Identification of Novel Direct Targets of Drosophila Sine Oculis by Integration of Genome-wide Data Sets

Abstract

Drosophila eye development is a complex process that involves many transcription factors (TFs) and interactions with their cofactors and targets. The TF Sine oculis (So) and its cofactor Eyes absent (Eya) are highly conserved and are both necessary and sufficient for eye development. Despite their many important roles during development, the direct targets of So are still largely unknown. Therefore the So-dependent regulatory network governing eye determination and differentiation is poorly understood. In this study, we intersected gene expression profiles of so or eya mutant eye tissue prepared from three different developmental stages and identified 1731 differentially expressed genes across the Drosophila genome. A combination of co-expression analyses and motif discovery identified a set of twelve putative direct So targets, including three known and nine novel targets. We also used our previous So ChIP-seq data to assess motif predictions for So and identified a canonical So binding motif. Finally, we performed in vivo enhancer reporter assays to test predicted enhancers from six candidate target genes and find that at least one enhancer from each gene is expressed in the developing eye disc and that their expression patterns overlap with that of So. In summary, we expand the set of putative So targets and show for the first time that the combined use of expression profiling of so with its cofactor eya is an effective method to identify novel So targets. Moreover, since So is highly conserved throughout the metazoa, our results provide the basis for future functional studies in a wide variety of organisms.

Introduction

Transcriptional regulation involves a variety of complex biological processes that are dependent on the interactions among transcription factors (TFs) and their target genes. TFs activate and/or repress downstream target genes through binding to their regulatory motifs. This regulation often involves feed-forward and feedback loops. Despite their vital regulatory roles during development, much remains to be learned about the underlying mechanisms, especially the direct targets of TFs. There are two main approaches utilized for gene regulatory network (GRN) mapping. First, ChIP-Seq, which combines chromatin immunoprecipitation (ChIP) with massively parallel DNA sequencing (Seq), can identify the binding regions of TFs genome-wide. The second is gene expression profiling using RNA sequencing (RNA-Seq) or microarrays, which rely on genetic perturbations of TF function (Walhout, 2011). However, both approaches have disadvantages. For example, ChIP-Seq can be technically challenging in the absence of high-specificity antibodies. On other hand, gene expression profiling does not provide cis-regulatory information. However, a combination of gene expression profiling and motif discovery is an effective way to discover direct targets of TFs. In Drosophila, this approach has been successfully used to identify targets for Atonal and Glass, two key TFs that regulate retinal differentiation (Aerts et al., 2010; Potier et al., 2014).

One of the pivotal transcription factors controlling retinal cell fate specification in Drosophila is Sine oculis (So). So recruits Eyes absent (Eya), a transcriptional coactivator and a protein phosphatase, to target DNA where they synergistically direct many developmental processes, including cell fate determination, differentiation, proliferation, and survival in organisms ranging from flies to humans (Abdelhak et al., 1997; Bonini et al., 1993; Cheyette et al., 1994; Dozier et al., 2001; Hoshiyama et al., 2007; Kawakami et al., 2000; Sahly et al., 1999; Seo et al., 1999; Serikaku and O'Tousa, 1994; Xu et al., 1997; Zimmerman et al., 1997). The role of the So/Eya complex during development was first revealed in Drosophila. Specifically, coexpression of eya and so greatly increases ectopic eye formation, both in size and frequency, compared with expression of either eya or so alone (Pignoni et al., 1997). Cell culture assays have shown that cotransfection of eya and so appreciably activate transcriptional reporters while either gene alone does not (Silver et al., 2003). More importantly, this synergy between eya and so is conserved in vertebrates. Loss- and gain-of-function studies with Eya1 and Six1 mutants in fish, frogs and chick demonstrate that they play critical roles in sensory placode development (Tadjuidje and Hegde, 2012; Xu, 2013). A clear synergistic activation of targets is observed with combinations of mouse Six and Eya in cell culture studies (Ohto et al., 1999). Moreover, Six1 and Eya1 synergistically regulate organogenesis in mice (Li et al., 2003). Additionally, recent findings have demonstrated that Sox2 physically interacts with Eya1 and Six1 to synergistically activate transcription of Eya1/Six1 targets during sensory cell development (Ahmed et al., 2012a) and neurogenesis (Ahmed et al., 2012b) in the inner ear. In addition to Eya, there are two other proteins, Groucho (Gro) and Fl(2)d, that physically interact with So during eye development (Anderson et al., 2014; Silver et al., 2003).

