the hierarchical enhancer model is that the co-regulatory factors can also be repressors, which could explain why key transcription factors, like Grh, have been given both repressive and activating roles^35,58. For many such factors, it was unclear how one TF could activate some target genes while repressing others at the same time. We show in this paper that Grh merely primes its target regions, making them accessible for other factors to bind. The open enhancers can then be subject to binding by either transcriptional activators or repressors that exert their effects on gene expression, explaining why Grh removal simultaneously leads to the reduced expression of some target genes, while increasing the expression of others.

Mammalian GRHL factors have recently been implicated in a number of human diseases, all involving issues with epithelial cell fate. In ovarian cancer for example, GRHL2 expression counteracts epithelial to mesenchymal transition and increases the overall survival of patients^59,60. Mutations in the GRHL factors are also linked to pulmonary fibrosis⁶⁶ and several craniofacial disorders⁶⁷. We found that mammalian GRHL homologs are required for setting up the accessible epithelial chromatin landscape, and that the biochemistry of GRHL binding and nucleosome displacement are highly analogous to Drosophila. Our findings could help to better understand the role Grh (GRHLs) in human disease.

In conclusion, we set out to determine how sequence information is linked to chromatin accessibility and enhancer activity in vivo. We identified Grh as a pioneer factor of the epithelial chromatin landscape in Drosophila, with likely conserved roles across metazoa. Furthermore, our results support a model in which pioneer factors, such as Grh, sit on top of regulatory hierarchies, establishing tissue-specific accessible chromatin landscapes upon which other factors can act.

Online Methods

Fly husbandry and genotypes

From the Drosophila Genetics Reference Panel, we have used these lines (Bloomington numbers): 25174, 25185, 25186, 25187, 25193, 25194, 25199, 25208, 28123, 28153, 28202, 28206, 28222, 28224, 28243, 28260, 28262, 28265, 28275, 29651, 29652, 55026, 55031. For the Grainyhead ChIP we used: w[1118]; PBac{y[+mDint2] w[+mC]=grh-GFP.FPTB}VK00033³⁹ (Bloomington 42272). For cell sorting we used the Optix2/3-GFP⁴⁸ transgenic line. For generating the grh mutant eye-antennal discs we recombined the grh^IM mutation⁵⁰ on the FRT42 chromosome to generate the following stock: ey^3.5-flp/ ey^3.5-flp; FRT42 grh^IM/CyO^--, mhc-RFP and crossed them to w; FRT42, ey-flp; FRT42 cell-lethal/CyO, mhc-RFP and ey^3.5-flp / ey^3.5-flp; FRT42 Ubi-GFP/CyO. For the ectopic grh expression in larval brain we crossed these lines: ELAV-gal4;tub-gal80^ts X UAS-grhN³⁴. All Drosophila lines were raised on a yeast based medium and kept in an incubator at 25°C. For the ectopic expression of Grh, crosses were raised for ~10 days at 18°C, until the larvae were early 3^rd instar, and then shifted to 29°C for 18 or 24 hours to inactivate Gal80, permitting Gal4 to activate the expression of UAS- grhN in ELAV expressing neurons.

Irradiation

3^rd instar larvae received a dose of 40 grey of ionizing radiation (160 KeV, 25 mA for 10 minutes) using the small animal research irradiator (RS-2000). Two hours later ATAC-seq was performed on these 15 irradiated DGRP lines (25174, 25185, 25186, 25193, 25194, 25208, 28123, 28202, 28222, 28260, 28262, 28265, 28275, 29652, 55031).

Whole tissue ATAC-seq

We applied exactly the same ATAC protocol for eye-antennal discs as previously described^6,70. For each sample, we dissected either 10 eye-antennal discs, 4 wing discs or 2 brains from wandering third instar larvae.

Cell sorted and single cell ATAC-seq

Eye-antennal discs were dissected in PBS and placed in SF900 medium, once 200 eye-antennal discs were dissected, the SF900 was removed and replaced with 400µl Trypsin, 0.05% EDTA for dissociation. The eye-antennal discs were then incubated at 37°C for 1 hour with agitation, samples were mixed every 20 minutes with a pipette. After dissociation, cells were pelleted by centrifugation at 800 xg for 5 minutes at 4°C, washed with PBS and re-pelleted using the same parameters. Finally, the cells were resuspended in 400 µl PBS and filtered using a 40µm cell strainer and stained with propium iodide (final conc 1µg/ml) to exclude dead cells. The cells were sorted on a BD Aria I, selecting against the presence of PI and for the presence of GFP.