so and eya are two key members of the Retinal Determination (RD) gene network, and are necessary and sufficient for eye formation (Pappu and Mardon, 2004). In Drosophila, loss of eya or so function causes massive apoptosis and the complete failure of eye development (Bonini et al., 1993; Cheyette et al., 1994; Zimmerman et al., 2000). This cell death phenotype resembles those seen in the ear and kidney primordia of Eya1 and Six1 mutant mouse embryos (Li et al., 2003; Xu et al., 1999). In addition, eya and so are required posterior to the morphogenetic furrow (MF) for normal retinal development (Atkins et al., 2013; Karandikar et al., 2014; Pignoni et al., 1997). In contrast to these loss-of-function phenotypes, ectopic overexpression of eya or so in other imaginal discs is sufficient to induce the formation of ectopic eyes (Bonini et al., 1997; Weasner et al., 2007). Although much has been learned from the genetic dissection of eya and so function during eye development, the direct targets of the So homeodomain transcription factor are largely unknown. To date, only a few direct So targets have been identified in Drosophila, including the RD network genes eyeless (ey) and dachshund (dac), as well as genes that regulate retinal differentiation and MF progression such as atonal (ato), lozenge (lz), prospero (pros), and hedgehog (hh) (Hayashi et al., 2008; Pappu et al., 2003; Pauli et al., 2005; Yan et al., 2003; Zhang et al., 2006). To increase our understanding of the So-dependent regulatory network during eye development, we first generated expression profiles from eye tissue mutant for so or eya at several stages of eye development. Co-expression analysis and motif discovery were then applied to screen candidate So targets. We also used our previous So ChIP-Seq data (Jusiak et al., 2014) to assess So motif predictions. Finally, to validate predicted So targets, we performed in vivo enhancer reporter assays and tested the expression levels of predicted direct So targets in so or eya conditional knockout eye discs using antibody staining. Together, these data expand the known regulatory landscape directly downstream of So and Eya.

Materials and Methods

Fly work

Fly stocks used for RNA-Seq experiments and the Janelia FlyLight enhancer-Gal4 lines used for enhancer reporter assays are listed in Table S7. All flies were maintained with standard corn meal and yeast extract medium at 25°C.

RNA isolation and RNA-Seq

30 eye discs were homogenized in 500 μl of TRIzol (Ambion) using a pestle. RNA was purified using the PureLink RNA Mini Kit (Ambion). We used the Illumina TruSeq protocol to build libraries after RNA extraction and sequenced them on the Illumina HiSeq 2000.

Immunohistochemistry

For immunohistochemistry, imaginal discs of wandering third instar larvae were dissected and processed as described in (Jin et al., 2013). Primary antibodies used were guinea pig anti-So (1:2000, gift from Dr. Ilaria Rebay), chick anti-GFP (1:1000, Abcam), rabbit anti-Pnt (1:500, gift from Dr. Jim Skeath), guinea pig anti-Rau (1:100, gift from Dr. Christian Klämbt), mouse anti-E(spl) (1:1, gift from Dr. Sarah Bray, E(spl) antibody recognize multiple E(spl)-complex members, including HLHmdelta (Cooper et al., 2000), and rabbit anti-E2f (1:100, gift from Dr. Nicholas Dyson). All secondary antibodies were made in goat and used at a final dilution of 1:500; Cy3, Cy5 and Alexa Fluor 488 secondary antibodies were obtained from Jackson ImmunoResearch and Life Technologies, respectively.

RNA-Seq preprocessing

RNA-Seq raw reads and splice junction sites were mapped to Drosophila melanogaster genome Dm3 with Bowtie (v0.12.7) and TopHat (v2.0.0), respectively. The reference Drosophila genome annotation file was downloaded from http://genome.ucsc.edu/. Read counts mapped to each gene were calculated by HTseq (http://www-huber.embl.de/users/anders/HTSeq/doc/overview.html). Normalization and differential expression analysis was performed through DESeq (Anders and Huber, 2010) (v1.20.0, R/Bioconductor package).

Identification of potential targets of eya and so

In order to select potential targets of eya and so, we performed differential expression contrasts of samples with altered “eya”, “so” and “eya and so” function versus “wild-type” at each of the different developmental stages, and another contrast with the three stages at once (DESeq using nbinomTest as the DE method). For each contrast, the standard DESeq workflow was followed, filtering out the 25% of genes with the lowest interquartile range (IQR). The samples included on the contrasts are shown in the Summary tab of Table S2. For “eya and so” vs. “wild-type”, a total of 6 samples (3 so samples and 3 eya samples) were put into the same group (since they function as a complex) for comparison to the corresponding 3 wild type samples from the same stage. Genes differentially expressed in any of the contrasts (p-value<0.05 and log₂ (Fold Change)>1), were selected as input for Gene Ontology enrichment analysis with GOrilla (http://cbl-gorilla.cs.technion.ac.il/) (Eden et al., 2009) and for GENIE3 (Huynh-Thu et al., 2010). As possible regulators (also required as input for running GENIE3), we used the 120 TFs (out of the 753 TFs list available in FlyTF (Adryan and Teichmann, 2006)) that were differentially expressed, plus eya. We selected the 112 genes that were targets of both eya and so within the top quartile of GENIE3 output (importance measure (IM) > 0.009482). We then used HTC clustering (Eisen et al., 1998) in TM4 MultiExperiment Viewer (MeV) (Saeed et al., 2003) to discard genes with inverse expression patterns to eya and so, and kept only the positively correlated genes (marked with the pink vertical bar in Figure 3). These genes were analyzed by gene ontology enrichment analysis with PANTHER (http://pantherdb.org/) (Mi et al., 2013). This provided a list of genes which were potential targets of eya and so based on co-expression. However, these co-expressed genes probably include many indirect downstream genes. To identify the potential direct targets, we looked for enriched TF binding motifs within the co-expressed genes using i-cisTarget (Imrichova et al., 2015). The source of the Six motif identified by i-cisTarget is the Homer Motif Database, which obtained the motif from a dataset from Six1 ChIP-on-Chip from myoblasts (GSE20150). Figure 3B and the interaction network were created in Cytoscape (Shannon et al., 2003) and the online software STRING (Szklarczyk et al., 2015), respectively.

Figure 3.

GO and regulatory network analyses of so/eya positively correlated genes. (A) Gene Ontology analysis, left: Biological processes, right: Molecular Function. The percentage of each section is proportional to the number of differentially expressed genes in GO categories. (B) Gene regulatory network showing predicted regulatory interactions of so/eya positively correlated genes obtained from hierarchical clustering analysis on the top 25% differentially expressed genes. The dashed links represent associations found by GENIE3, links that are confirmed by motif enrichment are indicated by solid lines. If the motif enrichment is significant (ranking higher than 775), arrows are red; otherwise they are blue. One exception is CG11145, which is ranked higher than 775 and therefore should be marked in red. However, its locus overlaps with so, which confounds interpretation. Therefore, we have marked CG11145 in blue. The node color intensity represents the fold change (FC) in differential expression. White nodes: not significant or lower FC than threshold in this contrast (“eya and so” vs. “wild-type”); however, these nodes may be significant in the individual contrasts “eya” vs. “wild-type” or “so” vs. “wild-type” (C) Logo of the SoSix1 motif.

Results

Transcriptome profiling of eya and so mutant eye tissue

So is known to function as a transcription factor with its coactivator Eya and the genetic and physical synergy between so and eya has been widely observed in multiple developmental processes from Drosophila to humans (Abdelhak et al., 1997; Bonini et al., 1993; Cheyette et al., 1994; Dozier et al., 2001; Hoshiyama et al., 2007; Kawakami et al., 2000; Sahly et al., 1999; Seo et al., 1999; Serikaku and O'Tousa, 1994; Xu et al., 1997; Zimmerman et al., 1997). To identify downstream target genes of So, we determined the gene expression profiles of wild-type and so, eya, and ato loss-of-function mutant or RNAi-knockdown eye discs at three distinct developmental stages. The first time point assayed was 68-70 hrs after egg laying (AEL), during which time eye discs are undergoing the late second and early third larval instar stage transition. At this time, so¹ and eya² eye discs are about to begin or have just started differentiation but massive apoptosis has not yet begun. Moreover, no So or Eya protein expression is observed in so¹ and eya² eye discs, respectively, and the size of eye discs is only slightly smaller compared to wild-type (Halder et al., 1998; Jin and Mardon, 2016). We also included expression profiles of ato¹ from the AEL stage to identify target genes expressed posterior to the MF. Since so¹ or eya² eye discs die by apoptosis after the early third instar larval stage, this genotype would not be appropriate for RNA-Seq posterior to the morphogenetic furrow (MF). Therefore, we used RNA interference (RNAi) to knock down eya and so posterior to the MF. Specifically, we dissected GMR>soRNAi and GMR>eyaRNAi eye discs at the white pre-pupal (WP) stage and 24 hrs after puparium formation (APF) for late-stage RNA-Seq. Genotypes of all samples are shown in Table S1. All samples were dissected and collected in three biological replicates and expression profiles of all replicates of a given genotype are closely correlated (Figure S1). To analyze relationships among samples at different developmental stages, principal component analysis (PCA) was applied. As expected, the expression profiles of 68-70 hrs AEL, WP, and 24 hrs APF samples are clearly separated (Figure 1), consistent with the idea that distinct developmental processes occur at these three time points.