For the sorted ATAC-seq samples, 50,000 cells were sorted into Eppendorfs containing 50 µl PBS. Cells were pelleted by centrifugation at 800 xg for 5 minutes at 4°C and ATAC-seq was performed as previously described^6,70. For single cell ATAC-seq, as many cells as possible were sorted into an Eppendorf, these were then pelleted by centrifugation at 800 xg for 5 minutes at 4°C and resuspended at a concentration of 1000 cells/µl. Single cell ATAC-seq was performed as previously described⁷¹, using the 5-10µm OpenApp IFCs on the Fluidigm C1 and with no cell washing step.

ChIPmentation

Grh-GFP (Bloomington stock 42272)³⁹ 3^rd instar larvae were dissected in ice-cold PBS and the carcasses with attached discs were fixed for 25 minutes at room temperature by mixing in 1 ml crosslinking solution. The carcasses were washed and the eye-antennal imaginal discs were dissected and sonicated until chromatin fragments reach an average size of 500 bp. The sonicated chromatin was spun at top speed for 10 minutes at 4°C. As a preclearing step, 20 µl of protein A/G magnetic beads (Merck, Millipore) was added, incubated for 1 hour at 4°C and removed by centrifugation at 3000 rpm for 2 minutes. The anti-GFP antibody (ab290, Abcam) was added to a fixed chromatin aliquot and incubated overnight at 4°C. Immunocomplexes were recovered by adding protein A/G magnetic beads to the sample followed by incubation for 3 hours at 4°C. The magnetic beads with precipitated chromatin were washed. Beads were resuspended in elution buffer, RNase was added to the immunoprecipitated chromatin and incubated for 30 minutes at 37°C. The immunoprecipitated DNA was purified. To incorporate sequencing adapters, we combined the purified cDNA with 4 µl of Nextera TD buffer (Illumina) and 1 µl of Nextera Tn5 enzyme (Illumina) on ice and incubated at 55°C for 5 min. The tagmented cDNA was purified again on a MinElute column and eluted in 20 µl EB buffer. To PCR amplify the fragments, we added 25 µl of NEBnext PCR master mix (Bioke), 5 µl of Nextera primer mix and incubated at 72°C for 5 min, then at 98°C for 30 seconds, followed by 15 cycles of 98°C for 10 seconds, 63°C for 30 secs and 72°C for 3 min. We purified the PCR amplicons with 55 µl AMPure beads (Analis).

ATAC and ChIPmentation-seq analysis

Adapted sequences were trimmed from the raw reads using fastq-mcf (default parameters using a list containing the common Illumina adapters). Cleaned reads went to a quality control step using FastQC from Babraham Bioinformatics.

We avoided mapping artefacts, as we mapped the ATAC-seq reads for each line to its personalized genome. All experiments were mapped to the Drosophila melanogaster Flybase r5.13 genome⁷², since this was the version used in the DGRP project to call the variance. In a first step the reads were mapped using bowtie2⁷³ onto the personalized genomes of each DGRP lines or on the standard Flybase r5.13 genome for the ChIPmentation and non-DGRP ATAC samples. For each of the 23 DGRP lines we adapted the consensus genome (r5.13) using seqtk mutfa⁷⁴, each time including their SNPs that were previously called form whole genome sequencing^28,29. After the first mapping round, additional SNPs were called on the ATAC reads using SAMtools⁷⁵ mpileup -B –f r5.13.fasta DGRP_lineX.bam | varscan.sh mpileup2snp --output-vcf 1. Newly called homozygous SNPs (several thousand per line) were added to the existing vcf files using VCFtools⁷⁶. The genomes were again updated to obtain a final personalized genome for every DGRP line, strongly reducing mapping errors and increasing the sensitivity of subsequent analyses. Cleaned reads were mapped onto the final genomes using bowtie2⁷³ again and Samtools⁷⁵ was used for sorting and indexing.