Figure 1.

Principal component analysis (PCA). PCA of RNA-Seq expression data from three distinct developmental stages: 68-70 hrs AEL, WP and 24 hrs APF.

Identification of candidate so and eya-regulated genes by co-expression analysis

The main approach used for identification of candidate So targets by co-expression analysis with so and eya was to use GENIE3, a co-expression network inference algorithm (Huynh-Thu et al., 2010). As input to GENIE3, we used all differentially expressed genes between any two relevant conditions, being “eya”, “so” or “eya and so” vs. “wild-type” from each of the three developmental stages separately, and one analysis using all stages at once. This resulted in 1731 genes that are differentially expressed in any of these contrasts (Table S2). We performed Gene Ontology (GO) analysis on these 1731 genes with GOrilla (Eden et al., 2009) and found that they are enriched in Biological Processes GO terms related to eye, photoreceptor, and neuronal development and are also very strongly enriched for TF activity (Table S3 and Figure S2). Notably, the top 10 terms include “sensory organ development” (GO:0007423; P-value 6.11*10^-14), “eye development” (GO:0001654; P-value 7.38*10^-14), “compound eye development” (GO:0048749, P-value 2.46*10^-12), and “regulation of transcription” (GO:0006355, P-value is 3.05E^-11).

We used the top 25% of the co-expression scores based on GENIE3 to select so and eya targets for further analysis. The intersection between the targets for so and eya, resulted in a set of 112 genes which are potential downstream targets for both so and eya (Table S4). Hierarchical clustering was performed on these 112 genes. We also intersected expression profiles of corresponding genes in 68-70 hrs AEL ato¹ mutant eye discs to identify genes that are specifically involved in differentiation (Zhang et al., 2006) and improve the accuracy of our predictions (Potier et al., 2014). As a result, we generated two distinct groups based on clustering, one positively correlated gene set (which follows the same overall expression pattern as eya and so, marked with a pink vertical bar in the heatmap, Figure 2) and one negatively correlated gene set consisting of all remaining genes. Since Eya acts as a coactivator of So and, together, they synergistically regulate downstream genes, the positively correlated genes in so and eya are likely to act downstream of So/Eya and play specific roles during Drosophila retinal development. This set has 34 genes (Table S4, marked in red), containing the positive control hh, so itself as an auto-regulatory target gene and eya, and is enriched in relevant developmental processes, such as neuron differentiation (P-value 2.36*10^-6), eye development (P-value 1.45*10^-5), sensory organ development (P-value 1.53*10^-5), and Bolwig's organ morphogenesis and development (P-value 2*10^-5) (Figure 3A and Table S5). Consistent with previous findings that most of the known So targets are transcription factors, positively correlated genes are over-represented in the Molecular Function GO terms “Transcription factor activity” (18.5%) and “Binding” (40.7%) (Figure 3A). The interconnectivity of So/Eya positively correlated candidates is shown in Figure 3B.

Figure 2.

Hierarchical clustering of the top 25% differentially expressed genes. The heatmap shows the expression profiles of 112 genes identified as potential targets of so and eya by GENIE3. The color scale illustrates relative expression levels across all samples: blue represents down-regulated genes and red represents up-regulated genes. The dendrograms on the left and top of the heatmap show the hierarchical clustering of the transcripts for gene names and for all different samples, respectively. Clusters of genes which positively correlate in so and eya mutant tissue are indicated by the pink vertical bar.