Peaks were called on the mapped reads using macs2⁷⁷ callpeak (-g dm –nomodel (for ATAC only) –keep-dup all –call-summits). The narrow peak files (bed format) for all the DGRP lines were merged into a single file that contained a total of 33771 regions, accessible in at least one DGRP line. After filtering out chrU, Uextra, Het, mitochondria and the chromosome borders, we end up with 30774 accessible regions for the eye-antennal discs. For every ATAC-seq sample we count the number of reads falling into each accessible region using featureCounts⁷⁸.

The raw counts matrices (see GEO processed files) were further processed in R version 3.2.2. DESeq2⁷⁹ was used to normalize the raw counts matrices using the sizeFactor for each sample. All differential accessible regions sets were obtained using DESeq2, following the standard procedure for differential mRNA. The contrasts were; 2 Optix positive vs 2 Optix negative sorted populations, 2 grh^IM / cell lethal mutants vs 2 wild type / cell lethal eye-antennal discs, 2 ectopic grh expressing brains vs control and 8 irradiated DGRP lines vs their non-radiated controls.

Identifying enriched motifs

We used i-cisTarget^41,42, a sequence based motif enrichment tool that takes conservation into account, to identify motifs that were enriched in our sets of accessible regions (online version, input bed files, genome dm3, 18k motif database). For every input set of region, the enrichment of each motif was calculated as an area under the recovery curve, observed over 136K candidate regulatory regions. We selected the motifs with the highest Normalized Enrichment Scores (NES)⁴² for our given peak sets.

For all 30774 accessible regions in the eye-antennal discs Trl/GAGA-factor motifs were the top enriched (NES=7.6) followed by Grh motifs (NES=5.7).

For all three GFP-Grh-ChIP(mentation) peak-sets the top enriched motifs were for Grh (or GRHL) transcription factors (NES=11.1-12.3). Grh motifs were also discovered de novo in the Grh-ChIP-seq peaks, using Homer⁸⁰ de novo motif (pval 10^-370) or RSAT⁸¹ peak-motifs (sig=12.58, eval=2.6*10^-13).

For the 434 Optix specific regions the top motifs were for sine occulis / Optix (NES=19.9).

For the larval brains, we analysed all accessible regions and the 1774 regions that specifically went up in the brains with ectopic grh expression, results see in Fig. 6f.

Single cell ATAC-seq analysis

Data alignment was performed using bowtie2⁷³ on the Drosophila's melanogaster genome version dm3 (r5.13). ScATAC data was deduplicated using picard tools markDuplicates⁸². Out of the 96 single cells, 28 were removed from the data set based on the reads' distributions around TSSs. ChromVAR⁸³ was used to quantify the number of fragments falling within each of the potentially accessible regions. The resulting matrix, (with cells as columns, regions as rows and values as counted fragments in a region within a cell) was binarized, taking a value of 1 if at least one fragment was quantified within the region in the cell and 0 otherwise.

Correlation and Variability between accessible regions

The counts matrix of the 30 DGRP ATAC samples and 2 Adult brain samples (public data obtained from GSE83975³⁰ and processed like our own ATAC data) containing a total of 34768 accessible regions (DGRP and Brain merged) was used. We used the corrplot⁸⁴ package in R to calculate and visualise the spearman correlations between every sample pair.

GLM and Chromatin Accessibility QTLs

The 30774 accessible regions were filtered before performing the GLM to reduce the effects of noise and mapping bias. 1453 regions with a high repeat content (> 25% RepeatMasker⁸⁵ (UCSC)) were removed using bedtools.2.26.0⁸⁶ intersectBed –v –f 0.25. Next, 1513 regions with a low coverage for every DGRP line were removed (coverage of the region below 0.2pb for every DGRP lines), ending up with 27808 accessible regions. For each region, we extracted the normalized ATAC-seq reads for these 30 DGRP lines (25174_IR, 25185_IR, 25186_IR, 25187, 25193_IR, 25194_IR, 25199, 25208_IR, 28123, 28123_IR, 28153, 28202_IR, 28206, 28222_IR, 28224, 28243, 28260, 28260_IR, 28262, 28262_IR, 28265, 28265_IR, 28275, 28275_IR, 29651, 29652, 29652_IR, 55026, 55031, 55031_IR) and linked each region to the annotated and additionally called SNPs for these lines (see ATAC-seq analysis). 297000 SNPs were assigned to their encompassing region using bedtools.2.26.0 intersectBed on the extended vcf file.