Motif discovery using known direct So targets

Although so plays an important role in Drosophila eye development as a core retinal determination gene, only six genes are currently known as direct targets. To check whether any So motif or Six family DNA binding motif is enriched among the known So target genes, we applied i-cisTarget, an integrative genomics tool to predict transcriptional regulatory modules, to this set of genes (Imrichova et al., 2015). This analysis reveals that a predicted Six1 motif is indeed present in the known So targets (Figure 3C), and has a normalized enriched score (NES) of 3.74. This motif, which we call SoSix1, closely matches the reported consensus motif for So, GTAANYNGANAYC/G (Y=C or T, N=A or C or G or T) (Pauli et al., 2005). This analysis of known direct So targets indicates that if a critical number of true direct So target genes are present among a set of candidate genes, it is possible to retrieve associated binding motifs.

Direct So target prediction by motif enrichment analysis

To screen for direct target genes of So, we performed motif enrichment analysis on the So/Eya positively correlated candidate targets, similarly as performed above on the set of previously known So targets. The SoSix1 motif again appears as significantly enriched at rank 6 (out of nearly 10,000 motifs), with a normalized enrichment score (NES) of 5.32. The optimal cut-off at position 775, defined by i-cisTarget, yields a subset of 21 So binding region predictions among the 34 putative So target genes (Table 1). As a result, we found 12 predicted direct target genes with this motif that are significantly differentially expressed and positively correlated with eya and so. We name this gene set as “confident hits” and are HLHmdelta, pointed (pnt), E2F transcription factor 1 (E2f), hh, so, eya, Matrix metalloproteinase 2 (Mmp2), CG10936, frizzled 2 (fz2), rau, doublesex cognate 73A (dsx-c73A), and betaTub97EF (Table 1). Nine of these are novel putative direct So targets.

Table 1.

List of confident hits. This table shows the rank, associated genes, SoSix1 motif locations predicted by i-cisTarget and ChIP-Seq data validation, as well as reporter assay results and corresponding human homologues. Gray highlights indicate motif predictions for which no corresponding Gal4 lines were available.

Rank	Associated genes	i-cis Target predicted So binding region	Overlap with So ChIP peak?	Reporter assay Positive?	Human homologue
14	HLHmdelta	chr3R: 21822368-21823453	YES	YES	N/A
17	Mmp2	chr2R: 5547784-5548876	YES	N/A	MMP
26	hh	chr3R: 18962365-18964702	YES	YES	SHH
48	so	chr2R: 3316058-3317633	NO	NO	SIX1-6
63	betaTub97EF	chr3R: 23801615-23803245	YES	N/A	TUBB1
85	so	chr2R: 3314710-3315493	YES	YES	SIX1-6
143	so	chr2R: 3304563-3305585	NO	NO	SIX1-6
144	eya	chr2L: 6549825-6550357	NO	YES (ocelli)	EYA1-4
180	eya	chr2L: 6544459-6546448	NO	NO	EYA1-4
228	CG10936	chr2R: 13515427-13516089	YES	N/A	N/A
266	fz2	chr3L: 19161223-19162160	YES	N/A	FZD1-10
278	rau	chr2L: 5869522-5870420	NO	N/A	N/A
279	eya	chr2L: 6541716-6543559	YES	NO	EYA1-4
307	Mmp2	chr2R: 5541749-5544247	YES	N/A	MMP
377	fz2	chr3L: 19176425-19177998	YES	N/A	FZD1-10
503	E2f	chr3R: 17471133-17472022	YES	YES	E2F1
529	dsx-c73A	chr3L: 16428779-16430005	NO	N/A	N/A
645	fz2	chr3L: 19200054-19200895	NO	N/A	FZD1-10
676	E2f	chr3R: 17475343-17476006	YES	YES	E2F1
732	pnt	chr3R: 19145518-19146845	YES	YES	ETS1-2
775	fz2	chr3L: 19140486-19141324	NO	N/A	FZD1-10

Predicted interaction network for So, Eya, and their direct target genes

To investigate predicted interactions for the 12 confident hits, and to assess if they are linked with each other and known So targets, we used the STRING database, which provides co-localization, as well as direct (physical) and indirect (functional) associations (Szklarczyk et al., 2015). Associations among proteins predicted by STRING are shown in Figure 4A. The interaction network shows So associations with most of the confident hits, either directly or indirectly, except for candidate genes dsx-c73A and CG10936. The network also shows a high degree of regulatory cross talk, with many genes regulated by more than one gene. Importantly, the pnt gene, which encodes a transcription factor, associates with so, eya, and all known So target genes. In addition, the STRING output confirms the previously reported physical and genetic interactions of Eya-So and Eya-Dac (Chen et al., 1997; Pignoni et al., 1997) and the genetic interactions of Ey with Eya, So, Dac and Hh; Pnt with E2f and Rau; and Ato with Hh (Bonini et al., 1997; Chen et al., 1999; Halder et al., 1998; Kango-Singh et al., 2003; Pappu et al., 2003; Sieglitz et al., 2013; Staehling-Hampton et al., 1999; White and Jarman, 2000).