For each region, we now have the normalized reads for each of the 30 lines as one vector and all SNPs called inside this region as a binary matrix for the 30 lines (present =1, absent=0, unknown=NA). We searched for correlating region-SNP vectors using the generalized linear model function in R (version 3.2.2)⁸⁷. The p-values were adjusted using the Benjamini-Hochberg procedure in R. With a FDR of 0.05 we identified 4289 highly correlating SNP-region pairs referred to as caQTLs (Chromatin Accessibility Quantitative Trait Loci).

Delta Motif scores

To single out motifs that correlate significantly with the open chromatin changes, a Delta motif score was calculated for every of the 18832 unique motifs in our collection. The sequence for each of the 2048 variable regions, that contained at least one caQTL, was extracted using bedtools getfasta. Next, we mutated these sequences with their encompassing caQTLs according to their effect on the open chromatin using seqtk mutfa. For each of the 2048 regions we now have 2 sequences, one for the accessible chromatin and one for the less accessible/closed chromatin. We scored every time both sequences with the 18832 motifs using cluster buster –m 0 –c 0 and retained for every motif the highest CRM score for each sequence. By subtracting the CRM score of the less accessible/closed region from the encompassing accessible region we obtained a delta motif score for that region. For every motif, we summed all delta scores from the 2048 regions to obtain a cumulative delta score for each motif (plotted on the x-axis of fig. 1c).

We calculated a Delta motif score, following the same procedure, on 66360 SNPs that were present in an accessible region but had no effect on chromatin accessibility (GLM FDR > 0.95). We then calculated for each motif whether it was significantly more affected (|Delta score| > 3) by caQTLs compared to the non-correlating SNPs, using the Fishers exact test⁸⁸ in R (log10 of the pval is plotted on the y-axis of fig. 1c).

AUROC statistics

In this method, used by the DREAM5 consortium⁸⁹, we calculated the AUROC statistics for motif matching between the 2048 accessible and inaccessible sequences (see above). For every one of the 18832 motifs, we first selected all the regions where the motif was present (CRM score > 4 in the accessible and/or inaccessible sequence in at least 50 regions) and then calculated AUROC statistics using the pROC⁹⁰ package in R. We then calculated for each motif its ability to discriminate accessible versus inaccessible sequences by comparing the AUROC to a random classifier (0.5). We calculated a p-value using stratified bootstrapping (n=2000) again with the pROC⁹⁰ package in R. For representing the data in Figure 1d we subtracted from every AUROC value 0.5 (random classifier) and log transformed the p-value.

DNA footprint

We merged all 30 ATAC-seq samples, resulting in a data set of more than 350 Mio mapped reads. We adapted the mapped reads to represent the actual binding of the transposons, according to the original ATAC-seq paper⁶. To obtain the footprint on the grh motif, we selected all Grh target regions (3246), centred each region on the best Grh motif and plotted the ATAC-seq signal, normalized by the number of regions.

Ranking Grainyhead regions using Order Statistics

Drosophila eye-antennal discs: We combined the publicly available Grh-ChIP-seq⁴⁰ data together with two new inhouse Grh-ChIPmentation datasets. Human: We combined 6 publicly available GRHL2-ChIP-seq^58,59 datasets from epithelial cell lines.

First, we score the entire Drosophila Melanogaster genome (dm3 r5.53) / Human reference genome (hg19) with the top enriched Grainyhead motif using Cluster-Buster⁹¹. We selected the top 10000 highest scoring motifs (none overlapping) and generated a bed file, extending the motif location by 300 base pairs both up- and downstream. Next, for each region in the bedfile we counted the number of reads coming from the three ChIP experiments, using bedtools multiBamCov⁸⁶. We obtained a final ranking base on the ChIP signal from all three experiments using order statistics. This ranking of regions with a Grainyhead motif was used to plot all other data in Fig. 3a (Grh-ChIPmentation in wing discs, ATAC in both eye-antennal and wing discs) using seqminer⁶⁹.