Figure 4.

Predicted interaction network and validation using ChIP-Seq data. (A) The associations shown were generated based on STRING analysis; stronger associations are represented by thicker lines. This network includes previously known So targets and predicted targets (confident hits) identified in this study. (B) Examples showing the overlap between SoSix1 motif predictions and ChIP-Seq peaks for the predicted target genes HLHmdelta, Mmp2, hh and so. Each panel includes ChIP-Seq data (2 replicates merged, middle line and negative control, bottom line) and the positions of each SoSix1 motif (in red).

Validation of motif predictions with ChIP-Seq data

To validate So motif predictions, we used our previous ChIP-Seq data for assessment (Jusiak et al., 2014). We first applied i-cisTarget on the top 10% of significantly enriched ChIP-Seq peaks (n=757) and found the most strongly enriched motif in these peaks is SoSix1, with a high significance score (rank=1, NES = 11.88) (Table S6). This finding strongly supports the proposal that SoSix1 is a bona fide So motif. By comparing locations of the SoSix1motif predictions with the ChIP-Seq peaks of confident hits, we found that 62% of motif predictions overlap with a So ChIP-Seq peak (13 out of 21) (Table 1). Examples in Figure 4B show the overlap between SoSix1 motif predictions and ChIP-Seq peaks for HLHmdelta, Mmp2, hh and so, the top four predicted target genes in Table 1.

Validation of predicted direct So targets in vivo

Each of the 12 predicted So target genes harbor a cluster of one or more So binding regions (Table 1). To test if any of these motif predictions overlap with retinal enhancer activity, we performed in vivo enhancer reporter assays using the Janelia collection of Gal4 lines (Pfeiffer et al., 2008). We found eleven Gal4 lines that cover a predicted motif, representing six target genes: HLHmdelta, hh, pnt (one Gal4 line each), E2f (two different Gal4 lines each), so, eya (three different Gal4 lines each). These lines were crossed to UAS-GFP for detection of enhancer activity. We also visualized So expression in the same eye discs. Remarkably, we found GFP expression in eye discs from at least one Gal4 for each of the six predicted target genes we tested (HLHmdelta, E2f, pnt, hh, so and eya), of which five are active in photoreceptor cells and overlap with So staining (Figure 5). One Gal4 line from the eya gene showed enhancer activity in the ocelli, which also colocalizes with So expression (Cheyette et al., 1994). Moreover, the SoSix motifs associated with each predicted So target genes are conserved (Figure S3). We furthermore confirmed that the predicted direct So targets Pnt, E2f, HLHmdelta and Rau, for which antibodies are available, are reduced in an so or eya post-mitotic knockout background (so^GMRcKO and eya^GMRcKO, respectively, unpublished data), which allows removing so or eya function in differentiating cells posterior to the MF (Figure 6). Overall, by intersecting genome-wide data from so and eya, we achieved a high success rate of So direct target gene prediction.

Figure 5.

Validation of predicted So target enhancers by in vivo enhancer-reporter assays. Predicted So binding motifs were tested for 3 predicted target genes (HLHmdelta, E2f and pnt) and 3 known target genes (hh, so and eya). (Left) UCSC Genome Browser screenshots showing the locations of So binding regions predicted by i-cisTarget (green horizontal bars) and the enhancer-GAL4 lines from the Janelia FlyLight collection (red horizontal bars, the enhancer-Gal4 lines tested in vivo and are positive in reporter assays are indicated by vertical black arrows). (Right) SoSix1 motif predictions were tested by crossing the Janelia FlyLight GAL4 lines indicated in red in the panels to the left with UAS-GFP. Third instar larvae from these crosses were stained using anti-GFP (green) and anti-So (red).