Random forest

We trained and evaluated random forest classifier⁵² models (ensemble of 151 decision trees), using the scikit-learn Python package⁹². The max_features parameter was set to the square root of the total number of features used to train the model. We evaluated the performance of our models by generating Precision Recall cures and calculating the AUC (Fig. 3f). This was each time done on left out data, to be precise we trained the model on half of the positive (650) and half of the negative (2000) sequences and scored the unseen other half (and vice versa).

Feature discovery was done by comparing the sequence of the top 1300 bound versus the bottom 4000 unbound regions (each time 612 basepair sequences, centered on the Grainyhead motif). We ran RSAT peak motif⁸¹ to discover oligo’s, motifs and dyads that were significantly enriched in the bound versus the unbound Grainyhead regions. We identified several repeats (Fig. 3d) that were significantly enriched in the bound regions around the Grh motif.

We took these repeat elements (GA, CAr, CAA, AAA, A, CAr_GA, CAr_GA_AAA) and used Cister⁹⁴, a tool for detecting cis-elements clusters with optimized parameters for the repeats (-a 0 -w 1000), to score the entire sequence (612 bp). We then integrated the Cister output over the sequence to obtain one score (feature) per repeat element. This way we can take the high variability of repeat lengths into account, longer repeats result in a higher feature score.

The score for the five Grainyhead motifs (Supplementary Fig. 5) was obtained using Cluster-Buster⁹¹, we used each of the five Position Weight Matrices to score the Grainyhead sequences and retained the top CRM score.

The GC fraction was calculated by selecting the DNA sequence one nucleosome up- and downstream of the Grainyhead motif (+-147 base pairs) and simply counting the occurrence of G’s and C’s divided by the total number of base pairs.

The final features that were used for training and evaluating the random forest were the following: (AUROC on 1300 positives and 1300 negatives (balances set) and AUPRC on 1300 positives and 4000 negatives (unbalanced set)):

• a)No training, random ordering (grey PR curve, AUROC = 0.5, AUPRC = 0.245, Padj = 1)
• b)One Grh motif (brown PR curve, AUROC = 0.613, AUPRC = 0.345, increase versus (a) Padj < 0.001)
• c)5 Grh motifs top CRM scores (red PR curve, AUROC = 0.766, AUPRC = 0.55, increase versus (b) Padj < 0.001)
• d)Integrated Cister score for the repeats (GA, CAr, CAA, AAA, A, CAr_GA, CAr_GA_AAA) over the entire region (612 bp) and closer to the Grainyhead motif (one nucleosome around the motif (306 bp) (green PR curve, AUROC = 0.82, AUPRC = 0.634, increase versus (c) Padj < 0.001)
• e)GC fraction for each region (one nucleosome (+-147 base pairs) around the Grh motif) (purple PR curve, AUROC = 0.842, AUPRC = 0.674, increase versus (c) Padj < 0.001)
• f)All previous features (Grh motifs, repeats and GC) combined (blue PR curve, AUROC = 0.87, AUPRC = 0.736, increase versus (d) Padj < 0.001, increase versus (e) Padj = 0.03)

Significance (Pval) of the improvement in AUROC or AUPRC was calculated using bootstrapping versus random or the specified previous RF result.

DNA shape and nucleosome predictions

Every time we compared the 1300 bound Grainyhead motifs versus the 4000 unbound motifs (extended by 300 base pairs each side). The DNA shape data (Helix Twist, Minor Groove Width, Roll and Propeller Twist) were downloaded from ftp://rohslab.usc.edu/dm3/⁹⁵ and an average score for the bound and unbound regions was calculated.

A similar analysis was performed for the nucleosome prediction data, this time extending 1kb around the motif. The nucleosome prediction data was obtained from https://genie.weizmann.ac.il/software/nucleo_genomes.html⁵⁵. For the Human nucleosome prediction we used the top 1000 GRHL2 bound regions versus bottom 1000 obtained after orderstatistics.