Figure 6.

so and eya are required for normal expression of putative direct So targets posterior to the MF. Antibody stainings for a set of predicted direct So targets in wild-type and post-mitotic knockout of so (so^GMRcKO) or eya (eya^GMRcKO) imaginal discs. So and Eya expression are eliminated posterior to the MF in so^GMRcKO and eya^GMRcKO third instar larval eye discs, respectively. Since third instar larval eye discs of so^GMRcKO flies are misshapen (Figure S4), disc fragments are shown to improve visualization of putative direct So targets expression (third column). Expression of Pnt (A) and E2f (B) are strongly reduced, and that of HLHmdelta (C) and Rau (D) are moderately reduced in both so^GMRcKO and eya^GMRcKO eye discs.

Discussion

The homeodomain transcription factor So regulates multiple aspects of eye development for a wide range of higher eukaryotes. However, the mechanisms underlying the So-dependent regulatory network during eye development largely remain unknown, in part since only a few direct So targets are currently identified. To end this, we used a combinatorial approach of expression profiling and motif discovery to identify 12 putative So direct targets, including 9 novel targets. Through in vivo enhancer reporter assays, we found at least one predicted enhancers from all six target genes tested is able to drive reporter expression in the developing eye disc. Our results expand the list of putative So direct targets in the eye, providing the basis for further functional studies.

Intersecting expression profiles of so and eya mutant eye tissue

Transcription factors usually regulate their downstream targets by binding cooperatively with their cofactors or other TFs (Hayashi et al., 2008; Zhang et al., 2006). To identify direct targets of So and infer a regulatory network, we performed co-expression analyses by intersecting expression profiles of so and its cofactor gene eya at three different developmental stages. The “AEL” samples are a mixture of eye discs that are immediately before, during and after differentiation. The RNAs from AEL eye discs were prepared in parallel (same time point of collection, same temperature and same manipulation) to minimize variation within a single genotype and make direct comparison of different genotypes meaningful. Moreover, we included AEL data from ato¹ mutant eye discs with that of so¹ and eya² to identify genes that are specifically involved in differentiation posterior to the MF. In addition to expression profiles at an early developmental stage, we included RNAi knockdown data of eya and so at two developmental stages after differentiation to gain a larger number of perturbations and minimize nonspecific candidate genes. In particular, numerous studies have suggested that co-expression analysis requires many related perturbations to achieve good prediction power (Hsu et al., 2015; Potier et al., 2014; Torkamani et al., 2010).

As expected, the 1731 genes that are differentially expressed in response to loss of eya or so function are enriched in GO categories that are relevant to eye development. For further analysis, 34 genes that positively correlate in so and eya mutant tissue were identified. Applying GO enrichment on this gene set, we found they are not only enriched in GO terms pertaining to the eye, but also rank high in the GO term “regulation of fibroblast growth factor receptor (FGFR) signaling pathway” (P-value 1.93*10^-3). FGFR signaling regulates the onset of retinal wrapping glial cell differentiation (Franzdottir et al., 2009), suggesting that So may regulate glial differentiation during eye development. By using in vivo enhancer reporter assays and validation of putative target expression in the absence of so or eya function posterior to the MF, we achieved a high rate of success for So direct target prediction. Our findings provide an example where intersection of expression profiling of a TF with its cofactor can increase specificity in subsequent analyses of RNA-Seq data, which should be generally applicable to target prediction for other TFs.

Integrative analysis of gene expression profiling and cis-regulatory analysis

Sequencing-based expression profiling, such as RNA-Seq, is widely used to obtain genome-wide quantitative gene expression levels in many organisms (Hong et al., 2011; McManus et al., 2010). RNA-Seq based on genetic perturbations provides a powerful approach to gain a deeper understanding of TF-regulated networks. However, this approach does not provide cis-regulatory information. To solve this problem, we applied regulatory motif discovery to putative target genes derived from RNA-Seq data and used our previous ChIP-Seq data to validate the locations of motif predictions. First, we identified the “SoSix1” motif as a putative So binding motif by using all known direct So targets as input. We then observed that this motif is over-represented across our predicted targets. The targetome of the SoSix1 motif comprises 12 out of the 34 putative So target genes, including the known So direct target hh and so itself. Finally, significantly enriched So ChIP-Seq peaks were used to assess motif predictions. We found that more than 62% of predicted motifs overlap with highly enriched ChIP-Seq peaks in the 12 genes studied. A similar strategy has been successfully employed to identify target genes for eye-specific TFs expressed within and posterior to the MF, including Atonal, Glass and other TFs in Drosophila (Aerts et al., 2010; Menoret et al., 2013; Oliva et al., 2015; Potier et al., 2014).