Drosophila Species and Branch Length Score

We performed ATAC-seq on the eye-antennal discs of 3^rd instar larvae from 10 related Drosophila species; Simulans, Sechellia, Yakuba, Erecta, Anasai, Pseudoobscura, Persimilis, Wilistoni, Mojavensis and Virilis.

The ATAC-reads were mapped to their respective genomes and peaks were called using macs2⁷⁷ standard parameters.

We used Kent tools liftover –minMatch=0.1 to obtain the bed coordinates of the Grainyhead enhancers in the other species. Next we used bedtools.2.26.0 intersectBed –wa -wb –f 0.1 –a specie.narrowpPeak –b specie.liftover.bed to obtain the list of conserved Grainyhead enhancers that are also accessible in their respective genomes. For every specie, the sequences of the conserved enhancers were obtained using the bedtools.2.26.0 getfasta command. All sequences were scored with our collection of 18832 motifs, using clusterBuster with the option –m 0 –c 0. For each motif, the highest CRM score per sequence was used to calculate the Branch Length Score according to⁴⁹. The Newick format was used for the phylogenetic tree data from http://flybase.org/maps/chromosomes/synteny_table.html. Branch length scores were summed over all conserved sequences to obtain a total score for every motif. As control, we shuffled all sequences (keeping same base pair compositions and sequence length) using shuffleseq from the EMBOSS package⁹⁶ and calculated the BLS score again for all motifs. The corrected Branch Length Score, used to evaluate the co-conservation of motifs, was obtained by subtracting the BLS score from the shuffled sequences from the BLS score calculated on the real sequences.

Co-conserved motif clustering

After calculating the BLS scores for our 18832 motifs, we selected those with a corrected BLS score over 700 (387 motifs or top ~2%). Next, to visualize these motifs we first clustered them using STAMP: a tool for DNA-binding motif similarities⁹⁷. We obtained 77 motifs clusters, from which we only retained those motifs that were directly annotated for a transcription factor that is expressed in the eye-antennal discs (supplementary table 5).

Gene Set Enrichment Analysis

Publicly available micro array data (Atonal gain of function: GSE16713⁴⁶, ionizing radiation: GSE37404⁹⁸ was used to generate the ranked gene list. Differential gene expression was calculated using GEO2R, the log fold change was used to rank the genes.

Regions were assigned to neighboring genes (5kb upstream or intronic) and these gene lists were used in the Gene Set Enrichment Analysis⁹⁹.

Generation of Transgenic lines

Genomic DNA was extracted from an adult fly of lines 25208, 28123, 28222.

The four enhancer-pairs of interest were obtained by genomic PCR from the specific lines using the primers listed in Supplementary table 6.

The PCR product was purified on an 1.2% agarose gel. The correct band was cut out and further purified using the Qiagen gel extraction kit. The constructs were cloned into an entry vector using the pENTR/D-TOPO cloning kit following the standard online protocol. The plasmids were transformed into chemically competent DH5alpha cells. The right construct was confirmed by sequencing. In a second cloning step, we set up the LR Gateway reaction (Invitrogen) between the entry clone and the modified pHStinger vector⁴⁶. The final constructs were stably and site specific (VK37(2L)22A3) integrated into the Drosophila Melanogaster genome by injecting it into embryos (done by GenetiVision) using the PhiC31 system.

Activity of Grh enhancers

Drosophila lines from the Janelia-Gal4 FlyLight¹⁴ enhancer project were selected that had a Grh-ChIP and ATAC peak + strong Grh motif (Sup table 3). These lines were crossed to a line with UAS-eGFP and eye-antennal and wing imaginal discs from 3^rd instar larvae were dissected, fixed and stained.