In this paper, we first applied this approach to So and identified a canonical So binding motif, which was extracted from a mouse Six1 ChIP-Seq dataset (GSE20150). This is not surprising since Six1/2 are the mammalian homologues of so (Kawakami and Kobayashi, 1998) and so, Six1/2, and Six4/5 share very similar DNA-binding sequences (Pauli et al., 2005). Moreover, previous protein-DNA interaction studies of Six and So found that a stretch of 13 nucleotides, GTAANYNGANAYC/G, is necessary for Six and So binding to DNA. Nucleotides G, A, A at positions 1, 4, 9, respectively, are most important for this interaction (Hazbun et al., 1997; Pauli et al., 2005; Rogers et al., 2005; Silver et al., 2003; Zhang et al., 2006). Our predicted SoSix1 motif closely matches this known consensus sequence. More importantly, the three nucleotides at positions 1, 4, and 9 are present and conserved within the SoSix1 motif.

Novel direct So targets in eye development

Through our analysis, we identified 12 predicted direct So target genes, including nine novel targets: HLHmdelta, pnt, E2f, Mmp2, CG10936, fz2, rau, dsx-c73A and betaTub97EF. Gene regulatory network analysis predicts a high degree of regulatory cross talk between novel and known So targets (including So and Eya). Interestingly, Mmp2, rau and pnt are involved in the EGFR and FGFR signaling pathways (Franzdottir et al., 2009). Specifically, Rau and its transcriptional activator Pnt act in a positive feedback loop downstream of FGFR- and EGFR-mediated signaling to regulate R7 photoreceptor development, while Mmp2 is implicated in cancer (Bossuyt et al., 2009; Sieglitz et al., 2013). These findings suggested that So may mediate FGFR and EGFR signaling pathways and cancer cell proliferation via its direct downstream targets. Additionally, Pnt and So directly regulate hh (Pauli et al., 2005; Rogers et al., 2005), providing an intricate model of cross-talk among the retinal determination genes and the EGFR, FGFR, and Hh signaling pathways.

HLHmdelta, pnt, E2f are three predicted targets which have FlyLight enhancer-Gal4 lines, all of which can drive reporter activation during eye development. HLHmdelta and E2f are involved in Notch signaling and are required for R3/R4 photoreceptor cell induction and cell proliferation in eye discs, respectively (Baonza and Freeman, 2005; Cooper and Bray, 1999). During Drosophila eye development, Notch signaling plays important roles in the growth of eye discs, dorsal/ventral boundary establishment, and MF progression (Bhattacharya and Baker, 2009; Dominguez and de Celis, 1998). Our work suggests that So regulates these processes by acting upstream of HLHmdelta and E2f in Notch signaling, consistent with the known roles of So during Drosophila eye development, including cell survival, MF initiation, and photoreceptor differentiation (Cheyette et al., 1994). Moreover, the putative So target Fz2 can bind Wingless (Wg) and act as a receptor for Wg in Drosophila (Chen and Struhl, 1999). Along with previous studies suggesting that the Wg signaling pathway regulates eya during ocelli development (Blanco et al., 2009) and our finding that a predicted enhancer of eya is able to drive reporter expression in ocelli, we hypothesize that So regulates this process through eya, fz2, and wg. In contrast, little is currently known about the function of CG10936, dsx-c73A and betaTub97EF, although they may play a role in sensory perception, cuticle development, and centrosome duplication, respectively. A detailed characterization of the regulatory interactions between these novel putative targets and So awaits further studies.

Conclusions

In conclusion, by integrating gene expression profiles of so and eya, we have obtained a high-quality set of so/eya target genes which show a strong GO term enrichment in eye development and TF regulation. Applying motif discovery to this gene set, we identified twelve predicted direct So target genes. Six of these were tested by in vivo enhancer reporter assays and at least one predicted enhancer from each gene is able to activate reporter expression in eye discs, suggesting that this approach for target prediction is robust. Since many eye-specific genes are highly conserved, our studies may yield new insights into the regulatory mechanisms underlying So-dependent eye development throughout the metazoa.

Highlights.

Determining the gene expression profiles of so or eya mutant eye tissues.

Applying co-expression analysis and motif discovery to differentially expressed genes.

A canonical So binding motif are identified from motif discovery and ChIP-seq data.

Enhancer reporter assays confirm predicted So binding motifs in vivo.