Immunohistochemistry and image analysis

Imaginal eye and wing discs from third instar larvae were fixed in 5% formaldehyde at RT for 30 min. Next, they were washed in PBT (PBS + 0.3% TritonX-100) and blocked in 5% normal Donkey Serum in PBT (PBNT) for 15 min. For testing the Grainyhead enhancers, the tissues were incubated with the primary antibody mix rabbit anti-GFP (Invitrogen), rat anti-ELAV (DSHB) at 4°C overnight. The secondary antibody Cy3-Donkey-anti rabbit (Molecular probes) and DAPI were added for 2h at RT, and then the samples were washed with PBNT, PBT, and PBS (3x10 min). Samples were postfixed 10 minutes in 4%Formaldehyde and washed (3x) with PBS before mounting the discs in Vectashield (Vector Laboratories). Rabbit-anti-Grh-Cterm (1:500) was stained similarly as for Dcp1 and visualized using Alexa647 conjugated Donkey-anti-Rabbit secondary antibody (Molecular probes).

For imaging the Olympus FV1200 confocal microscope was used (20x oil, z-stack). ImageJ (Bio-formats Importer plug-in) was used to merge and process the images. The quantification of the GFP positive fraction of cells in the discs was also done in ImageJ, manually selecting the threshold for each disc before measurement.

Knock-down of GRH-like transcription factors in human epithelial cell culture

Human epithelial, breast carcinoma MCF7 cells were cultured at 37°C and 5% CO2 in Dulbecco’s Modified Eagle’s Medium (ThermoFisher Scientific), supplemented with 10% fetal bovine serum (Lonza) and penicillin-streptomycin (ThermoFisher Scientific). Combined knockdown of GRHL1, GRHL2 and GRHL3 was performed using a mix of the ON-TARGETplus GRHL1-3 siRNA SMARTpools (Dharmacon) at a final concentration of 50nM for each pool in opti-MEM medium (ThermoFisher Scientific), we used qPCR to confirm the knockdown. Omni-ATAC was performed after 48 hours according the previously-published protocol¹⁰⁰.

ATAC-seq reads were mapped to the reference genome (hg19-Gencode v18) using Bowtie2⁷³ 2.2.6 (with the --sensitive-local parameter). Reads with mapping quality lower 4 were filtered out. Differential peaks were called using MACS2⁷⁷ algorithm (q < 0.01, --nomodel), with the NTC sample as treatment and GRH_kd sample as control.

The 1000 most differential peaks (lost in the GRHL knockdown) were used as input set for i-cisTarget⁴², a motif enrichment tool that found GRHL motifs as top enriched. We used the Fisher’s Exact Test for Count Data⁸⁸ in R to calculate for the regions that lost accessibility the overall enrichment between the GRHL target regions (4017) and all other accessible regions (13555) in MCF7 cells. f

Statistics and Reproducibility

caQTLs FDR: see GLM and Chromatin Accessibility QTLs.

The enrichment for caQTLs for each of the 18k motifs was calculated using the Fisher’s exact test (one sided) and pvals were corrected using Benjamini-Hochberg. The top scoring motif (Grh) had a Padj=6.75*10^-22 having 34 out of 19804 negatives versus 45 out of 2048 positives.

All comparisons between accessible regions (normalized ATAC-seq reads) were done using the two-sided Welch Two Sample t-test in R. Fig 2a: Grh regions n =3246, Other accessible regions n = 27528, t = 29.057, df = 3477, true difference in means is not equal to 0, p-value = 2.250596e-166. Fig. 6c: n = 3227, t = 5.5326, df = 6390.7, true difference in means is not equal to 0, p-value = 3.281e-08. Fig. 6e: n=3227, t = -6.3823, df = 6428.9, true difference in means is not equal to 0, p-value = 1.866e-10.

Fig. 8d Two-sided Fisher’s Exact Test (MCF7: 13555 accessible regions without GRHL motif of which 265 go down, 4017 with GRHL motif of which 285 go down) pval=2.11*10^-47.

Data availability and Accession Code Availability Statements

UCSC Genome Browser hub with ATAC-seq tracks for all DGRP lines, eye/wing injected lines ATAC-seq and ChIPmentation against Grh-GFP in eye-antennal and wing discs:

http://ucsctracks.aertslab.org/users/jjacobs/DGRP_injections/bigWig_norm/hub_norm.txt;

as well as a hub with all cross-species ATAC-seq: Evolution hub:

http://ucsctracks.aertslab.org/papers/evolution/hub.txt. To load these in the UCSC Genome Browser, go to My Data – Track Hubs.

The raw and processed data is available from Gene Expression Omnibus under accession number GSE102441.